A Sensor for Metabolic Profiling of
Traumatic Brain Injury
Victoria Helen Highland
Newnham College
University of Cambridge
This dissertation is submitted for the degree of
Doctor of Philosophy
January 2018

Declaration
This dissertation is the result of my own work and includes nothing which is the
outcome of work done in collaboration except as declared in the Preface and specified
in the text.
It is not substantially the same as any that I have submitted, or, is being con-
currently submitted for a degree or diploma or other qualification at the University
of Cambridge or any other University or similar institution except as declared in the
Preface and specified in the text. I further state that no substantial part of my disser-
tation has already been submitted, or, is being concurrently submitted for any such
degree, diploma or other qualification at the University of Cambridge or any other
University or similar institution except as declared in the Preface and specified in the
text.
It does not exceed the prescribed word limit for the relevant Degree Committee.
i

Abstract
A Sensor for Metabolic Profiling of Traumatic Brain Injury
Victoria Helen Highland
A Traumatic Brain Injury (TBI) occurs when brain tissue is damaged as a result
of an external mechanical force. It is a global health issue which, when serious, can
result in death or permanent disability. The damage caused by the initial trauma
continues to evolve in the days following the incident and therefore patients with a
serious TBI are closely monitored so that, if any change in the condition of the brain
is detected, an appropriate medical intervention can be made. Metabolic profiling of
extracellular brain fluid can be used to monitor brain health, as it provides data on
the efficiency of metabolic processes taking place within the brain. If metabolism is
disrupted, causing energy failure in the brain, it can lead to the destruction of brain
cells and a poor patient outcome.
D-glucose, L-lactate and pyruvate are the metabolic intermediates most commonly
monitored following TBI. Their respective concentrations are currently measured via
a two-stage process: the constituents of the extracellular brain fluid are first extracted
by microdialysis and are then analysed via a series of colorimetric assays which are
carried out using a commercially-available analyser. The use of this analyser presents
a number of drawbacks which have prevented the widespread use of metabolic profiling
in the management of TBI. For example, the process of manually loading samples into
the analyser is both time- and labour-intensive and therefore is carried out only in-
termittently by clinical staff. As a result, metabolic profiling data cannot be obtained
in real time and any necessary medical intervention may consequently be delayed.
This thesis describes the work undertaken to develop a spectroscopic method, and
an associated statistical model, for the detection of glucose, lactate and pyruvate
that could be directly integrated into the existing microdialysis setup and thus used
for continuous metabolic profiling of TBI patients. The techniques of Raman spec-
troscopy and Fourier Transform Infra-Red (FTIR) spectroscopy were investigated to
determine whether they could selectively detect glucose, lactate and pyruvate in solu-
tion and whether they possessed the sensitivity required to detect those compounds
at clinically-relevant concentrations. Consideration was also given to the practicalities
iii
iv
of adapting each technique to the specific requirements of the clinical setting in which
it would be used. These initial studies indicated that FTIR spectroscopy showed the
greater promise for development into a sensor and therefore subsequent work focused
only on this technique. A detailed study was carried out in order to compare vari-
ous FTIR spectroscopy methods. The data generated were used to build and test a
statistical model for the prediction of clinically-relevant glucose, lactate and pyruvate
concentrations from their FTIR spectra. The results of this work provide important
insights into the capabilities of currently-available FTIR techniques, when used in
combination with predictive statistical modelling, and will be used to guide the next
phase of sensor development.
Dedicated with love to Bill, Helen and John.

Acknowledgements
The last four years have been a significant learning curve, often in ways that I had
not expected. One of the most important things I have learned over the last four
years is that, despite what the declaration at the beginning of every thesis might say,
a PhD is very much a team effort. From scientific discussions and demonstrations,
to navigating the world of academia, to practical and emotional support, I have been
lucky enough to receive help from so many people. As a result, I have many, many
people to thank...
Firstly, I must thank my supervisor, Professor Stephen Elliott, for giving me the
opportunity to undertake this work and for his financial support during my final year.
I am particularly grateful for his help in shaping and proof-reading this thesis. I would
also like to thank Dr Tanya Hutter for her enthusiasm and support throughout this
project.
I would like to take the opportunity to say a huge ‘thank you’ to Dr Anna Hu¨fner
for her exceptional mentoring, which included a number of helpful discussions about
this project, MATLAB tutoring and wise advice about PhD life in general. Signifi-
cant thanks are also owed to to Dr Mike Casford, for providing access to the Spectrum
100 and for the many helpful discussions about FTIR spectroscopy, as well as to Dr
Yury Alaverdyan for his invaluable guidance with the Raman work. I am sincerely
grateful for their help. I am also grateful to the many other people who so generously
gave their time to offer advice and/or practical assistance with experiments, many
of which did not make it into this thesis. Thank you to Dr Matthew Capener, Dr
Thomas Mueller, Dr Jerome Charmet and Duncan Howe. I was also lucky to receive
excellent support from the superb technical services teams in the Chemistry Depart-
ment, particularly Manuel Wentscher and Wayne Bailey, as well as from Clive Tubbs
from the Department of Architecture.
Thank you very much to those members of the Department who offered such
significant support and guidance during my final year, in particular to Dr Deborah
Longbottom and Professor Markus Kalberer. I am also very grateful to have received
support from Dr Paul Barker and Rachel MacDonald, as well as financial support
from the Graduate Education Committee and Newnham College.
I am extremely grateful to Professor Liz Varga for her invaluable advice which
vii
viii
helped to shape this thesis.
Thank you also to the Elliott Group - Anna, Coco, Farah, Felix, Kostas, Joa˜o,
Marcin, Paul, Sarah, Taehoon, Tas and Yuchen for being such a great group of people
to work alongside.
Finally, the biggest thanks of all must go to my family and friends for their support
and for so many happy memories, which cover far more than just the last four years:
To my wonderful friends (you know who you are): you are all amazing. Thank
you for keeping me sane and for so many good times - I could not have done this
without you all. Special thanks to Becky for advising and supporting me through my
final year; to Megan, Charlotte and Sophie for their support while writing this thesis;
to team 32:52 and to Beth, forever my Head Honcho.
To Diane and Paul, the best in-laws I could ask for: thank you for being the source
of so many of the highlights of the last four years. I’m so happy to be a part of the
Ackroyd family.
I’m blessed to have a wonderful family who mean the world to me. To Nan, Pops,
Chris, Mel, Harry, Ollie, Nick, Corn, Olivia, Vee, Paul and Ted: thank you for your
love, support and belief in me.
To Tom and Katie: thank you for your support and friendship - I’m so lucky to
have you.
To Mum and Dad: I don’t know where to begin. Thank you so much for all you
have done and continue to do for me, for inspiring my love of science in the first place
and for your constant support.
Finally, to Mike: thank you for being by my side through it all. Here’s to the next
chapter - I love you.
List of abbreviations
ANOVA Analysis of Variance
ATP Adenosine Triphosphate
ATR Attenuated Total Reflectance
◦C Degrees Celsius
CCD Charge-Coupled Device
cm Centimetres
cm−1 Wavenumbers
DI Deionised
DNA Deoxyribonucleic Acid
FAD Flavin Adenine Dinucleotide
FRET Fo¨rster Resonance Energy Transfer
FTIR Fourier Transform Infra Red
g Grams
HATR Horizontal Attenuated Total Reflectance
ICP Intracranial Pressure
IR Infra-Red
k Force Constant
kB Boltzmann Constant
kDa Kilodaltons
LoB Limit of Blank
ix
xLoD Limit of Detection
LoQ Limit of Quantification
LPR Lactate:Pyruvate Ratio
µL Microlitre
µm Micrometre
mM Millimolar
mW Milliwatt
MCT Mercury Cadmium Telluride
MIP Molecularly Imprinted Polymer
nm Nanometre
PCA Principal Component Analysis
PCR Principal Component Regression
PDMS Polydimethylsiloxane
PET Polyethylene Terephthalate
PLSR Partial Least Squares Regression
QCL Quantum Cascade Laser
QCM Quartz Crystal Microbalance
RMSECV Root Mean Squared Error of Cross-Validation
RMSEP Root Mean Squared Error of Prediction
RSS Residual Sum of Squares
rsMD Rapid Sampling Microdialysis
s Second
SERS Surface-Enhanced Raman Spectroscopy
SPR Surface Plasmon Resonance
TBI Traumatic Brain Injury
TCA Tricarboxylic Acid
TMB 3,3’,5,5’-Tetramethylbenzidine
V Volt
Contents
1 Thesis Overview 1
1.1 Motivation & Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Experimental Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Note on Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Guide to Chapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Project Background 5
2.1 Introduction to TBI . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Mechanism of TBI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Monitoring TBI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3.1 Microdialysis in TBI . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.2 Compounds Recovered using Microdialysis . . . . . . . . . . . 9
2.3.3 ISCUS Flex Microdialysis Analyser . . . . . . . . . . . . . . . . 10
2.4 Metabolism in TBI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4.1 Cellular Respiration . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4.2 Glucose in TBI . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4.3 Lactate and Pyruvate in TBI . . . . . . . . . . . . . . . . . . . 14
2.5 Project Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Device Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 Theoretical Background 19
3.1 Molecular Vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.1.1 Classical Harmonic Oscillator . . . . . . . . . . . . . . . . . . . 19
3.1.2 Quantum-Mechanical Harmonic Oscillator . . . . . . . . . . . . 21
3.1.3 Anharmonic Oscillator . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Infra-Red Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1 Selection Rules of IR Spectroscopy . . . . . . . . . . . . . . . . 22
3.2.2 IR Spectroscopy Instrumentation . . . . . . . . . . . . . . . . . 23
3.2.3 Types of IR Spectroscopy . . . . . . . . . . . . . . . . . . . . . 24
3.2.3.1 Transmission-FTIR Spectroscopy . . . . . . . . . . . . 24
3.2.3.2 ATR-FTIR Spectroscopy . . . . . . . . . . . . . . . . 25
xi
xii CONTENTS
3.3 Raman Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.1 Raman Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.2 Selection Rules of Raman Spectroscopy . . . . . . . . . . . . . 28
3.3.3 Surface-Enhanced Raman Spectroscopy . . . . . . . . . . . . . 29
3.3.4 Raman Spectroscopy Instrumentation . . . . . . . . . . . . . . 31
3.4 Statistical Analysis and Modelling . . . . . . . . . . . . . . . . . . . . 33
3.4.1 Analysis of Variance . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.2 Statistical Modelling . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4.3 Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4.3.1 Principal Component Regression . . . . . . . . . . . . 36
3.4.3.2 Partial Least Squares Regression . . . . . . . . . . . . 36
3.4.3.3 Model Evaluation . . . . . . . . . . . . . . . . . . . . 37
4 Current State of the Art 39
4.1 Detection of Glucose, Lactate & Pyruvate . . . . . . . . . . . . . . . . 39
4.1.1 Glucose Detection . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.1.1 Electrochemical Sensing . . . . . . . . . . . . . . . . . 39
4.1.1.2 Optical Sensing . . . . . . . . . . . . . . . . . . . . . 42
4.1.1.3 Other Sensing Methods . . . . . . . . . . . . . . . . . 43
4.1.2 Lactate Detection . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.3 Pyruvate Detection . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 IR Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2.1 Detection of Metabolites by IR Spectroscopy . . . . . . . . . . 48
4.2.2 IR Spectroscopy & Microfluidics . . . . . . . . . . . . . . . . . 50
4.3 Raman Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.1 SERS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.2 Detection of Metabolites by Raman & SERS . . . . . . . . . . 52
4.3.3 Raman Spectroscopy & Microfluidics . . . . . . . . . . . . . . . 54
4.4 Neurochemical Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.5 Metabolic Profiling of TBI: Preliminary Work . . . . . . . . . . . . . . 57
5 Materials & Methods 61
5.1 Sample Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.1.1 Perfusion Fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.1.2 Synthetic Samples . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.1.3 Clinical Samples . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2 FTIR Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2.1 ATR-FTIR Protocol . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2.1.1 Single-Reflection ATR-FTIR Protocol . . . . . . . . . 63
5.2.1.2 Seven-Reflection ATR-FTIR Protocol . . . . . . . . . 64
CONTENTS xiii
5.2.2 Transmission-FTIR Protocol . . . . . . . . . . . . . . . . . . . 64
5.3 Raman Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.1 Instrument Specifications . . . . . . . . . . . . . . . . . . . . . 65
5.3.2 Instrument Operation . . . . . . . . . . . . . . . . . . . . . . . 66
5.4 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.4.1 ANOVA and Multiple Comparison Tests . . . . . . . . . . . . . 67
5.4.2 Statistical Modelling: Pre-Processing . . . . . . . . . . . . . . . 67
5.4.3 Statistical Modelling: PCR . . . . . . . . . . . . . . . . . . . . 68
5.4.4 Statistical Modelling: PLSR . . . . . . . . . . . . . . . . . . . . 68
6 IR Spectroscopy 69
6.1 IR Spectrum Characterisation . . . . . . . . . . . . . . . . . . . . . . . 70
6.1.1 FTIR Spectrum of Glucose . . . . . . . . . . . . . . . . . . . . 70
6.1.2 FTIR Spectrum of Lactate . . . . . . . . . . . . . . . . . . . . 74
6.1.3 FTIR Spectrum of Pyruvate . . . . . . . . . . . . . . . . . . . . 77
6.1.4 Mixtures of Glucose, Lactate and Pyruvate . . . . . . . . . . . 80
6.1.5 IR Spectrum Characterisation: Conclusions . . . . . . . . . . . 81
6.2 Interference from Competing Molecules . . . . . . . . . . . . . . . . . 84
6.2.1 Interference from Competing Molecules: Conclusions . . . . . . 94
6.3 FTIR Spectroscopy of Clinical Samples . . . . . . . . . . . . . . . . . 94
6.3.1 FTIR Spectroscopy of Clinical Samples: Conclusions . . . . . . 96
6.4 FTIR Spectroscopy Accessory Selection . . . . . . . . . . . . . . . . . 96
6.4.1 Crystal Selection for ATR-FTIR . . . . . . . . . . . . . . . . . 97
6.4.2 Path Length Selection for Transmission-FTIR . . . . . . . . . . 106
6.4.3 FTIR Spectroscopy Accessory Selection: Conclusions . . . . . . 109
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7 Assessment of Raman Spectroscopy 113
7.1 Raman Spectroscopy of Glucose . . . . . . . . . . . . . . . . . . . . . . 113
7.1.1 Raman Spectroscopy of Glucose: Conclusions . . . . . . . . . . 116
7.2 Proposed SERS Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2.1 Synthesis of Silver Nanoparticles . . . . . . . . . . . . . . . . . 117
7.3 Testing SERS Sensor Viability: Cysteine Study . . . . . . . . . . . . . 118
7.3.1 Raman Spectroscopy of Solid Cysteine . . . . . . . . . . . . . . 118
7.3.2 Raman Spectroscopy of Aqueous Cysteine . . . . . . . . . . . . 119
7.3.3 SERS of Aqueous Cysteine: Fixed Substrates . . . . . . . . . . 121
7.3.4 SERS of Aqueous Cysteine: Colloidal Substrates . . . . . . . . 124
7.3.5 Testing SERS Sensor Viability: Conclusions . . . . . . . . . . . 125
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
xiv CONTENTS
8 Mixture Analysis: ATR-FTIR Spectra 129
8.1 Mixture Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.1.1 Three-Component Mixtures . . . . . . . . . . . . . . . . . . . . 130
8.1.2 Two-Component Mixtures . . . . . . . . . . . . . . . . . . . . . 132
8.1.3 Pure Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.2 Mixture Selection: Conclusions . . . . . . . . . . . . . . . . . . . . . . 134
8.3 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.3.1 Data Acquisition: Conclusions . . . . . . . . . . . . . . . . . . 135
8.4 Spectrum Analysis Methods . . . . . . . . . . . . . . . . . . . . . . . . 135
8.4.1 Spectrum Analysis Methods: Conclusions . . . . . . . . . . . . 137
8.5 ATR-FTIR Spectrum Analysis: Comparison of Statistical Models . . . 137
8.5.1 PCR vs. PLSR: Ratio ATR-FTIR Spectra . . . . . . . . . . . . 138
8.5.2 PCR vs. PLSR: Single-Beam ATR-FTIR Spectra . . . . . . . . 140
8.5.3 Comparison of Statistical Models: Conclusions . . . . . . . . . 142
8.6 ATR-FTIR: Effect of Spectral Range . . . . . . . . . . . . . . . . . . . 142
8.6.1 Effect of Spectral Range: ATR-FTIR Ratio Spectra . . . . . . 143
8.6.2 Effect of Spectral Range: ATR-FTIR Single-Beam Spectra . . 144
8.6.3 Effect of Spectral Range: Conclusions . . . . . . . . . . . . . . 145
8.7 ATR-FTIR: Prediction of LPR . . . . . . . . . . . . . . . . . . . . . . 146
8.7.1 ATR-FTIR: Prediction of LPR from Ratio Spectra . . . . . . . 146
8.7.2 ATR-FTIR: Prediction of LPR from Single-Beam Spectra . . . 147
8.7.3 Prediction of LPR: Conclusions . . . . . . . . . . . . . . . . . . 147
8.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
9 Mixture Analysis: Transmission-FTIR Spectra 151
9.1 Mixture Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
9.2 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
9.2.1 Data Acquisition: Conclusions . . . . . . . . . . . . . . . . . . 154
9.3 Transmission-FTIR: Comparison of Statistical Models . . . . . . . . . 154
9.3.1 PCR vs. PLSR: Ratio Transmission-FTIR Spectra . . . . . . . 155
9.3.2 PCR vs. PLSR: Single-Beam Transmission-FTIR Spectra . . . 157
9.3.3 Comparison of Statistical Models: Conclusions . . . . . . . . . 159
9.4 Transmission-FTIR: Effect of Spectral Range . . . . . . . . . . . . . . 159
9.4.1 Spectral Range: Transmission-FTIR Ratio Spectra . . . . . . . 159
9.4.2 Spectral Range: Transmission-FTIR Single-Beam Spectra . . . 160
9.4.3 Effect of Spectral Range: Conclusions . . . . . . . . . . . . . . 161
9.5 Transmission-FTIR: Prediction of LPR . . . . . . . . . . . . . . . . . . 162
9.5.1 Prediction of LPR from Ratio Spectra . . . . . . . . . . . . . . 162
9.5.2 Prediction of LPR from Single-Beam Spectra . . . . . . . . . . 163
9.5.3 Prediction of LPR: Conclusions . . . . . . . . . . . . . . . . . . 163
CONTENTS xv
9.6 Transmission-FTIR: Analysis of Clinical Samples . . . . . . . . . . . . 164
9.6.1 Analysis of Clinical Samples: Ratio Spectra . . . . . . . . . . . 165
9.6.2 Analysis of Clinical Samples: Single-Beam Spectra . . . . . . . 166
9.6.3 Analysis of Clinical Samples: Conclusions . . . . . . . . . . . . 167
9.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
10 Discussion and Future Outlook 171
10.1 Summary of Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
10.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
10.3 Considerations for Future Device Development . . . . . . . . . . . . . 174
10.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Appendices 179
A Glucose, Lactate & Pyruvate Combinations 179
A.1 Three-Component Mixtures . . . . . . . . . . . . . . . . . . . . . . . . 179
A.2 Two-Component Mixtures . . . . . . . . . . . . . . . . . . . . . . . . . 182

List of Figures
2.1 Microdialysis Catheter . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Cellular Respiration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Project Aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.1 Harmonic and Anharmonic Potential Energy Curve . . . . . . . . . . . 20
3.2 Optical Path of a Typical FTIR Spectrometer . . . . . . . . . . . . . . 24
3.3 Attenuated Total Reflection FTIR . . . . . . . . . . . . . . . . . . . . 26
3.4 The Raman Effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Surface Plasmon Resonance . . . . . . . . . . . . . . . . . . . . . . . . 30
3.6 Optical Path of a Generalised Raman Spectrometer . . . . . . . . . . . 32
4.1 Reaction Between Glucose and Glucose Oxidase . . . . . . . . . . . . . 40
4.2 Electrode reactions in an Amperometric Glucose Sensor . . . . . . . . 40
4.3 Anode Reaction of Glucose Oxidase and Ferrocene . . . . . . . . . . . 41
4.4 Enzyme Reactions of rsMD . . . . . . . . . . . . . . . . . . . . . . . . 55
5.1 Perkin Elmer Spectrum 100 . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2 Single-Beam vs. Ratio Spectrum . . . . . . . . . . . . . . . . . . . . . 63
5.3 Accessories for ATR-FTIR . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.4 Pearl Liquid Transmission-FTIR Accessory . . . . . . . . . . . . . . . 65
5.5 Raman Sampling Accessories . . . . . . . . . . . . . . . . . . . . . . . 66
6.1 ATR-FTIR Spectrum: 100 mM Glucose . . . . . . . . . . . . . . . . . 72
6.2 Transmission-FTIR Spectrum: 100 mM Glucose . . . . . . . . . . . . 73
6.3 ATR-FTIR Spectrum: 100 mM Lactate . . . . . . . . . . . . . . . . . 75
6.4 Transmission-FTIR Spectrum: 100 mM Lactate . . . . . . . . . . . . . 76
6.5 ATR-FTIR Spectrum: 100 mM Pyruvate . . . . . . . . . . . . . . . . 78
6.6 Transmission-FTIR Spectrum: 100 mM Pyruvate . . . . . . . . . . . . 79
6.7 Contributions to Mixture Spectrum . . . . . . . . . . . . . . . . . . . . 80
6.8 ATR-FTIR Spectrum: Glu, Lac & Pyr Mixture . . . . . . . . . . . . . 82
6.9 Transmission-FTIR Spectrum: Glu, Lac & Pyr Mixture . . . . . . . . 83
6.10 ATR-FTIR Spectrum: Glycerol . . . . . . . . . . . . . . . . . . . . . . 86
xvii
xviii LIST OF FIGURES
6.11 Transmission-FTIR Spectrum: Glycerol . . . . . . . . . . . . . . . . . 87
6.12 ATR-FTIR Spectrum: Glutamate . . . . . . . . . . . . . . . . . . . . . 88
6.13 Transmission-FTIR Spectrum: Glutamate . . . . . . . . . . . . . . . . 89
6.14 ATR-FTIR Spectrum: Glu, Lac, Pyr, Gly & Glut Mixture . . . . . . . 92
6.15 Transmission-FTIR Spectrum: Glu, Lac, Pyr, Gly & Glut . . . . . . . 93
6.16 Transmission-FTIR Spectra of Clinical Samples . . . . . . . . . . . . . 95
6.17 Single-Reflection ATR-FTIR Spectra: Glucose Concentrations . . . . . 98
6.18 Normalised ATR-FTIR Spectrum of Glucose . . . . . . . . . . . . . . 99
6.19 ATR-FTIR: Glucose Concentration vs. Relative Peak Intensity . . . . 101
6.20 Results of Multiple Comparison Test: ATR-FTIR . . . . . . . . . . . . 102
6.21 ATR-FTIR: Glucose Concentration vs. Relative Peak Intensity . . . . 103
6.22 ATR-FTIR: Pyruvate Concentration vs. Relative Peak Intensity . . . . 104
6.23 Results of Multiple Comparison Test: ATR-FTIR of Lactate . . . . . 105
6.24 Transmission-FTIR Spectra: Glucose Concentrations . . . . . . . . . . 107
6.25 Comparison of Transmission-FTIR Spectra Noise Levels . . . . . . . . 108
6.26 Transmission-FTIR: Glucose Concentration vs. Relative Peak Intensity 109
6.27 Results of Multiple Comparison Test: Transmission-FTIR . . . . . . . 110
7.1 Raman Spectrum of 1M Glucose . . . . . . . . . . . . . . . . . . . . . 114
7.2 Change in Raman Intensity with Decreasing Glucose Concentration . 115
7.3 Proposed SERS Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.4 Raman Spectrum of Solid Cysteine . . . . . . . . . . . . . . . . . . . . 120
7.5 Raman Spectrum of 1M Cysteine . . . . . . . . . . . . . . . . . . . . . 121
7.6 SERS Spectra of 10 uM Cysteine on Silver Nanoparticles . . . . . . . 123
7.7 SERS Spectrum of 10 uM Cysteine on Colloidal Silver . . . . . . . . . 125
8.1 Histograms of Patient Data . . . . . . . . . . . . . . . . . . . . . . . . 131
8.2 ATR-FTIR Ratio Spectra of Mixtures: Examples . . . . . . . . . . . . 135
8.3 ATR-FTIR Single-Beam Spectra of Mixtures: Examples . . . . . . . . 136
8.4 PCR of ATR-FTIR Ratio Spectra: Overview . . . . . . . . . . . . . . 138
8.5 PLSR of ATR-FTIR Ratio Spectra: Overview . . . . . . . . . . . . . . 139
8.6 PCR of ATR-FTIR Single-Beam Spectra: Overview . . . . . . . . . . 140
8.7 PLSR of ATR-FTIR Single-Beam Spectra: Overview . . . . . . . . . . 141
8.8 Spectral Interval: RMSECV of ATR-FTIR Ratio Spectra . . . . . . . 144
8.9 Spectral Interval: RMSECV of ATR-FTIR Single-Beam Spectra . . . 145
8.10 Prediction of LPR from ATR-FTIR Ratio Spectra . . . . . . . . . . . 147
8.11 Prediction of LPR from ATR-FTIR Single-Beam Spectra . . . . . . . 148
9.1 Transmission-FTIR Ratio Spectra of Mixtures: Examples . . . . . . . 153
9.2 Transmission-FTIR Single-Beam Spectra of Mixtures: Examples . . . 154
9.3 PCR of Transmission-FTIR Ratio Spectra: Overview . . . . . . . . . . 155
LIST OF FIGURES xix
9.4 PLSR of Transmission-FTIR Ratio Spectra: Overview . . . . . . . . . 156
9.5 PCR of Transmission-FTIR Single-Beam Spectra: Overview . . . . . . 157
9.6 PLSR of Transmission-FTIR Single-Beam Spectra: Overview . . . . . 158
9.7 Spectral Interval: RMSECV of Transmission-FTIR Ratio Spectra . . . 160
9.8 Spectral Interval: RMSECV of Transmission-FTIR Single-Beam Spectra161
9.9 Prediction of LPR from Transmission-FTIR Ratio Spectra . . . . . . . 163
9.10 Prediction of LPR from Transmission-FTIR Single-Beam Spectra . . . 164
9.11 Analysis of Ratio Spectra of Clinical Samples: Whole Spectrum . . . . 165
9.12 Analysis of Ratio Spectra of Clinical Samples: Reduced Spectrum . . 166
9.13 Analysis of Single-Beam Spectra of Clinical Samples: Whole Spectrum 167
9.14 Analysis of Single-Beam Spectra of Clinical Samples: Reduced Spectrum167

List of Tables
2.1 Reactions of the ISCUS Flex Microdialysis Analyser . . . . . . . . . . 11
2.2 Analytical Performance of ISCUS Flex . . . . . . . . . . . . . . . . . . 12
4.1 Limits of Detection of Glucose, Lactate & Pyruvate . . . . . . . . . . 58
4.2 Prediction of Metabolite Concentrations in Microdialysate . . . . . . . 58
5.1 Perfusion Fluid Components . . . . . . . . . . . . . . . . . . . . . . . . 61
6.1 Unique FTIR Peaks of Glucose, Lactate and Pyruvate #1 . . . . . . . 84
6.2 Unique FTIR Peaks of Glucose, Lactate and Pyruvate #2 . . . . . . 94
6.3 Peak Assignments of FTIR Spectra of Clinical Samples . . . . . . . . 96
6.4 ATR Crystal Study Concentrations . . . . . . . . . . . . . . . . . . . . 97
6.5 Peaks Selected for ATR Crystal Analysis . . . . . . . . . . . . . . . . . 100
6.6 Results of ATR Crystal Analysis . . . . . . . . . . . . . . . . . . . . . 106
6.7 Results of Path Length Analysis . . . . . . . . . . . . . . . . . . . . . 107
8.1 Mixture Study Concentrations: 3-Component Mixtures . . . . . . . . . 133
8.2 Mixture Study Concentrations: 2-Component Mixtures . . . . . . . . . 133
8.3 Mixture Study Concentrations: Pure Solutions . . . . . . . . . . . . . 134
8.4 Mean and Standard Deviation of Mixture Spectra . . . . . . . . . . . 137
8.5 PCR vs. PLSR: ATR-FTIR Ratio Spectra . . . . . . . . . . . . . . . . 140
8.6 PCR vs. PLSR: ATR-FTIR Single-Beam Spectra . . . . . . . . . . . . 141
8.7 Achieved vs. Required Accuracy of PLSR Model . . . . . . . . . . . . 149
9.1 Mixture Study Concentrations: Pure Solutions . . . . . . . . . . . . . 152
9.2 PCR vs. PLSR: Transmission-FTIR Ratio Spectra . . . . . . . . . . . 156
9.3 PCR vs. PLSR: Transmission-FTIR Single-Beam Spectra . . . . . . . 158
9.4 Analysis of Clinical Samples: Model Comparison . . . . . . . . . . . . 168
9.5 Achieved vs. Required Accuracy of PLSR Model . . . . . . . . . . . . 169
9.6 Achieved vs. Required Accuracy of PLSR Model . . . . . . . . . . . . 169
A.1 Three-Component Mixtures . . . . . . . . . . . . . . . . . . . . . . . . 179
A.2 Two-Component Mixtures . . . . . . . . . . . . . . . . . . . . . . . . . 182
xxi

Chapter 1
Thesis Overview
1.1 Motivation & Aims
A Traumatic Brain Injury (TBI) is a highly heterogeneous condition in which brain
tissue is physically damaged as a result of an external mechanical force. Severe TBI
can lead to death or permanent disability and therefore patients with a serious injury
are closely monitored for signs of injury progression.
The concentrations of metabolic intermediates present in the extracellular brain
fluid provide an insight into the metabolic status, and hence the physical health, of
brain tissue. This fluid can be sampled using the technique of microdialysis, in which
a catheter possessing a permeable outer wall is inserted into the brain. Perfusion fluid
(a solution containing a combination of chloride salts) is continuously pumped through
the catheter and the resulting concentration gradient causes the constituents of the
extracellular brain fluid, including the metabolic intermediates, glucose, lactate and
pyruvate, to diffuse into the catheter. The resulting solution, known as microdialysate,
is pumped out of the catheter and collected in a vial for analysis.
This thesis describes an investigation to determine whether vibrational spectroscopy
could be used as the basis for an on-line sensor to monitor the concentrations of glu-
cose, lactate and pyruvate present in the microdialysate as it is pumped out of the
catheter. Two forms of vibrational spectroscopy were studied: Fourier Transform
Infra-Red (FTIR) spectroscopy and Raman spectroscopy.
FTIR spectroscopy involves the absorption of specific wavelengths of Infra-Red
(IR) radiation by a sample. The absorbed wavelengths correspond to the energy
required to increase the vibrational energy of certain bonds within a molecule and
hence can be used to determine which bonds are present.
Raman spectroscopy is also used to determine the nature of chemical bonds present
in a sample but is based upon the scattering, rather than absorption, of photons.
The technique involves the inelastic scattering of photons by a sample, wherein the
difference in energy between the incident and scattered photons corresponds to a
1
2 CHAPTER 1. THESIS OVERVIEW
difference in the vibrational energy of specific bonds within the molecule.
1.2 Experimental Work
During the first phase of this project, synthetic solutions of glucose, lactate and pyru-
vate in perfusion fluid were analysed via both FTIR and Raman spectroscopy.
The FTIR study was used to investigate two forms of FTIR spectroscopy: transmission-
FTIR and Attenuated Total Reflection FTIR (ATR-FTIR). Both techniques were used
to acquire FTIR spectra of pure solutions of glucose, lactate and pyruvate, as well
as mixtures of all three solutions. Spectra acquired using the same technique were
compared in order to identify absorbance peaks unique to each compound. FTIR spec-
tra of additional molecules likely to be present in microdialysate samples were also
acquired, in order to estimate the impact that these compounds might have on the
detection of glucose, lactate and pyruvate by FTIR spectroscopy. Transmission-FTIR
spectroscopy was also used to analyse a small number of clinical samples. Finally, a
quantitative study was carried out to determine which ATR-FTIR and transmission-
FTIR accessories yielded the lowest detection limits for glucose, lactate and pyruvate
and would thus be the most suitable for later, in-depth, studies of each technique.
The Raman study began with attempts to acquire spectra of pure solutions of
glucose, lactate and pyruvate. These results prompted a sub-project to determine
whether Surface-Enhanced Raman Spectrosopy (SERS) was a more viable technique
with which detect these compounds, using cysteine as a model compound.
The second phase of this project entailed systematic studies of the performances
of ATR-FTIR and transmission-FTIR spectroscopy. Both techniques were used to ac-
quire spectra of a series of synthetic mixtures containing glucose, lactate and pyruvate
at a range of clinically-relevant concentrations. The acquired spectra were used to de-
velop a statistical model which was optimised for the prediction of glucose, lactate
and pyruvate concentrations from FTIR spectra of mixtures of all three compounds.
This model was used to analyse the transmission-FTIR spectra of a small number of
clinical microdialysate samples.
1.3 Note on Terminology
Throughout this thesis, the terms ‘Limit of Detection’ (LoD) and ‘Limit of Quan-
tification’ (LoQ) are used to describe aspects of sensor performance. The LoD, as
used in this thesis, describes the smallest concentration of an analyte that produces
a measurable signal. The LoQ is the smallest concentration of an analyte that can
be both detected and differentiated from other concentrations of that analyte, to a
predetermined degree of accuracy.
1.4. GUIDE TO CHAPTERS 3
1.4 Guide to Chapters
The following paragraphs provide the reader with a brief guide to the contents of each
chapter in this thesis. In general, Chapters 2, 3 and 4 cover background information
and theory, whilst Chapters 5 – 10 discuss the methods, results and conclusions of the
work carried out.
Chapter 2 provides the clinical context for the work described in this thesis. An
introduction to the pathophysiology of TBI is given, as well as a description of the
current techniques used for metabolic profiling of patients and a discussion of the roles
played by glucose, lactate and pyruvate in TBI. The project aims are also discussed
in detail.
Chapter 3 describes the theory behind the tools and methods used in this work:
FTIR spectroscopy, Raman spectroscopy and statistical modelling.
Chapter 4 provides a review of the scientific literature, discussing recent advances
made in the research fields most closely related to this work.
Chapter 5 provides details of the reagents, instrumentation, experimental meth-
ods and statistical methods used routinely in this work.
Chapter 6 describes the preliminary studies carried out to assess the potential
utility of FTIR spectroscopy as a tool for metabolic profiling of TBI patients.
Chapter 7 describes the preliminary studies carried out to assess the potential
utility of Raman spectroscopy and SERS as tools for metabolic profiling of TBI pa-
tients.
Chapter 8 describes a detailed systematic study into the use of ATR-FTIR spec-
troscopy, combined with predictive statistical modelling, to predict the concentrations
of glucose, lactate and pyruvate when combined in mixtures at clinically-relevant con-
centrations.
Chapter 9 describes a detailed systematic study into the use of transmission-
FTIR spectroscopy, combined with predictive statistical modelling, to predict the con-
centrations of glucose, lactate and pyruvate when combined in mixtures at clinically-
relevant concentrations. The performance of the predictive model when applied to
clinical microdialysate samples is also discussed.
Chapter 10 discusses the conclusions that can be drawn from this work and its
application to the development of a sensor for metabolic profiling of TBI patients.
Possible next steps and areas of consideration for future sensor development are also
discussed.

Chapter 2
Project Background
The aim of this chapter is to provide context for the work described in this thesis.
It presents a brief introduction to the nature of traumatic brain injury and why it is
necessary to study the metabolic profiles of brain-injured patients, an explanation of
how this is currently achieved in the clinic and a discussion of the disadvantages of
the current system. Finally, it describes the aims of the work undertaken and lists
the desired properties of a sensor for metabolic profiling of traumatic brain injury.
2.1 Introduction to TBI
A Traumatic Brain Injury (TBI) is an injury to brain tissue caused by an external
mechanical force. This force may result from a variety of incidents including: violence,
road-traffic accidents, falls or sporting accidents and, as a result, TBI is a global
health issue which is prevalent in both low- and high-income countries and across all
age groups.
The outcome of TBI varies significantly between patients: some experience only a
mild concussion, from which they make a full recovery, whilst severe TBI can result
in death or permanent disability. In addition to the personal cost to the patient and
their family, severe TBI therefore poses a major socio-economic burden: in Europe,
TBI is responsible for the greatest number of years lived with disability resulting
from trauma and is among the top three causes of injury-related medical costs.1 The
global incidence of TBI has risen sharply in recent decades: in developing countries,
this is predominantly due to an increased use of motor vehicles,1 whilst in developed
countries there is an increasing incidence of TBI in the older population due to falls.2
Within the UK, a recent report by the National Institute for Health and Care Excel-
lence suggested that head injury is the most common type of trauma seen in hospital
emergency departments, accounting for 10% of all emergency department cases each
year.3 Effective management of TBI cases is therefore of critical importance, both
within the UK and worldwide.
5
6 CHAPTER 2. PROJECT BACKGROUND
2.2 Mechanism of TBI
The extent and type of damage to the brain following a TBI is dependent on a num-
ber of factors: the nature of the external force (for example: direct impact, rapid
acceleration or deceleration, penetration by an external object or blast waves from an
explosion) as well as its intensity, direction and duration all affect the presentation of
the injury. TBI is therefore a highly heterogeneous condition encompassing multiple
different types of brain damage, some or all of which may be present to varying de-
grees in a given individual. This heterogeneity is considered one of the most significant
barriers to finding effective therapeutic interventions for treating TBI.4
The progression of a TBI can be divided into two phases: the primary injury, which
occurs within minutes or hours of the traumatic event, and the secondary injury, which
evolves in the following hours and days.
Damage sustained by the brain during the primary injury can include: damage
to blood vessels, bruising of brain tissue and the shearing of bundles of axons that
connect different regions of the brain. The secondary injury is caused when these
early neurotraumatic events trigger additional pathophysiological mechanisms which
cause further damage to the brain. In many cases, these secondary mechanisms are
induced concurrently and with synergistic effect, thus increasing the damage caused.
One example of such a mechanism is the release of excessive amounts of excitatory
neurotransmitters, such as glutamate, into the interstitial fluid, which exacerbates
brain swelling, raises the intracranial pressure (ICP) and is ultimately a major cause
of early necrotic cell death. The extent of the secondary injury is determined by
the nature, severity and duration of these secondary mechanisms, combined with the
intensity of any systemic insults to the rest of the body.
2.3 Monitoring TBI
Secondary brain injury is the leading cause of in-hospital deaths that occur following a
TBI.5 It is therefore essential that patients are closely monitored for secondary brain
damage, which may not exhibit any obvious clinical symptoms, so that it can be
detected at the earliest possible stage and its progression studied. These observations
can then be used to inform treatment decisions, which may lead to an improved
outcome for the patient.
In severely brain-injured patients, the condition of the brain is monitored directly
via the insertion of up to three different probes into the brain tissue itself. The method
by which this is achieved and the region of the brain in which the probes are placed will
vary between patients, as these parameters are determined by institutional protocol
and the nature of the injury, respectively. The three properties monitored using these
probes are: ICP, cerebral tissue oxygenation and the concentrations of metabolites in
2.3. MONITORING TBI 7
the extracellular fluid.
Measurement of ICP, the pressure within the skull, provides information about the
blood flow to the brain, whilst measurement of cerebral tissue oxygenation provides
information about the oxygen saturation of the brain tissue itself. These parameters
are monitored in order to ensure that there is adequate blood flow to the brain and
that sufficient oxygen and glucose are delivered to ensure the preservation of any brain
tissue that has not been irreparably damaged by the initial physical insult. Monitoring
the concentration of brain metabolites provides information on the metabolic processes
occurring within the brain. It is known that the cellular damage caused by TBI can
induce changes in energy metabolism within the brain, via a number of different
pathways, which can ultimately lead to energy failure and the subsequent activation
of apoptotic and necrotic pathways.6 Knowledge of ICP, cerebral tissue oxygenation
and metabolic perturbations with the brain has significant implications for how TBI
patients are managed.
Patients may be monitored in this way for 3 – 5 days and, whilst the cerebral tissue
oxygenation and ICP probes allow for continuous automated measurement of these
parameters, the monitoring of extracellular brain chemistry is labour-intensive and
requires regular intervention by clinical staff. These factors are thought to limit the
number of neurocritical care centres able to study the extracellular brain chemistry
of brain-injured patients and therefore there is a clear clinical need to find a more
straightforward method of doing so, which this thesis aims to address. In order to put
this work in context, the following sections will discuss in more detail the method by
which extracellular brain chemistry is currently monitored.
2.3.1 Microdialysis in TBI
The probe used to monitor extracellular brain chemistry is a fine catheter consisting
of two concentric tubes capped with a gold end, which allows it to be visualised
via neuroimaging techniques. The outer tube is connected to a syringe pump that
delivers a solution of perfusion fluid at a constant flow rate of 0.3 µL min−1. The
lower portion of the wall of the outer tube is composed of a semi-permeable membrane,
which typically has a 100 kDa molecular weight cut-off, that allows the bi-directional
diffusion of molecules between the catheter and the surrounding extracellular fluid.
The continuous flow of perfusion fluid ensures that equilibrium across this membrane is
never reached, thus creating a permanent concentration gradient across the membrane
which facilitates the movement of molecules from the extracellular brain fluid into the
catheter. The tip of the catheter is sealed and therefore, upon reaching the end of
the outer tube, the solution is directed upwards through the inner tube where it flows
out of the catheter and into an external collection vial. This vial is typically changed
hourly and its contents, known as microdialysate, are currently analysed via enzyme-
8 CHAPTER 2. PROJECT BACKGROUND
catalysed colorimetric assays using a commercially-available analyser (Section 2.3.3).
This process of extracting the compounds present in extracellular brain fluid, known
as microdialysis, is summarised in Figure 2.1.
Figure 2.1: Illustration of the lower portion of a microdialysis catheter, showing the inner
and outer tubes and the gold-tipped closed end. The walls of the outer tube are made of
a semi-permeable membrane, allowing the exchange of substances between the catheter and
surrounding tissue (grey/red arrows). Perfusion fluid is pumped into the outer tube at a rate
of 0.3 µL/min (blue arrows), where it mixes with molecules that have entered the catheter
from the surrounding tissue. This mixture flows out of the catheter via the inner tube (red and
blue striped arrows) and is collected in a vial for analysis.
Microdialysis does not recover 100% of molecules present in the interstitial brain
fluid and therefore the measured concentration of a substance in the microdialysate is
not a measure of the true concentration of that substance in the extracellular space.
However, the measured concentration can be related to the true concentration by
the ‘relative mean recovery’ value: the microdialysate concentration expressed as
a percentage of the true interstitial concentration. The relative mean recovery is
affected by two main factors: the location of the catheter itself and the properties of
the perfusion fluid used.
The properties of the tissue into which the catheter is placed can affect the relative
mean recovery value by altering the amount of a substance available for collection via
microdialysis. For example, placement of the catheter into tissue which contains
leaking blood vessels, or which is otherwise damaged, reduces the availability of a
given substance near the catheter and therefore decreases the relative mean recovery
value.
Two properties of the perfusion fluid can affect the relative mean recovery: its flow
rate and its components. The flow rate of perfusion fluid through the catheter affects
the relative mean recovery as it determines the amount of time available for molecules
to diffuse across the catheter membrane: the slower the flow rate, the more time is
available for diffusion and the higher the relative mean recovery. The commonly-used
2.3. MONITORING TBI 9
flow rate of 0.3 µL min−1, in combination with a standard 10 mm length membrane,
generates an estimated mean recovery of 65−72% for the molecules that are typically
monitored in a clinical setting (see Section 2.3.2).
The constituents of the perfusion fluid affect the concentration gradients across
the catheter membrane and therefore have an effect on not only the relative mean
recovery but also the microenvironment of the extracellular fluid surrounding the
catheter. The perfusion fluid routinely used for brain microdialysis in the clinic is an
un-buffered solution containing salts of sodium, potassium, calcium and magnesium at
concentrations similar to those found in the interstitial fluid: 147 mM NaCl, 2.7 mM
KCl, 1.2 mM CaCl2 and 0.85 mM MgCl2. Sodium, potassium and calcium ions play
a major role in cell signalling and their inclusion in perfusion fluid at physiological
pH is intended to minimise their movement across the microdialysis membrane. Sim-
ilarly, magnesium ions play an important role in the functioning of certain glutamate
receptors, which are involved in the control of memory function and the modification
of synapse strength, and therefore they are included in perfusion fluid to ensure that
these receptors do not become activated as a result of magnesium depletion. In ad-
dition, the solution is un-buffered to prevent any significant shift of hydrogen ions
across the microdialysis membrane.7
Due to the variability in relative mean recovery values, the concentrations of brain
metabolites are typically reported as the values measured within the microdialysate,
without being adjusted to predict the true concentration in the interstitial fluid. Al-
though absolute measured concentrations of metabolites are used in the clinic, the
reporting of ratios of concentrations is preferred as these are independent of changes
in relative mean recovery and depend less on the properties of the tissue being mon-
itored. In addition, details of the catheter location, type and membrane length, as
well as the flow rate and constituents of the perfusion fluid, should ideally be reported
alongside any measured metabolite concentration.7
2.3.2 Compounds Recovered using Microdialysis
The membrane of the catheter used for microdialysis is selective only for size, as
mentioned in Section 2.3.1, and therefore any compound smaller than 100 kDa can
be collected, providing the appropriate concentration gradient is present. A number
of compounds meeting these criteria are present in the extracellular brain fluid of
TBI patients, in addition to metabolic intermediates. These include: drug molecules
such as anti-seizure medication, which is administered as part of the TBI treatment
programme for certain sub-types of brain injury;1 cytokines, which mediate the in-
flammatory response to injury;8 nitric oxide products, which are thought to play a role
in regulating blood flow;9 glutamate, a neurotransmitter, and glycerol, a component
of cell membranes.
10 CHAPTER 2. PROJECT BACKGROUND
In order for a compound to be selected for routine monitoring as part of a TBI
treatment programme, it must meet two main criteria. Firstly, there must exist a
detection method for that compound which is sufficiently selective and sensitive to
determine its concentration within the dialysate mixture at clinically-relevant levels.
In addition, the clinical significance of that compound must be understood, within
the context of TBI, in order that the measured concentration can be translated into
clinically-useful information regarding the status of the patient and potentially used
to initiate a clinical intervention.
Current guidelines, as set out in the Consensus Statement from the 2014 Interna-
tional Microdialysis Forum, suggest that the most important compounds to monitor
in TBI patients are the metabolic intermediates D-glucose, L-lactate and pyruvate.
Glycerol and glutamate may also be monitored, but their relationship to patient out-
come is less well understood.7
N.B: The terms ‘glucose’ and ‘lactate’ are henceforth used to refer to D-glucose
and L-Lactate respectively.
2.3.3 ISCUS Flex Microdialysis Analyser
The instrument typically used to analyse microdialysate samples is the ISCUS Flex
Microdialysis Analyser. This instrument is designed specifically for the analysis of
microdialysate samples and is optimised for use with small sample volumes of 0.2 -
1.0 µL. It can be used to measure the concentrations of glucose, lactate and pyruvate,
in addition to other small organic molecules such as urea, glycerol and glutamate.
The analysis method employed by the ISCUS Flex uses enzyme-mediated reac-
tions in combination with absorbance spectroscopy. Briefly, an enzyme that reacts
selectively with the analyte of interest is added to the microdialysate sample, along
with any other necessary reagents, and catalyses the conversion of the analyte into an-
other compound. This intermediate compound is typically used in at least one further
reaction, the product of which will be highly coloured, allowing it to be detected via
absorbance spectroscopy. The rate of formation of the coloured compound is directly
proportional to the concentration of the original analyte and therefore calculation of
the concentration of the coloured product, via absorbance spectroscopy, allows the
concentration of the original analyte to be determined. The enzymes and reactions
used by the ISCUS Flex to detect glucose, lactate and pyruvate are summarised in
Table 2.1.10
The LoD and LoQ of the ISCUS Flex, as determined in a 2011 study aimed at
validating its use in the analysis of microdialysate, are listed in Table 2.2. Despite
these encouraging results, the use of the ISCUS Flex poses a number of issues which
may be limiting the number of neurocritical care centres able to include metabolite
profiling as part of their TBI treatment programme. For example, the analyser is a
2.4. METABOLISM IN TBI 11
Table 2.1: Summary of the multi-step reactions reactions used by the ISCUS Flex Micro-
dialysis Analyser to generate coloured compounds which can be detected via absorbance spec-
troscopy and used to calculate the concentrations of D-glucose, L-lactate and pyruvate in a
sample of microdialysate.10 Pi = inorganic phosphate, TOOS = 3-(ethyl(m-tolyl)amino)-2-
hydroxypropane-1-sulfonic acid.
Analyte Reaction Enzyme
D-glucose
Step
1
D-glucose + O2 → gluconolactone + H2O2 GlucoseOxidase
Step
2
2H2O2 + phenol + 4-aminoantipyrine →
quinoneimine + 4H2O
Peroxidase
L-lactate
Step
1
L-lactate + O2 → pyruvate + H2O2 LactateOxidase
Step
2
H2O2 + 4-chlorophenol + 4-aminoantipyrine →
quinoneimine + 2H2O + HCl
Peroxidase
Pyruvate
Step
1
pyruvate + Pi + O2 → acetylphosphate + CO2 +
H2O2
Pyruvate
Oxidase
Step
2
2H2O2 +TOOS + 4-aminoantipyrine →
quinonediimine + 4H2O
Peroxidase
separate entity to the microdialysis setup and therefore samples must be manually
loaded into the instrument; a process which is both time- and labour-intensive. This
fact causes obvious logistical issues and means that samples are typically collected and
analysed of the order of once per hour. The resulting delay between the extraction
of metabolites from the brain and their analysis by the ISCUS Flex has a number
of clinical implications. Firstly, the lack of real-time data on brain-metabolite levels
means that any clinical intervention which may be recommended on the basis of this
information would also be delayed. An additional consequence of the delay is that the
microdialysate, and therefore its constituent metabolites, is collected in a single vial
and therefore the measured metabolite concentration is an average of that produced by
the brain over the course of the collection period. It is therefore difficult to identify
time-dependent trends in metabolite production or consumption, which may cause
further delays in the delivery of treatment or prognosis.
2.4 Metabolism in TBI
Glucose, lactate and pyruvate are key compounds in the process of cellular respira-
tion: the process by which glucose is metabolised to produce adenosine triphosphate
12 CHAPTER 2. PROJECT BACKGROUND
Table 2.2: A summary of the LoD and LoQ of the ISCUS Flex in the
measurement of glucose, lactate and pyruvate.11
Compound
Limit of
Detection/mM
Limit of
Quantification / mM
Glucose 0.02 0.06
Lactate 0.06 0.18
Pyruvate 0.0078 0.0191
(ATP), the source of chemical potential energy used by all cells for intracellular en-
ergy transfer. Within the brain, ATP is necessary for a variety of functions, including
the maintenance of cells and neurotransmitter production.12 This section provides a
brief overview of the roles of glucose, lactate and pyruvate in cellular respiration and
discusses how their concentrations in the brain may change following a TBI.
2.4.1 Cellular Respiration
The key reactions of cellular respiration are summarised in Figure 2.2.
Figure 2.2: Schematic highlighting the roles of glucose, lactate and pyruvate within the major
reactions of cellular respiration.
The first stage of cellular respiration is glycolysis; a series of enzyme-catalysed
reactions through which glucose is converted to pyruvate. This process produces a
net yield of two molecules of ATP and two molecules of pyruvate for each molecule of
glucose consumed.
2.4. METABOLISM IN TBI 13
The pyruvate produced via glycolyis is subsequently used as a substrate for one
of two different reaction pathways, depending upon the availability of oxygen. If
oxygen is present, pyruvate is used as the initial substrate for the reactions of aerobic
respiration: the Tricarboxylic Acid (TCA) cycle and the electron-transport system.
These processes can produce a net yield of up to 36 molecules of ATP for each molecule
of glucose originally metabolised to pyruvate.
In the absence of oxygen, pyruvate is converted to lactate via the anaerobic-
respiration pathway. This process is significantly less efficient than aerobic respiration,
yielding no further molecules of ATP.
The significance of glucose, lactate and pyruvate in cellular respiration allows
valuable information regarding the metabolic and redox states of the brain to be
inferred from the measured concentrations of these compounds within the extracellular
brain fluid. Clinicians are subsequently able to use this information to assess which
pathophysiological processes may be occurring within the brain, including: hypoxia,
ischemia and mitochondrial dysfunction. The following sections discuss these insights
in more detail.
2.4.2 Glucose in TBI
Brain-glucose levels are dictated by a balance between glucose supply and demand.
The current best estimate of glucose levels in the uninjured human brain, measured
during a microdialysis study of patients undergoing neurosurgery for benign lesions,
gives a mean glucose concentration of 1.7 ±0.9 mM.13 This study used a flow rate
of 0.3 µL min−1 and a 10 mm microdialysis membrane. Glucose levels in the brains
of TBI patients tend to deviate significantly from this healthy range as changes in
blood flow or metabolism, both of which are commonly seen in TBI, can have a
disproportionate effect on brain glucose.14
Low brain-glucose levels (defined as concentrations below 0.8 mM in microdialysate)
are commonly observed in TBI and can be caused by a decrease in glucose supply or
an increase in glucose demand. A decrease in the amount of glucose being delivered
to the brain is due to a deficiency in blood supply which, in turn, may be a result of
local capillary damage, elevated intracranial pressue, or failure of blood-flow autoreg-
ulation. In healthy individuals, the supply of glucose to the brain is also affected by
blood-glucose concentration, but this relationship may be lost in the injured brain.7
Alternatively, if an increase in glucose demand causes a sufficiently high rate of cellular
uptake, the rate of glucose supply is not enough to meet the demand and glucose levels
become depleted.6 Such an increase in glucose demand may result from a number of
factors, including: a requirement for anaerobic metabolism;5 a greater need for ATP
in order to restore ionic and neurochemical gradients which have been disturbed; or
an increased use of glucose for non-energy pathways, such as those required for repair
14 CHAPTER 2. PROJECT BACKGROUND
and biosynthesis.15 In addition to the causes discussed here, brain-glucose levels can
also be rapidly reduced by certain types of secondary injury, although the mechanisms
of these processes are beyond the scope of this work.7
Clinical observations have linked levels of brain glucose below 0.8 mM with an
unfavourable outcome from TBI. Additional evidence suggests that particularly high
brain-glucose levels may also be associated with unfavourable outcome, indicating that
there may be an optimal range within which brain-glucose levels should be maintained,
but there are currently insufficient data to define these parameters. The consensus
statement from the 2014 International Microdialysis Forum therefore proposes patho-
logical thresholds for low glucose levels only: brain tissue is considered to be ‘at risk’
if the glucose concentration is below 0.8 mM and intervention is recommended for
glucose concentrations below 0.2 mM.7
2.4.3 Lactate and Pyruvate in TBI
As discussed in Section 2.4.1, lactate and pyruvate are derived from glucose during
the process of cellular respiration. Their extracellular concentrations therefore partly
reflect the glycolytic activity, and relative levels of aerobic vs anaerobic respiration,
within a cell.16
Analysis of microdialysate derived from patients with no brain injury gave a mean
lactate concentration of 2.9 ±0.9 mM.13 Lactate levels in the microdialysate of TBI
patients are often higher than this and concentrations higher than 4 mM are gener-
ally linked to unfavourable patient outcomes.7 It is thought that raised lactate levels
indicate increased use of the anaerobic respiration pathway. However, the absolute
lactate concentration within the extracellular fluid is also influenced by the amount of
glucose that enters the TCA cycle and is therefore not in itself indicative of anaerobic
metabolism.17 As a result, the absolute lactate concentration found in microdialysate
samples is not usually used as a clinical marker in TBI.
The mean concentration of pyruvate in microdialysate from uninjured brain was
found to be 0.166 ± 0.047 mM.13 There is evidence to suggest that a high pyruvate
concentration is linked to favourable outcomes from TBI and yet, as with lactate,
pyruvate concentration alone is not usually used as a clinical marker.16
The relative proportion of lactate and pyruvate, known as the Lactate:Pyruvate
Ratio (LPR), is used as a marker of aerobic vs. anaerobic metabolism within the brain.
It is considered a more reliable indicator of patient outcome than the absolute con-
centrations of lactate and pyruvate, with a LPR value greater than 25 generally being
associated with unfavourable outcome.16,18 Knowledge of the absolute concentrations
of lactate and pyruvate is still of use, however, as they can be used to interpret the
causes of a high LPR. An increased LPR, in combination with low pyruvate levels
(and low oxygen, if being monitored) indicates a shift to anaerobic respiration due
2.5. PROJECT AIMS 15
to insufficient blood flow to the region of the brain being monitored. On the other
hand, an increased LPR with high lactate and high pyruvate (and normal oxygen, if
being monitored) indicates an increase in anaerobic respiration due to mitochondrial
dysfunction.7
Interpretation of the LPR is complicated by increasing evidence that the injured
brain is able to utilise lactate to produce ATP via the TCA cycle.19 It has been sug-
gested that the improved patient outcome associated with low extracellular lactate
could be due to lactate being taken up by neurons and utilised via the TCA cycle.
When neurons are too damaged to utilise this lactate, it accumulates in the extracel-
lular fluid, thus explaining the unfavourable outcome associated with high levels of
lactate.20
2.5 Project Aims
This work was completed as part of a wider collaborative project between two de-
partments at the University of Cambridge: the Department of Chemistry and the
Department of Clinical Neurosciences. The aim of the wider project is to develop an
alternative to the ISCUS Flex, for the analysis of glucose, lactate and pyruvate levels
in the microdialysate of TBI patients, which is capable of delivering a comparable
level of accuracy whilst improving upon its shortcomings.
It is envisioned that the developed sensor will take the form of a ‘lab-on-a-chip’
system, in which sample collection and analyte detection are both integrated and semi-
continuous. The proposed ‘sensing chip’ will be compatible with existing microdialysis
equipment used in the clinic, as well as being small enough to sit at the patient’s
bedside, such that microdialysate samples leaving the brain will flow directly into the
chip for analysis before being passed out the other side of the chip to be collected and
disposed of as clinical waste or retained for further analysis. The proposed setup is
illustrated in Figure 2.3.
Such a system is predicted to bring multiple advantages to the treatment of TBI.
By replacing the current hourly analysis with a continuous monitoring system, clin-
icians would have earlier warning of changes in a patient’s neurochemistry which, it
is hoped, would allow earlier intervention and lead to a more favourable outcome for
the patient. In addition, the setup would be more user-friendly than the existing
technology and would require less intervention from clinical staff, due to the direct
integration of the sampling and analysis components, making it both cost- and time-
effective when compared with the existing technology. It is hoped that these factors
would together incentivise and enable more neurocritical care units to adopt micro-
dialysis as a technique.
The aim of the work described in this thesis was to determine whether vibrational
spectroscopy would be a suitable technique to form the basis of the developed sen-
16 CHAPTER 2. PROJECT BACKGROUND
F
ig
u
re
2
.3
:
A
n
illu
stra
tio
n
o
f
th
e
a
im
o
f
th
e
D
epa
rtm
en
t
o
f
C
h
em
istry-D
epa
rtm
en
t
o
f
C
lin
ica
l
N
eu
ro
scien
ces
co
lla
bo
ra
tio
n
;
sh
o
w
in
g
th
e
p
ro
po
sed
m
eta
bo
lic
p
ro
fi
lin
g
setu
p
a
n
d
its
in
tegra
tio
n
in
to
cu
rren
tly-u
sed
T
B
I
m
o
n
ito
rin
g
system
s.
F
igu
re
a
d
a
p
ted
fro
m
a
n
illu
stra
tio
n
by
S
u
sa
n
G
io
rgi-C
o
ll.
Illu
stra
tio
n
co
p
yrigh
t
S
u
sa
n
G
io
rgi-C
o
ll
a
n
d
rep
rod
u
ced
h
ere
w
ith
h
er
perm
issio
n
.
2.6. DEVICE REQUIREMENTS 17
sor. Spectroscopy has the potential to operate via a lab-on-a-chip system, with the
additional advantage that it eliminates the need for consumables and reagents. This
would further decrease the amount of labour required to operate the sensor and has
the added benefit of reducing the device cost to a single initial purchase, with minimal
ongoing expenditure. The use of spectroscopy would also allow glucose, lactate and
pyruvate to be detected directly rather than relying upon chemical reactions and/or
the production of intermediates, as in the case of enzyme-based detection, which may
increase the reliability of the analysis. Vibrational spectroscopy in particular was
identified as a promising candidate due to its sensitivity and potential instrumental
simplicity, which provides greater potential for miniaturisation into a bedside device,
when compared to other forms of spectroscopy.
Two main factors were to be considered for each form of vibrational spectroscopy
studied: sensing performance (sensitivity and accuracy in detecting glucose, lactate
and pyruvate) and operation (suitability for the specific clinical environment in which
it would be used). Details of the specifications required of the final device are discussed
in Section 2.6.
2.6 Device Requirements
In order to fulfil the aim of developing a viable alternative to the ISCUS Flex, the
developed sensor must meet specific criteria. These were taken into consideration
throughout this work and are detailed below:
ˆ Performance requirements:
– Sensitivity - the ability to detect pyruvate, lactate and glucose within a
clinically relevant range:
* Pyruvate: 0.01 - 1.5 mM
* Lactate: 0.1 - 24 mM
* Glucose: 0.1 - 24 mM
– Selectivity: the ability to measure accurately the individual concentrations
of pyruvate, lactate and glucose within the microdialysate mixture
ˆ Operational requirements:
– The potential for miniaturisation, to allow the instrument to sit at the
patient’s bedside
– The ability to measure liquid flowing at 0.3 µL min−1
– The capacity for continuous measurement, defined here as taking a mea-
surement at least once every five minutes
18 CHAPTER 2. PROJECT BACKGROUND
– The ability to measure multiple analytes simultaneously
– User-friendliness: the device must be straightforward to set up for each
new patient and require minimal intervention from clinical staff once it is
operational
Chapter 3
Theoretical Background
This chapter provides the theoretical background behind the tools and techniques used
in this work. It details the theory and operating principles behind both infra-red and
Raman spectroscopy, provides an introduction to statistical modelling and describes
the methods used to analyse and model spectroscopic data.
3.1 Molecular Vibrations
Molecular vibrations are motions of atoms in a molecule which alter the distance
between two atoms (a stretch); the angle between two atoms (a bend); or the angle
between projections of two bonds onto a plane orthogonal to a common bond (a
torsional/twisting motion) within the molecule. The number of vibrational modes of
a molecule is dictated by the number of internal degrees of freedom it possesses.
Each atom within a molecule has three degrees of freedom of motion (the x, y
and z directions) and therefore a molecule of N atoms has 3N internal degrees of
freedom. Three of these are translational motions in the x, y or z directions and
rotational motion accounts for a further two (linear molecules) or three (non-linear
molecules) degrees of freedom. The remaining degrees of freedom are vibrational
motions. A linear molecule therefore possesses 3N − 5 vibrational degrees of freedom,
whilst a non-linear molecule possesses 3N−6 vibrational degrees of freedom. Examples
of vibrational motion include symmetric and antisymmetric stretching modes and
scissoring, rocking, wagging or twisting bending modes.
3.1.1 Classical Harmonic Oscillator
The simplest model for describing a molecular vibration is the simple harmonic os-
cillator, in which a diatomic molecule is considered as two point masses connected
by a massless spring. The stretching/compressive dispacement from the equilibrium
position of the two masses varies periodically with time as a sine or cosine function,
19
20 CHAPTER 3. THEORETICAL BACKGROUND
whilst the molecular centre of mass remains stationary. Although different masses
oscillate along the axis with different amplitudes, they do so with the same frequency
and therefore go through their equilibrium positions simultaneously. The potential
energy of such a molecule is given by:
PE =
1
2
kX2 (3.1)
where X is the distance between the two point masses and k is the force constant. The
latter is a function of the energy of a bond between two nuclei and is related to bond
strength: the stronger the bond, the higher the value of k. A plot of the potential
energy as a function of X gives a parabola that is symmetric about the equilibrium
internuclear distance (Figure 3.1(a)).
Figure 3.1: Illustration summarising the potential-energy curves (dark blue) and vibrational
energy levels (light blue) of the (an)harmonic oscillator models. (a) The potential energy of
a simple harmonic oscillator as a function of distance. (b) The potential energy curve and
quantised vibrational energy levels of a diatomic molecule modelled as a quantum-mechanical
harmonic oscillator, showing the equal spacing of the vibrational energy levels. (c) The po-
tential energy curve and vibrational energy levels of a diatomic molecule modelled as an an-
harmonic oscillator, showing the convergence of the quantised vibrational energy levels with
increasing vibrational quantum number.
The vibrational frequency (ν), in wavenumbers, of a diatomic molecule modelled
by the harmonic oscillator is given by:
ν˜ =
1
2pic
√
k
µ
(3.2)
where c is the speed of light and µ is the reduced mass of the molecule. The reduced
mass allows both nuclei to be represented by a single mass which reflects the relative
3.1. MOLECULAR VIBRATIONS 21
masses of each nucleus. It is calculated using the equation:
1
µ
=
1
m1
+
1
m2
(3.3)
where m1 and m2 are the point masses of the two nuclei.
3.1.2 Quantum-Mechanical Harmonic Oscillator
A quantum-mechanical treatment of the simple harmonic oscillator dictates that the
vibrational energy of a diatomic molecule is quantised. The allowed values of the
vibrational energy are given by:
Ev = (v +
1
2
)hν (3.4)
In this expression, h is Planck’s constant, ν is the classical vibrational frequency of
the oscillator (given in Equation 3.2) and v is a quantum number that may only take
integer values. When v = 0, the molecule retains a vibrational energy of 12hν which
corresponds to the fundamental vibrational energy of the system, known as the zero-
point energy, and cannot be removed from the molecule. Solving this equation for a
series of consecutive values of v gives solutions which are separated by hν, showing
that the vibrational energy levels are equally spaced (Figure 3.1(b)).
Vibrational spectroscopy probes transitions between these vibrational energy lev-
els.
3.1.3 Anharmonic Oscillator
The anharmonic oscillator model reflects the fact that, if the internuclear distance
between two bonded nuclei becomes too great, the bond will eventually break. The
Morse expression for the allowed energy levels in the anharmonic oscillator model is:
Ev = (v +
1
2
)hν − (v + 1
2
)2hνχ (3.5)
where χ is the anharmonicity constant. The solutions to this equation show that con-
secutive vibrational energy levels in the anharmonic oscillator become closer together,
as the value of v increases, and eventually converge at the dissociation limit. The
energy difference between the zero-point energy and the convergence limit is the dis-
sociation energy of the bond. Figure 3.1(c) shows the potential energy and vibrational
energy levels of a bond described by the anharmonic oscillator model.
It can be seen from Figure 3.1 that the simple harmonic and anharmonic oscillators
are similar for low vibrational energy levels. As a result, whilst the anharmonic
oscillator is a significantly more accurate model of molecular vibrations at energy
levels close to the dissociation limit, the simple harmonic oscillator may be used as a
22 CHAPTER 3. THEORETICAL BACKGROUND
satisfactory model of low-energy molecular vibrations.
3.2 Infra-Red Spectroscopy
IR spectroscopy is a technique which probes the vibrational energy levels of a molecule
via the absorption of IR radiation.
The energy required to promote a molecule from one vibrational energy level to
a higher energy level of the quantum-mechanical harmonic oscillator corresponds to
wavelengths of light in the IR region of the electromagnetic spectrum. It can be
seen from Equations 3.4 and 3.2 that the energy of a vibrational level is dictated
by the mass of the nuclei within the molecule and the strength of the bond between
them. Bonds of different type (such as single vs. double bonds) or those involving
different types of atoms will therefore have different vibrational energy levels which
will correspond to different wavelengths of IR radiation.
IR spectroscopy takes advantage of these phenomena to enable the identification
of the functional groups present in a sample. In a typical IR experiment, a sample
is irradiated with a polychromatic IR source. The sample absorbs specific wave-
lengths from the incident radiation, corresponding to the vibrational transitions of
the bonds present in the sample, whilst the remaining wavelengths are transmitted
directly through the sample. The resulting IR spectrum shows which wavelengths of
light have been absorbed by the sample and these can then be related to the functional
groups present.
3.2.1 Selection Rules of IR Spectroscopy
According to the quantum-mechanical oscillator model of molecular vibrations and
IR absorption, there are are two selection rules which must be obeyed in order for
incident IR radiation to cause a vibrational transition in a molecule and thus show a
peak in an IR spectrum.
The first selection rule states that a molecular vibration must change the dipole
moment of the molecule in order for it to be IR-active. The larger the change in the
dipole moment, the stronger the intensity of the band in the resulting IR spectrum.
The static dipole moment, µ, is given by:
µ = qx (3.6)
where ±q is the magnitude of the charge and x is the distance between the charges.
This selection rule places limits on the types of molecule and vibrational motion
that can be detected using IR spectroscopy. For example, homonuclear diatomic
molecules cannot be detected via IR spectroscopy as they have a dipole moment of zero
and this does not change with vibrational motion. However, a polyatomic molecule
3.2. INFRA-RED SPECTROSCOPY 23
with no overall dipole moment may possess some IR-active vibrational modes. One
example of such a molecule is the linear CO2 molecule, which is composed of two
symmetrical polar bonds and therefore has an overall equilibrium dipole moment of
zero. During the symmetric stretch of CO2, during which the two C=O bonds stretch
and compress simultaneously, the dipole moment of each bond changes by the same
magnitude and therefore the overall molecular dipole moment does not change. The
antisymmetric stretch of the molecule causes one C=O bond to be compressed while
the other is stretched and, as a result, the dipole moments of each bond change by
different amounts and generate an overall dynamic dipole moment of the molecule.
The antisymmetric stretch of CO is therefore IR-active, whilst the symmetric stretch
is IR-inactive.
The second selection rule restricts the allowed change in vibrational energy level
to ∆v = ±1 in the simple harmonic approximation. The distribution of molecules
across the differing energy levels can be predicted using the Boltzmann distribution:
Ni
Nj
= exp(− ∆E
kBT
) (3.7)
where NiNj is the ratio of the number of molecules in two different energy levels,
Ni and Nj , ∆E is the difference in energy between the levels, kB is the Bolzmann
constant and T is the temperature. The Boltzmann distribution predicts that, at 298
K, the majority of molecules are in their vibrational ground state because ∆E is much
larger than kBT . Consequently, the vibrational transition most often studied by IR
absorption spectroscopy is that from v = 0 to v = 1.
3.2.2 IR Spectroscopy Instrumentation
Routine mid-IR spectroscopy measurements, using light of 2.5 – 10 µm in wavelength,
are typically carried out using a FTIR spectrometer.
The principle behind a generalised FTIR spectrometer is shown in Figure 3.2. An
IR light source emits electromagnetic radiation which is split into two beams by a
semi-transparent beam-splitter. One beam is directed towards a stationary mirror,
where it is reflected back towards the beam-splitter, whilst the other is reflected by
a moving mirror. The motion of the moving mirror produces a difference in path
length between the two beams and therefore upon recombination there is constructive
and destructive interference between them, which generates an interferogram. The
recombined beam is directed towards the sample.
The sample absorbs certain wavelengths of the recombined beam, as discussed in
Section 3.2, thus removing them from the interferogram. The remaining wavelengths
of light are transmitted to a detector which converts the optical signal into a measur-
able electrical signal, such as voltage. The variation in intensity as a function of time
24 CHAPTER 3. THEORETICAL BACKGROUND
Figure 3.2: An illustration of the optical path of a typical FTIR spectrometer. Light emitted
by an IR source is split by a beam-splitter and the resulting two beams are reflected by different
mirrors, one of which is moving. Upon recombination, the beams form an interferogram. The
recombined beam is passed through the sample, which absorbs certain wavelengths, and the
remaining wavelengths are directed towards a detector.
is thus recorded for each wavelength of light that reaches the detector.
A computer connected to the instrument performs a Fourier transform for each
point of the interferogram to allow the intensity to be expressed as a function of
frequency, making it possible to see which frequencies are missing from the interfero-
gram due to their absorption by the sample. These data are typically plotted to show
intensity vs. inverse wavelength (in cm−1) to produce a conventional IR spectrum.
3.2.3 Types of IR Spectroscopy
The basic principle of IR spectroscopy has been developed into multiple different
techniques. This section provides details of the two techniques utilised in this work.
3.2.3.1 Transmission-FTIR Spectroscopy
Transmission-FTIR is the most simple form of IR spectroscopy. In transmission-FTIR
experiments, the sample is held between two windows, which must be composed of
an IR-transparent material, and placed directly in the light path of the spectrometer.
Incident light enters the cell through one window, passes through the sample and exits
3.2. INFRA-RED SPECTROSCOPY 25
the cell via the second window.
Two main experimental parameters must be considered when planning a transmission-
FTIR experiment. The first of these parameters is the choice of material for the sample
cell window: the material must be transparent to IR radiation over the wavelength
range of interest and must also be inert to the consituents of the sample. The second
experimental parameter to consider is the distance between the two windows, which
dictates the path length of the sample. Longer path lengths increase the amount
of interaction between the IR beam and the sample and therefore lead to increased
absorption and stronger peaks in the IR spectrum.
The major advantage of transmission-FTIR is that it is a simple and well-established
technique. However, the long path lengths used make it unsuitable for the study of
strongly IR-absorbing samples, such as those with a high water concentration, as the
signal intensity produced by such molecules can saturate the detector and thus the
spectrum. In addition, the study of solid samples by transmission-FTIR can be com-
plicated, as the sample must be ground up and prepared as a suspension in oil in
order to allow transmission of the IR beam.
3.2.3.2 Attenuated Total Reflectance FTIR Spectroscopy
ATR-FTIR spectroscopy is based upon the principle of total internal reflection. This
phenomenon occurs when a propagating wave travels from a material of a higher
refractive index to a material of a lower refractive index. If the wave strikes the
boundary between the two materials at an incident angle which is greater than a
critical angle, which is measured with respect to the normal to the surface, it will
be totally internally reflected within the first material. The critical angle, θc, is a
function of the refractive indices of the two materials and is given by:
θc = sin
−1n1
n2
(3.8)
where n1 and n2 are the refractive indices of the first and second materials, respec-
tively.
ATR-FTIR spectroscopy uses a crystal of high refractive index (typically diamond,
Ge, Si or ZnSe) which is placed in direct contact with the sample. Total internal
reflection of the incident IR beam propagating in the ATR crystal occurs at the
crystal-sample interface, generating an evanescent wave outside the crystal which
extends into the sample (Figure 3.3). The evanescent wave will be attenuated at the
specific wavelengths absorbed by the sample, and thus the IR absorption properties of
the sample can be determined as in conventional FTIR spectroscopy. In many ATR
crystals, the light is totally internally reflected multiple times, allowing the absorption
of IR radiation by the sample at each point of reflection.
The intensity of the evanescent wave decays exponentially with distance from the
26 CHAPTER 3. THEORETICAL BACKGROUND
Figure 3.3: Schematic of the principle behind ATR-FTIR. IR radiation entering the crys-
tal at a particular incident angle (θ) is totally internally reflected inside the ATR crystal,
generating an evanescent wave which penetrates into the sample.
surface of the ATR crystal. The depth of penetration, dp, is defined as the distance
from the crystal/sample interface at which the intensity of the evanescent wave decays
to 1e of its original value. It is calculated from:
dp =
λ
2pi(n21sin
2θ − n22)
1
2
(3.9)
where λ is the wavelength of the incident radiation and θ is the angle of incidence
of the IR beam. This distance is typically of the order of 2 - 15 µm and therefore
ATR-FTIR spectroscopy is generally considered to be a surface-based technique that
is insensitive to sample thickness. As a result, ATR-FTIR is ideal for studying samples
which strongly absorb IR radiation, such as dilute aqueous solutions, or are too thick
to study by transmission-FTIR. It is also a useful technique for studying material
surfaces.
Equation 3.9 highlights the dependency of the penetration depth on wavelength.
This relationship has an effect on the resulting ATR-FTIR spectrum, causing a de-
crease in relative band intensity with increasing wavenumber compared to a transmission-
FTIR spectrum of the same sample. As a result, ATR and transmission-FTIR spectra
cannot be directly compared. However, the experimental set up of an ATR-FTIR
experiment can be approximately compared with that of a transmission-FTIR exper-
iment via a property known as ‘effective path length’. This conversion considers two
key properties of an ATR crystal, its penetration depth and the number of reflections
it allows, and converts them to a distance which can be compared with the path-length
3.3. RAMAN SPECTROSCOPY 27
property of transmission-FTIR cells. It is calculated by:
Effective Path Length = dp ×Number of Reflections (3.10)
These parameters, and therefore the absorbance intensity of the resulting spectrum,
can be controlled to some extent by experimental design. The penetration depth
is affected by both the choice of ATR crystal material and the angle of incidence.
A material with a high refractive index leads to a lower penetration depth and thus
reduces the effective path length. Similarly, the higher the angle of incidence, the lower
the penetration depth. The angle of incidence also affects the number of reflections,
with a high angle of incidence leading to fewer reflections and hence a shorter effective
path length and lower absorbance intensity.
Additional experimental parameters which should be considered when designing
an ATR-FTIR experiment include the quality of sample contact, which should be as
high as possible due to the short penetration depth of the evanescent wave, in addition
to the IR-transmission range and chemical reactivity of the ATR crystal material.
3.3 Raman Spectroscopy
Raman spectroscopy, like IR spectroscopy, is a technique which can be used to probe
the vibrational energy levels of a molecule. However, unlike IR spectroscopy it is based
on the scattering, rather than the absorption, of photons. The scattering processes
are much weaker than absorption processes and therefore Raman spectra generally
contain bands of much weaker intensity than IR spectra.
3.3.1 Raman Effect
A photon can interact with a molecule via two main mechanisms: it can be absorbed
by the molecule or it can be scattered. The majority of incident photons that undergo
a scattering mechanism are scattered elastically: both the photon and the molecule
involved possess the same energy after the collision as they did prior to it, a process
known as Rayleigh scattering. However, a small number of these collisions (approxi-
mately 1 in 108)21 are inelastic, such that the energy of the scattered photon differs
from that of the incident photon. This process is known as the Raman effect and is
the principle upon which Raman spectroscopy is based.
The scattering of a photon by a molecule can be described as an excitation of a
molecule to a virtual energy state, followed by a relaxation that leads to the release
of a photon. When the molecule relaxes, it may return to its original vibrational
energy level (Rayleigh scattering) or to a different vibrational energy level (Raman
scattering). If the molecule is left in a higher vibrational energy level than before
the interaction, releasing a photon of lower energy than the incident photon, the
28 CHAPTER 3. THEORETICAL BACKGROUND
process is known as Stokes Raman scattering. Conversely, if the molecule is left
in a lower vibrational energy level, releasing a photon of higher energy than the
incident photon, it is known as Anti-Stokes Raman scattering (Figure 3.4). In both
cases, the magnitude of the energy difference between the incident and the scattered
photon corresponds to the change in energy of the molecule and therefore reflects the
difference in energy between the two vibrational states. Raman spectroscopy uses a
monochromatic light source, which may be in the UV, visible or near-IR range, so
that the difference in photon energy can easily be measured.
Figure 3.4: Diagram showing the molecular excitation and corresponding change in vibra-
tional energy level induced by Stokes and anti-Stokes Raman scattering, as well as Rayleigh
scattering.
The intensity of Stokes scattering is much greater than that of anti-Stokes scatter-
ing at room temperature due to the relative populations of vibrational energy levels,
as discussed in Section 3.2.1. As a result, Raman spectra are generally plotted to
show only the signals generated by Stokes scattering.
3.3.2 Selection Rules of Raman Spectroscopy
Photon scattering arises when an external electric field, such as that of electromagnetic
radiation, induces a dipole moment in a molecule. The size of the induced dipole
moment, P , is given by:
P = αE (3.11)
where α is the polarisibility of the molecule, a measure of how easily the electron cloud
of a molecule can be distorted, and E is the electric field. The induced dipole causes
light to be scattered from the molecule at the optical frequency of the incident light
wave (Rayleigh scattering). Raman scattering occurs when the polarisibility of the
3.3. RAMAN SPECTROSCOPY 29
molecule is altered, causing a change in the size of the induced dipole and hence in
the frequency of the emitted radiation. The polarisibility of a molecule can be altered
by molecular vibrations, leading to the first selection rule of Raman spectroscopy:
for a vibrational mode to be Raman-active, it must change the polarisibility of the
molecule.
The intensity of the scattering is proportional to the change in the polarisibil-
ity of a molecule. If a vibration does not significantly change the polarisibility, the
scattering intensity and thus the intensity of the Raman band will be low. The vibra-
tions of highly polar species are generally weak and therefore do not greatly alter the
polarisibility of the molecule. As a result, an external electric field cannot induce a
large change in the dipole moment and therefore these species have a low scattering
intensity and show weak bands in their Raman spectra.
One example of a highly polar bond which has a low scattering intensity is the O-
H bond. Consequently, water molecules have a low scattering intensity and therefore
Raman spectroscopy is a useful technique with which to study aqueous solutions.
The second selection rule dictates that the allowed change in vibrational energy
level is restricted to ∆v = ±1, as for IR spectroscopy.
3.3.3 Surface-Enhanced Raman Spectroscopy
The main limitation of Raman spectroscopy is the low probability with which Raman
scattering occurs, which makes it a relatively insensitive technique when compared
with IR spectroscopy. SERS is a variation on traditional Raman spectroscopy, in
which the intensity of Raman scattering is enhanced through the use of nanostructured
noble-metal surfaces.
The increase in Raman scattering intensity observed in SERS is thought to occur
via two different mechanisms. The mechanism which provides the greater contribu-
tion to the enhancement, generating an average enhancement factor of 105, operates
through the phenomenon of Surface Plasmon Resonance (SPR). SPR is made possible
due to the high electron density present at the surface of a metal. Whilst the con-
duction electrons of the bulk metal are held in a lattice by the positive charge at the
metal ions, the surface electrons experience a positive charge only on the side closest
to the metal bulk and are therefore less tightly bound to the metal. As a result, the
electron density extends a considerable distance beyond the surface of the metal and
the electrons are able to move freely in a lateral direction along it.
Irradiation of the metal surface with electromagnetic radiation of the correct fre-
quency causes these electrons to oscillate as a collective group across the face of the
metal. This process is SPR and is illustrated in Figure 3.5. If SPR is induced on a
nanostructured surface, the resulting surface plasmons become localised in the val-
leys. Excitation of these localised surface plasmons with visible light leads to the
30 CHAPTER 3. THEORETICAL BACKGROUND
Figure 3.5: Diagram illustrating the generation of surface plasmon resonance localised about
metal nanoparticles. Free conduction electrons at the metal surface oscillate collectively as a
result of coupling with incident electromagnetic radiation.
generation of significant electric fields. Electromagnetic SERS enhancement relies
on Raman-active molecules being confined within these electromagnetic fields, which
cause a significant increase in the magnitude of their induced dipole and therefore an
increase in the scattering intensity.
SERS can also operate via a chemical mechanism. This mechanism has a much
smaller impact on the Raman signal than the electromagnetic mechanism and is
thought to contribute an average enhancement factor of 102.22 Chemical enhancement
of Raman scattering involves the formation of a bond between the analyte and the
surface of the SERS substrate and a subsequent charge-transfer mechanism from the
metal to the analyte. The gain in charge increases the polarisibility of the molecule and
therefore increases the intensity of Raman scattering. A molecule must be chemically
adsorbed to a roughened surface in order to experience the chemical enhancement,
and hence the process is both site-specific and and analyte-dependent.
The most critical aspect of a SERS experiment is the choice of noble-metal sub-
strate: the SERS intensity relies largely upon the excitation of localised surface plas-
mons, it is necessary to control all of the factors influencing the localised SPR in order
3.3. RAMAN SPECTROSCOPY 31
to maximise signal strength and ensure reproducibility. The frequency of electromag-
netic radiation needed to induce SPR is determined by the type of metal, as well as
the size and shape of features on the metal surface. Silver and gold plasmons oscillate
at frequencies in the visible region, making them compatible for use with the visible
and near-IR lasers used in Raman spectroscopy, and hence these metals are the most
commonly used SERS substrates. Common substrates include colloidal nanoparticles
or surface-confined nanostructures.
In addition to the obvious advantage of an increase in Raman scattering intensity,
SERS has found use in the study of surfaces: the abrupt decay of the enhanced
electromagnetic field, which decays as 1
r3
from the surface,23 as well as the requirement
for chemisorption of an analyte in order to induce chemical enhancement, ensures
that only molecules located on or close to the substrate are probed. Conversely, these
factors also mean that SERS has limited applicability when the molecule of interest
is not adsorbed directly onto the substrate. The most significant limitation of SERS
is its lack of reliability, as minor variations in fabrication processes, or the instability
of nanostructures, can lead to surfaces which exhibit spatially inconsistent optical
properties and widely varying enhancement factors.
One notable property of SERS is that it can result in the appearance of bands
not seen in the conventional Raman spectrum of the same analyte. This occurs
when the interaction between the analyte molecule and the SERS substrate, which is
necessary for the enhancement to occur, alters the molecule symmetry and affects its
polarisibility.
3.3.4 Raman Spectroscopy Instrumentation
Raman spectrometers may take a number of different configurations and specifica-
tions. This section will discuss the configuration of a generalised dispersive Raman
microscope, which is the type of instrument used in this work. A Raman microscope is
an instrument in which a microscope forms an integral part of the Raman spectrome-
ter and the sample is placed under the microscope objective. This type of instrument
enables the study of extremely small sample sizes or trace amounts of material, despite
the weakness of Raman scattering.
Figure 3.6 shows the optical path of a generalised dispersive Raman microscope.
Light from a laser source is directed towards a notch filter which removes any unwanted
radiation, such as weak emission lines outside the main exciting line of the laser, and
thus increases the monochromaticity of the exciting beam. The light is then passed
through a lens to produce a parallel beam which is directed into the optics of the
microscope. The laser light is focused onto the sample via the microscope objective.
Light scattered or reflected by the sample is also collected by the objective and
is returned to the spectrometer via the same optics. In a confocal microscope ar-
32 CHAPTER 3. THEORETICAL BACKGROUND
Figure 3.6: An illustration of the optical path of a generalised dispersive Raman microscope.
Laser light is monochromated using a notch filter and passed through a lens to produce a
parallel beam. Microscope optics are used to focus the incident light on the sample and collect
light scattered by the sample. The scattered light is passed through a notch filter to remove
elastically-scattered Rayleigh light and the remaining wavelengths are focused onto a CCD
detector by a grating, which separates the beam into its constituent wavelengths.
rangement, the microscope contains a pinhole in its focal plane which enables only
light scattered from the focal plane to be collected efficiently. As a result, confocal
microscopes can allow excellent control over depth profiling within the sample.
Once returned to the spectrometer, the scattered light is passed through a notch
filter which selectively absorbs all light of the frequency of the incident laser light.
This step removes specularly reflected and elastically scattered light, which is a nec-
essary step due to the higher relative intensity of this light compared to inelastically-
scatttered light. The remaining scattered light is then directed onto a grating which
spatially separates the beam into its constituent wavelengths. These wavelengths are
focused onto a Charge-Coupled Device (CCD) detector. Each photon incident on the
surface of the CCD generates an electrical charge which allows the signal from the
scattered light to be transferred into digital signals.
Raman intensities are measured in units of counts per second. The measured
intensity value depends on many apparatus-specific parameters and therefore the or-
dinate in a Raman spectrum is typically expressed in terms of arbitrary units, or the
scale is omitted entirely. The abscissa represents the frequency difference between the
3.4. STATISTICAL ANALYSIS AND MODELLING 33
incident photon and the scattered photon and thus should strictly be quoted in units
of ∆ cm−1 but is often expressed simply as cm−1.
3.4 Statistical Analysis and Modelling
This section provides an overview of the theory behind the statistical methods used
in this work.
3.4.1 Analysis of Variance
Analysis Of Variance (ANOVA) is a statistical technique which is used to determine
whether there are any statistically significant differences between the means of three
or more independent groups. It is used to analyse discrete data sets in which the
independent variable is categorical and the dependent variable is continuous. The
ANOVA test can be applied to data sets which have any number of independent
variables, with each independent variable having multiple different categories.
One-way ANOVA is used for data sets containing only one independent variable.
It compares the mean values of each category, or group, of the independent variable
and determines whether any of those means are statistically significantly different
from each other. Specifically, it tests the null hypothesis:
H0 = µ1 = µ2 = µ3 = ... = µk (3.12)
where µ is the group mean and k is the number of groups. The null hypothesis states
that there is no statistically significant difference between the group means. If a one-
way ANOVA returns a statistically significant result, the alternative hypothesis is
accepted, which states that there are at least two group means that are statistically
significantly different from one another.
Three assumptions must be met in order for the ANOVA test to produce a reliable
result. Firstly, the residuals (the differences between the mean value and the measured
values) of the dependent variable are normally distributed within each group that is
being compared. Secondly, the population variances are equal for each group. Finally,
it is assumed that observations of the dependent variable are made independently of
one another.
One-way ANOVA can only determine whether there are statistically significant
differences between the means of any of the groups tested: it cannot identify which
specific groups are statistically significantly different from one another. In order to
determine which of the groups are statistically significantly different, a multiple com-
parison post-hoc test must be used.
34 CHAPTER 3. THEORETICAL BACKGROUND
3.4.2 Statistical Modelling
Statistical modelling is the generation of a mathematical formula (the ‘model’) which
can be used to describe a set of observed data in which the observations are subject to
uncertainty, or randomness, and are therefore regarded as having been drawn from one
or more probability distributions. The model is used to estimate the variables, and
the relationships between them, that resulted in the observed values of the dataset. It
can thus be considered as an approximation of the data-generating process, allowing
it to be used to gain deeper insight into the observed dataset or to make predictions
about a new set of data.
A statistical model, by definition, provides only an estimation of the data-generating
process. As a result, the performance and reliability of a given model is significantly
affected by the way in which the model is constructed, with regards to both the for-
mula and the sample data used. The following sections describe some of the most
widely-used statistical models.
3.4.3 Regression Analysis
Regression analysis is used to construct relationships, of the form y = f(X), between
a single dependent variable (y) and one or more explanatory variables (X). The
dependent variable represents the value or process that is being explained or predicted,
whilst the explanatory variables are those used to model, or predict, the values of the
dependent variables. It should be noted that the statistical relationship between the
explanatory and dependent variables does not imply any causation of the dependent
variable by the explanatory variable, but represents only an association in the data.
The regression analysis techniques used in this work, Principal Component Re-
gression (PCR) and Partial Least Squares Regression (PLSR), are based upon linear
models which specify a linear relationship between the dependent and explanatory
variables. This relationship is expressed as a linear function of the form:
y = β0 + β1X1 + β2X2 + ...βnXn + ε (3.13)
where βi are regresssion coefficients, which take a different value for each explanatory
variable, and ε is the random error term. Regression coefficients are used to quantify
the strength and the nature of the relationships between the explanatory variables
and the dependent variable. The magnitude of the regression coefficient indicates
the strength of the relationship between the explanatory variable and the dependent
variable, with a large magnitude indicating a strong relationship, whilst the sign of the
regression coefficient indicates whether the association is positive or negative. β0 is the
regression intercept, which represents the expected value of the dependent variable if
all the explanatory variables were zero. As mentioned previously, any statistical model
3.4. STATISTICAL ANALYSIS AND MODELLING 35
is only an approximation of the data-generating process and, as a result, the predicted
values of the dependent variable will rarely match the observed values exactly. The
random error term represents the portion of the dependent variable that cannot be
explained by the model.
The development of a regression model begins with a set of known y values, which
are often referred to as observed values. These observed values constitute the training
dataset and are used to build, or calibrate, the regression model. Prior to devel-
oping the model, the data in the training set may be tailored in order to simplify
the subsequent calculations. This is often achieved by mean-centering the values of
the explanatory variables, a process in which the average value for each explanatory
variable is calculated and subtracted from the corresponding variable. During the
calibration process, the model can be applied to the training dataset to determine the
accuracy with which the identified explanatory variables and regression coefficients
can predict the observed values of the dependent variable within the dataset. This
process of internal validation is often carried out using the ‘leave-one-out’ method of
cross-validation, in which the model is trained using n−1 datasets and is then used to
predict the value of the dependent variable in the remaining dataset. The procedure
is carried out iteratively until the value of the dependent variable has been predicted
for each dataset. Following cross-validation, the performance of the model may be
quantified by the Root Mean Square Error of Cross-Validation (RMSECV), which is
a measure of the error in the predictive performance of the model during the cross-
validation procedure. It provides an estimate of the likely uncertainty that could be
expected if the model was used to predict the values of the dependent variable from
new datasets. The lower the RMSECV, the more accurate the predictive ability of
the model is likely to be.
The calibrated model may then be applied predictively in order to estimate the
values of the dependent variable from a new dataset. In order to achieve this, the ex-
planatory variables of one or more samples must be obtained for use as a test dataset.
The explanatory variables of the test set, together with the optimised regression co-
efficients identified during the calibration step, are then used to predict the values of
the dependent variable within the test set. This process may be considered as external
validation of the model. In this case, the performance of the model may be quantified
by the Root Mean Squared Error of Prediction (RMSEP), which is a measure of the
uncertainty that can be expected when predicting the values of the dependent variable
from new datasets. It is calculated by:
RMSEP =
√
Σ(yi − yˆi)2
n
(3.14)
where yi and yˆi are the observed and predicted values of the dependent variable,
respectively, and n is the number of values. The lower the RMSEP, the more accurate
36 CHAPTER 3. THEORETICAL BACKGROUND
the predictive ability of the model.
3.4.3.1 Principal Component Regression
PCR is a form of regression analysis in which the explanatory variables in a dataset
are analysed via Principal Component Analysis (PCA) prior to the regression analysis
itself.
PCA is a data-reduction technique which is used to reduce the number of explana-
tory variables to a few linear combinations of the data. Each linear combination is
referred to as a ‘principal component’. The goal of PCA is to explain the maximum
possible amount of variance in the data (the variation of a set of values from their
mean) with the fewest principal components.
The first principal component is the linear combination of explanatory variables
that has the maximum variance, compared with the other possible linear combina-
tions, and therefore accounts for as much variation in the values of the explanatory
variables as possible. The second principal component is the linear combination of
explanatory variables that accounts for as much of the remaining variation as possi-
ble, under the constraint that the correlation between the first and second principal
components is zero. All subsequent principal components follow this pattern: they are
linear combinations of the data that account for as much of the remaining variation
in the data as possible and are not correlated with the previously-defined principal
components.
Following PCA, the values of the dependent variable are regressed on a subset
of the identified principal components of the explanatory variables, rather than on
the explanatory variables themselves. This process generates a number of estimated
regression coefficients equal to the number of principal components. Finally, the
regression coefficients are transformed to match the scale of the explanatory variables,
in order to obtain values which adequately characterise the original dataset.
3.4.3.2 Partial Least Squares Regression
PLSR is a method of constructing regression equations that involves the generation
of new explanatory variables, known as components, which are fewer in number than
the original explanatory variables.
The PLSR method is comparable with PCR, as both techniques involve the pro-
jection of the explanatory variables into a lower-dimensional space in order to generate
new explanatory variables. However, the two methods differ significantly in the man-
ner in which the new explanatory variables are selected. As described previously,
the construction of principal components for PCR is based upon accounting for as
much of the variation in the values of the explanatory variables as possible. However,
this criterion does not guarantee that the principal components, which ‘explain’ the
3.4. STATISTICAL ANALYSIS AND MODELLING 37
explanatory variables optimally, will be relevant for the prediction of the dependent
variable. PLSR addresses this issue by incorporating information from the values of
the dependent variable when constructing the components to be used in the regression
equation.
Each component generated in PLSR is a linear combination of explanatory vari-
ables. The first component is selected by finding the linear combination of explanatory
variables that explains the maximum variance in the values of the dependent variable.
The second component is constructed to explain as much of the remaining variation
on the values of the dependent variable as possible, under the constraint that the
correlation with the previously-identified component is zero. This process is repeated
for the construction of all subsequent components, which are linear combinations of
the explanatory variables that account for as much of the remaining variation in the
data as possible whilst being uncorrelated to the previously-identified components.
A standard regression analysis is then performed on the components, rather than
on the explanatory variables.
3.4.3.3 Model Evaluation
A number of parameters may be used to evaluate the performance of a regression
model. This section will discuss the parameters considered in this work.
The R2 value is a statistic, derived from the regression equation, which indicates
the degree to which the developed model is able to explain the observed values of the
dependent variable. This value can range from 0.0 – 1.0, with higher values indicating
that the model is able to explain a greater proportion of the observed data. However,
a model with an extremely high R2 value may be heavily biased towards the training
data, rendering it less applicable to new datasets, a situation known as model over-
fitting. Over-fitting is often caused by the generation of a complex model containing
a large number of explanatory variables and poses a problem for models that are
intended for predictive use and therefore require generalisation to unknown datasets.
For predictive models, an assessment of the RMSECV or RMSEP often provides a
more useful indicator of model performance. The RMSEP is generated following the
application of the model to a test dataset and is therefore a more accurate reflection
of the predictive performance of the model as applied to unknown datasets than
the RMSECV, which is generated via internal cross-validation and is therefore an
estimate of the predictive performance of the model. However, the RMSECV may be
more convenient to use on relatively small datasets, as it allows all available data to be
used in the calibration of the model. Conversely, calculation of the RMSEP requires
a dataset to be split into a training set and a test set, which necessarily reduces the
amount of data available for model calibration. For small datasets, this reduction in
data may result in the generation of a less accurate model which is more strongly
38 CHAPTER 3. THEORETICAL BACKGROUND
affected by the presence of outlying data points.
Chapter 4
Current State of the Art
This chapter describes the current state of the art, as determined by a review of
the scientific literature, in four key areas relevant to this thesis: detection of glucose,
lactate and pyruvate; IR spectroscopy; Raman spectroscopy; and neurochemical mon-
itoring. Preliminary work into the analysis of microdialysate samples via ATR-FTIR
spectroscopy, carried out by a previous member of the Elliott group, is also described.
4.1 Detection of Glucose, Lactate & Pyruvate
The ISCUS Flex is the ‘gold-standard’ commercially-available instrument for analysing
glucose, lactate and pyruvate concentrations in microdialysate samples, as discussed
in Section 2.3.3. However, the development of alternative methods for detecting these
compounds, within a variety of contexts, remains an active area of research. The
following sub-sections will provide an overview of the latest developments in this field,
as described in the scientific literature, for each compound in turn.
4.1.1 Glucose Detection
The development of glucose sensors is a major field of research due to the significant
role of glucose detection in the management of diabetes mellitus. The research en-
compasses a range of transducer types, each of which will be addressed separately
here.
4.1.1.1 Electrochemical Sensing
The majority of glucose sensors rely upon an electrochemical transducing system due
to the rapid respose time it affords.24 Electrochemical glucose sensors are typically
based on amperometric measurements in which the presence of glucose causes a change
in current between two electrodes which have an applied potential across them. These
sensors are often based upon the glucose oxidase-catalysed oxidation of glucose, a
39
40 CHAPTER 4. CURRENT STATE OF THE ART
reaction which requires the flavin adenine dinucleotide (FAD) cofactor and involves
the concurrent reduction of molecular oxygen to H2O2. The process is summarised
in Figure 4.1. The use of glucose oxidase in these devices also confers the benefit of
selectivity towards glucose.
Figure 4.1: Summary of the reaction between glucose and glucose oxidase, showing the role
of FAD and the concurrent reduction of molecular oxygen to H2O2.
The glucose oxidase enzyme is immobilised on the anode of an electrochemical cell,
allowing the production of H2O2 to be detected by the generation of a current resulting
from the re-oxidation of H2O2 at the electrode (Figure 4.2). The amplitude of the
current is proportional to the concentration of H2O2 and hence to the concentration
of glucose in the sample.
Figure 4.2: Summary of the reactions which occur at the electrodes of a glucose oxidase-based
amperometric glucose sensor.
A microfabricated amperometric glucose sensor based on this principle has been
incorporated into a contact lens for the in situ monitoring of tear glucose levels.25
Glucose oxidase was immobilised in a titania sol-gel layer positioned between the
working and counter electrodes, which were fabricated on a polyethylene terephthalate
(PET) film. The device exhibited a response time of 20 seconds and demonstrated a
linear response to the range of glucose concentrations found in tears (0.1–0.6 mM).
A recent paper describes a graphene-based dermal patch containing an amper-
ometric glucose sensor and an array of polymeric microneedles, which can be ther-
mally activated to deliver drugs transcutaneously, for the simultaneous monitoring
and treatment of diabetes mellitus.26 The sensing element of the device, which is used
to monitor the glucose concentration in the sweat of the wearer, is based upon the
amperometric measurement of H2O2 as described above. It has been used to mea-
sure the sweat glucose levels of two healthy volunteers and the results were compared
4.1. DETECTION OF GLUCOSE, LACTATE & PYRUVATE 41
with those obtained using a commercially-available glucose sensor. The two set of
measurements were correlated with an R2 value of 0.89.
One drawback of devices which operate on this principle is their dependency on
the availability of dissolved molecular oxygen at the anode. This dependency can be
avoided through the use of chemical mediators, such as ferrocenes, to transfer the
electrons directly from the glucose oxidase/FAD complex to the anode as illustrated
in Figure 4.3.
Figure 4.3: Generation of a current at the anode of an amperometric glucose sensor using
glucose oxidase and a ferrocene mediator.
Some studies report the development of novel electrodes which are able to remove
electrons directly from the glucose oxidase enzyme, negating the need for a separate
electroactive species. Recently-reported electrodes include a gold electrode coated
with chitosan nanoparticles27 and a palladium/platinum nanocube electrode coated
with chitosan.28
Other research efforts have focused on the development of electrodes for non-
enzymatic amperometric glucose detection. These devices measure the current pro-
duced from the direct electro-oxidation of glucose by the electrode. Recently-developed
electrodes used for this purpose include: a ZnO nanorod array coated with nickel
nanoparticles and covered with graphene oxide nanosheets;29 multi-walled carbon
nanotubes coated with nanoparticles possessing a gold-palladium core and a CuO
outer shell;30 a Ni(OH)2/Al(OH)
−
4 film
31 and CuWO4 microrings.
32
Non-enzymatic amperometric glucose sensors can suffer from a lack of selectivity
due to the absence of a glucose-specific receptor. This problem has been addressed
through the integration of molecularly-imprinted polymers (MIPs) with the working
electrode of such a sensor.33 A CuCo electrode, capable of the direct electro-oxidation
of glucose, was coated with a polymer imprinted with glucose molecules to enable
shape-specific recognition of glucose. The resulting device exhibited a dynamic range
of 0.001–25 mM and was successfully used to measure the glucose concentration in
whole-blood samples taken from healthy human volunteers.
Electrochemical glucose detection has been achieved using non-amperometric tech-
niques. For example, one group has reported the development of a chemi-resistive sen-
sor based on single-walled carbon nanotubes.34 The device contained gold electrodes
42 CHAPTER 4. CURRENT STATE OF THE ART
coated with a polymer-nanotube composite which was functionalised with glucose ox-
idase. This material exhibited a measurable decrease in electrical resistance upon
exposure to glucose, attributed to p-doping of the nanotubes caused by the H2O2 pro-
duced during the reaction between glucose and glucose oxidase. The sensor exhibited
a linear response to glucose concentrations between 0.08–2.2 mM.
4.1.1.2 Optical Sensing
A number of studies report the optical detection of glucose via colorimetric assays.
Many of these assays are based upon the oxidation of glucose by glucose oxidase:
H2O2 produced during the reaction is used to oxidise 3,3’,5,5’-tetramethylbenzidine
(TMB), in a reaction catalysed by a peroxidase mimic, to produce a blue colour that
can be visually detected. The intensity of the colour, which is typically measured by
absoprtion spectroscopy, is proportional to the concentration of glucose in the sample.
Multiple peroxidase mimics have been trialled within this system, including chitosan-
gold nanoparticles;35 cobalt oxyhydroxide nanoflakes;36 metal-organic frameworks;37
platinum nanoclusters38 and carbon-coated Fe3O4 nanoparticles.
39 This method is
typically suitable for detection of glucose concentrations within the millimolar range.
A recent study describes a colorimetric glucose assay based on an alternative
mechanism, in which MnO2 nanosheets were rapidly decomposed by H2O2 produced
during the glucose oxidase-catalysed oxidation of glucose. The decrease in the MnO2
absorption band at 370 nm was proportional to the concentration of glucose in the
sample and exhibited a linear relationship for concentrations of glucose ranging from
0.5–50 µM.40
Other studies have used the generation or quenching of luminescence as a trans-
duction method for the detection of glucose.41–43 For example, one study describes the
synthesis of aluminium-doped ZnO thin films which exhibit stong photoluminescence
at 372 nm and 564–775 nm.44 Immersing the thin film in a solution containing glucose
and glucose oxidase resulted in a quenching of luminescence which is believed to be
caused by charge transfer from H2O2, produced during the reaction between glucose
and glucose oxidase, to the film. This method was successfully used to detect glucose
concentrations ranging from 0.02–20 mM.
Similarly, some studies have reported the detection of glucose via a change in
fluorescence intensity.45–47 For example, the fluorescence intensity of blue-emitting
gold nanoclusters capped with haemoglobin was found to increase in the presence of
H2O2.
48 This property was exploited for the detection of glucose in human serum
samples: the samples were incubated with glucose oxidase for ten minutes at 37 ◦C,
before being added to a solution of gold nanoclusters. The increase in fluorescence
intensity, caused by the H2O2 produced during the reaction between glucose and glu-
cose oxidase, was compared against a calibration curve which had been created using
4.1. DETECTION OF GLUCOSE, LACTATE & PYRUVATE 43
synthetic samples of known glucose concentration.This method was able to predict
the glucose concentration to within 10% of the actual concentration, although it is
noted that the total time required for detection was 40 minutes.
4.1.1.3 Other Sensing Methods
Electrochemical sensing and optical sensing are the most widely-used techniques for
the detection of glucose. However, other transducer types have been reported in the
literature. For example, a quartz-crystal microbalance (QCM), which detects the
adsorption of molecules via a change in the resonance frequency of a quartz-crystal
resonator, has been used for the mechanical detection of glucose.49 The QCM was
functionalised with cyclic peptides which were designed to mimic the glucose binding
site of glucose binding protein, conferring a degree of selectivity for glucose. However,
the nonspecific binding of proteins to the sensor surface interfered with the analysis of
glucose in human serum samples. The effect of nonspecific binding could be reduced
by a ten-fold dilution of the serum samples in phosphate-buffered saline.
A very different study describes the detection of glucose by a DNAzyme, a sequence
of deoxyribonucleic acid (DNA) which possesses catalytic properties, in combination
with glucose oxidase.50 The DNAzyme used in this study is a DNA-cleaving molecule
which requires a combination of H2O2 and Cu
2+ as cofactors. H2O2, produced by
the glucose oxidase-catalysed oxidation of glucose, therefore induced a self-cleavage
reaction of the DNAznyme in the presence of Cu2+. The degree of DNA cleavage was
observed via gel electrophoresis and could be related to the concentration of glucose
present in the sample. This device was successfully used for the analysis of glucose
levels in human tears and human saliva.
4.1.2 Lactate Detection
Lactate detection has two significant clinical uses. The first of these is the monitoring
of patients with a number of medical conditions, including TBI, as lactate concen-
tration is closely linked to both oxygen availability and metabolic status. This rela-
tionship between lactate concentration, oxygen availability and metabolic status leads
to the second clinical use of lactate: the assessment of athletic performance. Moni-
toring of tissue-lactate levels during exercise enables the calculation of the anaerobic
threshold of an individual, which is considered a key measure of fitness.
Reported detection methods for lactate fall into into two main categories: optical
detection and electrochemical detection. Of the studies utilising optical-detection
methods, the majority are based on monitoring changes in fluorescence intensity. For
example, a recent study used a boron-dipyrromethene derivative, modified with an
ester group for optimal binding of hydroxyl compounds, which exhibited an increase
in emission upon forming a complex with lactate.51 The modified compound was
44 CHAPTER 4. CURRENT STATE OF THE ART
designed to emit light in the near-IR region to allow it to be used for monitoring
lactate within biological tissue, athough it is noted that this compound would exhibit
the same behaviour upon binding to any hydroxyl-containing compound and thus the
detection method is not lactate-specific.
Lactate-specific fluorescent detection has been achieved by introducing the lactate
oxidase enzyme, which undergoes a similar set of reactions to glucose oxidase, into the
analytical setup. One such study used a tetraphenylethylene-based fluorophore which
was readily oxidised by H2O2, generated from the lactate oxidase-catalysed oxidation
of lactate, to produce a compound which released a strong emission at 530 nm.52 The
intensity of the emission was proportional to the concentration of H2O2 produced and
hence to the amount of lactate in the sample. However, this method had an upper
limit of 0.02 mM, beyond which the relationship between emission intensity and lactate
concentration was non-linear. Another study, which also utilised the reaction between
lactate and lactate oxidase, incorporated cupric oxide nanoparticles as an additional
reagent.53 The nanoparticles catalysed the decompositon of H2O2 and the resulting
hydroxyl radicals oxidised terephthalic acid to form a fluorescent product with an
emission of 422 nm. This method was used for the quantitative detection of lactate
in human serum samples, within the range 2.5–3.5 mM, although it is noted that
proteins were filtered out of the sample prior to analysis in order to reduce possible
interference.
Colorimetry has been used as an alternative optical-detection method for lactate.54
This method was based upon a two-step reaction. In the first step, lactate oxidase was
used for the analyte-specific oxidation of lactate. The H2O2 produced by this reaction
was then used to oxidise TMB, in a reaction catalysed by horseradish peroxidase, to
form a coloured product. The intensity of the colour change was proportional to the
concentration of lactate in the sample. This device was produced as a paper-based
sensor in which each reagent was immobilised in separate layers atop a paper substrate.
Lactate concentrations were determined via reflectance measurements taken using the
camera and flash of a smartphone. The group also developed an accessory which was
used to ensure consistent positioning of the functionalised paper surface, at the correct
distance from the smartphone camera, and to reduce interference from ambient light.
The setup was successfully used to monitor lactate levels within saliva samples taken
from healthy human volunteers at various stages of a running-based exercise routine,
with a limit of detection of 0.1 mM, and required a total of one minute for each
measurement.
Multiple studies report the detection of lactate via electrochemical methods. Many
of these involve amperometric measurements in which lactate oxidase is immobilised
on the working electrode: flexible graphene nanowalls55 and carbon electrodes coated
with gold, copper or platinum nanoflakes56 have both been trialled as working elec-
4.1. DETECTION OF GLUCOSE, LACTATE & PYRUVATE 45
trodes in such experiments. One interesting aperometry-based study used printable
electrodes placed on a mouthguard for the detection of lactate from saliva.57 A screen-
printing method was used to produce an Ag/AgCl reference electrode and a prussian
blue working electrode, which was functionalised with lactate oxidase, on a PET film.
A current was produced at the working electrode via the prussian blue-catalysed
reduction of hydrogen peroxide, which was produced through the reaction between
lactate oxidase and lactate. The PET film was then applied to a mouthguard using
double-sided adhesive and connected to an analyser via wires attached to Ag/AgCl
contacts on the sensor. This system was used to measure the lactate concentration of
human saliva samples spiked with known concentrations of lactate, but has not been
tested in vivo.
Screen-printed electrodes have also been used to manufacture an electrochemical
biosensor that affixes to the skin, in the form of a temporary tattoo, for the mea-
surement of lactate levels in sweat.58 This system used a carbon working electrode,
functionalised with lactate oxidase tethered to multiwalled carbon nanotubes, and an
Ag/AgCl reference electrode. Both electrodes were screen-printed using inks which
included a dispersion of carbon fibres, in order to improve the tensile strength of the
tattoo, and were covered with a chitosan overlayer which served as the tattoo base.
Tattoos were applied on the deltoid of a human subject and connected to an electro-
chemical analyser via stainless steel wires. The temporal sweat lactate profiles of a
total of ten subjects were collected via the tattoo, using amperometric measurements,
over the course of a cycling exercise session. In each case, the performance of the
device was tested against control tattoo sensors, which contained no lactate oxidase,
in order to confirm the specificity of the sensor towards lactate. The sensor readout
demonstrated a linear response to increasing lactate concentration up to 20 mM.
An alternative electrochemical lactate sensor used MIPs in place of lactate oxidase
to achieve specific detection of lactate.59 The sensor consisted of a planar, screen-
printed, carbon electrode modified with poly(3-aminophenylboronic acid) which was
molecularly imprinted with lactate to create a shape-specific recognition site. The
electrode was used to evaluate the lactate content of human sweat samples via im-
pedence measurements, in which the resistance of the film was found to decrease in
response to increasing lactate concentration. Lactate concentrations from 10–40 mM
were measured using the device, which exhibited an average response time of 2–3
minutes.
4.1.3 Pyruvate Detection
A review of the scientific literature found far fewer studies relating to pyruvate detec-
tion than to the detection of glucose or lactate. An early study used a two-stage re-
action which began with the pyruvate decarboxylase-catalysed conversion of pyruvate
46 CHAPTER 4. CURRENT STATE OF THE ART
to acetaldehyde and CO2. The resulting CO2 increased the acidity of a pH-sensitive
fluorescent indicator, causing a change in emission at 520 nm. A slice of corn tissue,
integrated directly into the sensor, was used as the source of pyruvate decarboxylase.
The sensor was assembled in a vertical configuration: the corn tissue was placed at
the bottom of the assembly, in direct contact with a gas-permeable membrane placed
above it. A reservoir of indicator solution was positioned atop the membrane and was
enclosed by a glass slide above it. The tip of an optical fibre, the light source of the
sensor, was placed above the glass slide in direct contact with its surface. Finally,
the whole assembly was enclosed within a pipette tip, allowing the sensor to be op-
erated by dipping its lower end into a buffer solution containing the analyte. This
setup was used to detect pyruvate in buffer solution at concentrations of 0.2 –3 mM.
However, the response time of the sensor was slow, requiring between 7 – 35 minutes
for detection, depending upon the concentration of pyruvate under investigation.
Another study employed Fo¨rster Resonance Energy Transfer (FRET) sensing
for the detection of pyruvate.60 This technique involves a transfer of energy, via
dipole-dipole coupling, from an excited-state donor chromophore to an acceptor chro-
mophore. The efficiency of the transfer is correlated to the distance between donor
and acceptor, being inversely proportional to the sixth power of the distance between
them, and can therefore be used to monitor small changes in the separation between
the two. The FRET study utilised PdhR protein, a transcriptional regulator which
binds to pyruvate in vivo, which was modified to contain a donor and acceptor chro-
mophore at either end of the molecule. Upon binding to pyruvate, the conformation
of the protein was altered and hence the distance between the chromophores would
change. This allowed pyruvate binding events to be detected as a variation in the
ratio of the fluorescence intensities of the two chromophores. The modified protein
was introduced to living cells, including neurons, and used in this way to perform
real-time quantitative measurements of pyruvate production and consumption in in-
dividual cells.
Pyruvate has also been detected electrochemically.61 In this study, the working
electrode consisted of a platinum-coated 0.25 mm gold-wire tip which was coated
with a layer of recombinant pyruvate oxidase immobilised in an electron-conducting
hydrogel. The recombinant enzyme catalysed the conversion of pyruvate and or-
thophosphate to acetyl phosphase, with concurrent production of CO2 and H2O2,
at the working electrode. The working electrode was held at 0.4 V compared to an
Ag/AgCl reference electrode and this setup was able to detect pyruvate concentrations
as low as 0.03 mM against a background of calf serum.
4.2. IR SPECTROSCOPY 47
4.2 IR Spectroscopy
IR spectroscopy has successfully been used for a wide range of applications, including
the detection of HCN in exhaled breath;62 monitoring CO2 production in human cell
cultures;63 investigating monolayers of membrane proteins,64 the differentiation of
mammalian cell types65 and the characterisation of wine vinegars to ensure quality
control.66
A recent study has proposed the use of FTIR spectroscopy as a tool for the forensic
analysis of deaths following TBI.67 The study focuses on traumatic axonal injury, a
common secondary brain injury, which involves the shearing and stretching of axons
and cannot be detected using conventional neuroimaging techniques: it is usually
only diagnosed via histopathological staining during post-mortem examinations. An
ability to determine the time interval between the onset of traumatic axonal injury
and death may be useful in forensic analysis for estimating the time of the initial
traumatic incident. This study used FTIR spectroscopy to analyse differences in the
brain tissue of rats euthanased at different time points (12, 24 or 72 hours) following
a controlled TBI. The spectra of tissue taken from animals euthanased at different
times post-injury were distinguishable from each other, and from tissue taken from a
control group of animals with no TBI, due to changes in absorption peaks associated
with proteins, lipids and carbohydrates. The spectra were analysed via hierarchical
cluster analysis, which produces a diagram that shows the degree of similarity between
spectra, and via PLSR in order to produce a calibration model that was successfully
used to predict the injury interval of traumatic axonal injury in an unknown sample
dataset.
FTIR spectroscopy has been investigated as an alternative to histopathological
staining for a number of other tissues,68–71 including the oesophagus. A recent study
describes the use of FTIR spectroscopy mapping to develop an automated histology
tool for the discrimination of normal, pre-malignant and malignant cells of the oe-
sophagus.72 It is hoped that this automated method could provide a more objective
method of assessing histopathology, as the interpretation of conventional histopatho-
logical staining can vary between indiviuals carrying out the examination. A total
of 22 tissue samples were biopsied from human patients, frozen and cut into sections
8 µm thick. Frozen sections were mounted on CaF2 slides and thawed prior to the
acquisition of FTIR spectra. Sections immediately adjacent to those examined by
FTIR spectroscopy were analysed using conventional histological staining techniques
to provide reference data. A total of 737,280 FTIR spectra were acquired, across all
tissue samples, using a FTIR spectrometer connected to a microscope. A series of
data-analysis steps were carried out on the spectral region 1800–950 cm−1 in order
to create a diagnostic classification model based on the differences between tissue
spectra. After normalisation and mean-centring of the dataset, PCA was performed
48 CHAPTER 4. CURRENT STATE OF THE ART
to analyse the variance in the data. An ANOVA test was carried out on the first
50 principal components identified by PCA in order to find the components which
provided the greatest discrimination between pathology groups. Linear discriminant
analysis was then carried out on the principal components selected by the ANOVA
analysis and used to assign samples into groups to build a diagnostic classification
model. Linear discriminant analysis is a technique used to find a linear combination
of features which separates two or more classes of object and is used for datasets in
which the independent variable is continuous and the dependent variable is catagor-
ical. The categorical predictions for each spectrum were then combined to generate
a prediction for the whole tissue sample. In samples where multiple pathologies were
found, it was determined that each sample would be categorised as having the worst
pathology if at least 30% of the spectra exhibited that pathology. The resulting model
was tested using leave-one-out cross-validation. Classification of normal vs. abnormal
tissue samples using this method was found to have 100% sensitivity and specificity.
4.2.1 Detection of Metabolites by IR Spectroscopy
IR spectroscopy has also been used for quantitative analysis of two of the compounds
relevant to this work: glucose and lactate. One early study attempted to measure
exercise-induced changes to blood glucose and lacate levels in human subjects.73
Three replicates of 50 µL blood samples were taken from human subjects at rest,
following medium-intensity exercise and following high-intensity exercise, producing
a total of 198 samples. Two samples from each set of replicates were analysed us-
ing well-establised techniques to determine their glucose and lactate concentrations,
while the third was studied via FTIR spectroscopy. Prior to the acquisition of FTIR
spectra, samples were diluted with water, homogenised and then dried. Spectra were
acquired via transmission-FTIR spectroscopy using 32 accumulations to generate each
spectrum. Glucose and lactate concentrations were determined by analysing a single
absorbance peak for each compound: the 1033 cm−1 glucose peak and the 1127 cm−1
lactate peak. A calibration curve was created for each peak, using solutions of known
concentrations of glucose and lactate, and used to estimate glucose and lactate concen-
trations from the FTIR spectra of the test samples. This approach was complicated
by the fact that the lactate peak of interest partially overlapped with a glucose ab-
sorbance peak and, as a result, required the subtraction of the glucose absoption
spectrum from the overall spectrum to allow an analysis of the lactate concentration.
Despite this adjustment, large prediction errors were found in lactate concentrations
below 0.5 mM or above 15 mM.
A later study also investigated exercise-induced changes in metabolite levels in the
blood, but focused only on the detection of lactate. This work used near-infrared spec-
troscopy, covering the 11500–4000 cm−1 spectral range, to measure changes in human
4.2. IR SPECTROSCOPY 49
blood-lactate levels before, immediately following and ten minutes after exercise.74
The analysis was carried out on whole-blood samples, which were collected into tubes
and immediately treated to arrest glycolysis. Spectra were acquired via transmission-
FTIR using a 1 mm optical path length. In this study, the spectra were analysed
using statistical methods rather than using a calibration curve. Prior to analysis, the
spectra were mean-centred, to emphasise subtle variations in spectra, and the second
derivative of each spectrum was calculated in order to enhance the spectral varia-
tion of lactate over the background spectrum and minimise baseline variation. PLSR
using the ‘leave-N-out’ method of cross-validation was used to analyse the spectral
range from 4962–4167 cm−1, which was selected to minimise interference from water,
whilst reference concentrations were determined by enzymatic measurements using a
commercially-available analyser. The final PLSR model contained seven components
and was able to predict blood-lactate concentrations ranging from 0.3–25 mM with a
standard error of cross validation of 0.65 mM.
An alternative analytical approach was taken in a 2008 study which aimed to
measure glucose levels in whole-blood samples taken from healthy human volunteers.75
Samples were initally analysed by ATR-FTIR spectroscopy, using a ZnSe crystal,
but it was found that the spectra changed with time as they were affected by the
sedimentation of red blood cells onto the crystal surface. An alternative method
was devised which utilised transmission-FTIR spectroscopy in combination with a
dual-wavelength CO2 laser tuned to the isosbestic point at 950 cm
−1, for background
correction, and to the glucose absorption band at 1080 cm−1. Whole blood was
placed directy into an optical cell possessing a path length of 50 µm and the ratio
of the two wavelengths of light detected after passing through the optical cell was
used to determine the glucose concentration. These measurements could be carried
out in milliseconds and were therefore unaffected by red blood cell sedimentation.
Reference measurements were carried out separately on each sample and a comparison
of predicted vs. refererence values, for glucose concentrations ranging from 0.2 - 2 mM,
gave an R2 value of 0.997.
Mid-IR spectroscopy has been used for noninvasive measurements of blood-glucose
levels in human subjects by focusing IR light directly at the skin.76 A mid-IR, broadly
tunable, quantum cascade laser (QCL) was coupled with a hollow-core fibre to de-
liver IR radiation to a region of skin between thumb and forefinger. Light was able
to penetrate to the dermis layer of the skin, where the interstitial glucose concen-
tration has the closest correlation to blood-glucose levels.77 Backscattered light was
collected via a bundle of 6 fibres, identical in composition and size to the delivery
fibre, which was coupled directly to a commercially-available detector. Measurements
were taken on three healthy volunteers before, during and after the consumption of
food. The process was repeated multiple times over a period of weeks in order to
50 CHAPTER 4. CURRENT STATE OF THE ART
build a sufficiently large data set for PLSR analysis. Calculated values of the glucose
concentration were compared with measurements of blood-glucose acquired using a
commercially-available electrochemical blood-glucose meter immediately before and
after each optical reading. The PLSR model was able to predict blood-glucose con-
centrations between 0.5 and 0.9 mM to clinical accuracy, defined as being within 20%
of the actual value, for 84% of measurements.
4.2.2 IR Spectroscopy & Microfluidics
An on-line spectroscopy-based sensor for metabolic profiling of TBI patients may
be required to detect glucose, lactate and pyruvate within flowing microdialysate
samples, which will be of low volume as a result of the low flow rate (0.3 µL min−1)
used in microdialysis, depending upon the design of the sensor chip. It is therefore
encouraging to note a number of examples in the literature describing the successful
integration of IR spectroscopy with microfluidics.78
A recent example of such a study combined ATR-FTIR spectroscopy with mi-
crofluidics to analyse pharmaceutical formulations.79 The study used microfluidic
chips that were created by casting Polydimethylsiloxane (PDMS) over a 3D printed
acrylic mould, producing a device with microfluidic channels imprinted on its lower
face. This imprinted surface was placed directly against a commercially-available
ATR-FTIR crystal and secured using a perspex top-plate, creating sealed channels in
which the upper and side walls formed by the PDMS chip and the lower wall of the
channel was the ATR crystal itself. This setup was used to study the dissolution of
micro-formulations of ibuprofen and polyethylene glycol under flowing conditions via
ATR-FTIR spectroscopy. The technique was also used to monitor the recrystallisation
of ibuprofen upon contact with acidic solutions, using a flow rate of 10 µL min−1.
4.3 Raman Spectroscopy
Raman spectroscopy has been used in multiple applications, including the histologi-
cal imaging of a human polyp sample,80 differentiation of prostatic adenocarcinoma
cell lines81 and determining the level of carbonation in cement paste with a view to
monitoring the status of concrete structures.82
A number of studies have described the development of accessories for the acqui-
sition of Raman spectra which extend the range of potential applications for which it
can be used. For example, a hand-held contact Raman spectroscopy probe, capable
of identifying cancer cells that have invaded healthy brain tissue at cellular resolu-
tion, has been developed and used intraoperatively during the resection of grade 2–4
gliomas.83 The probe contains fibre-optic cables, fitted with micrometre scale in-line
filters for the attenuation of elestically-scattered light, which are connected to a CCD
4.3. RAMAN SPECTROSCOPY 51
and PC to enable the visualisation of spectra in real time. A 785 nm laser was focused
to a spot size of 0.2 mm2 and used to measure the Raman signal at several points
in the surgical cavity. The total acquisition time used for the examination of a site
was 0.2 seconds, which corresponded to the acquisition of three Raman spectra and
a background measurement in which there was no laser excitation. A corresponding
tissue biopsy was taken from each interrogation site and examined via histological
staining to allow verification of the results of the Raman study. A total of 161 Ra-
man measurements were taken across 17 patients and the spectra were analysed by
the ‘boosted trees’ machine-learning algorithm, which was used to classify each re-
gion of tissue as ‘normal’, ‘dense cancer’ or ‘normal tissue invaded by cancer cells’,
based upon its Raman spectra. Comparison of the algorithm results with histological
staining indicated that the method could be used to classify brain tissue with a sen-
sitivity of 93% and a specificity of 91%. It is hoped that this method could facilitate
the intraoperative identification of tumour boundaries, leading to improved patient
outcome and reducing the need for repeated operations.
Another study reports the development of a multifibre Raman probe that fits
inside the bore of a hypodermic needle, enabling its use for subcutaneous Raman
spectroscopy of tissues.84 The probe consists of a central excitation fibre surrounded
by six collection fibres, composed of low-OH silica, and has a total diameter of 720 µm
which allows it to be easily inserted into a 19-gauge disposable hypodermic needle.
Crucially, the expensive optical components of the system are positioned outside the
needle probe, meaning that the portion of fibre within the needle is relatively inexpen-
sive and can therefore be disposable. The properties of the probe were tested using
a two-layer tissue phantom and an excised human lymph node, using an excitation
wavelength of 830 nm and excitation powers ranging from 20–100 mW. A comparison
with data generated in a previous study by the same group suggests that the signal-
to-noise ratio demonstrated by the probe is 3–4 times higher than that required for
discrimination between benign lymph nodes, primary lymph node malignancies and
secondary lymph node malignancies.85 These results suggest that the developed probe
has potential for the in vivo diagnosis of Hodgkin’s and non-Hodgkin’s lymphomas.
4.3.1 SERS
Much of the current research into SERS is focused on the development of novel SERS
substrates offering improved enhancement factors. This often involves the creation
of new nanostructures for the generation of localised surface plasmons. Previously-
reported nanostructures include: vertical silver nanorod zig-zag arrays containing
nanogaps at the joins between rods;86 vertically-oriented ZnO nanorod arrays coated
with a 30 nm gold layer;87 a three-dimensional carbon foam coated with silver nanosheets;88
flower-like silver mesoparticles;89 silver nanowires coated with silver nanoparticles;90
52 CHAPTER 4. CURRENT STATE OF THE ART
porous ZnO microspheres coated with silver nanoparticles91 and gold nanostar ar-
rays coated with a layer of graphene followed by a layer of gold nanoparticles.92 One
group has reported the development of a ‘smart’ SERS substrate consisting of silver
nanoparticles encapsulated in polyacryloyl hydrazide.93 Polyacryloyl hydrazide is a
pH-responsive polymer which swells at high pH and shrinks as the pH decreases. This
property allowed the spacing of the silver nanoparticles, and hence the characteristics
of the local surface plasmons, to be tuned by controlling the pH of the substrate.
Many groups report the functionalisation of SERS substrates with analyte-specific
receptors to enable SERS-based detection of a compound of interest. Analyte-specific
aptamers (nucleotide or peptide sequences which enable shape-specific recognition
through their three-dimensional structure) have been used in the detection of a num-
ber of substances, including Hg2+ ions;94–96 circulating tumour cells97 and ATP.98
Sequence-specific DNA probes have also been coupled with SERS for the detection
of nucleic acids, such as micro-ribonucleic acid99 and hepatitis B viral DNA,100 at
femtomolar concentrations.
4.3.2 Detection of Metabolites by Raman Spectroscopy & SERS
A number of studies describe the detection of metabolic intermediates by Raman
spectroscopy or SERS, with the majority of studies focusing on glucose detection.
The van Duyne group, in particular, has carried out a significant amount of work
on the development of a SERS substrate for the detection of glucose and lactate.
The substrate, known as a ‘film-over-nanosphere’ surface, is chemically modified in
order to concentrate the analyte molecules within 2 nm of the substrate surface to
ensure that they are exposed to the high electric fields necessary for SERS.101 It
consists of silica nanospheres arranged in a hexagonal close-packed array on a silicon
or titanium substrate and coated with a 200 nm layer of silver or gold. The metallic
layer is then functionalised with a mixed self-assembled monolayer of 1-decanethiol
and mercaptohexanol to which glucose and lactate are able to adsorb. This system was
initially used to detect aqueous glucose concentrations from 0–25 mM using a 633 nm
excitation source, with a power at the sample of 3.25 mW, and an acquisition time of 30
seconds. Attempts to detect aqueous glucose using an unfunctionalised substrate were
unsuccessful, indicating that the 1-decanethiol/mercaptohexanol monolayer plays a
significant role in glucose detection.
This method has also been applied to the quantitative in vitro detection of lactate,
at concentrations ranging from 0.5–22 mM, using a 532 nm excitation source with a
power of 10 mW at the sample.102 The substrate was incubated in lactate solution
for two minutes prior to performing measurements, to ensure adsorption onto the
monolayer, and a 2 minute acquisition time was used for each spectrum. Significant
spectral pre-processing was carried out prior to data analysis, including: smoothing,
4.3. RAMAN SPECTROSCOPY 53
cosmic-ray removal and the removal of background changes that occurred over time.
PLSR analysis of the resulting data set, in combination with leave-one-out cross-
validation, yielded a RMSEP of 3.4 mM.
A silver film-over-nanosphere substrate has been used to demonstrate in vivo,
transcutaneous, monitoring of glucose in live rats.103 The substrate was implanted
subcutaneously in a rat and regular measurements were taken using surface-enhanced
spatially offset Raman spectroscopy, a variation of SERS in which the excitation of
molecules and collection of scattered light are spatially separated. This technique is
useful for measuring layered samples in which the analyte of interest lies below an
opaque external layer, such as the skin. It requires the acquisition of multiple spec-
tra, one in which light is collected at the source and others where light is collected
at an offset position typically millimetres away from the excitation point, and subse-
quent calculation of the differences between spectra allows the spectra of sub-surface
components to be revealed. Using this technique, combined with PLSR analysis and
leave-one-out calibration, the group were able to monitor the glucose concentration
of the animal for 17 days with a RMSEP of 0.8 mM. Beyond this time, the quality of
the spectrum was found to deteriorate significantly. A 785 nm excitation source was
used for all measurements, generating a surface power of 50 mW, and each spectrum
was acquired over two minutes.
Recently, the group have developed an alternative chemical modification for film-
over-nanosphere substrates based on an analogue of the 4-amino-3-fluorophenylboronic
acid molecule.104 Boronic acid derivatives are able to covalently bond, through their
diol functional groups, to saccharide molecules to form cyclic boronate esters. The
particular analogue used by the van Duyne group has been shown to exhibit a degree
of chemical selectivity by preferentially binding to glucose over fructose in an equimo-
lar mixture of the two. Gold film-over-nanosphere substrates functionalised with this
molecule were used, in combination with a 785 nm excitation source, to detect glucose
in phosphate buffer at physiologically-relevant concentrations ranging from 1 mM–10
mM. PCA was performed on the resulting data set and a hierarchical cluster analy-
sis was carried out on the first six principal components of the data. With no prior
knowledge of the glucose concentrations involved, this procedure was able to accu-
rately group the spectra into distinct categories corresponding to the level of glucose:
no glucose, hypoglycaemia, normal blood glucose and hyperglycaemia. No prediction
of absolute glucose concentrations from the spectra was attempted. It should also
be noted that substrates were incubated in the appropriate glucose solutions for an
undisclosed period of time prior to measurement, indicating that the system is not
currently capable of real-time measurement.
Another group has used conventional Raman spectroscopy for the non-invasive,
transcutaneous, determination of blood-glucose levels in human volunteers.105 An 830
54 CHAPTER 4. CURRENT STATE OF THE ART
nm excitation source was focused directly on the skin, producing a spot size of 1 mm2
and a power of 300 mW, and backscattered light was collected by a mirror and trans-
ferred to a spectrometer via a notch filter and an optical-fibre bundle. Each spectrum
was acquired via 90 two-second acquisitions and were taken every five minutes for a
period of 2-3 hours following food consumption. A total of 461 Raman spectra in
the range 355–1545 cm−1 were collected from 17 human volunteers and the data set
from each volunteer was analysed individually via PLSR combined with the leave-one-
out cross-validation method. The measured blood-glucose levels ranged from 3.8–2.4
mM and the mean error in prediction, as compared to reference values obtained by
standard capillary blood analysis, across all 17 subjects was 7.8%.
4.3.3 Raman Spectroscopy & Microfluidics
Raman spectroscopy and SERS have been used in combination with microfluidics to
study analytes under continuous flow.106 One such study is notable for not only car-
rying out SERS analysis, but also the synthesis of the SERS substrate used, using
a microfluidic system.107 The system used a square glass capillary with an internal
diameter of 200 µm. A silver-nanoparticle based SERS substrate was synthesised
through the continuous injection of a solution containing silver nitrate and sodium
citrate into the capillary, at a flow rate of 3.3 µL s−1, whilst a laser beam with a
wavelength of 515 nm was focused on the lower inner face of the capillary. Exposing
the solution to the laser beam for 40s resulted in the in situ photoinduced synthesis
of silver nanoparticles in a compact 7 µm spot on the capillary surface. The synthe-
sised SERS substrate was then used to study a 0.01 mM solution of malachite green,
an antimicrobial agent controversially used in aquaculture in certain countries, which
was continuously injected into the capillary at a rate of 5 µL s−1. Saturation of the
strongest malachite green SERS peak at 1617 cm−1 occurred 40 seconds after begin-
ning the injection of the solution. Desorption of malachite green and regeneration of
the SERS substrate was monitored by reintroducing the silver nitrate/sodium citrate
solution through the capillary at a rate of 3.3 µL s−1 and observing a decrease in the
band at 1617 cm−1.
4.4 Neurochemical Monitoring
Multiple studies report attempts to improve the time resolution of neurochemical mon-
itoring. The most notable of these is the development of rapid sampling microdialysis
(rsMD), a technique that combines clinical microdialysis with an online enzyme-based
electrochemical assay system for the simultaneous detection of glucose and lactate.108
The rsMD system connects directly to the microdialysis tubing. Microdialysate
from the patient is fed into a valve where it is mixed with a solution containing
4.4. NEUROCHEMICAL MONITORING 55
ferrocene monocarboxylate. Every 30 seconds, the valve injects the sample into two
parallel flow streams which each feed into a differenct enzyme reactor: one for glucose
detection and one for lactate detection. Each assay utilises an enzyme bed consisting
of two immobilised enzymes, horseradish peroxidase and either glucose oxidase or
lactate oxidase. The assay consists of four main steps: firstly, any glucose and lactate
present in the sample reacts with their respective oxidase enzyme to produce H2O2 as a
by-product. H2O2 is then converted to water by horseradish peroxidase. This reaction
involves the simultaneous oxidation of ferrocene monocarboxylate which regenerates
the horseradish peroxidase enzyme. The resulting charged ferrocene species is then
reduced by a glassy carbon electrode, located downstream of the enzyme bed, and
the resulting change in current with respect to an Ag—AgCl reference electrode is
detected. This process is summarised in Figure 4.4. The working electrode is held
at -100 mV relative to the reference electrode. The amplitude of the current peak is
related to the concentration of analyte by a calibration process in which solutions of
known concentration are manually injected into the device. Using a flow rate of 2 µL
min−1, the device is able to sample the microdialysate at 30 second intervals.
Figure 4.4: Summary of the reactions of rsMD. The substrate is typically glucose or lactate
and SPx is the substrate oxidase enzyme. HRP is the horseradish peroxidase enzyme and Fc
is a ferrocene species.109
The device has been used in a number of clinical applications, including the mon-
itoring of TBI patients in a critical care unit and during surgery to repair a brain
aneurysm.108 The high time resolution of this system, when compared with conven-
tional analysis of microdialysate samples, has enabled its use in the study of spreading
depolarizations in injured brain tissue. Spreading depolarisations are an acute event,
which can occur spontaneously in the injured brain, involving a large dynamic change
in the distribution of ions between intracellular and extracellular compartments within
the brain. These events typically last 2–5 minutes and therefore a monitoring system
with high time resolution is required to study metabolic changes that occur during
this process. Time-stamped rsMD data were correlated with data generated through
electrocorticography, an electrophysiological monitoring system which used electrodes
placed direcctly on the surface of the brain to record electrical activity, to measure
the metabolic profile of spreading depolarisations.110
56 CHAPTER 4. CURRENT STATE OF THE ART
Another study used the rsMD to monitor the health of tissue used in free-flap
reconstruction survey, a technique used to transfer healthy tissue from one region
of the body to another.111 The associated disconnection and reconnection of blood
supply to the transferred tissue is critical to the success of the procedure. rsMD
was used to assess the metabolic status of the tissue, which could be used to infer
the quality of its blood supply, both during surgery and in the post-operative phase.
In this set-up, the working electrodes were held at 0 V compared to the reference
electrodes.
The rsMD setup has recently been miniaturised in order to further improve the
time resolution of the system.112 The new device is based upon a microfluidic PDMS
chip contaning a 50 µm platinum disc microelectrode coated in a polyethylene glyc-
erol hydrogen matrix containing entrapped glucose oxidase or lactate oxidase, for
glucose or lactate detection respectively, in addition to an Ag—AgCl reference micro-
electrode. This setup utilises molecular oxygen, which is readily accessible due to the
permeability of PDMS to O2, as a mediator for the oxidation of H2O2 produced by the
substrate-substrate oxidase reactions. A potassium sensor, based on a miniaturised
ion-sensing electrode which selectively binds to the ion of interest, is also incorporated
into the system to detect the onset of spreading depolarisation, which is associated
with an increase in potassium levels in the microdialysate. Whilst in use, calibration
of the instrument is carried out every three hours via the introduction of solutions of
known concentration. This device has been used in conjunction with electrocorticog-
raphy for in vivo monitoring of spreading depolarisation in a rat model112 and in a
study of human TBI patients.113 The time resultion of this system, including the time
taken for the sample to reach the sensor chip, is estimated to be 68 seconds from in
vitro experiments.
Although this method has undoubtedly improved the time resolution of micro-
dialysate monitoring, it does not address all of the disadvantages of the ISCUS Flex.
For example, the setup still requires the use of reagents and consumables, including
enzymes, which must be purchased on an on-going basis and have a limited lifespan:
the enzyme beds are estimated to be stable for 1-2 months.108 In addition, glucose
and lactate are indirectly detected, as in the ISCUS Flex, whereas a spectroscopy-
based masurement would allow the direct detection of analytes. It is also noted that
the current rsMD setup enables the monitoring of only two analytes simultaneously,
meaning that pyruvate detection is omitted. Moreover, the reported detection limit
of 0.025 mM for glucose and lactate would not be low enough to encompass the com-
plete clinical range of pyruvate.111 On a more speculative note, it is also possible that
biofouling of the electrodes could present a problem after prolonged use.
An alternative approach to electrochemical on-line monitoring of microdialysate
samples utilised carbon-film electrodes coated onto a glass plate.114 The device con-
4.5. METABOLIC PROFILING OF TBI: PRELIMINARY WORK 57
tained four electrodes: two working electrodes, which were coated with horseradish
peroxidase followed by a layer of either glucose oxidase or lactate oxidase, and two
silver-coated reference electrodes. A 1 mm × 23 µm flow channel contained the
two working electrodes, which were arranged in parallel to reduce chemical cross-talk
between them, with a micro-separator between them. The assembly was sealed by
placing a second glass plate on top of the first and the completed device could then be
connected to a standard microdialysis probe. During device operation, the working
electrode potentials were held at -50 mV, compared to their respective reference elec-
trodes, and flow rates of 0.5–4 µL min−1 could be used. The device was successfully
used for continuous electrochemical monitoring of glucose and lactate levels within
the brain of a live rat.
Microdialysis sampling has also been combined with fluorescence-based detection
methods for in vivo on-line monitoring of rat-brain extracellular glucose and lac-
tate.115 The study used a 3D-printed flow reactor containing separate reaction cham-
bers for glucose and lactate detection, each of which was internally coated with the
appropriate oxidase enzyme. The introduction of microdialysate samples to a reaction
chamber resulted in the oxidation of glucose or lactate, leading to the production of
H2O2 as a byproduct. This reacted with Amplex UltraRed, a H2O2 molecular probe,
to produce a fluorescence-emitting product which was detected by a downstream flu-
orescence spectrophotometer. The assay was designed to ensure that microdialysate
passed through only one of the reaction chambers at a given point, so that H2O2
production could be correctly correlated with either glucose or lactate detection. Mi-
crodialysate loading into each chamber occurred at 4 minute intervals and was pumped
through the system at a rate of 1 µL min−1. The working concentration range of the
device was 0.1–5.0 mM.
4.5 Metabolic Profiling of TBI: Preliminary Work
A previous student in the Elliott group carried out a study into the feasibility of
detecting glucose, lactate and pyruvate, at concentrations relevant to TBI, using ATR-
FTIR spectroscopy in combination with PLSR.116 All measurements were carried out
using a Perkin Elmer Spectrum 100, coupled with a Specac Heated Golden Gate ATR
Accessory containing a single-reflection diamond ATR crystal, at room temperature.
The first stage of the study involved the analysis of synthetic solutions of glucose,
lactate and pyruvate in perfusion fluid. The data from each experiment was analysed
in the same way: spectral data were randomly organised into two groups. The first
group was used to build a predictive PLSR model using cross-validation and the
predictive properties of the model were then tested on the second data set. A plot
of the actual vs. predicted concentration of analyte was then extrapolated to predict
the limit of detection of that compound.
58 CHAPTER 4. CURRENT STATE OF THE ART
This method was initially used to study pure solutions of glucose, lactate and
pyruvate. Each compound was analysed using a different selection of PLSR models,
each of which examined a different spectral range. Three models were used for the
analysis of glucose spectra, three models were used for the analysis of lactate spectra
and a single model covering a wider spectral range was used for the analysis of pyruvate
spectra. Each model used to analyse the glucose and lactate spectra was found to
possess a different predicted limit of detection. The lowest predicted limits of detection
for each compound are listed in Table 4.1.
Mixtures of glucose and lactate in perfusion fluid were analysed in the same way.
The lowest predicted limits of detection of each compound in the mixture are also
reported in Table 4.1.
Table 4.1: The limits of detection of glucose, lactate and pyruvate in synthetic
solutions, using single-reflection ATR-FTIR spectroscopy, as predicted from the
extrapolation of results from a predictive PLSR model.116
Analyte Sample
Limit of
Detection / mM
Glucose Glucose in perfusion fluid 0.3
Lactate Lactate in perfusion fluid 0.5
Pyruvate Pyruvate in perfusion fluid 0.09
Glucose Glucose & lactate in perfusion fluid 0.7
Lactate Glucose & lactate in perfusion fluid 0.4
Subsequent work focused on the analysis of clinical microdialysis samples via ATR-
FTIR spectroscopy. ATR-FTIR spectra of ten microdialysate samples, all of which
were obtained from the same patient, were acquired as described previously. The
PLSR models developed in the first phase of the study were used to predict indi-
vidually the concentrations of glucose, lactate and pyruvate in the samples. The
predictions generated by each model were compared with the actual concentrations of
glucose, lactate and pyruvate in each sample, as measured using the ISCUS Flex Mi-
crodialysis Analyser. Table 4.2 lists the RMSEP of the best-performing PLSR model
for each compound.
Table 4.2: The RMSEP of the best-performing PLSR models used to predict the
concentrations of Glucose, Lactate and Pyruvate in clinical microdialysate sam-
ples.116
Analyte Spectral Range of Model / cm−1 RMSEP / mM
Glucose 1600–750 0.28
Lactate 1600–750 1.07
Pyruvate 1800–1000 0.023
This study provided encouraging results to suggest that ATR-FTIR spectroscopy
may be a suitable technique upon which to base a sensor for metabolic profiling of
4.5. METABOLIC PROFILING OF TBI: PRELIMINARY WORK 59
TBI patients. However, many outstanding questions remained following this work.
Firstly, spectra were acquired using only a single-reflection diamond ATR crystal. It
is possible that a crystal which provides a longer effective path length would yield a
lower limit of detection, due to a greater interaction of the IR beam with the sample,
and therefore it would be interesting to compare the performances of different crystals.
Similarly, transmission-FTIR was not used in this study and it is not clear how its
performance would compare to that of ATR-FTIR.
This study did not attempt to detect glucose, lactate and pyruvate using Raman
spectroscopy or SERS. It would be interesting to compare the performance of these
techniques with FTIR spectroscopy, in order to analyse the relative merits of the two
forms of vibrational spectroscopy when applied to the specific circumstances of TBI
monitoring.
It is also noted that only a single type of statistical model was applied in this
work. In addition, the performance of the model across different regions of the FTIR
spectrum was not studied systematically. Ideally, the performance of multiple sta-
tistical models should be compared and the chosen model should be optimised in a
systematic manner.
Finally, the analysed microdialysate samples were obtained from a single patient
and therefore did not cover the full concentration range of glucose, lactate and pyru-
vate expected in clinical samples. It is therefore unclear how the technique would
perform at the extreme ends of the clinical range.

Chapter 5
Materials & Methods
This chapter provides details of the materials and methods used for experiments and
procedures that were carried out routinely during this work. Any materials or tech-
niques used only for select experiments are described in the relevant results chapters.
5.1 Sample Preparation
5.1.1 Perfusion Fluid
Stock solutions of perfusion fluid were prepared to the same recipe as that used for
microdialysis of TBI patients used in the clinic.19 The compounds listed in Table 5.1
were dissolved in the appropriate volume of deionised (DI) water at room temperature
and mixed thoroughly. The resulting solution was stored at 4 ◦C.
Table 5.1: Components of lab-made perfusion fluid
Compound Supplier
Concentration
/ mM
NaCl Breckland Scientific 147
KCl Sigma Aldrich 2.7
CaCl2 Sigma Aldrich 1.2
MgCl2 Sigma Aldrich 0.85
5.1.2 Synthetic Samples
D-(+)-glucose, sodium L-lactate and sodium pyruvate used in the preparation of syn-
thetic samples were purchased from Sigma Aldrich.
A pure solution of each compound at a concentration of either 1 M or 100 mM was
freshly prepared at the outset of each experiment. The required mass of compound
was weighed to an accuracy of ±0.0001 g and dissolved in lab-made perfusion fluid at
room temperature. Solutions of lower concentration were prepared by serial dilution
61
62 CHAPTER 5. MATERIALS & METHODS
of the master solution with perfusion fluid. All solutions were stored at 4 ◦C.
5.1.3 Clinical Samples
Microdialysate samples from brain-injured patients were collected by clinical staff
at the Neurosciences Critical Care Unit at Addenbrooke’s Hospital over a period of
months. Samples were stored at -80 ◦C and kept on the Unit until required for use in
this project. All patient data remained anonymous.
5.2 FTIR Spectroscopy
FTIR spectra were acquired using a Perkin Elmer Spectrum 100 FTIR spectrometer,
which was operated via the Perkin Elmer ‘Spectrum’ software. The spectrometer
was equipped with a SiC hot spot-stabilised IR light source, a KBr beamsplitter, a
Michelson interferometer and a liquid N2-cooled mercury cadmium telluride (MCT)
detector. An accessory appropriate to the FTIR technique being used was fitted into
the connection port located in the rear of the sample compartment (Figure 5.1).
Figure 5.1: The Perkin Elmer Spectrum 100 IR Spectrometer. Accessories for performing
variations of IR spectroscopy were fitted into the sample compartment, via the connection port
on its rear wall, and aligned with the IR beam, as indicated by the red arrow.
The N2 reservoir of the MCT detector was filled approximately 15 minutes prior to
commencing spectrum acquisition to allow the temperature of the detector to stabilise.
It was later re-filled at 4-hourly intervals to ensure that a contant temperature was
maintained at the detector. The required accessory was cleaned and fitted to the
system and a sample was loaded (see Sections 5.2.1.1, 5.2.1.2 and 5.2.2 for details).
Wherever possible, samples were measured in order of increasing concentration to
5.2. FTIR SPECTROSCOPY 63
minimise the contamination of consecutive measurements. The relevant sample-holder
was cleaned between each sample using deionised (DI) water which was deemed to be
a suitable solvent due to the high water solubility of the compounds being measured.
Mid-IR spectra were collected over the range 4000 - 450 cm−1. Each spectrum was
generated via the accumulation of 150 scans with a resolution of 4 cm−1. Two types of
spectra were routinely generated: single-beam spectra and ratio spectra. Single-beam
spectra displayed the Fourier transform of the raw spectrum. Ratio spectra were
generated by dividing the spectrum of a test sample by the spectrum of a suitable
chemical background, enabling the absorption peaks of the sample to be visualised,
as illustrated in Figure 5.2. The work described in this thesis used perfusion fluid as
the chemical background for the acquisition of FTIR ratio spectra.
5001000150020002500300035004000
Wavenumber / cm -1
0
50
100
150
200
Ar
bi
tra
ry
 U
ni
ts
Single-Beam Spectrum
5001000150020002500300035004000
Wavenumber / cm -1
97
97.5
98
98.5
99
99.5
100
100.5
%
 T
ra
ns
m
iss
io
n
Ratio Spectrum
Figure 5.2: The single-beam and ratio FTIR spectrum of a 24 mM solution of lactate in
perfusion fluid. The ratio spectrum was acquired using perfusion fluid as a background sample.
Both spectra were acquired via ATR-FTIR spectroscopy using a seven-reflection ZnSe crystal.
5.2.1 ATR-FTIR Protocol
Three different ATR-FTIR crystals were used in this work: a single-reflection dia-
mond crystal and two seven-reflection crystals, one composed of Ge and the other
composed of ZnSe. The accessories used to integrate these crystals into the Perkin
Elmer Spectrum 100 are shown in Figure 5.3.
5.2.1.1 Single-Reflection ATR-FTIR Protocol
Single-reflection ATR-FTIR measurements were conducted using the Spectrum 100
coupled with a Specac Heated Golden Gate ATR Accessory containing a single-
reflection diamond ATR crystal (Figure 5.3(a)).
The diamond was cleaned thoroughly at the outset of an experiment and between
samples. Firstly, any liquid sample already present was soaked up using a clean lint-
free wipe. 20 µL DI water was then deposited on the surface of the diamond, left for
64 CHAPTER 5. MATERIALS & METHODS
Figure 5.3: Accessories used for ATR-FTIR measurements. (a) Heated Golden Gate ATR
Accessory containing a single-reflection diamond ATR crystal. (b) Horizontal ATR Sampling
Accessory (shown containing a trough plate fitted with a seven-reflection ZnSe ATR crystal).
30-60 seconds and wiped away. This process was repeated twice. Liquid samples were
loaded by dropping 20 µL of sample directly onto the surface of the diamond using a
micropipette and the anvil was left in the upright position during measurements.
5.2.1.2 Seven-Reflection ATR-FTIR Protocol
Seven-reflection ATR-FTIR measurements were carried out using the Spectrum 100
coupled with a Horizontal ATR (HATR) Sampling Accessory from Perkin Elmer (Fig-
ure 5.3(b)). A trough plate containing either a ZnSe or germanium crystal, both of
which had an angle of incidence of 45◦, was mounted into a slot on the upper face of
the accessory.
Identical protocols were used for the measurement and cleaning of both top plates.
The crystal was cleaned thoroughly before beginning an experiment and between
samples. Any residual liquid in the sample trough was removed using a pipette and
the trough was then filled with approximately 1 ml deionised water which was left
for 30-60 seconds before being removed. This process was repeated twice. A sample
volume of 1 ml was used for all measurements and samples were loaded into the trough
using a micropipette.
5.2.2 Transmission-FTIR Protocol
Transmission-FTIR spectra were acquired using a Specac Pearl Liquid Transmission
Accessory coupled with a Specac Oyster Lift and Tilt Cell Assembly (Figure 5.4). The
Oyster cell is the sample holder for this accessory and is manufactured with a fixed
5.3. RAMAN SPECTROSCOPY 65
path length. Each Oyster cell used in this work was fitted with ZnSe windows. The
bottom window of each cell was fixed at an angle of 0.1◦ relative to the top window to
minimise the generation of interference fringes in the IR spectra. The Oyster cell was
held in a horizonal configuration within the Pearl unit, which contained the optics
necessary to direct the IR beam through the cell and into the detector.
Figure 5.4: The Pearl Liquid Transmission Accessory used for transmission-FTIR measure-
ments. Images show the accessory with an open sample tray containing an open Oyster Lift
and Tilt Cell, ready for sample loading (a); an open sample tray with a closed Oyster Lift and
Tilt Cell containing a liquid sample (b); a closed sample tray (c).
In order to load a sample into the accessory, an open Oyster cell was placed on the
Pearl sample tray (Figure 5.4(a)) and a micropipette was used to load a 10 µL sample
onto the bottom window of the cell. The cell was then closed, causing the sample to
spread evenly between the two windows (Figure 5.4(b)) and the sample tray pushed
into the Pearl unit (Figure 5.4(c)).
The Oyster cell was cleaned before use and between samples. Initially, any residual
sample was soaked up using lint-free wipes. A pipette was then used to cover both
windows of the cell with DI water, which was left for approximately 30-60 s before
being soaked up using a lint-free wipe. This was repeated twice.
5.3 Raman Spectroscopy
5.3.1 Instrument Specifications
Raman spectra were acquired using a dispersive HORIBA LabRAM HR Evolution Ra-
man microscope which was operated via LabSpec 6 software (HORIBA Jobin Yvon).
The system was equipped with laser light sources of four different wavelengths (785
nm, 633 nm, 532 nm, 473 nm) and an air-cooled CCD detector. An upright confocal
microscope was integrated into the instrument and its objectives were surrounded by
66 CHAPTER 5. MATERIALS & METHODS
a Class 1 enclosure to control ambient light interference.
The system contained two options for mounting samples within the microscope
enclosure (Figure 5.5). Solid or liquid samples could be placed on a glass slide and
mounted on the motorised XY stage beneath the microscope objectives. Liquid sam-
ples could alternatively be contained within a cuvette and mounted in the ’macro
sampler’, an accessory that connected directly to the microscope nose piece and con-
tained the optics necessary to direct laser light from the instrument through the
cuvette and reflect scattered light back into the instrument.
Figure 5.5: Raman spectroscopy sample-mounting options. Samples could be placed under a
microscope objective (a) or within the macro sampling accessory (b). The blue auto-calibration
objective is also visible in Figure (a).
5.3.2 Instrument Operation
The instrument was calibrated prior to the start of a series of measurements and was
re-calibrated each time there was a change to the laser or grating being used. Auto-
calibration was carried out using a proprietary calibration objective from HORIBA,
the Standard PRO (seen in Figure 5.5(a)). The objective contained a disc of polystyrene
which was used as a reference sample. This was fixed at a constant distance from the
objective lens and therefore increased the consistency of calibration measurements by
eliminating the need to manually focus on the sample. The system was calibrated to
a Raman peak of polystyrene at 1001 cm−1, corresponding to the breathing mode of
its aromatic ring.
Samples mounted on the XY stage were illuminated using a white-light source and
imaged using a camera via the LabSpec 6 software. A positioning joystick was used to
alter both the XY stage position and the distance of the microscope objective from the
sample, until the desired region of the sample was in focus. Liquid samples mounted
in the macro sampler did not require any further adjustment. The parameters used
for spectrum acquisition, including acquisition time, number of accumulations, laser
5.4. DATA ANALYSIS 67
power and grating, were then set using the LabSpec 6 software and acquisition was
initiated.
Acquisition parameters were varied with each experiment and are detailed with
the corresponding experimental results in Chapter 7. Data files were saved in .txt
format for importing to MATLAB and also in LabSpec 6 format in order to preserve
the experimental details.
5.4 Data Analysis
All data analysis was carried out using MATLAB version R2016b.
5.4.1 ANOVA and Multiple Comparison Tests
One-way ANOVA tests were performed using the ‘anova1’ function of the Statistics
and Machine Learning Toolbox—. The multiple comparison tests were carried out
using the ‘multcompare’ function, also included in the Statistics and Machine Learning
Toolbox—.
5.4.2 Statistical Modelling: Pre-Processing
Two pre-processing steps were carried out on spectral data used for statistical mod-
elling. Firstly, the transmission data were converted into values of absorbance using
the equation:
A = log10
100
%T
(5.1)
where A is absorbance and %T is percentage transmission. Absorbance data were
used in the modelling process because they have a linear relationship to concentration
(the independent variable in the model) as defined by the Beer-Lambert law:
log
I0
I
= εcl (5.2)
where I0 is the intensity of the incident radiation, I is the intensity of the transmitted
radiation, c is the sample concentration, l is the path length of the sample and ε is
the extinction coefficient, which is a constant for a given transition within a particular
molecule.
Secondly, the absorbance data were mean-centred. The mean value of the ab-
sorbance at each wavenumber was subtracted from the value of the absorbance at the
corresponding wavenumber, resulting in a dataset which had a mean absorbance value
of zero at each wavenumber.
68 CHAPTER 5. MATERIALS & METHODS
5.4.3 Statistical Modelling: PCR
A PCA was performed on the absorbance data using the ‘pca’ function of the Statistics
and Machine Learning Toolbox—. A linear regression of the absorbance data was then
performed on the principal components using the ‘regress’ function. The PCR results
were subsequently transformed to regression coefficients for the original variables of
absorbance and concentration for interpretation.
The mean-squared error in prediction of the model was calculated using the leave-
one-out method of cross-validation, in which the model was constructed using the
data from all but one of the spectra within the dataset and the resulting model was
used to predict the concentrations of glucose, lactate and pyruvate represented by
the remaining ‘test’ spectrum. This process was repeated until each spectrum in the
dataset had been used as a test spectrum.
5.4.4 Statistical Modelling: PLSR
A PCA was performed on the absorbance data prior to carrying out the PLSR in order
to compress the dataset and therefore speed up the procedure. The regression analysis
was carried out using the ‘plsregress’ function of the Statistics and Machine Learning
Toolbox—and was calibrated using the leave-one-out method of cross-validation.
Chapter 6
IR Spectroscopy of Glucose,
Lactate & Pyruvate
This chapter describes the preliminary work undertaken to determine whether mid-
FTIR spectroscopy could be used to detect glucose, lactate and pyruvate in clini-
cal microdialysate samples, within the context of the device requirements set out in
Section 2.6. Two methods of FTIR spectroscopy were evaluated: ATR-FTIR and
transmission-FTIR. ATR-FTIR was selected for investigation as it is relatively insen-
sitive to the effect of strongly IR-absorbing samples, such as water, and the final device
will be required to detect glucose, lactate and pryuvate in highly dilute aqueous solu-
tions. Transmission-FTIR was chosen, despite its sensitivity to strongly IR-absorbing
samples, due to its instrumental simplicity which renders it likely to be the most
straightforward of the FTIR techniques to develop into a miniaturised bedside device.
Both techniques were used to acquire the FTIR spectra of glucose, lactate and
pyruvate, as individual solutions and as a mixture, which were qualitatively anal-
ysed to identify spectral peaks unique to each compound. Differences between the
ATR-FTIR and transmission-FTIR spectra of each compound were also noted. Con-
sideration was given to compounds other than the analytes of interest which are likely
to be present in perfusion fluid and, where possible, their IR spectra were analysed
to determine regions of overlap with the spectra of glucose, lactate and pyruvate.
Transmission-FTIR spectroscopy was subsequently used to analyse six clinical micro-
dialysate samples in order to establish whether any compound-specific peaks identified
from this preliminary work could be observed. Finally, quantitative studies were car-
ried out in order to determine which ATR-FTIR crystal and transmission-FTIR path
length should be used in the next phase of work.
69
70 CHAPTER 6. IR SPECTROSCOPY
6.1 FTIR Spectra of Glucose, Lactate & Pyruvate
The aim of this section of work was to characterise the mid-FTIR spectra of glucose,
lactate and pyruvate, both individually and in a mixture, in order to find peaks
specific to each compound which are not compromised by the presence of the other two
analytes. Each compound was prepared both as a pure 100 mM solution and as part
of a mixture of glucose, lactate and pyruvate in which all compounds were present at
a concentration of 100 mM: this concentration was chosen to ensure that absorbance
peaks would be clearly visible whilst not saturating the measured spectrum. All
solutions were prepared with perfusion fluid, to better reflect the conditions under
which these compounds would be detected in the clinic, and a spectrum of pure
perfusion fluid was used as a background reference. The following sections discuss,
in turn, the FTIR spectra of glucose, lactate, pyruvate and a mixture of all three
compounds. All spectra shown are ratio spectra, which have been corrected for the
presence of perfusion fluid, in order to facilitate the identification of IR absorption
peaks corresponding to the target molecules.
6.1.1 FTIR Spectrum of Glucose
The ATR-FTIR and transmission-FTIR spectra of a 100 mM solution of glucose in
perfusion fluid, in addition to their associated peak assignments, are shown in Figure
6.1 and Figure 6.2, respectively. The most distinctive feature of both spectra is a
cluster of strong peaks in the region 1152 – 994 cm−1, which correspond to vibrations
of the C–O and C–C bonds within the glucose molecule. Literature values indicate
that these are the most significant peaks for the identification of glucose via mid-FTIR
spectroscopy.73
A clear difference between the ATR-FTIR and transmission-FTIR spectra is the
presence in the latter of a high level of noise in the regions 3700 – 3000 cm−1, 1700
– 1600 cm−1 and 900 – 450 cm−1. This is caused by the strong absorption of IR
radiation by water which, due to the significant interaction between the IR beam and
the sample, has saturated the detector of the instrument. The ATR-FTIR spectrum
does not exhibit the same level of noise in these water-absorption regions, as the
effective path length used to acquire the spectrum is significantly shorter than that
used in transmission-FTIR spectroscopy (Section 3.2.3.2). It should be noted that a
proportion of the noise seen in the 800 – 450 cm−1 region of both spectra is due to
the limited IR transmission range of ZnSe, the material from which the transmission
cell windows and ATR crystal material are composed, which extends from 20,000 –
650 cm−1.
The water-induced noise in the transmission-FTIR spectrum of glucose leads to
the loss of approximately 40% of the peaks seen in the ATR-FTIR spectrum of the
6.1. IR SPECTRUM CHARACTERISATION 71
molecule. However, the majority of these are weak bands seen at the extreme ends of
the spectrum and the strongest peaks in the region 1152 – 994 cm−1 are still visible.
An additional consequence of the short effective path length of ATR-FTIR, in
comparison with transmission-FTIR, is a decrease in the intensity of all IR absorption
bands in the ATR-FTIR spectrum when compared with transmission-FTIR. It can
be seen from Figures 6.1 and 6.2 that the strongest IR absorption peak at 1035 cm−1
is approximately 30% stronger in the transmission-FTIR spectrum of glucose than in
the ATR-FTIR spectrum.
72 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
88
90
92
94
96
98
100
102
%
 T
ra
ns
m
is
sio
n
Peak Position /
cm−1
Peak
Description
Assignment
3531 Weak, Broad O–H Stretch
2924 Medium C–H Stretch
2873 Weak O–H Stretch
2858 Weak O–H Stretch
1433 Medium CH2 Bend
1364 Medium O-H bend
1319 Weak C-C (Alicyclic) Vibration
1287 Weak C-C (Alicyclic) Vibration
1260 Weak C-C (Alicyclic) Vibration
1204 Weak C-C (Alicyclic) Vibration
1152 Medium C-C (Alicyclic) Vibration
1107 Medium C-C (Alicyclic) Vibration
1081 Strong C–O Stretch
1035 Strong C–O Stretch
994 Medium C-C (Alicyclic) Vibration
919 Weak C–O–C Stretch
901 Weak C–O–C Stretch
864 Weak C–O–C Stretch
842 Weak C–O–C Stretch
Figure 6.1: The ATR-FTIR ratio spectrum and associated peak assignments117 of a 100
mM solution of glucose in perfusion fluid, acquired using a seven-reflection ZnSe crystal.
6.1. IR SPECTRUM CHARACTERISATION 73
5001000150020002500300035004000
Wavenumber / cm -1
60
70
80
90
100
110
%
 T
ra
ns
m
is
sio
n
Peak Position /
cm−1
Peak
Description
Assignment
2892 Medium, Broad O–H Stretch
1429 Medium CH2 Bend
1364 Medium C–O–H bend
1287 Weak C–C (Alicyclic) Vibration
1262 Weak C–C (Alicyclic) Vibration
1203 Weak C–C (Alicyclic) Vibration
1151 Medium C-C (Alicyclic) Vibration
1107 Medium C–C (Alicyclic) Vibration
1080 Strong C–O Stretch
1035 Strong C–O Stretch
994 Medium C-C (Alicyclic) Vibration
Figure 6.2: The transmission-FTIR ratio spectrum and associated peak assignments117 of a
100 mM solution of glucose in perfusion fluid, acquired using a 50 µm path length.
74 CHAPTER 6. IR SPECTROSCOPY
6.1.2 FTIR Spectrum of Lactate
The ATR-FTIR and transmission-FTIR spectra, and associated peak assignments, of
a 100 mM solution of lactate in perfusion fluid are shown in Figures 6.3 and 6.4. Both
spectra are characterised by three distinct regions: a variable number of strong peaks
at approximately 1576 cm−1, corresponding to the stretching of the CO−2 functional
group, and two groups of medium-intensity peaks at approximately 1456–1315 cm−1
and 1125–1041 cm−1. The IR absorption peak centred around 1576 cm−1 is not a
specific indicator of the presence of lactate when used in isolation, despite producing
the strongest peaks, as all molecules with a carbonyl group are known to strongly
absorb IR radiation in that region of the spectrum. Literature values suggest that the
peak at 1125 cm−1, which corresponds to a C–O stretching vibration, is therefore the
most useful in the identification of lactate.73
A comparison of the ATR-FTIR and transmission-FTIR spectra of lactate yields
broadly similar conclusions to those drawn from the comparison of glucose FTIR
spectra. Significant water-induced noise is seen in the transmission-FTIR spectrum,
but not in the ATR-FTIR spectrum, resulting in the loss of peaks at the extreme ends
of the transmission-FTIR spectrum. In addition, the transmission-FTIR spectrum
yields more intense peaks than the ATR-FTIR spectrum, with the strongest peaks
appearing approximately 30% more intense in the transmission-FTIR spectrum.
One further difference between the ATR-FTIR and transmission-FTIR spectra
of lactate is the appearance of additional peaks in the transmission-FTIR spectrum.
These peaks are seen at 1580 cm−1, 1567 cm−1 and 1556 cm−1 and overlap very
strongly with the central peak, which is attributed to an asymmetric stretch of the
CO−2 functional group, at 1572 cm
−1.
The additional peaks are also thought to correspond to asymmetric stretching
modes of the CO−2 group. It is throught that their absence from the ATR-FTIR
spectrum is due to an orientation effect caused by the preferential adsorption of the
lactate ion in a particular orientation on the ZnSe crystal surface. Orientation effects
in ATR-FTIR spectroscopy arise due to the fact that the electric field of incident IR
radiation has components both parallel to and perpendicular to the plane of incidence.
The effective path length of a ZnSe crystal is known to be higher for the component
of the electric field which is parallel to the plane of incidence, increasing the relative
intensity of absorption peaks that correspond to vibrational modes with a dipole
change in the plane of incidence.118,119 If the lactate ion exhibits prefential adsorption
in a particular orientation on the crystal surface, it would result in certain vibrational
modes always having a dipole change perpendicular to the plane of incidence and
therefore their associated absorption peaks would be of significantly lower relative
intensity. Alternatively, the preferential adsorption of the lactate ion in a particular
orientation on the ZnSe ATR crystal surface may cause the restriction of specific
6.1. IR SPECTRUM CHARACTERISATION 75
vibrational modes of the CO−2 group or render them IR-inactive. These phenomena
would not affect transmission-FTIR spectra as the bulk of the solution is sampled,
causing any orientation effects to be averaged out, and thus the absorption peaks of
all IR-active vibrational modes can be seen.
5001000150020002500300035004000
Wavenumber / cm -1
90
92
94
96
98
100
%
 T
ra
ns
m
iss
io
n
Peak Position /
cm−1
Peak
Description
Assignment
2984 Weak O–H Stretch
2941 Weak O–H Stretch
1575 Strong CO−2 Stretch (Asym.)
1456 Medium CH3 Bend
1417 Medium CO−2 Stretch (Sym.)
1367 Medium CH3 Bend
1315 Medium C–O–H Bend
1125 Medium C–O Stretch
1086 Weak C–O Stretch
1041 Medium C–O Stretch
927 Weak C–C (Aliphatic) Vibration
854 Weak C–C (Aliphatic) Vibration
Figure 6.3: The ATR-FTIR ratio spectrum and associated peak assignments117 of a 100 mM
solution of lactate in perfusion fluid, acquired using a seven-reflection ZnSe crystal. Asym. =
asymmetric vibration, Sym. = symmetric vibration.
76 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
60
65
70
75
80
85
90
95
100
105
%
 T
ra
ns
m
is
sio
n
Peak Position /
cm−1
Peak
Description
Assignment
2945 Medium, broad O–H Stretch
1580 Strong CO−2 Stretch (Asym.)
1572 Strong CO−2 Stretch (Asym.)
1567 Strong CO−2 Stretch (Asym.)
1556 Medium CO−2 Stretch (Asym.)
1456 Medium CH3 Bend
1417 Medium CO−2 Stretch (Sym.)
1365 Medium CH3 Bend
1316 Medium C–O–H Bend
1125 Medium C–O Stretch
1087 Weak C–O Stretch
1041 Medium C–O Stretch
Figure 6.4: The transmission-FTIR ratio spectrum and associated peak assignments117 of a
100 mM solution of lactate in perfusion fluid, acquired using a 50 µm path length. Asym. =
asymmetric vibration, Sym. = symmetric vibration.
6.1. IR SPECTRUM CHARACTERISATION 77
6.1.3 FTIR Spectrum of Pyruvate
The ATR-FTIR and transmission-FTIR spectra of a 100 mM solution of pyruvate in
perfusion fluid can be seen in Figures 6.5 and 6.6, respectively. Peak assignments for
both spectra are also given.
The ATR-FTIR and transmission-FTIR spectra of pyruvate are both distinguished
by four groups of medium/strong intensity peaks, which are approximately equidistant
from each other, in the regions 1710 cm−1, 1600 cm−1, 1425 – 1356 cm−1 and 1176
cm−1. However, a clear difference is observed in the appearance of these peaks between
the two spectra: in the ATR-FTIR spectrum, the peaks are of varying intensity, with
that at 1708 cm−1 being the strongest, whilst in the transmission-FTIR spectrum,
the peaks at 1711 cm−1, 1583 cm−1, and 1356 cm−1 are of similar intensity and the
peak at 1176 cm−1 is notably stronger.
There also exists a difference in the number and position of peaks between the
ATR-FTIR and transmission-FTIR spectra of pyruvate. The ATR-FTIR spectrum
peak at 1600 cm−1 is absent in the transmission-FTIR spectrum and is replaced by
two peaks at 1583 cm−1 and 1572 cm−1. All peaks are attributed to the asymmetric
stretching of the CO−2 functional group. It is theorised that these variations in peak
intensity, number and position are a result of the orientation effect observed in the
FTIR spectra of lactate. Preferential adsorption of the pyruvate ion in a specific ori-
entation on a ZnSe surface may cause a relative loss of intensity of peaks arising from
vibrational modes which have a dipole change perpendicular to the plane of incident
IR radiation. Alternatively, orientation-specific adsorption may restrict certain vibra-
tional modes of the molecule or render them IR-inactive. The effect of this would
be noticeable in an ATR-FTIR spectrum, which is acquired through analysing the
sample surface, but in transmission-FTIR the effect would be negligible as the bulk
of the sample is analysed.
Differences between the ATR-FTIR and transmission-FTIR spectra due to water
absorption, as observed for glucose and lactate, are also seen in the spectra of pyruvate.
Water-induced noise is present in the 3700 – 3000 cm−1, 1700 – 1600 cm−1 and 900
– 450 cm−1 regions of the transmission-FTIR spectrum, resulting in the loss of peaks
at either end of the spectrum.
78 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
90
92
94
96
98
100
102
%
 T
ra
ns
m
is
sio
n
Peak Position /
cm−1
Peak
Description
Assignment
3511 Weak, broad O-H Stretch
1708 Medium C=O stretch
1600 Strong CO−2 Stretch (Asym.)
1423 Medium CO−2 Stretch (Sym.)
1397 Medium CO−2 Stretch (Sym.)
1356 Medium CH3 Bend
1176 Strong C–C (Aliphatic) Vibration
1021 Weak C–C (Aliphatic) Vibration
982 Weak CH3 Bend
950 Weak C–C (Aliphatic) Vibration
836 Weak C–C (Aliphatic) Vibration
Figure 6.5: The ATR-FTIR ratio spectrum and associated peak assignments117 of a 100 mM
solution of pyruvate in perfusion fluid, acquired using a seven-reflection ZnSe crystal. Asym.
= asymmetric vibration, Sym. = symmetric vibration.
6.1. IR SPECTRUM CHARACTERISATION 79
5001000150020002500300035004000
Wavenumber / cm -1
75
80
85
90
95
100
105
110
%
 T
ra
ns
m
is
si
on
Peak Position /
cm−1
Peak
Description
Assignment
1711 Strong C=O stretch
1583 Strong CO−2 Stretch (Asym.)
1572 Medium CO−2 Stretch (Aym.)
1424 Medium CO−2 Stretch (Sym.)
1398 Medium CO−2 Stretch (Sym.)
1356 Strong CH3 Bend
1176 Strong C–C (Aliphatic) Vibration
1021 weak C–C (Aliphatic) Vibration
Figure 6.6: The transmission-FTIR ratio spectrum and associated peak assignments117 of a
100 mM solution of pyruvate in perfusion fluid, acquired using a 50 µm path length. Asym.
= asymmetric vibration, Sym. = symmetric vibration.
80 CHAPTER 6. IR SPECTROSCOPY
6.1.4 Mixtures of Glucose, Lactate and Pyruvate
A mixture containing 100 mM glucose, 100 mM lactate and 100 mM pyruvate in per-
fusion fluid was analysed via both ATR-FTIR and transmission-FTIR spectroscopy.
The aim of these measurements was to allow a comparison of the mixture spectra with
those of the pure compounds, in order to identify any compound-specific peaks which
remain visible in the mixture spectra, as well as those which are lost due to overlap
with the peaks of another compound. In this way, it was possible to determine key
peaks which could be used to identify each individual compound from a mixture of
the three.
The contribution of each compound to the overall mixture spectrum is most easily
visualised by overlaying the spectra of individual compounds with that of the mixture.
This is exemplified in Figure 6.7 which, for clarity, shows only the fingerprint region
of each spectrum.
800900100011001200130014001500160017001800
Wavenumber / cm -1
84
86
88
90
92
94
96
98
100
102
%
 T
ra
ns
m
iss
io
n
Glucose Lactate Pyruvate Mixture
Figure 6.7: The ATR-FTIR ratio spectrum of a mixture containing 100 mM glucose, 100
mM lactate and 100 mM pyruvate in perfusion fluid, overlaid witth the ATR-FTIR spectra of
pure 100 mM solutions of each compound to show the contribution of each individual compound
to the mixture spectrum. All spectra were acquired using a seven-reflection ZnSe crystal.
The ATR-FTIR and transmission-FTIR spectra of the mixture are shown in Fig-
ures 6.8 and 6.9, respectively, in addition to an estimate of the molecular origin of
each peak.
The resulting spectra indicate that a number of molecule-specific peaks remain
6.1. IR SPECTRUM CHARACTERISATION 81
visible when glucose, lactate and pyruvate are combined in solution. Within the ATR-
FTIR spectrum, peaks unique to glucose can be found at 3524 cm−1, 2926 cm−1, 1266
cm−1, 1153 cm−1, 1081 cm−1, 993 cm−1 and 900 cm−1. All but two of these peaks
are retained in the transmission-FTIR spectrum, with only the peaks at 3524 cm−1
and 900 cm−1 being lost. The strongest of the glucose-specific peaks are those at
1081 cm−1 and 993 cm−1, corresponding respectively to a C–O stretching vibration,
and a C–C vibration. The peak at 1035 cm−1, which is one of those identified in
the literature as being important for glucose identification, overlaps strongly with the
1041 cm−1 peak corresponding to a C–O stretching mode of lactate. This overlap
results in a single strong peak, positioned at 1038 cm−1, in both the ATR-FTIR
and the transmission-FTIR spectra of the mixture. The data in Figure 6.7 suggests
that, when both compounds are present at equal concentrations, glucose makes a
more significant contribution to this peak, yet nevertheless it cannot be considered
glucose-specific.
Lactate-specific peaks can be seen in the ATR-FTIR spectrum at 2982 cm−1,
1456 cm−1, 1316 cm−1, 925 cm−1 and 853 cm−1. The strongest of these peaks are
those at 1456 cm−1 (corresponding to a CH3 bend) and at 1316 (corresponding to
a C–O–H bend). An analysis of the transmission-FTIR spectrum shows that the
peaks at 2982 cm−1, 925 cm−1 and 853 cm−1 are lost due to water-induced spectral
noise, but additional strong peaks are seen at 1567 cm−1 and 1556 cm−1. These
peaks correspond to asymmetric stretching vibrations of the CO−2 group and represent
two of the additional peaks seen in the transmission-FTIR spectrum of pure lactate.
Additional peaks corresponding to vibrations of the CO−2 functional group are present
in both the ATR-FTIR and the tranmsission-FTIR spectra of the mixture, but these
overlap entirely with peaks corresponding to the vibration of the same moiety within
the pyruvate molecule and hence are not lactate-specific.
There are 3 pyruvate-specific peaks present in both the ATR-FTIR and the transmission-
FTIR spectra: one corresponding to a C=O stretch (approximately 1710 cm−1), one
corresponding to a bending mode of the CH3 group (1357 cm
−1) and one correspond-
ing to a C–C vibration (1176 cm−1). All of these are of medium/strong intensity. The
ATR-FTIR spectrum contains an additional weak pyruvate-specific peak at 842 −1
which is lost in the transmission-FTIR spectrum due to water-induced spectral noise.
6.1.5 Mid-FTIR Spectra of Glucose, Lactate & Pyruvate: Conclu-
sions
The key conclusion of this section of work is that both the ATR-FTIR and transmission-
FTIR spectra of glucose, lactate and pyruvate exhibit peaks which allow each com-
pound to be individually identified from the FTIR spectra of an equimolar mixture
of all three compounds. These peaks occur primarily in the fingerprint region of the
82 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
84
86
88
90
92
94
96
98
100
102
%
 T
ra
ns
m
iss
io
n
Peak
Position /
cm−1
Peak
Description
Assignment
3524 Weak, Broad O–H Stretch: Glu
2982 Weak, Broad O–H Stretch: Lac
2926 Medium, Broad C–H Stretch: Glu
1709 Medium C=O Stretch: Pyr
1583 Strong CO−2 Stretch (Asym.): Lac & Pyr
1456 Medium CH3 Bend: Lac
1418 Medium, broad CO−2 Stretch (Sym.): Lac & Pyr
1357 Medium CH3 Bend: Pyr
1316 Medium C–O–H Bend: Lac
1266 Weak C–C (Alicyclic) Vibration: Glu
1176 Medium C–C (Aliphatic) Vibration: Pyr
1153 Weak C–C (Alicyclic) Vibration: Glu
1123 Medium, Broad C–O Stretch: Lac & Glu
1081 Strong C–O Stretch: Glu
1038 Strong C–O Stretch: Glu & Lac
993 Medium C–C (Alicyclic) Vibration: Glu
925 weak C–C (Aliphatic) Vibration: Lac
900 Weak C–O–C Stretch: Glu
853 Weak C–C (Aliphatic) Vibration: Lac
842 Weak C–C (Aliphatic) Vibration: Pyr
Figure 6.8: The ATR-FTIR ratio spectrum and associated peak assignments117 of a mixture
containing 100 mM glucose, 100 mM lactate and 100 mM pyruvate in perfusion fluid, acquired
using a seven-reflection ZnSe crystal. Glu = glucose, Lac = lactate, Pyr = pyruvate, Asym.
= asymmetric vibration, Sym. = symmetric vibration.
6.1. IR SPECTRUM CHARACTERISATION 83
5001000150020002500300035004000
Wavenumber / cm -1
60
70
80
90
100
110
%
 T
ra
ns
m
is
si
on
Peak
Position /
cm−1
Peak
Description
Assignment
2923 Medium, Broad C–H Stretch: Glu
1711 Medium C=O Stretch: Pyr
1580 Strong CO−2 Stretch (Asym.): Lac & Pyr
1572 Strong CO−2 Stretch (Asym.): Lac & Pyr
1567 Strong CO−2 Stretch (Asym.): Lac
1556 Medium CO−2 Stretch (Asym.): Lac
1455 Medium CH3 Bend: Lac
1419 Strong, Broad CO−2 Stretch (Sym.): Lac & Pyr
1357 Strong CH3 Bend: Pyr
1316 Medium C–O–H Bend: Lac
1266 Weak C–C (Alicyclic) Vibration: Glu
1176 Strong C–C (Aliphatic) Vibration: Pyr
1151 Medium C–C (Alicyclic) Vibration: Glu
1123 Strong, Broad C–O Stretch: Lac & Glu
1081 Strong C–O Stretch: Glu
1038 Strong C–O Stretch: Glu & Lac
994 Medium C–C (Alicyclic) Vibration: Glu
Figure 6.9: The transmission-FTIR ratio spectrum and associated peak assignments117 of a
mixture containing 100 mM glucose, 100 mM lactate and 100 mM pyruvate in perfusion fluid,
acquired using a 50 µm path length. Glu = glucose, Lac = lactate, Pyr = pyruvate, Asym. =
asymmetric vibration, Sym. = symmetric vibration.
84 CHAPTER 6. IR SPECTROSCOPY
spectrum and are listed in Table 6.1.
Table 6.1: FTIR peaks of glucose, lactate and pyruvate that allow the distinction
of each compound when combined in an equimolar mixture.
Compound Spectrum Type Unique Peaks / cm−1
Glucose
ATR-FTIR
3524, 2926, 1266, 1153, 1081,
993, 900
Transmission-FTIR 2923, 1266, 1151, 1081, 994
Lactate
ATR-FTIR 2982, 1456, 1316, 925, 853
Transmission-FTIR 1567, 1556, 1455, 1316
Pyruvate
ATR-FTIR 1709, 1357, 1176, 842
Transmission-FTIR 1711, 1357, 1176
This work has also highlighted differences between the ATR-FTIR and transmission-
FTIR spectra of a given compound. The ATR-FTIR spectrum is less affected by the
presence of water, meaning that a greater number of analyte peaks can be seen, but
the overall peak intensity is lower than that of transmission-FTIR. Conversely, the
transmission-FTIR spectrum exhibits a higher peak intensity at the expense of a
higher level of water-induced noise, causing the loss of peaks at the extreme ends of
the spectrum. In addition, the transmission-FTIR spectra of lactate and pyruvate
exhibited differences in the number and relative intensity of peaks in the fingerprint
region of the spectrum, which have been attributed to orientation-specific adsorption
of the CO−2 group on the surface of the ATR-FTIR crystal.
6.2 Interference from Competing Molecules
The clinical microdialysate samples to be analysed by the developed sensor are likely
to contain compounds other than glucose, lactate and pyruvate, as discussed in Section
2.3.2. The aim of the work in this section was to consider the ‘competing molecules’
which are likely to be present in clinical microdialysate samples and assess the degree
to which they may affect the performance of a mid-FTIR spectroscopy sensor for
glucose, lactate and pyruvate.
Of the compounds discussed in Section 2.3.2, the effect of drug molecules on the
performance of the sensor is the most difficult to estimate. Patients with TBI often
have multiple medical needs and may be treated with a variety of different drugs,
in addition to the prophylactic anti-seizure medication mentioned in Section 2.3.2.
Furthermore, the concentration of a given medication in the microdialysate will vary
between patients, depending upon the dose administered and the rate at which it is
metabolised. As a result, it was concluded that the only way to account for the effect
of drug molecules on the performance of the sensor would be to carry out extensive
6.2. INTERFERENCE FROM COMPETING MOLECULES 85
measurements on multiple clinical samples. This was not possible in this work due to
the scarcity of clinical samples and is recommended as a priority during the clinical
testing phase of the final device.
Cytokines are a group of proteins which mediate the inflammatory response to
injury or infection and are small enough to be recovered by microdialysis. However,
these molecules are reported to be present in microdialysate at picomolar concentra-
tions,8 which are approximately 106 times lower than the lowest expected concentra-
tion of any of the analytes of interest, and are therefore not considered likely to have
a significant impact on sensor performance.
Nitric oxide is used within the body to regulate blood flow. It is highly reactive
and is therefore unlikely to be present in significant concentrations in microdialysate,
yet its products (nitrates and nitrites, collectively termed NOx) have been successfully
collected by microdialysis. A pilot study which measured the total concentration of all
NOx products recovered using microdialysis, across 13 catheters placed in 12 patients,
established a mean concentration of 32.7 ± 16.8 µM.9 This concentration is 3-fold
lower than the minimum expected concentrations of glucose and lactate in perfusion
fluid but is within the range of expected pyruvate concentrations, although at the
lower end. However, it is difficult to investigate the impact that NOx may have on the
performance of an IR-based sensor, as the measured mean concentration encompasses
a large number of different compounds. It is not clear which specific compounds are
present, nor the level of contribution they make to the overall concentration of NOx.
It was therefore concluded that, as with drug molecules, the effect of NOx on the final
sensor would be best established during the clinical testing phase.
Glycerol is a component of cell membranes and therefore its presence in the extra-
cellular fluid has been implicated as an indicator of cellular breakdown. The reported
range of observed glycerol concentrations in the microdialysate of TBI patients ex-
tends from approximately 10 – 400 µM.16,120,121 This range overlaps with the expected
clinical pyruvate concentration range, and with the lower end of the expected glucose
and lactate concentration ranges, and it was thus considered appropriate to investigate
the FTIR spectrum of glycerol to identify regions of possible overlap with the spectra
of pyruvate, lactate and glucose. The ATR-FTIR and transmission-FTIR spectra of
a 100 mM solution of glycerol in perfusion fluid are shown in Figures 6.10 and 6.11,
respectively.
Glutamate, a neurotransmitter, is also found in the microdialysate of TBI patients.
A 2011 analysis of microdialysate samples from 223 patients reported glutamate con-
centrations of 1 – 13 µM.16 The presence of glutamate in microdialysate samples may
therefore affect pyruvate detection, as the upper end of this range overlaps with the
lower end of the pyruvate concentration range expected in microdialysate samples.
The detection of glucose and lactate is unlikely to be affected, as the reported glu-
86 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
93
94
95
96
97
98
99
100
101
%
 T
ra
ns
m
is
sio
n
Peak
Position /
cm−1
Peak
Description
Assignment
2945 Weak CH2 Stretch
2889 Weak CH2 Stretch
1461 Weak CH2 Bend
1419 Weak C–H Bend
1334 Weak C–H Bend
1215 Weak C–C (Aliphatic) Vibration
1113 Medium C–O Stretch
1045 Strong C–O Stretch
996 Medium C–C (Aliphatic) Vibration
925 Weak C–C (Aliphatic) Vibration
852 Weak C–C (Aliphatic) Vibration
Figure 6.10: The ATR-FTIR ratio spectrum and associated peak assignments117 of a 100
mM solution of glycerol in perfusion fluid, acquired using a seven-reflection ZnSe crystal.
tamate concentration range is 10 – 100-fold lower than the lowest expected levels of
glucose and lactate in microdialysate. The ATR-FTIR and transmission-FTIR spec-
tra of a solution of L-glutamic acid in perfusion fluid are shown in Figures 6.12 and
6.13, respectively.
6.2. INTERFERENCE FROM COMPETING MOLECULES 87
5001000150020002500300035004000
Wavenumber / cm -1
80
85
90
95
100
105
%
 T
ra
ns
m
is
si
on
Peak
Position /
cm−1
Peak
Description
Assignment
2943 Medium CH2 Stretch
2888 Medium CH2 Stretch
1461 Weak CH2 Bend
1424 Weak C–H Bend
1415 Weak C–H Bend
1337 Weak C–H Bend
1217 Weak C–C (Aliphatic) Vibration
1114 Medium C–O Stretch
1045 Strong C–O Stretch
996 Medium C–C (Aliphatic) Vibration
Figure 6.11: The transmission-FTIR ratio spectrum and associated peak assignments117 of
a 100 mM solution of glycerol in perfusion fluid, acquired using a 50 µm path length.
88 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
95.5
96
96.5
97
97.5
98
98.5
99
99.5
100
100.5
%
 T
ra
ns
m
iss
io
n
Peak
Position /
cm−1
Peak
Description
Assignment
2947 Strong, Broad CH2 Stretch
2648 Weak, Broad C–H Stretch
1996 Weak, Broad NH+3 Bend (Overtone)
1720 Medium C=O Stretch
1715 Medium C=O Stretch
1600 Strong CO−2 Stretch (Asym.)
1520 Medium NH+3 Bend (Sym.)
1451 Weak CH2 Bend
1408 Strong CH2 Bend
1353 Medium CO−2 Stretch (Sym.)
1323 Weak N–C–H Bend
1269 Weak C–O Stretch
1226 Medium C–O Stretch
1155 Weak C–O Stretch
1083 Weak C–O Stretch
1000 Weak C–C (Aliphatic) Vibration
930 Weak C–C (Aliphatic) Vibration
Figure 6.12: The ATR-FTIR ratio spectrum and associated peak assignments117 of a 50 mM
solution of L-glutamic acid in perfusion fluid, acquired using a seven-reflection ZnSe crystal.
Asym. = asymmetric vibration, Sym. = symmetric vibration.
6.2. INTERFERENCE FROM COMPETING MOLECULES 89
5001000150020002500300035004000
Wavenumber / cm -1
75
80
85
90
95
%
 T
ra
ns
m
is
sio
n
Peak
Position
/ cm−1
Peak
Description
Assignment
2949 Strong, Broad CH2 Stretch
2940 Strong, Broad CH2 Stretch
2658 Weak, Broad C–H Stretch
1998 Weak, Broad NH+3 Bend (Overtone)
1723 Strong C=O Stretch
1720 Strong C=O Stretch
1584 Medium CO−2 Stretch (Asym.)/NH
+
3 Bend (Asym.)
1573 Medium CO−2 Stretch (Asym.)/NH
+
3 Bend (Asym.)
1557 Medium CO−2 Stretch (Asym.)/NH
+
3 Bend (Asym.)
1536 Medium NH+3 Bend (Sym.)
1528 Strong NH+3 Bend (Sym.)
1524 Medium NH+3 Bend (Sym.)
1518 Strong NH+3 Bend (Sym.)
1512 Medium NH+3 Bend (Sym.)
1450 Weak CH2 Bend
1437 Weak CO−2 Stretch (Sym.)
1409 Strong CH2 Bend
1353 Medium CO−2 Stretch (Sym.)
1323 Weak N–C–H Bend
1227 Medium C–O Stretch
1156 Weak C–O Stretch
1084 Weak C-O Stretch
Figure 6.13: The transmission-FTIR ratio spectrum and associated peak assignments117 of
a 50 mM solution of L-glutamic acid in perfusion fluid, acquired using a 50 µm path length.
Asym. = asymmetric vibration, Sym. = symmetric vibration.
90 CHAPTER 6. IR SPECTROSCOPY
ATR-FTIR and transmission-FTIR spectra of a mixture containing equimolar
amounts of glucose, lactate, pyruvate, glycerol and L-glutamic acid were also ac-
quired (Figures 6.14 and 6.15, respectively). L-glutamic acid is less soluble in aqueous
media than the other compounds under investigation and therefore all compounds
were present in the mixture at 50 mM to account for the lower aqueous solubility of
L-glutamate. Peak assignments for the mixture spectra were based on those already
described for glucose, lactate, pyruvate, glycerol and glutamate. This proved to be
difficult due to the overlap of peaks and the potential for multiple compounds to con-
tribute to a single peak. Where it was not possible to ascertain the origin of a peak,
the possible contributing compounds are indicated.
Both the ATR-FTIR and the transmission-FTIR spectra of the glucose, lactate,
pyruvate, glycerol and glutamate mixure show peaks specific to glucose, lactate and
pyruvate.
Two weak glucose-specific peaks are observed in the ATR-FTIR spectrum of the
mixture, at 3524 cm−1 and 1265 cm−1. The former is lost in the transmission-FTIR
spectrum, due to noise, but the 1265 cm−1 peak still remains. However, it is possible
that the peak at 1081 cm−1, which is present in both spectra, would be the most useful
peak for glucose detection despite the fact that it is attributed to the C–O stretches
of both glucose and glutamate in Figures 6.14 and 6.15. This is due to the fact that
glutamate levels in clinical microdialysate samples are likely to be 10 – 100-fold lower
than the lowest expected concentration of glucose and are thus expected to have a
negligible impact upon the FTIR spectrum of glucose. This assumption is particularly
applicable when considering the 1081 cm−1 peak, as it is one of the strongest in the
glucose spectrum.
The ATR-FTIR spectrum of the mixture does not show any peaks specific to lac-
tate. However, two peaks have been attributed to both lactate and glutamate: one
at 1454 cm−1, corresponding to the CH3 bend of lactate and the CH2 bend of gluta-
mate, and the other at 1319 cm−1 (corresponding to the C–O–H bend of lactate and
the N–C–H bend of glutamate). Using the same argument as that presented for the
1081 cm−1 peak of glucose, it could be concluded that these peaks may be considered
lactate-specific for the analysis of microdialysate samples. The transmission-FTIR
spectrum contains a single peak which is clearly lactate-specific: a strong peak at
1567 cm−1, corresponding to one of the CO−2 stretching modes that is present in the
transmission-FTIR spectrum of a pure solution of lactate but absent from its ATR-
FTIR spectrum. In addition to this, the transmission-FTIR spectrum contains three
peaks which are attributed to both lactate and glutamate and so may be considered
to be lactate-specific for the analysis of clinical microdialysate samples: 1555 cm−1,
1453 cm−1, 1321 cm−1.
Both the ATR-FTIR and the transmission-FTIR spectra of the mixture contain
6.2. INTERFERENCE FROM COMPETING MOLECULES 91
three peaks which have been attributed solely to pyruvate, at 1710 cm−1, 1356 cm−1
and 1176 cm−1. All of the peaks are of medium intensity and are therefore clearly
visible in the spectra.
The presence of these peaks indicates that it may be possible to detect glucose,
lactate and pyruvate against a background of additional compounds recovered via
microdialysis. Although this is not a comprehensive experimental study of all possible
competing molecules, it is encouraging to note that these analyte peaks have been
detected from an equimolar mixture of all compounds, yet in clinical microdialysate
samples many of the additional competing compounds are likely to be present at
concentrations significantly lower than those of glucose, lactate and pyruvate.
92 CHAPTER 6. IR SPECTROSCOPY
5001000150020002500300035004000
Wavenumber / cm -1
90
92
94
96
98
100
102
%
 T
ra
ns
m
is
sio
n
Peak
Position /
cm−1
Peak
Description
Assignment
3524 Weak, Broad O–H Stretch: Glu
2941 Medium, Broad CH2 Stretch: Glut / CH2 Stretch: Gly
2892 Medium, Broad O–H Stretch: Lac / CH2 Stretch: Gly
1710 Medium C=O Stretch: Pyr
1593 Strong CO−2 Stretch: Glut
1454 Weak CH2 Bend: Glut / CH3 Bend: Lac
1409 Strong CH2 Bend: Glut
1356 Medium CH3 Bend: Pyr
1319 Weak C–O–H Bend: Lac / N–C–H Bend: Glut
1265 Weak C–C (Alicyclic) Vibration: Glu
1233 Weak C–O Stretch: Glut
1176 Medium C–C (Aliphatic) Vibration: Pyr
1153 Weak
C–C (Alicyclic) Vibration: Glu / C–O
Stretch: Glut
1115 Medium C–O Stretch: Gly
1081 Strong C–O Stretch: Glu & Glut
1039 Strong C–O Stretch: Glu & Lac
994 Weak C–C (Alicyclic) Vibration: Glu & Gly
925 Weak C–C (Aliphatic) Vibration: Lac & Gly
Figure 6.14: The ATR-FTIR ratio spectrum and associated peak assignments117 of a mixture
of glucose, lactate, pyruvate, glycerol and glutamate in perfusion fluid. All compounds were
present at a concentration of 50 mM. The spectrum was acquired using a seven-reflection ZnSe
crystal. Glu = glucose, Lac = lactate, Pyr = pyruvate, Asym. = asymmetric vibration, Sym.
= symmetric vibration.
6.2. INTERFERENCE FROM COMPETING MOLECULES 93
5001000150020002500300035004000
Wavenumber / cm -1
70
75
80
85
90
95
100
105
%
 T
ra
ns
m
iss
io
n
Peak
Position
/ cm−1
Peak
Description
Assignment
2940 Strong, Broad CH2 Stretch: Glut / Gly
2891 Strong, Broad CH2 Stretch: Gly
1712 Medium C=O: Pyr
1581 Strong CO−2 Stretch: Glut, Lac, Pyr / NH
+
3 : Glut
1573 Strong CO−2 Stretch: Glut, Lac, Pyr / NH
+
3 : Glut
1567 Strong CO−2 Stretch: Lac
1555 Medium CO−2 Stretch: Glut, Lac / NH
+
3 : Glut
1453 Medium CH2 Bend: Glut / CH3 Bend: Lac
1409 Strong CH2 Bend: Glut
1356 Strong CH3 Bend: Pyr
1321 Weak C–O–H Bend: Lac / N–C–H Bend: Glut
1266 Weak C–C (Alicyclic) Vibration: Glu
1233 Weak C–O Stretch: Glut
1176 Medium C–C (Aliphatic) Vibration: Pyr
1152 Weak
C–C (Alicyclic) Vibration: Glu / C–O
Stretch: Glut
1116 Medium C-O Stretch: Gly
1081 Strong C–O Stretch: Glut / Glu
1040 Strong C–O Stretch: Glu / Lac
994 Weak C-C: Glu / Gly
Figure 6.15: The transmission-FTIR ratio spectrum and associated peak assignments117 of a
mixture of glucose, lactate, pyruvate, glycerol and glutamate in perfusion fluid. All compounds
were present at a concentration of 50 mM. The spectrum was acquired using a seven-reflection
ZnSe crystal. Glu = glucose, Lac = lactate, Pyr = pyruvate, Asym. = asymmetric vibration,
Sym. = symmetric vibration.
94 CHAPTER 6. IR SPECTROSCOPY
6.2.1 Interference from Competing Molecules: Conclusions
This section has identified a number of molecules which are likely to be present in
clinical microdialysate samples in addition to glucose, lactate and pyruvate. It was
concluded that several of these molecules cannot be accounted for in the current work
and their investigation must be a priority when carrying out clinical testing on the
final device. It has also been demonstrated that absorbance peaks unique to glucose,
lactate and pyruvate can be identified from transmission-FTIR spectra of the three
compounds in an equimolar mixture with two key competing molecules: glycerol and
glutamate. ATR-FTIR spectra of the same mixture yield unique peaks for glucose and
pyruvate but not for lactate. However, the concentrations of glycerol and glutamate
in clinical microdialysate samples are expected to be significantly lower than those of
lactate and thus certain lactate absorbance peaks may be considered unique to lactate
under these circumstances. All unique peaks identified are listed in Table 6.2.
Table 6.2: FTIR peaks of glucose, lactate and pyruvate that allow the distinction
of each compound when combined in an equimolar mixture containing glycerol and
glutamate. Peaks marked with an asterisk are not unique to that compound when
in an equimolar mixture of glycerol and glutamate, but are considered likely to be
unique in clinical samples due to the expected low concentrations of glycerol and
glutamate.
Compound Spectrum Type Unique Peaks / cm−1
Glucose
ATR-FTIR 3524, 1081*, 1265
Transmission-FTIR 1081*, 1266
Lactate
ATR-FTIR 1454*, 1319*
Transmission-FTIR 1567, 1454*, 1319*
Pyruvate
ATR-FTIR 1710, 1356, 1176
Transmission-FTIR 1712, 1356, 1176
6.3 FTIR Spectroscopy of Clinical Samples
Single microdialysate samples taken from six different patients were analysed via
transmission-FTIR spectroscopy only: the samples could not be measured using ATR-
FTIR spectroscopy due to the small sample volumes available. The aim of this study
was to determine whether any of the compound-specific peaks identified in Sections
6.1 and 6.2 were visible in the spectra and hence whether any of the compounds of
interest could be detected directly from a clinical sample using transmission-FTIR
spectroscopy. Prior to the acquisition of FTIR spectra, each sample was analysed
using the ISCUS Flex to determine the concentrations of glucose, lactate and pyruvate
present. Concentration data for glycerol, glutamate and other competing molecules
6.3. FTIR SPECTROSCOPY OF CLINICAL SAMPLES 95
present in the sample were unavailable. The ISCUS Flex measurements were carried
out by Dr Susan Giorgi-Coll.
9001000110012001300140015001600
Wavenumber / cm -1
92
94
96
98
100
102
104
%
 T
ra
ns
m
iss
io
n
(1) G:1.11 L:4.34 P:0.13
(2) G:0.96 L:1.67 P:0.04
(3) G:1.20 L:2.35 P:0.04
(4) G:1.17 L:1.85 P:0.06
(5) G:1.41 L:1.93 P:0.05
(6) G:1.52 L:1.82 P:0.07
Analyte Concentration / mM
Figure 6.16: The transmission-FTIR ratio spectra of six clinical microdialysate samples,
acquired using a 50 µm path length. G = glucose, L = lactate, P = pyruvate.
Transmission-FTIR spectra were acquired using the same protocol as that used
for synthetic samples, described in Section 5.2.2, using a 50 µm path length. The
spectra have been shown in the region 1650–900 cm−1, which is where the majority
of the peaks of interest are located.
All spectra show weak peak intensities as a result of the very low concentrations
of glucose, lactate and pyruvate present and consequently only a few could be distin-
guished from the spectral noise. Table 6.3 highlights only those peaks which could be
unequivocally identified and assigned to a particular compound without further data
processing.
In each spectrum, there are four peaks which are clearly visible: 1365 cm−1, 1124
cm−1, 1081 cm−1 and 1042 cm−1. The peak at 1365 cm−1 has possible contributions
from both the CH3 bend of lactate and the C–O–H bend of glucose, which may help to
explain why it is the strongest band in the spectrum. The remaining strong bands are
due to C–O stretches of lactate (1124 cm−1 and 1043 cm−1) and glucose (1081 cm−1).
In Section 6.2 it was noted that, while the C–O stretch of glutamate may contribute
to the peak at 1081 cm−1 in the FTIR spectrum of equimolar amounts of glucose and
glutamate, the contribution of glutamate to this peak in the FTIR spectrum of clinical
samples is likely to be negligible due to its expected low clinical concentration range.
As a result, the peak has been attributed solely to glucose in this analysis. There
96 CHAPTER 6. IR SPECTROSCOPY
Table 6.3: Peak assignments for the transmission-FTIR spectra of clinical micro-
dialysate samples.
Peak
Position /
cm−1
Sample
Number
Assignment
1365 1, 2 , 3, 4, 5, 6 CH3 Bend: Lac / C–O–H Bend: Glu
1315 1, 2, 3, 6 C–O–H Bend: Lac
1124 1, 2, 3, 4, 5, 6 C–O Stretch: Lac
1086 1 C–O Stretch: Lac
1081 1, 2, 3, 4, 5, 6 C–O Stretch: Glu
1043 1, 2, 3, 4, 5, 6 C–O Stretch: Lac
1034 2, 4, 6 C–O Stretch: Glu
992 1, 2, 3, 4, 5, 6 C–C (Alicyclic) Vibrations: Glu
are no visible peaks corresponding to vibrations of the pyruvate molecule, which is
reflective of the very low concentration of this compound in comparison to glucose
and lactate.
It is noted that the locations of certain peaks in the clinical microdialysate samples
are different to those in the equimolar mixtures studied in Sections 6.1 and 6.2. For
example, the FTIR spectra of an equimolar mixture of glucose, lactate and pyruvate
show a peak at 1038 cm−1, which is attributed to an overlap of the 1035 cm−1 peak
due to the C–O stretch of glucose and the 1041 cm−1 peak corresponding to the C–O
stretch of lactate. In Figure 6.16, these vibrations have been resolved as two separate
peaks, possibly due to the difference in concentration between glucose and lactate in
the clinical microdialysate samples.
6.3.1 FTIR Spectroscopy of Clinical Samples: Conclusions
This section has indicated that transmission-FTIR spectroscopy can be used to detect
glucose and lactate directly from clinical microdialysate samples via the identification
of compound-specific peaks. However, this method was not able to detect pyruvate
directly from the samples without further processing of the data, indicating that a
device based on FTIR spectroscopy would most likely require a statstics-based method
for analysing spectral data.
6.4 FTIR Spectroscopy Accessory Selection
This section describes two quantitative studies that were carried out to determine
which FTIR accessories would be used for an in-depth analysis of the capability of
FTIR to detect glucose, lactate and pyruvate at concentrations relevant to TBI mon-
itoring. The studies were used to select an ATR-FTIR crystal for future ATR-FTIR
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 97
studies and a path length for future transmission-FTIR studies.
6.4.1 Crystal Selection for ATR-FTIR
The performances of three ATR-FTIR crystals were compared by carrying out an
identical series of ATR-FTIR measurements with each crystal and analysing the re-
sulting spectra to estimate the limits of detection of glucose, lactate and pyruvate
afforded by each one. The crystals available for comparison were a single-reflection
diamond, a seven-reflection Ge crystal and a seven-reflection ZnSe crystal.
Pure solutions of glucose, lactate and pyruvate, at ten different concentrations
per compound, were prepared in perfusion fluid. Three separate samples of each
solution were measured using ATR-FTIR spectroscopy. Two spectra of each sample
were acquired: one ratio spectrum, in which perfusion fluid was used as a reference
sample, and one raw spectrum. The concentrations used in this study were selected
to cover the full clinical ranges of glucose, lactate and pyruvate that the developed
sensor will be required to detect (0.1 –24 mM for glucose and lactate, 0.01 – 1.5 mM
for pyruvate) and are listed in Table 6.4.
Table 6.4: Concentrations of glucose, lactate and pyruvate used in the performance
comparison of ATR-FTIR crystals.
Glucose Solutions
/ mM
Lactate Solutions
/ mM
Pyruvate Solutions
/ mM
0 0 0
0.05 0.05 0.005
0.1 0.1 0.01
0.19 0.19 0.05
0.4 0.4 0.10
0.75 0.75 0.15
1.5 1.5 0.75
3.0 3.0 1.0
6.0 6.0 1.5
12.0 12.0 7.5
24.0 24.0 15.0
The limits of detection of glucose, lactate and pyruvate afforded by each crystal
were determined by identifying the lowest concentration of each compound that could
be detected from the spectrum via an analysis of peak intensities: this method was
chosen in preference to a statistical modelling approach due to the large number of
spectra required for a robust statistical model. A comparison of peak intensities
was assumed to be a valid method of determining the limit of detection due to the
known relationship between peak intensity, the absorption of IR radiation and the
concentration of an analyte. The peak intensity is proportional to the amount of IR
radiation absorbed by a species which, in turn, is related to the concentration of that
98 CHAPTER 6. IR SPECTROSCOPY
species via the Beer-Lambert law (Section 5.4.2).
The spectra acquired in this study all exhibited relatively flat baselines. However,
many spectra were found to have a significant vertical offset with respect to other
spectra acquired using the same crystal, meaning that a comparison of the absolute
peak intensities generated by each concentration would not yield meaningful results.
These features can be observed in Figure 6.17 which provides an example of the range
of spectra generated by a single compound, at multiple concentrations, using a single
type of ATR-FTIR crystal.
1000150020002500300035004000
Wavenumber / cm -1
97.5
98
98.5
99
99.5
100
100.5
101
101.5
%
 T
ra
ns
m
is
si
on
   0
0.05
 0.1
0.19
 0.4
0.75
 1.5
   3
   6
  12
  24
Conc. / mM
Figure 6.17: The ATR-FTIR ratio spectra of glucose in perfusion fluid, at a range of con-
centrations, acquired using a single-reflection diamond crystal. Spectra have flat baselines but
are vertically offset with respect to one another.
As a result of this offset, relative peak intensities were calculated and compared for
each spectrum. The relative peak intensity is the measured difference in transmission
between a peak maximum and an isosbestic point on the spectrum. This concept can
be illustrated graphically by normalising the spectra to an isosbestic point, as shown
in Figure 6.18.
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 99
1050110011501200
Wavenumber / cm -1
99.5
99.6
99.7
99.8
99.9
100
100.1
%
 T
ra
ns
m
is
si
on
   0
0.05
 0.1
0.19
 0.4
0.75
 1.5
   3
   6
  12
  24
Conc. / mM
(a) The unmodified ATR-FTIR ratio spectrum of glucose in perfusion fluid. For clarity,
only the 1200 – 1050 cm−1 region of the spectrum is shown.
1050110011501200
Wavenumber / cm -1
99.5
99.6
99.7
99.8
99.9
100
100.1
%
 T
ra
ns
m
iss
io
n
   0
0.05
 0.1
0.19
 0.4
0.75
 1.5
   3
   6
  12
  24
Conc. / mM
(b) The ATR-FTIR ratio spectrum of glucose in perfusion fluid, normalised to 1180
cm−1. For clarity, only the 1200 – 1050 cm−1 region of the spectrum is shown.
Figure 6.18: An illustration of spectrum normalisation on the ATR-FTIR ratio spec-
tra of glucose at various concentrations, acquired using a single-reflection diamond ATR
crystal. In the unmodified spectrum (Figure 6.18(a)) the absolute peak intensity at 1081
cm−1 does not correlate well with the glucose concentration. After normalisation to the
isosbestic point at 1180 cm−1 (Figure 6.18(b)), the absolute peak intensity at 1081 cm−1
increases with increasing glucose concentration.
100 CHAPTER 6. IR SPECTROSCOPY
The peaks used in this study were selected from those identified in Section 6.1 as
being unique to each compound when combined in a mixture of glucose, lactate and
pruvate. The two strongest unique peaks of each compound were chosen, whilst isos-
bestic points were determined from an examination of the spectra of each compound.
The same isosbestic point was used in the calculation of the relative intensity of both
glucose absorption peaks, as was the case for lactate. However, a single isosbestic
point which yielded satisfactory measurements of the relative intensity of both pyru-
vate absorption peaks could not be found and so different isosbestic points were used
for each peak. The spectral positions used to calculate the relative intensities of each
peak are summarised in Table 6.5.
Table 6.5: The spectral peak positions of glucose, lactate and pyruvate selected for calculation
of relative peak intensity.
Analyte
Position of Peak
Maximum / cm−1
Position of Isosbestic
Point / cm−1
Glucose
1081 1180
993 1180
Lactate
1456 1280
1316 1280
Pyruvate
1357 1340
1176 1200
An analysis of relative peak intensity with changing concentration is shown in
Figure 6.19. This analysis was performed on ATR-FTIR spectra of glucose, acquired
using the single-reflection diamond ATR crystal, and is focused on the absorption peak
at 1081 cm−1. The relationship between relative peak intensity and concentration
was assumed to be linear and therefore a line of best fit is superimposed on the
data. For this reason, data from the spectra containing only perfusion fluid (a glucose
concentration of zero) was omitted from the analysis as it would not have fitted a
linear regime. The data for each peak, acquired using each ATR-FTIR crystal, were
treated separately and all datasets were analysed in the same way.
Two parameters derived from this analysis were used to assess crystal performance:
the gradient of the line of best fit, which indicates the degree of response of the
mean relative peak intensity to an increase in concentration, and the coefficient of
determination (R2) which quantifies the linearity of the relationship. The values of
these parameters, calculated for all datasets, are summarised at the end of this section
in Table 6.6.
A one-way ANOVA test was performed on each dataset to determine whether any
of the mean values of the relative peak intensity were statistically significantly different
from one another. The results of the test confirmed that all datasets contained at least
two mean values of the relative peak intensity which were statistically significantly
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 101
0 5 10 15 20 25
Concentration / mM
-0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
R
el
at
iv
e 
Pe
ak
 In
te
ns
ity 0.05
 0.1
0.19
 0.4
0.75
 1.5
   3
   6
  12
  24
Conc. / mM
Figure 6.19: A plot showing the relationship between glucose concentration and the relative
intensity of the glucose absorption peak at 1081 cm−1 in ATR-FTIR spectra acquired using
a single-reflection diamond crystal. The results of each repeat are shown here. Gradient =
0.018, R2 = 0.978.
different. A post-hoc multiple comparison test was subsequently performed in order to
identify which of the mean values of the relative peak intensity within a dataset were
statistically significantly different from one another. The test compared the mean
relative peak intensity at a given concentration with those at all other concentrations.
This process was repeated for each concentration and hence the mean relative peak
intensities at all concentrations were systematically compared with one another. The
resulting interactive graphical output was used to determine which concentrations
could be detected and differentiated from other concentrations and hence provided
a measure of the LoQ, for the range of concentrations investigated, for each peak
and each crystal. An example of this is shown in Figure 6.20, which compares the
mean relative peak intensities of the 1081 cm−1 peak of glucose, acquired using the
single-reflection diamond crystal, at 12 mM and 6 mM. In this example, the mean
relative peak intensity at 12 mM is statistically significantly different to those at all
other concentrations, but the mean relative peak intensity at 6 mM is not. The LoQ
of glucose, as detected using a single-reflection diamond ATR crystal combined with
an analysis of peak intensities, is therefore between 6 mM and 12 mM. The results of
the post-hoc multiple comparison test performed on each dataset are given at the end
of this section in Table 6.6.
102 CHAPTER 6. IR SPECTROSCOPY
-0.1 0 0.1 0.2 0.3 0.4 0.5
Mean Relative Peak Intensity
24.0
12.0
6.00
3.00
1.50
0.75
0.40
0.19
0.10
0.05
Co
nc
en
tra
tio
n 
/ m
M
(a) The mean relative peak intensity at 12 mM (highlighted in blue) is statistically
significantly different to those at all other concentrations (highlighted in red).
-0.1 0 0.1 0.2 0.3 0.4 0.5
Mean Relative Peak Intensity
24.0
12.0
6.00
3.00
1.50
0.75
0.40
0.19
0.10
0.05
Co
nc
en
tra
tio
n 
/ m
M
(b) The mean relative peak intensity at 6 mM (highlighted in blue) is statistically
significantly different to those concentrations highlighted in red. However, it is not
statistically significantly different to that at 3 mM, as indicated by the grey marker.
Figure 6.20: The output of the multiple comparison test performed on the relative peak
intensity vs. concentration data of the 1081 cm−1 peak of glucose, acquired using the single-
reflection diamond crystal. The selected concentration is highlighted in blue. Concentrations
with mean relative peak intensities which are statistically significantly different from those of
the selected concentration are highlighted in red. Those which are not statistically significantly
different are shown in grey.
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 103
The magnitude of the gradient of the line of best fit between the concentration of
a compound and the relative intensity of a given peak indicates the strength of the
relationship between the two variables: the greater the magnitude of the gradient, the
greater the increase in peak intensity with increasing concentration. A high gradient
is thus desirable for a sensor based on FTIR spectroscopy. This value was highest
for spectra acquired using the seven-reflection ZnSe, for every peak and compound
analysed. The highest gradient (0.093) was observed for the line of best fit between
the concentration of glucose and the relative intensity of the glucose absorption peak
at 1081 cm−1. This data is shown in Figure 6.21.
0 5 10 15 20 25
Concentration / mM
-0.5
0
0.5
1
1.5
2
2.5
R
el
at
iv
e 
Pe
ak
 In
te
ns
ity 0.05
 0.1
0.19
 0.4
0.75
 1.5
   3
   6
  12
  24
Conc. / mM
Figure 6.21: A plot showing the relationship between glucose concentration and the relative
intensity of the glucose absorption peak at 1081 cm−1 in ATR-FTIR spectra acquired using a
seven-reflection ZnSe crystal. The results of each repeat are shown here. The gradient of the
line of best fit is 0.093, the highest seen for any of the peaks or crystals studied.
The R2 value is a measure of the linearity of the relationship between peak height
and relative peak intensity. It can range from 0 – 1, with values close to one indicating
a highly linear relationship. A linear relationship is highly desirable for a FTIR-based
sensor as it renders the prediction of glucose, lactate and pyruvate concentrations
more straightforward. This value was consistently lower for spectra acquired using
the seven-reflection Ge crystal than for spectra acquired using the single-reflection
diamond crystal or the ZnSe crystal. The diamond and ZnSe crystals exhibited com-
parable R2 values for all peaks and compounds studied. The highest R2 value (0.995)
was observed for the relationship between the intensity of the 1357 cm−1 absorption
104 CHAPTER 6. IR SPECTROSCOPY
peak of pyruvate, as calculated from spectra obtained using the seven-reflection ZnSe
crystal, and the concentration of pyruvate. This relationship is shown in Figure 6.22.
0 5 10 15
Concentration / mM
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
R
el
at
iv
e 
Pe
ak
 In
te
ns
ity 0.01
0.01
0.05
 0.1
0.15
0.75
   1
 1.5
 7.5
  15
Conc. / mM
Figure 6.22: A plot showing the relationship between pyruvate concentration and the relative
intensity of the pyruvate absorption peak at 1357 cm−1 in ATR-FTIR spectra acquired using
a seven-reflection ZnSe crystal. The results of each repeat are shown here. The gradient of
the line of best fit is 0.034 and the R2 value is 0.995, the highest seen for any of the peaks or
crystals studied.
The seven-reflection ZnSe crystal gave the lowest LoQ of glucose and lactate, at all
analysed peak positions, and equalled the LoQ of pyruvate achieved by the other two
crystals. A low limit of quantitation, within the expected clinical concentration range
of glucose, lactate and pyruvate, is a necessary reqirement of the developed sensor.
The lowest calculated LoQ (0.75–1.5 mM) was obtained for lactate, using an analysis
of the 1316 cm−1 absorption peak of spectra acquired using the seven-reflection ZnSe
crystal. The results of the multiple comparison test carried out on these spectra are
shown in Figure 6.23.
Overall, it was concluded that the seven-reflection ZnSe crystal was the best-
performing ATR-FTIR crystal of those tested and would be used for a future analysis
of the performance of ATR-FTIR spectroscopy. This superior performance was ex-
plained by the fact that it was the crystal with the longest effective path length: it
allowed a greater number of reflections of the IR beam than the diamond crystal (seven
reflections vs. one reflection, respectively) and, for a given IR wavelength, exhibited
a greater penetration depth than the seven-reflection germanium crystal.122 A longer
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 105
-0.1 0 0.1 0.2 0.3 0.4 0.5 0.6
Mean Relative Peak Intensity
24.0
12.0
6.00
3.00
1.50
0.75
0.40
0.19
0.10
0.05
Co
nc
en
tra
tio
n 
/ m
M
Figure 6.23: The output of the multiple comparison test performed on the relative peak
intensity vs. concentration data of the 1316 cm−1 peak of lactate, acquired using a seven-
reflection ZnSe crystal. The mean relative peak intensity at 1.5 mM (highlighted in blue) is
statistically significantly different from those of all other concentrations analysed, as indicated
by the data points highlighted in red, showing that the LoQ of lactate using this system is below
1.5 mM. This was the lowest LoQ observed for any of the peaks or crystals studied.
effective path length allows for increased interaction of the IR beam with the sample
and hence a higher absorption of IR radiation, leading to more intense peaks in the
ATR-FTIR spectrum. With this in mind, the poor performance of the seven-reflection
Ge crystal in comparison with the single-reflection diamond crystal is surprising. It is
possible that the performance of the Ge crystal was negatively affected by its physical
condition, as its surface was observed to be significantly more scratched than that of
the other crystals.
106 CHAPTER 6. IR SPECTROSCOPY
Table 6.6: The summarised results of the ATR-FTIR crystal study. The following properties
of each data set are given: the gradient of the line of best fit from a plot of concentration
vs. relative peak intensity; the R2 value of the linear relationship between concentration and
relative peak intensity; and the LoQ within the range of analysed concentrations.
Parameter Crystal
Glucose Lactate Pyruvate
1081 993 1456 1316 1357 1176
Gradient
Diamond 1R 0.018 0.007 0.004 0.003 0.007 0.011
Germanium 7R 0.027 0.004 0.004 0.005 0.008 0.016
ZnSe 7R 0.093 0.036 0.020 0.023 0.034 0.072
R2
Diamond 1R 0.978 0.833 0.769 0.985 0.972 0.993
Germanium 7R 0.860 0.017 0.503 0.943 0.959 0.976
ZnSe 7R 0.973 0.939 0.771 0.980 0.995 0.971
LoQ /
mM
Diamond 1R 6–12 12–24 12–24 6–12 1.5–7.5 1.5–7.5
Germanium 7R 12–24 >24 >24 12–24 1.5–7.5 1.5–7.5
ZnSe 7R 1.5–3 6–12 6–12 0.75–1.5 1.5–7.5 1.5–7.5
6.4.2 Path Length Selection for Transmission-FTIR
All of the transmission cells used in this work were fitted with ZnSe windows. This
material was chosen based upon four properties: broad IR transmission range (20,000
– 454 cm−1), lack of IR absorption bands within the mid-IR spectral region, chemical
inertness and insolubility in water.123
An optimum path length for use in a future transmission-FTIR study was selected
by carrying out an identical series of transmission-FTIR measurements using each
path length and analysing the peak intensities of the resulting spectra to estimate
the limit of detection of glucose afforded by each one. The path lengths available for
comparison were 25 µm, 50 µm and 100 µm. These accessories were acquired on a
short-term loan from the manufacturer, Specac, and thus time limitations restricted
the number of compounds and sample concentrations that could be tested.
Glucose was selected as a test analyte as its IR absorbance peaks are more thor-
oughly characterised in the literature than those of lactate and pyruvate. Pure so-
lutions of glucose in perfusion fluid were prepared at five concentrations, which were
selected to cover the full clinical range that the final device will be required to detect:
0.1 mM, 0.4 mM, 1.5 mM, 6 mM and 24 mM.
The transmission-FTIR spectra of glucose at all analysed concentrations, acquired
using a path length of 100 µm, are shown in Figure 6.24. These spectra exhibit
relatively flat baselines but are vertically offset with respect to one another, as was
the case with the spectra gathered using ATR-FTIR, and therefore relative peak
intensites were once again calculated and compared for each spectrum.
The glucose absorption peaks at 1081 cm−1 and 993 cm−1, which were analysed
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 107
5001000150020002500300035004000
Wavenumber / cm -1
60
80
100
120
140
160
%
 T
ra
ns
m
is
sio
n
  0
0.1
0.4
1.5
  6
 24
Conc. / mM
Figure 6.24: The transmission-FTIR spectra of glucose in perfusion fluid, at a range of
concentrations, acquired using a path length of 100 µm. Spectra have flat baselines but are
vertically offset with respect to one another.
for the ATR-FTIR crystal study, were once again selected for analysis in this study.
The isosbestic point at 1180 cm−1, which was used in the ATR-FTIR crystal study
to calculate the relative intensities of each peak, was also found to be effective when
calculating the relative intensities of the peaks in the transmission-FTIR spectrum
and was therefore used in this study.
The methodology used to compare the performace of each transmission-FTIR path
length was the same as that used to compare the performance of ATR-FTIR crystals
and thus the details of the analysis will not be repeated here. The results of the
analysis are presented in Table 6.7.
Table 6.7: A summary of the results of the study conducted to analyse the perfor-
mance of different path lengths of a transmission-FTIR cell with ZnSe windows.
Path Length
/ µm
Peak /
cm−1 Gradient R
2 LoQ /
mM
25
1081 0.201 0.993 1.5–6
993 0.072 0.967 6–24
50
1081 0.402 0.999 0.4–1.5
993 0.101 0.854 6–24
100
1081 0.471 0.893 6–24
993 0.148 0.434 >24
108 CHAPTER 6. IR SPECTROSCOPY
The effect of increasing concentration on the increase in mean relative peak inten-
sity, as measured by the gradient of the line of best fit through a plot of concentration
vs. mean relative peak intensity, is proportional to the path length used. This result is
expected, as an increased path length allows greater interaction of the IR beam with
the sample and thus the absorption of radiation will be more strongly affected by the
amount of analyte molecules present. However, the linearity of this relationship, as
measured by the R2 value, and the LoQ do not follow this pattern. The R2 values do
not have a clear relationship with the path length. This is due to the fact that the
water-induced noise increases with longer path lengths. As the level of noise increases
the degree of saturation of the detector of the instrument also increases, causing a
reduction in the size of the region of the spectrum unaffected by noise. This was ob-
served when comparing the spectra of a 24 mM solution of glucose in perfusion fluid,
acquired using all three tested path lengths, as seen in Figure 6.25. In regions of the
spectrum which are close to saturation, the relationship between peak intensity and
concentration becomes non-linear, as observed in these results.
80010001200140016001800
Wavenumber / cm -1
%
 T
ra
ns
m
iss
io
n 
/ A
rb
itr
ar
y 
Un
its
25 / m
50 / m
100 / m
Path Length
Figure 6.25: The transmission-FTIR spectra of a 24 mM solution of glucose in perfusion
fluid, acquired using different path lengths. For clarity, each spectrum has been plotted with a
vertical offset with respect to the other spectra. When the shorter path length (25 µm) is used,
the 1600 – 900 cm−1 region of the spectrum, where the majority of glucose absorption peaks
are located, is free from water-induced noise. However, when the longest path length (100 µm)
is used, water-induced noise is visible in the 1600 – 900 cm−1 region of the spectrum.
According to the data presented in Table 6.7, the balance between these com-
6.4. FTIR SPECTROSCOPY ACCESSORY SELECTION 109
peting factors is optimised at a path length of 50 µm, which yields the lowest limits
of detection. This path length was therefore selected for use in a later study of
transmission-FTIR spectroscopy. A plot showing the relationship between relative
peak intensity and glucose concentration for transmission-FTIR spectra acquired us-
ing this path length, as measured for the 1081 cm−1 absorption peak of glucose, is
shown in Figure 6.26. The output of the multiple comparison plot carried out on the
same dataset is shown in Figure 6.27. This plot highlights the fact that the LoQ of
glucose using this system is below 1.5 mM, the lowest achieved by any of the tested
path lengths.
0 5 10 15 20 25
Concentration / mM
-2
0
2
4
6
8
10
12
R
el
at
iv
e 
Pe
ak
 In
te
ns
ity
  0
0.1
0.4
1.5
  6
 24
Conc. / mM
Figure 6.26: A plot showing the relationship between glucose concentration and the relative
intensity of the glucose absorption peak at 1081 cm−1 in transmission-FTIR spectra acquired
using a 50 µm path length. The results of each repeat are shown here. Gradient = 0.402, R2
= 0.999.
6.4.3 FTIR Spectroscopy Accessory Selection: Conclusions
The aim of the work described in this section was to select an ATR-FTIR crystal,
to use in a future study of ATR-FTIR spectroscopy, and a path length to use in a
future study of transmission-FTIR spectroscopy. A methodology allowing quantitative
analysis of the performance of a selection of ATR-FTIR crystals and a selection of
transmission-FTIR path lengths, based on an analysis of relative peak intensities, has
been presented. The results of this analysis indicated that a seven-reflection ZnSe
110 CHAPTER 6. IR SPECTROSCOPY
-2 0 2 4 6 8 10
Mean Relative Peak Intensity
24.0
6.00
1.50
0.40
0.10
Co
nc
en
tra
tio
n 
/ m
M
Figure 6.27: The output of the multiple comparison test performed on the relative peak
intensity vs. concentration data of the 1081 cm−1 peak of glucose, acquired using a 50 µm
path length. The mean relative peak intensity at 1.5 mM (highlighted in blue) is statistically
significantly different from those of all other concentrations analysed, as indicated by the data
points highlighted in red, showing that the LoQ of glucose using this system is below 1.5 mM.
ATR-FTIR crystal should be used for the study of ATR-FTIR spectroscopy and a 50
µm path length should be used for the study of transmission-FTIR spectroscopy.
The results achieved using ATR-FTIR spectroscopy cannot be directly compared
to those obtained using transmission-FTIR spectroscopy. This is due to the fact that
the two studies were carried out using a different series of glucose concentrations, which
has a significant impact on the post-hoc multiple comparison test used to calculate
the LoQ. However, it is noted that none of the ATR-FTIR or transmission-FTIR
accessories tested were able to achieve a LoQ close to that required for the developed
sensor.
6.5 Conclusions
This chapter describes an analysis of the performance of two forms of FTIR spec-
troscopy, ATR-FTIR and transmission-FTIR, for the detection of glucose, lactate and
pyruvate in solution. All data were considered within the context of the requirements
of a sensor for the metabolic profiling of TBI patients, in order to assess the suitability
of FTIR for use in such a sensor.
6.5. CONCLUSIONS 111
The results presented here suggested that both ATR-FTIR and transmission-FTIR
spectroscopy showed promise for use in the metabolic profiling of TBI patients. Both
techniques yielded spectra which allowed absorption peaks unique to glucose, lac-
tate or pyruvate to be identified from an equimolar mixture of all three compounds.
Some of these peaks remained visible when the two techniques were used to analyse
an equimolar mixture containing all three compounds in combination with glycerol
and glutamate, two competing molecules which are likely to be present in clinical
microdialysate samples.
Encouragingly, specific absorption peaks of glucose and lactate were shown to be
identifiable from transmission-FTIR spectra of clinical microdialysate samples.
Several outstanding questions remained following this work. The limits of per-
formance of ATR-FTIR spectroscopy and transmission-FTIR spectroscopy were un-
known: in particular, the accuracy with which each technique could be used to deter-
mine the concentrations of pyruvate, lactate and glucose when combined in a mixture
was not known. In a typical clinical microdialysate sample, the concentration of pyru-
vate is ten-fold lower than those of glucose and lactate: the effect of this difference on
the accuracy with which the levels of each compound could be determined from an
FTIR spectrum of the sample was also unknown.
The optimum method for determining the concentrations of glucose, lactate and
pyruvate from their FTIR spectra was an additional unknown. Some of the results
presented in this chapter had suggested that an analysis based on the calculation of
peak intensities would not yield a sufficiently low limit of detection, but no alternative
methods had been investigated.
Finally, it was not clear which of the two techniques would be most suitable for
development into a sensor for metabolic profiling of TBI patients. Although differences
between the two techniques had been observed, the degree to which they would affect
the detection of glucose, lactate and pyruvate was unknown.
It was therefore concluded that a more detailed study of each technique was re-
quired in order to address the questions presented here. Two quantitative studies
were therefore carried out: one to determine the most suitable ATR-FTIR crystal to
be used in a study of ATR-FTIR spectroscopy and another to determine the most
suitable path length to use for a study of transmission-FTIR spectroscopy. These
studies concluded that a seven-reflection ZnSe crystal should be used for a study of
ATR-FTIR spectroscopy and a 50 µm path length should be used for a study of
transmission-FTIR spectroscopy.

Chapter 7
Assessment of a Raman
Spectroscopy-Based Sensor
This chapter describes the work undertaken to assess whether Raman spectroscopy
could be used to detect glucose, lactate and pyruvate in microdialysate samples from
TBI patients, within the constraints of the device requirements set out in Section 2.6.
Raman spectroscopy was selected for investigation due to its insensitivity to water,
which would be of significant benefit in the analysis of microdialysate samples as
they are highly dilute. Furthermore, Raman spectroscopy has previously been used in
clinical applications, as discussed in Section 4.3, and has potential for miniaturisation
into a bedside device.
Attempts were initally made to acquire the Raman spectra of pure solutions of
glucose in perfusion fluid. Following this attempt, a SERS-based system for the on-
line detection of glucose, lactate and pyruvate in perfusion fluid was proposed. The
viability of the proposed system was subsequently assessed using cysteine as a model
analyte.
7.1 Raman Spectroscopy of Glucose
It was assumed that the acquisition of Raman spectra of glucose, lactate or pyruvate
would be non-trivial as they are small molecules with low polarisibility and thus exhibit
a low Raman-scattering intensity. Glucose was considered to be the most polarisible of
these compounds, due to the increased electron density caused by its ring structure,
and was thus expected to exhibit the strongest Raman-scattering intensity. Initial
work therefore focused on the acquisition of a Raman spectrum of glucose.
A concentrated 1M solution of glucose in perfusion fluid was initially analysed, to
account for the low Raman-scattering intensity of glucose. Spectral acquisition was
focused on the 750 – 1450 cm−1 range, which contains three peaks that are considered
to be the ‘Raman fingerprints’ of glucose.124 These peaks are found at approximately
113
114 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
910 cm−1, 1060 cm−1 and 1125 cm−1 and are attributed to C–C and C–O vibrations
within the molecule. A 785 nm laser was used to acquire the spectrum in order
to minimise the effect of photoluminescence. The resulting Raman spectrum and
associated peak assignments are shown in Figure 7.1.
800 900 1000 1100 1200 1300 1400
Raman Shift / cm -1
150
200
250
300
350
400
450
In
te
ns
ity
 / 
Ar
bi
tra
ry
 U
ni
ts
Raman Shift
/ cm−1
Peak
Intensity
Assignment
796 Medium C–C (Alicyclic) Vibration
844 Medium C–C (Alicyclic) Vibration/C–O–C Stretch
897 Medium C–C (Alicyclic) Vibration/C–O–C Stretch
913 Medium C–C (Alicyclic) Vibration/C–O–C Stretch
1061 Strong C–O Stretch
1126 Strong C–C (Alicyclic) Vibration
1240 Weak C–C (Alicyclic) Vibration
1366 Weak C-O-H Bend/CH2 Bend
Figure 7.1: The Raman spectrum and associated peak assignments of a 1M solution of glucose
in perfusion fluid mounted in the macro sampler. Characteristic Raman peaks of glucose can
be seen at 913 cm−1, 1061 cm−1 and 1125 cm−1. Laser wavelength = 785 nm, power at
sample = 100 mW, acquisition time = 1 × 600 s.
The spectrum shown in Figure 7.1 shows the expected peaks at 913 cm−1, 1061
cm−1 and 1125 cm−1, indicating that a Raman spectrum of glucose was successfully
acquired. However, an acquisition time of 600 s was required in order to obtain the
spectrum. This is significantly longer than the desired temporal resolution of a sensor
for metabolic profiling of TBI patients, which will require a measurement to be taken
7.1. RAMAN SPECTROSCOPY OF GLUCOSE 115
at least every five minutes (Section 2.6). A test was therefore carried out to determine
whether the characteristic Raman peaks of glucose could be seen when an acquisition
time appropriate to the device was used. Solutions of glucose in perfusion fluid at
various concentrations were mounted in the macro sampler and Raman spectra were
acquired using a 785 nm laser with a power at the sample of approximately 100 mW
and an acquisition time of 120 seconds. This acquisition time was selected as it is
comparable with the time taken to acquire the FTIR spectra discussed in Chapter 6.
The resulting spectra are shown in Figure 7.2.
800 900 1000 1100 1200 1300 1400
Raman Shift / cm -1
40
50
60
70
80
90
100
In
te
ns
ity
 / 
Ar
bi
tra
ry
 U
ni
ts
500
50
5
0.5
Conc. / mM
Figure 7.2: The Raman spectra of glucose in perfusion fluid at various concentrations, ob-
tained using a single acquisition of 120 s. Characteristic glucose peaks at 913 cm−1, 1061
cm−1 and 1125 cm−1 were observed when a glucose concentration of 500 mM was used, but
could not be seen at lower concentrations. Samples were mounted in the macro sampler for
measurement. Laser wavelength = 785 nm, power at sample = 100 mW.
An acquisition time of 120 s yielded characteristic glucose peaks at 913 cm−1,
1061 cm−1 and 1125 cm−1 in the Raman spectrum of 500 mM glucose. However,
a qualitative analysis of the spectra shown in Figure 7.2 suggested that these peaks
could not be detected in the Raman spectra of glucose at concentrations of 50 mM
or lower. A peak at approximately 1060 cm−1 is visible in the spectra of glucose at
concentrations of 50 mM, 5 mM and 0.5 mM. However, this peak does not appear to
vary in intensity as the glucose concentration is varied and therefore it was assumed
not to be a ‘true’ Raman peak. These results suggested that, if an acquisition time
of 120 s was used with the available Raman setup, the detection limit of glucose via
116 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
Raman spectroscopy was between 50 and 500 mM.
7.1.1 Raman Spectroscopy of Glucose: Conclusions
This section describes the successful acquisition of Raman spectra of glucose, as ev-
idenced by the presence of characteristic Raman peaks of glucose at 913 cm−1, 1061
cm−1 and 1125 cm−1. However, when using an acquisition time compatible with the
requirements of a sensor for metabolic profiling of TBI patients, the detection limit of
glucose using the available Raman spectroscopy setup was predicted to be between 50
and 500 mM. This is significantly higher than the maximum expected concentration
of glucose in clinical microdialysate (24 mM). Furthermore, detection of lactate and
pyruvate via Raman spectroscopy was expected to yield higher detection limits due
to their lower polarisibility.
Raman scattering intensity is proportional to 1
λ4
and thus it may have been possible
to obtain a stronger Raman signal using a light source of shorter wavelength. However,
the use of such a light source may also have increased the intensity of the fluorescence
visible in the Raman spectrum and therefore would not have provided sufficient im-
provement in the signal to allow the detection of clinically-relevant concentrations
of glucose, lactate or pyruvate. It was therefore concluded that clinically-relevant
concentrations of the compounds of interest would only be detected via Raman spec-
troscopy if surface-enhancement techniques were used. A decision was thus made to
investigate the possibility of developing a SERS-based sensor for glucose, lactate and
pyruvate.
7.2 Proposed SERS Sensor
When considering the possible designs of a SERS-based sensor for on-line monitoring
of glucose, lactate and pyruvate in microdialysate samples from TBI patients, it was
concluded that a fixed nanostructured surface would not be suitable as a SERS sub-
strate. This was due to the fact that SERS substrates are typically not re-usable due
to instability of the nanostructures. There were additional concerns that the nanos-
tructures may prevent the complete removal of analyte molecules adsorbed on the
substrate surface, resulting in a build up of analyte molecules over time and therefore
a false reading of the analyte concentration. If a fixed SERS substrate was incorpo-
rated into a sensor, these factors would result in a need to regularly replace the SERS
substrate. This process would be labour-intensive and would not be compatible with
the requirement that the developed sensor be capable of continuous on-line operation,
without the need for regular interference from clinical staff.
These concerns led to the proposal of an alternative SERS-based sensor design
in which colloidal nanoparticles were used to provide surface enhancement. Colloids
7.2. PROPOSED SERS SENSOR 117
composed of aggragated nanoparticles have highly nanostructured surfaces and their
use as SERS substrates has previously been reported.125,126 The proposed sensor
design involved a setup in which microdialysate from the patient was combined with
a solution of silver colloids prior to entering the sensing chip, to allow time for the
analyte molecules to adsorb on the colloid surface. Upon entering the sensing chip,
analyte molecules adsorbed on silver colloids would be detected via SERS. The colloid-
analyte complexes would then be flushed from the chip and disposed of, allowing new
silver colloids to enter the system and effectively regenerate the SERS substrate. This
concept is summarised in Figure 7.3.
Figure 7.3: A diagram showing the set up of the proposed SERS sensor. Microdialysate
from the catheter is mixed with fresh silver colloids prior to entering the sensing chip. Analyte
molecules adsorb on the surface of the colloids and are detected via SERS. Colloid particles
with adsorbed analyte are then flushed from the sensing chip and disposed of.
It was decided that silver colloids would be used for an initial investigation of the
viability of this sensing approach, due to the high enhancement factor provided by
silver nanoparticles and the ease with which they can be synthesised.127
7.2.1 Synthesis of Silver Nanoparticles
Silver nanoparticles for aggregation were synthesised using a method based on the
reduction of AgNO3 by sodium citrate.
128 Briefly, 50 mL of an aqueous solution of
1 mM AgNO3 was heated to 90
◦C, with constant stirring, and a 0.1 mL aliquot of
aqueous 1% sodium citrate was added. The temperature of the resulting mixture was
118 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
maintained at 90 ◦C whilst a further three aliquots of 0.3 mL of aqueous 1% sodium
citrate solution were added at 15 minute intervals. When all of the sodium citrate
had been added, the temperature of the resulting solution was maintained at 90 ◦C
for a further 30 minutes, after which it was cooled to room temperature. The silver
nanoparticles were protected from light and stored at 4◦C.
7.3 Testing SERS Sensor Viability: Cysteine Study
The detection of glucose, lactate and pyruvate via SERS was predicted to be non-
trivial due to the low Raman-scattering intensities exhibited by these compounds. As
the proposed SERS sensor was an unfamiliar system, it was decided that it would
be advantageous to assess the viability of the system first using a model analyte so
that any problems with the setup could be unambigously identified. It was felt that if
glucose, lactate or pyruvate were used to characterise the system, it would be unclear
whether any difficulties in spectrum acquisition were due to problems with the setup
or due to the inherent low Raman-scattering intensity of the analyte.
It was concluded that the model analyte should exhibit comparable properties
to glucose, lactate and pyruvate to ensure that the results of the viability study
were relevant. L-cysteine, a naturally-occurring amino acid, was selected as a suitable
candidate because it has a similar molecular structure to glucose, lactate and pyruvate.
Its molecular weight is 121.15 g mol−1, compared to molecular weights of 180.06 g
mol−1, 89.07 g mol−1 and 87.06 g mol−1 for glucose, lactate and pyruvate, respectively,
and it possesses a simple non-aromatic structure, suggesting that it would exhibit
similar Raman-scattering properties. However, cysteine contains a thiol functional
group and was therefore expected to form a bond with a silver substrate, which would
allow chemical enhancement of its Raman signal and thus make it easier to detect via
SERS than glucose, lactate or pyruvate.
This study was conducted in two phases. Initially, the Raman spectra of the solid
and aqueous phases of cysteine were characterised. The SERS behaviour of aqueous
cysteine on silver was then analysed using a fixed SERS substrate comprised of a glass
surface coated with a dense layer of silver nanoparticles. Finally, SERS of aqueous
cysteine was attempted using silver colloids as a SERS substrate. All spectra were
acquired using a 785 nm laser due to the lower background fluorescence observed at
longer laser wavelengths.
7.3.1 Raman Spectroscopy of Solid Cysteine
The Raman spectrum of solid cysteine crystals, mounted on a standard 76 mm × 25
mm × 1.2 mm glass slide, was acquired using a 785 nm laser, a power at the sample of
approximately 0.45 mW and an acquisition time of 300 s. The resulting spectrum is
7.3. TESTING SERS SENSOR VIABILITY: CYSTEINE STUDY 119
shown in Figure 7.4. A low laser power was used in the acquisition of this spectrum in
order to minimise laser-induced damage to the sample. The 200 – 1600 cm−1 region
of the spectrum is displayed, as this is where the majority of the Raman peaks of solid
cysteine are located.
The Raman spectrum of cysteine is well-characterised in the literature, allowing
the peaks in the acquired spectrum to be compared with literature values.129–131 The
acquired spectrum was found to correlate with these literature values and the observed
peaks were assigned to specific vibrational modes of each functional group, based on
the results of a comprehensive study of the Raman spectrum of solid cysteine, as seen
in Figure 7.4.131
A broad bump in the spectrum baseline, centred at a Raman shift of approxi-
mately 1400 cm−1, is observed in the acquired spectrum. This feature was observed
in all Raman spectra of solid cysteine and is thought to indicate thermally-induced
molecular damage that has resulted in the breaking of C–C bonds.
7.3.2 Raman Spectroscopy of Aqueous Cysteine
Raman spectroscopy of aqueous cysteine was predicted to be challenging due to the
low polarisability and consequent low Raman-scattering intensity of the molecule.
A concentrated 1M solution of L-cysteine in DI water was therefore analysed. The
sample was mounted in the macro sampler and a weak spectrum of aqueous cysteine
was successfully acquired using a total acquisition time of 1500 s, comprised of three
acquisitions of 500 s each, and a power at the sample of 100 mW. The spectrum and
associated peak assignments are shown in Figure 7.5.
The Raman spectrum of aqueous cysteine was studied in the same spectral region
as that investigated for solid cysteine to allow the spectra of the two phases to be
compared. However, the acquired spectrum of aqueous cysteine displayed a low sig-
nal:noise ratio and Raman peaks were observed only in the region 550–1500 cm−1.
The spectrum displayed in Figure 7.5 has been limited to this region to allow the
peaks to be visualised. Observed peaks in this region are consistent with previous
studies reporting Raman spectroscopy of aqueous cysteine.130,132
The Raman spectrum of aqueous cysteine retains approximately 30% of the peaks
observed in the Raman spectrum of the solid. The resolution of the aqueous-phase
spectrum also appears to be lower than that of the solid-phase spectrum. This is
evidenced by a broad peak with a Raman shift of 782 cm−1 that is thought to be a
combination of two peaks observed in the Raman spectrum of solid cysteine at 772
cm−1 and 806 cm−1, which respectively correspond to a CH2 rocking vibration and
CO2 scissoring vibration.
120 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
200 400 600 800 1000 1200 1400 1600
Raman Shift / cm -1
0
200
400
600
800
1000
1200
In
te
ns
ity
 / 
Ar
bi
tra
ry
 U
ni
ts
Raman Shift
/ cm−1
Peak
Intensity
Assignment
210 Medium CH2–CH torsion
265 Medium C–C (Aliphatic) Vibration
366 Medium S–H bend
411 Strong C–C (Aliphatic) Vibration
536 Strong CH2–CH–N Bend
639 Strong CH–CO2 Stretch
692 Weak C–S Stretch
772 Medium CH2 Rock
806 Weak CO2 Scissors
823 Medium C–C (Aliphatic) Vibration
868 Medium CO2 Wag
940 Medium N–CH Stretch
1001 Weak CH2–CH Stretch
1065 Medium S–H Bend
1106 Weak NH3 Rock
1142 Weak NH3 Rock
1199 Medium CH2 Twist
1293 Weak C–H Bend
1302 Weak C–H Bend
1343 Weak CO2 Symmetric Stretch
1353 Weak CO2 Symmetric Stretch
1399 Weak CH2 Scissors
1424 Weak CH2 Scissors
1579 Weak NH3 Asymmetric Bend
Figure 7.4: The Raman spectrum and associated peak assignments131 of solid cys-
teine. Cysteine crystals were mounted on a glass slide and the spectrum was acquired
using a 785 nm laser, a power at the sample of approximately 0.45 mW and an ac-
quisition time of 1 × 300 s.
7.3. TESTING SERS SENSOR VIABILITY: CYSTEINE STUDY 121
600 700 800 900 1000 1100 1200 1300 1400 1500
Raman Shift / cm -1
70
80
90
100
110
120
130
140
150
160
In
te
ns
ity
 / 
Ar
bi
tra
ry
 U
ni
ts
Raman Shift
/ cm−1
Peak
Intensity
Assignment
687 Weak C–S Stretch
782 Weak CH2 Rock/CO2 Scissors
822 Weak C–C (Aliphatic) Vibration
939 Medium N–CH Stretch
1063 Weak S–H Bend
1202 Medium CH2 Twist
1303 Weak C–H Bend
1351 Weak CO2 Symmetric Stretch
1419 Weak CH2 Scissors
Figure 7.5: The Raman spectrum and associated peak assignments of a 1M solution of
aqueous cysteine mounted in the macro sampler. Laser wavelength = 785 nm, power at sample
= 100 mW, acquisition time = 3 × 500 s. A software-based multi-spectrum spike filter, which
eliminated any sharp peaks that were only seen in one of the three spectra, was used to remove
the presumed influence of cosmic rays from the spectrum.
7.3.3 SERS of Aqueous Cysteine: Fixed Substrates
The SERS behaviour of cysteine was initially investigated using a fixed substrate
comprised of a glass surface coated with a dense layer of silver nanoparticles. This
intermediate step in the viability study was taken because it was assumed that the
acquisition of a SERS spectrum of cysteine bound to a fixed substrate would be more
straightforward than that of cysteine bound to a colloid in suspension.
122 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
A silver nanoparticle-coated SERS substrate was produced using 25 mm × 25 mm
× 1 mm glass coverslips. Coverslips were cleaned prior to functionalisation by soni-
cation in acetone, followed by sonication in isopropanol, for five minutes each. Clean
coverslips were placed in individual 10 ml glass beakers containing a solution of 10 %
(3-aminopropyl)trimethoxysilane (APTMS) in methanol and incubated for 24 hours
at room temperature. The size of the beaker allowed the slides to stand diagonally,
ensuring that both faces of the slide were exposed to the APTMS/methanol solution.
Following incubation, the slides were rinsed thoroughly with methanol and were visu-
ally tested for hydrophobicity. 100 µL of silver nanoparticle solution was drop-cast on
the slide and incubated at room temperature for one hour. The functionalised surface
was then gently rinsed with methanol and 100 µL of aqueous cysteine was drop cast
on the area of the slide coated with nanoparticles, which was identifiable as a dark
grey spot on the glass surface. After incubation at room temperature for 30 minutes,
the slide was rinsed with methanol and used immediately.
Substrates were mounted on a standard 76 mm × 25 mm × 1.2 mm glass slide and
covered with 100 µL DI water during SERS measurements and a 50X water-immersion
objective was used to focus the laser beam on the sample. A SERS spectrum of 10
µM aqueous cysteine obtained in this way is presented in Figure 7.6. Spectra were
acquired in the region 400 – 1600 cm−1, as this was the region in which peaks were
observed in the Raman spectra of solid and aqueous cysteine.
A comparison of the Raman peaks observed in the spectrum shown in Figure 7.6
with previously-acquired Raman spectra or literature values was not straightforward,
as SERS spectra of a given compound are known to vary, depending upon the par-
ticular vibrational modes that are enhanced. However, a sufficient number of peaks
correlated with reported literature values of the SERS spectrum of cysteine on silver
substrates to be confident that a SERS spectrum of cysteine had been successfully
acquired.130,133 In addition, the observed peak intensity in the spectrum is approxi-
mately 100× higher than that observed in the Raman spectrum of aqueous cysteine,
despite the concentration of cysteine being 105 lower than that used to acquire the
Raman spectrum, indicating that surface enhancement has taken place.
Many of the peaks in the acquired SERS spectrum are within 20 cm−1 of those
acquired in the Raman spectra of solid and aqueous cysteine. Where this is the case,
the peak assignment given to the Raman spectrum was applied, as denoted by a ‘**’
symbol. Peaks which correspond exactly to quoted literature values of the SERS peaks
of cysteine are denoted by a ‘*’ symbol and an assignment based on that reported in
the literature was given. The peaks correspponding to Raman shifts of 603 and 1267
cm−1 do not closely correspond to reported literature values or previously-observed
Raman shifts and were assigned using standard correlation tables.134
Generation of a SERS spectrum of cysteine was not straightforward due to the sig-
7.3. TESTING SERS SENSOR VIABILITY: CYSTEINE STUDY 123
400 600 800 1000 1200 1400 1600 1800
Raman Shift / cm -1
2000
3000
4000
5000
6000
7000
8000
9000
10000
In
te
ns
ity
 / 
Ar
bi
tra
ry
 U
ni
ts
(a)
(b)
Raman Shift
/ cm−1
Peak
Intensity
Assignment
547 Medium CH2–CH–N Bend**
603 Weak C–C (Aliphatic) Vibration
642 Weak CH–CO2 Stretch**
731 Weak C–S Stretch*
814 Medium CH2 Wag / H–C–S Bend*
855 Weak CO−2 Wag**
920 Strong N–CH Stretch**
997 Weak CH2–CH Stretch* **
1076 Weak S–H Bend**
1128 Weak NH3 Rock**
1181 Weak CH3 Twist**
1267 Weak C–C (Aliphatic) Vibration
1293 Weak CH2 Wag / C–C–H Bend* **
1364 Medium CO2 Symmetric Stretch**
1436 Weak CH2 Scissors**
1569 Medium NH3 Asymmetric Bend**
1607 Weak NH3 Bend*
Figure 7.6: SERS spectra of 10 µM cysteine on a layer of silver nanoparticles. The
same sample and acquisition settings were used to acquire spectra (a) and (b): laser = 785
nm, power at sample = 10 mW, acquisition time = 1 × 15 s. Differences in the spectra
are attributed to inadvertent collection of scattered light from different focal planes in
the sample. Peak assignments correspond to Spectrum (a). * = peaks corresponding to
literature spectra,130,133 ** = peaks corresponding to observed Raman spectra of cysteine.
nificant fluorescent background exhibited by the silver nanoparticles. Spectra acquired
using silver nanoparticle-covered fixed SERS substrates were anecdotally observed to
124 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
be highly variable. Significant differences between SERS spectra were observed, de-
pending upon how accurately the laser beam was focused on the surface of the sam-
ple. An example of these differences is shown in Figure 7.6, in which Spectrum B is
shown to exhibit high fluorescence and a poorer signal-to-noise ratio than Spectrum
A, despite both spectra being acquired using the same sample and with identical ac-
quisition parameters. It is thought that the difference between spectra was caused by
the inadvertent selection of different focal planes within the sample. This effect was
predictable, as the SERS effect is known to exponentially decay with distance from
the substrate surface.
The successful detection of cysteine at 10 µM is significant, as this is the lowest
analyte concentration that a sensor for metabolic profiling of TBI patients would be
required to detect. However, although the spectrum was acquired whilst the substrate
surface was submerged in water, it is noted that a 30 minute incubation of analyte
molecules on the substrate surface prior to SERS analysis would not be practical to
adopt in a clinical application.
7.3.4 SERS of Aqueous Cysteine: Colloidal Substrates
Silver nanoparticles were aggregated into colloids by mixing in a 1:1 ratio with per-
fusion fluid. Aqueous cysteine was then added to the aggregated nanoparticles and
incubated for 30 minutes. Following the incubation, a 50µL sample was removed from
the solution and placed on a standard 76 mm × 25 mm × 1.2 mm glass slide for
analysis via SERS.
The acquisition of a SERS spectrum of cysteine on the surface of silver colloids
proved to be very difficult. As discussed in Section 7.3.3, the successful acquisition
of a SERS spectrum requires accurate focus of the laser beam on the surface of the
sample. This was challenging to achieve due to the motion of the colloids within the
perfusion fluid. It is also possible that the high levels of fluorescence observed in the
spectra masked underlying peaks. Examples of the spectra acquired from a sample
of 10 µM cysteine, and their associated acquisition parameters, are shown in Figure
7.7. Once again, these spectra were acquired in the region 400 – 1600 cm−1 to allow
comparison with previously-acquired spectra.
None of the acquired spectra could be confirmed as being a SERS spectrum of
cysteine. All spectra were characterised by relatively high fluorescent backgrounds
and contained few, or no, Raman peaks as exemplified in Figure 7.7. Some spectra
displayed two peaks in the 650 – 750 cm−1 region of the spectrum, as seen in two
of the spectra shown in Figure 7.7. One of these peaks is centred around 658 cm−1,
which does not correlate with the peaks observed in any of the previously-acquired
Raman or SERS spectra of cysteine. The second peak is centred at approximately 738
cm−1. This loosely correlates with a peak observed in the SERS spectrum of cysteine
7.3. TESTING SERS SENSOR VIABILITY: CYSTEINE STUDY 125
400 600 800 1000 1200 1400 1600
Raman Shift / cm -1
0.5
1
1.5
2
2.5
3
In
te
ns
ity
 / 
Ar
bi
tra
ry
 U
ni
ts
104
Power at Sample: 0.5 mW, Acquisition Time: 1 x 120 s
Power at Sample: 0.5 mW, Acquisition Time: 1 x 30 s
Power at Sample: 1.5 mW, Acquisition Time: 3 x 100 s
Power at Sample: 5.0 mW, Acquisition Time: 1 x 120 s
Acqusition Parameters
Figure 7.7: SERS spectra of 10 µM cysteine in a mixture with silver colloids suspended
in perfusion fluid, acquired using a variety of acquisition parameters. The spectra show
relatively few features, aside from two peaks with Raman shifts betwen 650 – 750 cm−1
which appear in two of the spectra. Laser wavelength = 785 nm for all spectra.
on a fixed SERS substrate, at 731 cm−1, and may be attributed to a C–S stretching
vibration.
7.3.5 Testing SERS Sensor Viability: Conclusions
This section has described the characterisation of the Raman spectra of cysteine in
solid and aqueous phases and the assessment of two different substrates for SERS of
cysteine. A fixed substrate of densely-packed silver nanoparticles on a glass surface
was successfully used to acquire a SERS spectrum of cysteine at a concentration of
10 µM, the lowest analyte concentration that a sensor for metabolic profiling of TBI
patients would be required to detect. However, significant variability was observed in
the quality of the acquired spectra. It was also noted that the 30 minute incubation
time allowed for the binding of cysteine to the substrate surface would not be practical
to adopt in a clinical application.
Finally, attempts were made to acquire the SERS spectrum of cysteine adsorbed
on the surface of silver colloids in suspension, as suggested in the proposed SERS
sensor design. This proved to be challenging, due to the difficulty of focusing the
laser beam on the surface of a moving colloid, and the results of these attempts were
126 CHAPTER 7. ASSESSMENT OF RAMAN SPECTROSCOPY
inconclusive.
Although it would have been possible to further refine the process of obtaining
SERS spectra from molecules adsorbed on the surface of the colloidal silver substrates,
experimental work was concluded at this point. The reasons for this are explained in
the following section.
7.4 Conclusions
This chapter presents an exploration of the use of Raman spectroscopy in an on-line
sensor for continuous monitoring of glucose, lactate and pyruvate concentrations in
the microdialysate of TBI patients. Initial attempts to acquire the Raman spectrum of
aqueous glucose suggested that detection of clinical concentrations of glucose, lactate
and pyruvate by Raman scattering-based techniques would only be possible using
surface enhancement. A sensor design, based on the use of silver colloids as SERS
substrates, was proposed and investigated using cysteine as a model analyte. These
studies showed some promising results, including the detection of 10 µM cysteine
adsorbed on a fixed substrate of densely-packed silver nanoparticles.
The use of SERS appeared to successfully overcome the problem of the inherent
insensitivity of Raman spectroscopy and the low scattering intensity exhibited by
glucose, lactate and pyruvate. Following this work, it seemed possible that SERS
could be used to generate Raman peaks corresponding glucose, lactate and pyruvate,
even at the lower limit of the clinical concentration range. In addition to this, the
proposed SERS sensor design addressed the re-usability issues associated with fixed
SERS substrates. However, this study also highlighted serious problems with the
proposed SERS sensor. The significant differences in spectra, depending upon the
focus of the laser beam, and the associated difficulty in obtaining a spectrum of
cysteine on the surface of a silver colloid in suspension, were of particular concern,
as it was assumed that these factors would significantly increase the complexity of
device operation. In addition to this, it was noted that the development of a method
for measurement of glucose, lactate and pyruvate concentrations from SERS spectra
would be very difficult, as the spectra can vary significantly, depending upon the
location and orientation of an analyte molecule with respect to the surface plasmon.
It was noted that, for all Raman and SERS spectra acquired, an increase in signal
intensity may have been achieved via the use of a light source of shorter wavelength.
However, the use of shorter wavelengths can also increase the amount of fluorescence
present in the spectrum and therefore the overall increase in signal:noise ratio was
unlikely to have been significant enough to allow the detection of clinically-relevant
concentrations of glucose, lactate or pyruvate via Raman spectroscopy alone. This
factor was deemed less relevant to the consideration of SERS as a potential sensor
technology, as the difficulties associated with the technique were unrelated to sensor
7.4. CONCLUSIONS 127
sensitivity.
This work was conducted in parallel with the FTIR spectroscopy studies described
in Chapter 6. When compared with the technical issues described here, it was clear
that FTIR spectroscopy showed significantly more promise as an on-line sensor for
metabolic profiling of TBI patients. The decision was thus taken to cease work on the
proposed SERS-based sensor and focus on the development of an FTIR sensor.

Chapter 8
Analysis of Glucose, Lactate and
Pyruvate Mixtures via
ATR-FTIR Spectroscopy
This chapter describes the selection and optimisation of a statistical model for the
prediction of glucose, lactate and pyruvate concentrations from ATR-FTIR spectra
of solutions containing one, two or all three compounds. The primary aim of this
study was to determine the accuracy with which clinically-relevant concentrations of
glucose, lactate and pyruvate could be predicted from ATR-FTIR spectra, in order to
assess the feasibility of using ATR-FTIR for metabolic profiling of TBI patients.
A decision was made not to use clinical microdialysate samples for this study or
for the transmission-FTIR study described in Chapter 9. This was due to the fact
that clinical samples were relatively scarce and of low volume. In order to carry out
a comprehensive study, a large number of samples would have been needed and it
was considered more important to conserve clinical samples for use in testing the final
sensor-chip design. This problem was exacerbated by the fact that the seven-reflection
ATR-FTIR crystal used in this work required sample volumes of 1 ml, meaning that
5 – 10 clinical samples would need to be combined to obtain a single spectrum.
Instead, pure solutions of glucose, lactate and pyruvate in perfusion fluid were pre-
pared at a number of different, clinically-relevant, concentrations and then combined
systematically to produce a series of solutions containing one, two or three com-
pounds. ATR-FTIR spectra of these solutions were acquired using a seven-reflection
ZnSe ATR crystal and analysed in a number of ways. Initially, an analysis based on
the measurement of relative peak intensity, as used in Chapter 6, was considered. It
was concluded that predictive statistical modelling was likely to be a more robust
method for the prediction of analyte concentrations from spectra and therefore two
commonly-used predictive statistical models, PLSR and PCR, were employed. The
129
130 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
PLSR model produced more promising results and thus the remainder of the work
focused on optimising this model to minimise its error in prediction.
8.1 Mixture Selection
The aim of this study was to analyse mixtures containing glucose, lactate and pyru-
vate. This section describes the analysis and decision-making process used to deter-
mine which mixtures should be investigated.
8.1.1 Three-Component Mixtures
Two factors were considered to ensure that the mixtures used in this study were as rel-
evant as possible to the clinical context in which a developed ATR-FTIR sensor would
be used. The first factor to be addressed was the ratio of glucose:lactate:pyruvate that
should be used in order to most closely mimic those found in the microdialysate of
TBI patients. The relationship between the relative amounts of glucose, lactate and
pyruvate present in the extracellular brain fluid of TBI patients is non-trivial to deter-
mine and is a subject of active research. It was therefore decided that the most logical
approach would be to analyse every possible combination of the glucose, lactate and
pyruvate concentrations selected for use in the study, to ensure that a range of ratios
was systematically examined.
The second factor was the selection of glucose, lactate and pyruvate concentra-
tions that were considered most relevant to the monitoring of TBI. It was considered
impractical to conduct a thorough study covering the whole range of glucose, lactate
and pyruvate concentrations observed in the clinic and therefore a subset of concen-
trations with particular clinical significance was selected for investigation. This subset
was partially identified via an analysis of anonymised patient data, collected by the
Department of Clinical Neurosciences at the University of Cambridge over a period of
six years (2009 – 2015), which listed the measured glucose, lactate and pyruvate con-
centrations in the microdialysate of 50 TBI patients at approximately hourly intervals
for periods of 1 – 5 days. The resulting dataset contained approximately 24,000 data
points. These data were analysed by Dr Anna Hu¨fner (Department of Chemistry)
to determine the number of times that each concentration of glucose, lactate and
pyruvate had been observed. A summary of this analysis is shown in Figure 8.1.
The histograms shown in Figure 8.1 were used to identify a concentration range
for each compound which encompassed the majority of observed concentrations. The
identified concentration ranges were:
ˆ Glucose 0.1–3 mM
ˆ Lactate 0.1–6 mM
8.1. MIXTURE SELECTION 131
Figure 8.1: Histograms summarising the frequency with which various concentrations of
glucose, lactate and pyruvate were observed in an analysis of patient data. Figures are adapted
from plots generated by Dr Anna Hu¨fner.
ˆ Pyruvate 0.01–0.2 mM
The highest concentration of each range was selected as the maximum concentration
that would be used to form a mixture. It was assumed that accurate prediction of
analyte concentrations from ATR-FTIR spectra would be easier to achieve at higher
concentrations and therefore it was not considered necessary to place additional focus
on concentrations higher than this. It was noted that the maximum concentrations
of glucose, lactate and pyruvate that a developed sensor would be required to detect
(24 mM, 24 mM and 1.5 mM, respectively) were not recorded in any of the patient
samples.
The lowest clinical concentrations of glucose, lactate and pyruvate that a devel-
oped sensor would be required to detect (0.1 mM, 0.1 mM and 0.01 mM, respectively)
were also used to form three-component mixtures. The ability to accurately measure
these concentrations was considered to be the greatest challenge in sensor development
and therefore it was considered necessary to place additional focus on detecting these
compounds by including them in the mixtures. For pyruvate only, a concentration
below this lower limit (0.005 mM) was included in the study. This additional con-
centration was chosen to acknowledge the fact that the clinical concentration range
of pyruvate is approximately ten-fold lower than those of glucose and lactate and
was therefore likely to be the most difficult compound to detect accurately. It was
decided that additional focus should be placed on the lower limit of clinical pyruvate
concentrations to determine the accuracy with which it could be predicted. In order
to do this, it was necessary to include a concentration below the concentration limit,
to assess whether concentrations at that limit could be accurately differentiated from
concentrations that are lower still.
The conclusions of the Consensus Statement from the 2014 International Micro-
132 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
dialysis Forum were also taken into account when selecting clinically significant con-
centrations for inclusion in mixtures. These conclusions identified pathological thresh-
olds for the concentrations of glucose, lactate and pyruvate, as discussed in Sections
2.4.2 and 2.4.3. It was decided that the ability to accurately differentiate concentra-
tions lying close to these boundaries would be a key indicator of sensor performance
and therefore concentrations close to these thresholds were included in this study.
A glucose concentration below 0.8 mM is considered an indicator of poor patient
outcome in TBI and therefore glucose concentrations of 0.7 mM and 1.0 mM were
included in this study. The pathological threshold of lactate is 4 mM and thus lactate
concentrations of 2 mM and 4 mM were selected for use in this study.
No pathological threshold for the absolute concentration of pyruvate is given in
the Consensus Statement. Instead, the remaining concentrations of pyruvate to be
included in the three-component mixtures were selected by evenly dividing the key
concentration range of pyruvate, identified from the patient histograms (Figure 8.1), in
order to cover the entire range. It was hoped that the small differences between these
concentrations would lead to more accurate predictions of the pyruvate concentration
and hence a higher resolution of the sensor.
The concentrations of glucose, lactate and pyruvate used to form the three-component
mixtures analysed in this study are summarised in Table 8.1.
The concentrations of glucose, lactate and pyruvate listed in Table 8.1 were system-
atically mixed, in every possible combination which included all three compounds, to
produce a total of 96 three-component mixtures. An additional three three-component
mixtures, considered ‘special cases’, were also studied:
ˆ Glu: 0.05 mM, Lac: 0.05 mM, Pyr: 0.005 mM - All three compounds
below the expected clinical range.
ˆ Glu: 0.05 mM, Lac: 2.00 mM, Pyr: 0.005 mM - Glucose and pyruvate
below the clincal concentration range, lactate within it.
ˆ Glu: 1.00 mM, Lac: 0.05 mM, Pyr: 0.005 mM - Lactate and pyruvate
below the clinical concentration range, glucose within it.
A full list of all three-component mixtures analysed in this study can be found in
Appendix A.1.
8.1.2 Two-Component Mixtures
It was theorised that it may also be useful to analyse spectra containing just two of
the three compounds. It was also decided that a slightly different set of concentrations
to those used in the three-component mixtures should be used for the two-component
mixtures, in order to allow a greater range of concentrations to be studied overall. An
8.1. MIXTURE SELECTION 133
Table 8.1: Concentrations of glucose, lactate and pyruvate used
to form the three-component mixtures analysed via ATR-FTIR
spectroscopy.
Glucose / mM Lactate / mM Pyruvate / mM
0.1 0.1 0.005
0.7 2.0 0.01
1.0 4.0 0.05
3.0 6.0 0.10
0.15
0.20
additional concentration of 0.05 mM was included for both glucose and lactate, in order
to allow an assessment of the accuracy with which the lowest clinical concentrations of
those compounds (0.1 mM for both) could be detected. This decision meant that, for
practical reasons, the upper concentration limits used in the three-component mixture
study were not used. One pyruvate concentration (0.1 mM) was also removed for
practical reasons. The concentrations of glucose, lactate and pyruvate used to form
two-component mixtures are listed in Table 8.2.
Table 8.2: Concentrations of glucose, lactate and pyruvate used
to form the two-component mixtures analysed via ATR-FTIR spec-
troscopy.
Glucose / mM Lactate / mM Pyruvate / mM
0.05 0.05 0.005
0.1 0.1 0.01
0.7 0.5 0.05
1.0 4.0 0.15
0.2
A list of two-component mixtures included in this study is shown in Appendix
A.2.
8.1.3 Pure Solutions
Spectra of pure solutions of glucose, lactate and pyruvate were also included in this
study. The concentrations used in the two- and three-component mixtures were all
analysed as pure solutions. Additional concentrations, close in value to those used in
the mixtures, were added in an attempt to improve the resolution in the prediction of
concentrations. At a later date, the data generated using the ZnSe ATR-FTIR crystal
during the ATR-FTIR crystal study, described in Section 6.4.1, were also included in
the analysis to increase the size of the dataset. Table 8.3 lists the concentrations of
glucose, lactate and pyruvate studied as pure solutions, including those used in the
ATR-FTIR crystal study.
134 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
Table 8.3: Concentrations of glucose, lactate and pyruvate stud-
ied as pure solutions via ATR-FTIR spectroscopy, including those
from the ATR-FTIR crystal study.
Glucose / mM Lactate / mM Pyruvate / mM
0.05 0.05 0.005
0.10 0.10 0.01
0.19 0.19 0.02
0.20 0.20 0.05
0.40 0.40 0.10
0.70 0.50 0.15
0.75 0.75 0.20
1.00 1.50 0.75
1.50 2.00 1.00
3.00 3.00 1.50
5.00 4.00 7.50
6.00 6.00
10.0 10.0
12.0 12.0
24.0 24.0
8.2 Mixture Selection: Conclusions
This section has described the reasoning used to select solutions of glucose, lactate
and pyruvate for use in a study to determine the feasibility of using ATR-FTIR
spectroscopy as the basis of an on-line sensor for metabolic profiling of TBI patients.
8.3 Data Acquisition
The chosen solutions were prepared in perfusion fluid and their spectra acquired using
a Perkin Elmer Spectrum 100 FTIR Spectrometer, coupled with a seven-reflection
ZnSe HATR Sampling Accessory, as described in Section 5.2. The seven-reflection
ZnSe ATR-FTIR crystal was that identified in Section 6.4.1 as exhibiting the best
performance for the detection of pure solutions of glucose, lactate and pyruvate (6.4.1).
Three aliquots of each solution were analysed via ATR-FTIR spectroscopy. Two
spectra were acquired of each aliquot: one ratio spectrum, in which perfusion fluid
was used as the background, and one single-beam spectrum. Spectra of pure perfusion
fluid were also acquired. A total of 16 aliquots of perfusion fluid were analysed and
two spectra were acquired of each aliquot: one ratio spectrum in which perfusion fluid
was also used as the background and one single-beam spectrum.
The spectra of three-component mixtures, two-component mixtures and pure so-
lutions of glucose, lactate and pyruvate were combined into two large datasets, one
consisting of all of the acquired ratio spectra and one consisting of all the acquired
single-beam spectra. A total of 643 ratio and 643 single-beam spectra were acquired.
8.4. SPECTRUM ANALYSIS METHODS 135
Each dataset was used in its entirety for the analysis described in this chapter, without
differentiating between the spectra of one-, two- and three-component solutions.
A sample of the spectra from each dataset can be seen in Figures 8.2 and 8.3. The
spectra plotted in this figure were selected by choosing every 37th spectrum from a
list of all spectra ordered by increasing glucose concentration.
5001000150020002500300035004000
Wavenumber / cm -1
98
98.5
99
99.5
100
100.5
101
%
 T
ra
ns
m
iss
io
n
   0           0           0
   0           0         0.1
   0        0.05           0
   0         0.1         0.2
   0           3           0
0.05           0           0
0.05           4           0
 0.1         0.1         0.1
 0.1           4        0.01
 0.2           0           0
 0.7         0.1         0.1
 0.7           4        0.01
   1           0           0
   1         0.1        0.15
   1           4        0.05
   3           0           0
   3           4        0.01
   5           0           0
Conc. / mM (Glu  Lac  Pyr)
Figure 8.2: A randomly-selected sample of the ratio ATR-FTIR spectra acquired
for the ATR-FTIR mixture study. Concentrations are listed in the horizontal order:
glucose - lactate - pyruvate.
8.3.1 Data Acquisition: Conclusions
This section has described the process used to generate two large datasets of ATR-
FTIR spectra, one consisting of ratio spectra and one consisting of single-beam spec-
tra, for use in this study of ATR-FTIR spectroscopy.
8.4 Spectrum Analysis Methods
The solutions of glucose, lactate and pyruvate selected in Section 8.1 were chosen
with the intention of analysing their spectra via statistical modelling rather than an
analysis of peak intensities. This decision was made following the results of the ATR-
FTIR crystal study described in Section 6.4.1, in which the LoQ of each compound,
as determined from the relationship between peak intensity and concentration, were
found to be significantly higher than those required of the developed sensor. These re-
sults indicated that an analysis of peak intensities would not be a sufficiently sensitive
method of analysing the acquired spectra.
136 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
5001000150020002500300035004000
Wavenumber / cm -1
0
50
100
150
200
Ar
bi
tra
ry
 U
ni
ts
   0           0           0
   0           0         0.1
   0        0.05           0
   0         0.1         0.2
   0           3           0
0.05           0           0
0.05           4           0
 0.1         0.1         0.1
 0.1           4        0.01
 0.2           0           0
 0.7         0.1         0.1
 0.7           4        0.01
   1           0           0
   1         0.1        0.15
   1           4        0.05
   3           0           0
   3           4        0.01
   5           0           0
Conc. / mM (Glu  Lac  Pyr)
Figure 8.3: A randomly-selected sample of the single-beam ATR-FTIR spectra ac-
quired for the ATR-FTIR mixture study. Concentrations are listed in the horizontal
order: glucose - lactate - pyruvate.
The use of peak intensity measurements to analyse the spectra of mixtures of
glucose, lactate and pyruvate raised additional concerns. Although peaks unique to
glucose, lactate and pyruvate had been identified from a mixture of all three com-
pounds (Section 6.1) there was concern that, if any of these peaks were positioned
close to an absorption peak of another compound in the sample, the intensity of the
unique peak may be affected by the shoulder of the adjacent peak. This effect was
addressed quantitatively through a statistical analysis of the ratio ATR-FTIR spectra
acquired in this study.
The statistical analysis was conducted on spectral positions corresponding to two
unique glucose peaks, 993 cm−1 and 1081 cm−1. The mean and standard deviation
of the percentage transmission were calculated at each position. Initially, the mean
and standard deviation of the ATR-FTIR ratio spectra of perfusion fluid (‘blank
spectra’) at both 993 cm−1 and 1081 cm−1 were calculated. This process was then
repeated for three additional subsets of the ATR-FTIR ratio spectra dataset: spectra
of pure solutions of lactate, spectra of pure solutions of pyruvate and spectra of two-
component mixtures containing lactate and pyruvate. None of the analysed spectra
represented solutions containing glucose. It was hypothesised that, if the glucose peaks
at 993 cm−1 and 1081 cm−1 were unaffected by the presence of pyruvate or lactate
in the solution, the mean and standard deviation of the percentage transmission at
each peak position of the blank spectra would be similar to those of the spectra with
additional compounds added. The results of this analysis are shown in Table 8.4.
8.5. ATR-FTIR SPECTRUMANALYSIS: COMPARISONOF STATISTICALMODELS137
Table 8.4: The mean and standard deviation (SD) of spectra at the positions of two
unique glucose peaks.
Compound
1081 cm−1 Peak 993 cm−1 Peak
Mean SD Mean SD
(Blank) 100.003 0.0186 100.001 0.0264
Lactate 99.946 0.0599 100.014 0.0445
Pyruvate 100.001 0.0241 100.000 0.0214
Lactate & Pyruvate 99.983 0.0250 100.002 0.0201
It can be seen from Table 8.4 that the mean and standard deviation at both peak
positions increased from their blank spectra values when lactate was present in the
sample. The effect was greater in spectra of solutions containing lactate only than in
spectra of solutions containing both lactate and pyruvate. It is thought that this is
due to the fact that higher concentrations of lactate were used in single-component
solutions than in two-component solutions. This provided an indication that glucose
peak intensity may be affected by the presence of lactate.
8.4.1 Spectrum Analysis Methods: Conclusions
The evidence presented in this chapter, in combination with results described in Chap-
ter 6.4.1, was considered sufficient to conclude that an ATR-FTIR-based sensor for
glucose, lactate and pyruvate which incorporated an analysis of peak intensities to
determine the levels of each compound would not be suitable for use in the clinic. It
was therefore decided that the use of predictive statistical modelling in the analysis
of the ATR-FTIR spectral dataset would be assessed.
8.5 ATR-FTIR Spectrum Analysis: Comparison of Sta-
tistical Models
The first stage in the analysis of the acquired ATR-FTIR spectral dataset was the
selection of a predictive statistical model. It was initially assumed that the data were
linear and could therefore be analysed using a linear regression model, which would be
more straightforward to implement than a non-linear model. Two of the most widely-
used linear regression models are PCR and PLSR, and it was therefore decided that
the performance of these two models should be compared.
Both statistical models were individually applied, in turn, to the ATR-FTIR ratio
spectra and ATR-FTIR single-beam spectra datasets. Little was known about the
potential of the data to be modelled and therefore it was deemed logical to apply
the model to as much of the spectral range as possible. Although all spectra were
acquired over the range 4000 – 450 cm−1, noise was seen below 650 cm−1 due to the
IR transmission cut-off of the ZnSe crystal, as seen in Figure 8.2. Attemping to model
138 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
this noise would worsen the performance of both models, as it is by definition random,
and therefore both models were applied over the spectral range 4000 – 650 cm−1.
Three criteria were used to evaluate the performance of the two models. The first
criterion was the number of model components required to reach a maximum R2 value,
with the use of fewer components considered desirable as it reduced the likelihood of
model over-fitting. A maximum limit of 30 components was imposed for this analysis.
The predictive capacity of each model was the feature of most interest to the de-
velopment of a sensor based on ATR-FTIR spectroscopy. This was assessed by calcu-
lating the RMSECV in the prediction of glucose, lactate and pyruvate concentrations,
as determined from the results of a leave-one-out cross-validation. The RMSECV was
selected as a measure of predictive accuracy, rather than the RMSEP, in order to re-
tain the size of the dataset and reduce the effect of any outlying values. The number
of model components required to minimise the RMSECV was one way in which the
model was assessed, with fewer components once again considered preferable. In ad-
dition, the RMSECV in the prediction of the concentration of each compound, using
a model with 20 components, was measured.
8.5.1 PCR vs. PLSR: Ratio ATR-FTIR Spectra
A summary of the performance of a PCR model applied to the dataset of ATR-FTIR
ratio spectra is shown in Figure 8.4.
0 5 10 15 20 25 30
Number of PCR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PCR Components
0
0.5
1
1.5
2
2.5
3
R
M
SE
CV
 o
f P
CR
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 8.4: A summary of the performance of a PCR model applied to the dataset of ATR-
FTIR ratio spectra of glucose, lactate and pyruvate. The change in R2 with an increasing
number of model components and the change in the RMSECV in concentration predictions,
calculated using leave-one-out cross-validation, are shown for each compound.
It can be seen from Figure 8.4 that the model required more components to ac-
curately model pyruvate than to model glucose or lactate. This was expected, as
8.5. ATR-FTIR SPECTRUMANALYSIS: COMPARISONOF STATISTICALMODELS139
pyruvate is present at concentrations approximately ten-fold lower than those of glu-
cose and lactate. The R2 values of the models of all three compounds appear to
converge at a value very close to 1.0 when more than 25 components are used, which
is a possible sign of over-fitting. Conversely, the RMSECV for each compound does
not appear to converge, even when 30 model components are used. The prediction
of glucose concentrations generates the highest absolute error, although the error in
pyruvate prediction is proportionally higher.
A summary of the performance of a PLSR model on the same dataset of ATR-
FTIR ratio spectra is shown in Figure 8.5. The results generated by this model
appear to follow broadly similar trends to those generated by the PCR model: more
components were required to accurately model pyruvate than to model glucose or
lactate, though R2 values converged when a large number of components were used,
and the prediction of glucose concentrations generated the highest absolute error.
However, a quantitative comparison of the data, as shown in Table 8.5, showed that
there were distinct differences between the performances of the two models.
0 5 10 15 20 25 30
Number of PLSR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PLSR Components
0
0.5
1
1.5
2
2.5
3
R
M
SE
CV
 o
f P
LS
R
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 8.5: A summary of the performance of a PLSR model applied to the dataset of ATR-
FTIR ratio spectra of glucose, lactate and pyruvate. The change in R2 with an increasing
number of model components and the change in the RMSECV in concentration predictions,
calculated using leave-one-out cross-validation, are shown for each compound.
The data summarised in Table 8.5 show that, by all of the assessed criteria, the
PLSR model performed better than the PCR model. The PLSR model required fewer
components to accurately model the lactate and pyruvate data, as measured by the
R2 value, and used an equal number of components to model the glucose data. In
addition, the PLSR model required fewer components to minimise the RMSECV in
predicted glucose, lactate and pyruvate concentrations and showed a lower RMSECV
for all compounds when a model with 20 components was used to analyse the data.
140 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
Table 8.5: A comparison of the performances of PCR and PLSR statistical models when
applied to the same set of ATR-FTIR ratio spectra.
Compound Parameter PCR PLSR
Glucose
Number of components for max. R2 Value 11 11
Number of components for min. RMSECV 18 21
RMSECV with 20 Components 0.41 0.27
Lactate
Number of components for max. R2 Value 13 8
Number of components for min. RMSECV 22 19
RMSECV with 20 Components 0.26 0.15
Pyruvate
Number of components for max. R2 Value 20 13
Number of components for min. RMSE 21 18
RMSECV with 20 Components 0.173 0.087
8.5.2 PCR vs. PLSR: Single-Beam ATR-FTIR Spectra
The performances of a PCR model and a PLSR model applied to the dataset of ATR-
FTIR single-beam spectra are shown in Figures 8.6 and 8.7, respectively. PLSR was
once again the better-performing model, based on the assessed criteria, although the
differences between the performances of the two models were observed to be more
significant when applied to single-beam spectra. The PLSR model was able to accu-
rately model the data with significantly fewer components than the number required
by the PCR model, which required more than 30 components to achieve an R2 value
greater than 0.8. Similarly, the RMSECV decreased more rapidly with an increasing
number of model components when PLSR was used to model the data.
0 5 10 15 20 25 30
Number of PCR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PCR Components
0.5
1
1.5
2
2.5
3
R
M
SE
CV
 o
f P
CR
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 8.6: A summary of the performance of a PCR model applied to ATR-FTIR single-
beam spectra of glucose, lactate and pyruvate. The change in R2 with an increasing number
of model components and the change in the RMSECV in concentration predictions, calculated
using leave-one-out cross-validation, are shown for each compound.
8.5. ATR-FTIR SPECTRUMANALYSIS: COMPARISONOF STATISTICALMODELS141
0 5 10 15 20 25 30
Number of PLSR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PLSR Components
0
0.5
1
1.5
2
2.5
3
R
M
SE
CV
 o
f P
LS
R
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 8.7: A summary of the performance of a PLSR model applied to ATR-FTIR single-
beam spectra of glucose, lactate and pyruvate. The change in R2 with an increasing number
of model components and the change in the RMSECV in concentration predictions, calculated
using leave-one-out cross-validation, are shown for each compound..
The differences between the performance of the two models are quantified in Table
8.6. It is also noted that the performances of both models were poorer when applied
to single-beam spectra than when they were applied to ratio spectra. This is thought
to be due to the fact that single-beam spectra contain influences from multiple factors,
such as the instrument optics and environment. These influences are removed from
ratio spectra, as they are created by the division of two single-beam spectra. It is
therefore likely that more information is contained within the single-beam spectra and
hence a greater number of components would be needed to account for the variation
in the data.
Table 8.6: A comparison of the performances of PCR and PLSR statistical models when
applied to the same set of ATR-FTIR single-beam spectra.
Compound Parameter PCR PLSR
Glucose
Number of components for max. R2 Value >30 21
Number of components for min. RMSECV 29 26
RMSECV with 20 Components 1.9 0.63
Lactate
Number of components for max. R2 Value 29 20
Number of components for min. RMSECV 30 26
RMSECV with 20 Components 1.05 0.53
Pyruvate
Number of components for max. R2 Value 29 25
Number of components for min. RMSECV >30 26
RMSECV with 20 Components 1.095 0.601
142 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
8.5.3 Comparison of Statistical Models: Conclusions
This section describes a comparison of the performances of PCR and PLSR statistical
models when applied to the same set of ATR-FTIR spectra. PLSR was the best-
performing of the two models, when applied to both ATR-FTIR ratio spectra and
ATR-FTIR single-beam spectra, with respect to the modelling of data and predictive
accuracy. These results may be explained by the differing mechanisms by which the
two models operate: PCR is an unsupervised technique that does not consider the
dependent variable when constructing principal components to explain the variation
in the data whilst PLSR, as a supervised technique, does consider the dependent
variable.
These results suggested that predictive statistical modelling could successfully be
used to analyse the spectra generated by an ATR-FTIR spectroscopy-based sensor for
glucose, lactate and pyruvate. The results presented in this section led to the selection
of the PLSR model for further refinement and testing.
8.6 ATR-FTIR: Effect of Spectral Range
Having chosen to analyse the acquired ATR-FTIR spectra using a predictive PLSR
model, attempts were made to optimise its performance. The intention behind the
use of the model was to predict the concentrations of glucose, lactate and pyruvate
directly from the ATR-FTIR spectra of solutions containing one, two or three of those
compounds. As a result, it was concluded that the most important parameter of the
PLSR model was the degree of error in its prediction of glucose, lactate and pyruvate
concentrations. This section describes the work undertaken to minimise that error.
One factor that was assumed to have a significant impact on the error in prediction
of the model was the spectral range over which the model was applied. A systematic
study was therefore carried out to determine which spectral interval yielded the lowest
RMSECV. This parameter was again selected over the RMSEP in order to retain the
size of the training dataset. Minimisation of the RMSECV was achieved by assessing
two spectral parameters: the size of the interval and the region of the spectrum in
which the interval was located. The assessed interval sizes were: 100, 200, 300, 400,
500, 600, 700, 800, 900 and 1000 cm−1. The region of the spectrum over which the
interval was located was defined by the ‘interval start point’, which marked the lowest
end of the interval range (in cm−1). Interval start points occurred every 100 cm−1.
Intervals of all selected sizes were assessed at each interval start point. For exam-
ple, for an interval start point of 1000 cm−1, the following intervals were assessed:
ˆ 1000 - 1100 cm−1 (Interval size = 100 cm−1)
ˆ 1000 - 1200 cm−1 (Interval size = 200 cm−1)
8.6. ATR-FTIR: EFFECT OF SPECTRAL RANGE 143
ˆ 1000 - 1300 cm−1 (Interval size = 300 cm−1)
ˆ 1000 - 1400 cm−1 (Interval size = 400 cm−1)
ˆ 1000 - 1500 cm−1 (Interval size = 500 cm−1)
ˆ 1000 - 1600 cm−1 (Interval size = 600 cm−1)
ˆ 1000 - 1700 cm−1 (Interval size = 700 cm−1)
ˆ 1000 - 1800 cm−1 (Interval size = 800 cm−1)
ˆ 1000 - 1900 cm−1 (Interval size = 900 cm−1)
ˆ 1000 - 2000 cm−1 (Interval size = 1000 cm−1)
A PLSR model was applied to every selected spectral interval and was calibrated
using the leave-one-out method of cross-validation. At each interval, the model was
iterated multiple times using a different number of components, up to a maximum of
20 components, until the overall RMSECV for all three compounds was minimised.
The minimised RMSECV, and the number of components needed to achieve it, were
recorded for each interval. This procedure was carried out separately on the datasets
of ratio ATR-FTIR spectra and single-beam ATR-FTIR spectra and was applied to
the spectral range 4000 - 650 cm−1.
8.6.1 Effect of Spectral Range: ATR-FTIR Ratio Spectra
The results of the interval study carried out on ATR-FTIR ratio spectra are sum-
marised in Figure 8.8. The centre of each coloured rectangle in the figure can be
mapped to an interval start point (x-axis) and an interval size (y-axis), whilst the
colour of the rectangle reflects the lowest-achievable RMSECV of the model when it
is applied to the specified spectral interval using a maximum of 20 components. It
can be seen that the overall RMSECV for all compounds was lowest for intervals with
a starting point between 700 – 1400 cm−1. This could be predicted from the work
described in Chapter 6, in which the majority of absorbance peaks in the ATR-FTIR
spectra of glucose, lactate and pyruvate were located in the fingerprint region of the
spectrum.
The interval yielding the lowest overall RMSECV had a size of 600 cm−1 and was
located at the 900 – 1500 cm−1 region of the spectrum. The overall RMSECV of this
interval was 0.226 mM, whilst the RMSECV for the prediction of the concentrations,
in mM, of each individual compound over this interval were:
ˆ Glucose: 0.20
ˆ Lactate: 0.08
144 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
All Compounds
700 1100 1500 1900 2300 2700
Interval Start / cm -1
100
200
300
400
500
600
700
800
900
1000
In
te
rv
al
 S
ize
 / 
cm
-
1
0.5
1
1.5
2
2.5
3
3.5
RMSECV / mM
Figure 8.8: A colourmap showing the effect of spectral interval location and size on the
RMSECV of a PLSR model applied to ATR-FTIR ratio spectra.
ˆ Pyruvate: 0.068
A 20-component model was required in order to achieve these RMSECV values.
These errors were lower than those achieved by applying the PLSR model over the
whole spectrum, as seen in Section 8.5.1, and therefore this study was successful
in improving the predictive performance of PLSR when applied to ATR-FTIR ratio
spectra.
8.6.2 Effect of Spectral Range: ATR-FTIR Single-Beam Spectra
The results of the interval study carried out on ATR-FTIR single-beam spectra are
summarised in Figure 8.9.
These results show a general increase in the overall RMSECV across all intervals
when compared with the analysis of ATR-FTIR ratio spectra. This trend is in agree-
ment with the observation in Section 8.5.2 that statistical modelling of single-beam
ATR-FTIR spectra is more complex than that of ATR-FTIR ratio spectra. It can be
seen from the figure that the intervals yielding the lowest overall RMSECV are those
between 200 – 400 −1 in length which begin at 900 cm−1 or 1000 cm−1.
The interval which produced the lowest overall RMSECV had a size of 400 cm−1
and was located at the 900 – 1300 cm−1 region of the spectrum. A 20-component
8.6. ATR-FTIR: EFFECT OF SPECTRAL RANGE 145
All Compounds
700 1100 1500 1900 2300 2700
Interval Start / cm -1
100
200
300
400
500
600
700
800
900
1000
In
te
rv
al
 S
ize
 / 
cm
-
1
0.5
1
1.5
2
2.5
3
3.5
RMSECV / mM
Figure 8.9: A colourmap showing the effect of spectral interval location and size on the
RMSECV of a PLSR model applied to ATR-FTIR single-beam spectra.
model was required in order to achieve an overall RMSECV for this interval of 0.306
mM. The RMSECV for the prediction of the concentrations, in mM, of each individual
compound over this interval were:
ˆ Glucose: 0.26
ˆ Lactate: 0.12
ˆ Pyruvate: 0.112
Once again, the RMSECV for the prediction of the concentration of each individual
compound was lower than that achieved when PLSR was applied over the whole
spectral range, as seen in Section 8.5.2, meaning that this study was successful in
improving the predictive performance of the model when applied to ATR-FTIR single-
beam spectra.
8.6.3 Effect of Spectral Range: Conclusions
This section has described the work undertaken to lower the RMSECV of a PLSR
model, applied to the prediction of glucose, lactate and pyruvate concentrations from
ATR-FTIR spectra, by identifying an optimum spectral range over which to apply
146 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
the model. The values of the RMSECV generated for both ratio and single-beam
ATR-FTIR spectra were successfully lowered from those obtained by applying the
same model over the whole spectrum. The model achieved a lower overall RMSECV
value, as well as lower RMSECV values for the prediction of all three compounds,
when applied to ratio spectra than when applied to single-beam spectra.
8.7 ATR-FTIR: Prediction of LPR
The LPR is an important indicator of patient outcome following TBI, as discussed in
Section 2.4.3, with a value of 25 considered to be the pathological threshold.7 Accurate
calculation of the LPR would therefore be an important indicator of the performance
of an ATR-FTIR-based sensor for metabolic profiling of TBI patients.
The ability of the optimised PLSR models to accurately detect the LPR from ATR-
FTIR spectral data was analysed by randomly dividing the spectra into ’training’ and
’test’ datasets. A test dataset size of 64 spectra, corresponding to approximately
10% of the dataset, was chosen in order to maximise the size of the training set.
PLSR was carried out on the training dataset, over the spectral region identified in
Section 8.6 as yielding the lowest overall RMSECV, and calibrated using leave-one-
out cross-validation. The calibrated model was then applied to the test dataset for
the prediction of glucose, lactate and pyruvate concentrations and the predicted LPR
was determined by dividing the predicted concentration of lactate by the predicted
concentration of pyruvate for each sample. The actual LPR was calculated by dividing
the observed concentrations of lactate and pyruvate for each sample in the dataset.
This process was used to predict the LPR from both ratio and single-beam spectra.
8.7.1 ATR-FTIR: Prediction of LPR from Ratio Spectra
The PLSR model used to predict the LPR from the dataset of ATR-FTIR ratio
spectra was applied over the spectral range 1500 - 900 cm−1 and used 20 components,
the combination identified in Section 8.6 as yielding the lowest overall RMSECV. A
comparison of the observed and predicted values of the LPR can be seen in Figure
8.10(a).
It is clear from the figure that there was a significant error in the prediction of LPR
values. Most notably, a number of the predicted values are negative, which indicates
that the PLSR model had predicted negative concentration values for either lactate or
pyruvate. The R2 value for a linear relationship between the observed and predicted
values of the LPR was 0.10, which confirms a significant lack of agreement between
the two sets of values and hence a large error in the prediction of the LPR. The
magnitude of the error in LPR prediction is explained by the fact that the individual
errors in the prediction of lactate and pyruvate were compounded by the division of
8.7. ATR-FTIR: PREDICTION OF LPR 147
0 200 400 600 800 1000 1200
Observed LPR
-400
-200
0
200
400
600
800
1000
1200
1400
1600
Pr
ed
ict
ed
 L
PR
(a)
0 50 100 150
Observed LPR
0
50
100
150
Pr
ed
ict
ed
 L
PR
(b)
Figure 8.10: Scatter plots showing the observed vs. predicted LPR. Predicted values of the
LPR were derived from the concentrations of lactate and pyruvate predicted from ATR-FTIR
ratio spectra using a PLSR model. Plot (a) shows all results, whilst (b) shows only the region
containing the most data points.
the two values to generate the predicted LPR.
However, it was noted that when the scale of Figure 8.10(a) was adjusted to show
only the region of the plot in which the majority of the data points were located,
the relationship between the observed and predicted LPR appeared did appear to be
qualitatively more linear (Figure 8.10(b)).
8.7.2 ATR-FTIR: Prediction of LPR from Single-Beam Spectra
A 20-component PLSR model was used to predict the LPR from ATR-FTIR single-
beam spectra and was applied over the spectral range identified in Section 8.6 as
yielding the lowest overall RMSECV (1300 - 900 cm−1). A comparison of the observed
and predicted values of the LPR is shown in Figure 8.11. Once again, there was a
significant error in the prediction of the LPR, as confirmed by an R2 value for a linear
relationship of the data of 0.04. In addition, several of the predicted LPR values were
negative, indicating that the PLSR model had again generated negative values of the
concentrations of lactate or pyruvate.
The prediction of the LPR from ATR-FTIR ratio spectra was more accurate than
that obtained using ATR-FTIR single-beam spectra. This was expected due to the
higher RMSECV exhibited by a PLSR model when applied to single-beam spectra.
8.7.3 Prediction of LPR: Conclusions
This section describes an attempt to predict the LPR from ATR-FTIR spectra via a
straightforward division of the concentrations of lactate and pyruvate predicted using
148 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
0 200 400 600 800 1000 1200
Observed LPR
-80
-60
-40
-20
0
20
40
60
80
Pr
ed
ict
ed
 L
PR
(a)
0 20 40 60 80 100 120
Observed LPR
0
20
40
60
80
100
120
Pr
ed
ict
ed
 L
PR
(b)
Figure 8.11: Scatter plots showing the observed vs. predicted LPR. Predicted values of the
LPR were derived from the concentrations of lactate and pyruvate predicted from ATR-FTIR
single-beam spectra using a PLSR model. Plot (a) shows all results, whilst (b) shows only the
region containing the most data points.
a PLSR model. It has been shown that the RMSEP currently achieved by the model,
particularly with respect to the prediction of pyruvate concentrations, was too high
to allow accurate prediction of the LPR using this method. Of particular concern was
the prediction of negative LPR values, for both ratio and single-beam spectra, which
indicate that the PLSR models had calculated negative values for the concentration of
either lactate or pyruvate. This generation of invalid predictions shows that, in their
current form, the models are not suitable for use in the clinic. However, it is noted
that the RMSEP may have been affected by the reduction in the size of the training
datasets of ratio and single-beam spectra, which was caused by the separation of the
datasets into ‘training’ and ‘test’ datasets.
8.8 Conclusions
This chapter has described a systematic study of the potential of ATR-FTIR to be used
as the basis of a sensor for the analysis of glucose, lactate and pyruvate concentrations
present in microdialysate from TBI patients. A range of synthetic solutions of glucose,
lactate and pyruvate, contaning one, two or three compounds, was rationally selected
using patient data and reported literature values. Prepared samples were analysed
via ATR-FTIR spectroscopy, using a seven-reflection ZnSe crystal, to yield two large
datasets of 643 ratio spectra and 643 single-beam spectra.
Two statistical models were applied to each dataset: PCR and PLSR. It was
found that PLSR performed better at both the modelling and prediction of data and
8.8. CONCLUSIONS 149
therefore this model was selected for optimisation. A systematic study identified an
optimum spectral range over which the model should be applied, which minimised the
overall RMSECV, of 1500 – 900 cm−1 for ATR-FTIR ratio spectra and 1300 – 900
cm−1 for single-beam spectra. In both cases, 20 components were required to achieve
the minimum RMSECV.
The application of the optimised PLSR model to ATR-FTIR ratio spectra was
found to yield a lower RMSECV, both overall and for each compound studied, than
that obtained when the model was applied to single-beam spectra. This was explained
by the additional information contained within single-beam spectra, which are affected
by instrument optics and the sample environment, which is not present in ratio spectra
due to the way in which it is derived.
The viability of using ATR-FTIR spectroscopy, in combination with the optimised
PLSR model, for metabolic profiling of TBI patients can be assessed by comparing the
RMSECV of the optimised PLSR model with the value of the lowest concentration
of each compound that a developed sensor would be required to detect. This is due
to the fact that, in order to accurately quantify a given concentration, the error in
measurement must be lower than that concentration. It can therefore be assumed that
the RMSECV for the prediction of each compound must be lower than the minimum
possible clinical concentration of that compound in order for a developed sensor to
perform accurately over the whole clinical range. A comparison of these values for the
optimised PLSR model applied to ATR-FTIR ratio spectra is shown in Table 8.7.
Table 8.7: A comparison of the RMSECV of the optimised PLSR model, as
applied to the detection of glucose, lactate and pyruvate from ATR-FTIR ratio
spectra, with the lowest concentration of each compound that a developed sensor
would be required to detect.
Compound RMSECV / mM Minimum Concentration / mM
Glucose 0.2 0.1
Lactate 0.08 0.1
Pyruvate 0.068 0.01
It is shown in Table 8.7 that the RMSECV for the prediction of lactate concentra-
tions was below the lowest limit that a developed ATR-FTIR sensor woud be required
to detect. This error was therefore assumed to be acceptable. The RMSECV for the
prediction of glucose concentrations is double the lowest concentration that a devel-
oped sensor would be required to detect. This is considered a promising result which
may be improved upon with further model refinement, such as that achieved with a
larger dataset. However, the RMSECV in the prediction of pyruvate concentrations
is six times higher than the required limit of detection, which is an unacceptable level
of accuracy. It is not clear whether model optimisation alone would be sufficient to
improve this value or whether changes to the experimental setup would be necessary.
150 CHAPTER 8. MIXTURE ANALYSIS: ATR-FTIR SPECTRA
This is discussed further in Chapter 10.
Chapter 9
Analysis of Glucose, Lactate and
Pyruvate Mixtures via
Transmission-FTIR Spectroscopy
This chapter describes the selection and optimisation of a statistical model for the
prediction of glucose, lactate and pyruvate concentrations from transmission-FTIR
spectra of solutions containing one, two or all three compounds. This study was
designed to be compared directly with the corresponding study of ATR-FTIR spectra
described in Chapter 8. The aims of this study were thus two-fold: one aim was
to determine the accuracy with which clinically-relevant concentrations of glucose,
lactate and pyruvate could be predicted from transmission-FTIR spectra, in order to
assess the feasibility of using transmission-FTIR spectroscopy for on-line metabolic
profiling of TBI patients. The second aim of this study was to compare its results
with those generated by the ATR-FTIR spectroscopy study to allow a comparison of
the two techniques and determine which one holds the most promise for development
into an on-line sensor.
As discussed in Chapter 8, clinical microdialysate samples were not used for the
identification or optimisation of the statistical model used in this work. Instead,
identical synthetic samples to those used in the ATR-FTIR spectroscopy study were
freshly prepared and analysed via transmission-FTIR spectroscopy using a 50 µm
path length. Subsequent analysis of the acquired spectra followed the same methods
as those described in Chapter 8. The performances of two linear regression models
(PCR and PLSR) were compared and PLSR was found to generate the most promising
results. This model was subsequently optimised and used to estimate the LPR of the
spectral data. Finally, the optimised PLSR model was applied predictively to the
transmission-FTIR spectra of six clinical samples. This latter experimental step was
performed only during this transmission-FTIR study, and not during the ATR-FTIR
151
152 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
study, due to the fact that the volumes of clinical samples were significantly lower
than the sample volumes required for the acquisition of ATR-FTIR spectra.
9.1 Mixture Selection
The two- and three component mixtures selected in Chapter 8 were freshly prepared
again for use in this study. The pure solutions prepared specifically for use in the
ATR-FTIR study were also freshly prepared and used in this work. However, the
total number of single-component spectra analysed in this study was slightly different
to the number analysed in Chapter 8, as a result of the later addition to the ATR-
FTIR study of data acquired during the ATR-FTIR crystal study. The equivalent
transmission-FTIR accessory work was conducted to identify the most suitable path
length for use in this transmission-FTIR study, as described in Section 6.4.2. It
was carried out only on solutions of glucose, and used fewer concentrations than the
ATR-FTIR crystal study, meaning that the addition of the data to that acquired in
this study did not increase the size of the dataset to the same degree. The single-
component solutions analysed in this study, including those analysed as part of the
path length study, are therefore summarised in Table 9.1.
Table 9.1: Concentrations of glucose, lactate and pyruvate stud-
ied as pure solutions via transmission-FTIR spectroscopy, includ-
ing those from the transmission-FTIR path length study.
Glucose / mM Lactate / mM Pyruvate / mM
0.05 0.05 0.005
0.10 0.10 0.01
0.20 0.20 0.02
0.40 0.50 0.05
0.70 2.00 0.10
1.00 4.00 0.15
1.50 6.00 0.20
3.00 10.00 1.00
5.00
6.00
10.0
24.0
9.2 Data Acquisition
Data acquisition followed a very similar protocol to that used for the ATR-FTIR mix-
ture study. The main difference between the two protocols was the fact that spectra
acquired for this study were obtained using the Specac Pearl Liquid Transmission
Accessory for the Perkin Elmer Spectrum 100 FTIR Spectrometer, coupled with an
9.2. DATA ACQUISITION 153
Oyster cell of 50 µm path length. The 50 µm path length had previously been iden-
tified, from a selection of path lengths, as exhibiting the best performance in the
detection of pure solutions of glucose, lactate and pyruvate (Section 6.4.2).
Three aliquots of each solution were analysed and two transmission-FTIR spectra
were acquired of each aliquot: one ratio spectrum, in which perfusion fluid was used
as the background, and one single-beam spectrum. Spectra of perfusion fluid only,
in which the concentrations of glucose, lactate and pyruvate were zero, were also
acquired. A total of 8 aliquots of perfusion fluid were analysed and two spectra were
acquired of each aliquot: one ratio spectrum, in which perfusion fluid was also used
as the background, and one single-beam spectrum.
This process was used to acquire two large datasets of 575 spectra: one containing
transmission-FTIR ratio spectra and one containing transmission-FTIR single-beam
spectra. As with the ATR-FTIR study, each dataset was used in its entirety for the
analysis described in this chapter, without differentiating between the spectra of one-,
two- and three-component solutions.
A sample of the spectra from each dataset can be seen in Figures 9.1 and 9.2. The
spectra plotted in this figure were selected by choosing every 35th spectrum from a
list of all spectra ordered by increasing glucose concentration.
5001000150020002500300035004000
Wavenumber / cm -1
92
94
96
98
100
102
104
106
108
110
%
 T
ra
ns
m
iss
io
n
   0           0           0
   0        0.05        0.01
   0         0.2           0
   0           4        0.15
0.05        0.05        0.01
 0.1           0         0.2
 0.1           2        0.05
 0.1           6        0.01
 0.7           0        0.05
 0.7           2        0.01
 0.7           4         0.2
   1           0        0.05
   1         0.5           0
   1           4        0.15
   3         0.1        0.01
   3           4        0.01
   5           0           0
Conc. / mM (Glu  Lac  Pyr)
Figure 9.1: A randomly-selected sample of the transmission-FTIR ratio spectra acquired
for the transmission-FTIR mixture study. Concentrations are listed in the horizontal order:
glucose - lactate - pyruvate.
154 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
5001000150020002500300035004000
Wavenumber / cm -1
0
10
20
30
40
50
60
70
80
90
Ar
bi
tra
ry
 U
ni
ts
   0           0           0
   0        0.05        0.01
   0         0.2           0
   0           4        0.15
0.05        0.05        0.01
 0.1           0         0.2
 0.1           2        0.05
 0.1           6        0.01
 0.7           0        0.05
 0.7           2        0.01
 0.7           4         0.2
   1           0        0.05
   1         0.5           0
   1           4        0.15
   3         0.1        0.01
   3           4        0.01
   5           0           0
Conc. / mM (Glu  Lac  Pyr)
Figure 9.2: A randomly-selected sample of the transmission-FTIR single-beam spectra ac-
quired for the transmission-FTIR mixture study. Concentrations are listed in the horizontal
order: glucose - lactate - pyruvate.
9.2.1 Data Acquisition: Conclusions
This section has described the process used to generate two large datasets of transmission-
FTIR spectra, one containing ratio spectra and one containing single-beam spectra,
both of which were acquired using the same set of samples.
9.3 Transmission-FTIR: Comparison of Statistical Mod-
els
The transmission-FTIR spectral datasets were initially analysed using the two linear
regression models employed during the first stage of the ATR-FTIR study: PCR and
PLSR. As with the ATR-FTIR study, it was considered logical to apply the model to
the maximum possible spectral range, whilst avoiding regions of significant spectral
noise. Transmission-FTIR spectra of aqueous solutions contain higher levels of noise
than ATR-FTIR spectra, as discussed in Chapter 6, due to the strong absorption of
IR radiation by water molecules. It can be seen in Figure 9.1 that the largest regions
of noise in transmission-FTIR spectra lie between 3700 – 3000 cm−1 and 800 – 450
cm−1. The models were therefore applied to the spectral range 3000 – 800 cm−1
thoughout this work, in order to avoid these regions.
Both models were applied in turn to each of the datasets and a maximum limit
9.3. TRANSMISSION-FTIR: COMPARISON OF STATISTICAL MODELS 155
of 30 components was imposed. The performance of the models was compared using
the same criteria as those used in Chapter 8: the number of components needed to
achieve the maximum R2 value, the number of components required to minimise the
RMSECV, and the RMSECV in prediction of glucose, lactate and pyruvate concen-
trations achieved when a PLSR model of 20 components was used.
9.3.1 PCR vs. PLSR: Ratio Transmission-FTIR Spectra
A summary of the performance of a PCR model applied to the transmission-FTIR
ratio spectra dataset is shown in Figure 9.3.
0 5 10 15 20 25 30
Number of PCR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PCR Components
0
0.5
1
1.5
2
2.5
R
M
SE
CV
 o
f P
CR
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 9.3: A summary of the performance of a PCR model applied to transmission-FTIR
ratio spectra of glucose, lactate and pyruvate. The change in R2 with an increasing number
of model components (a) and the change in the RMSECV in predicted concentrations with an
increasing number of model components (b), calculated using leave-one-out cross-validation,
are shown for each compound.
It can be seen in Figure 9.3 that, using PCR, both glucose and lactate were
accurately modelled with fewer than ten components. The RMSECV of glucose and
lactate were also minimised using a model of fewer than ten components. However,
the most striking feature of the data was the extremely poor performance of PCR
with respect to the modelling and prediction of pyruvate concentrations. The R2
value of the PCR model of pyruvate was close to zero, even when a 30-component
model was used, and the RMSECV in prediction of pyruvate concentrations appears
to have been unaffected by the number of PCR components used. It was initially
thought that these effects may be due to an error in the data, but no such error could
be found. Finally, it was concluded that the noise in the spectral region over which
the model was applied, particularly at 1600 cm−1 where multiple pyruvate peaks are
based, was sufficient to mask the contribution of pyruvate to the overall spectrum.
156 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
A summary of the performance of a PLSR model applied to the dataset of transmission-
FTIR ratio spectra is shown in Figure 9.4
0 5 10 15 20 25 30
Number of PLSR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PLSR Components
0
0.5
1
1.5
2
2.5
R
M
SE
CV
 o
f P
LS
R
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 9.4: A summary of the performance of a PLSR model applied to transmission-FTIR
ratio spectra of glucose, lactate and pyruvate. The change in R2 with an increasing number
of model components (a) and the change in the RMSECV in predicted concentrations with an
increasing number of model components (b), calculated using leave-one-out cross-validation,
are shown for each compound.
Once again, the poor performance of the model when applied to the modelling
and prediction of pyruvate concentrations was the most striking feature of Figure 9.4.
The trends exhibited by the PLSR model for each compound were broadly similar to
those observed for the PCR model, athough the PLSR model generated better results
according to the predefined criteria for model comparison. A quantitative comparison
of the PCR and PLSR models, when applied to transmission-FTIR ratio spectra, is
provided in Table 9.2.
Table 9.2: A summary of the performance of PCR and PLSR statistical models when applied
to transmission-FTIR ratio spectra.
Compound Parameter PCR PLSR
Glucose
Number of components for max. R2 Value 8 7
Number of components for min. RMSECV 17 14
RMSECV with 20 components 0.20 0.14
Lactate
Number of components for max. R2 Value 8 6
Number of components for min. RMSECV 14 14
RMSECV with 20 components 0.30 0.20
Pyruvate
Number of components for max. R2 Value >30 >30
Number of components for min. RMSECV 1 1
RMSECV with 20 Components 0.098 0.096
The data presented in Table 9.2 confirmed the general trends observed from the
9.3. TRANSMISSION-FTIR: COMPARISON OF STATISTICAL MODELS 157
graphical data. PLSR was the better performing model for glucose and lactate by all
measured criteria. Interestingly, there was very little difference in the performance of
the two models when applied to pyruvate, suggesting again that the contribution of
pyruvate to the spectral dataset was minimal.
9.3.2 PCR vs. PLSR: Single-Beam Transmission-FTIR Spectra
A summary of the performances of the PCR and PLSR models applied to the dataset of
transmission-FTIR single-beam spectra are shown in Figures 9.5 and 9.6, respectively.
Once again, the most noticeable feature of the data was the equally poor performance
of both models when applied to pyruvate.
0 5 10 15 20 25 30
Number of PCR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PCR Components
0
0.5
1
1.5
2
2.5
R
M
SE
CV
 o
f P
CR
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 9.5: A summary of the performance of a PCR model applied to transmission-FTIR
single-beam spectra of glucose, lactate and pyruvate. The change in R2 with an increasing
number of model components (a) and the change in the RMSECV in predicted concentrations
with an increasing number of model components (b), calculated using leave-one-out cross-
validation, are shown for each compound.
The PLSR model was able to accurately model the glucose and lactate data with
fewer components than the number required by the PCR model. The difference be-
tween the predictive performance of the two models when applied to glucose and
lactate was less stark, but PLSR was again the better-performing model by the cri-
teria considered. One interesting feature of the PCR model was that the relationship
between the number of PCR components and the RMSECV in glucose and lactate
concentrations appeared to show less correlation than had previously been observed.
In the data gathered thus far, an increase in the number of PCR components would
lead to a decrease in the RMSECV of the model until the predictive limit of the model
was reached. However, in Figure 9.5 it can be seen that, although the general trend
was in agreement with that usually observed, there were instances where an increase
158 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
0 5 10 15 20 25 30
Number of PLSR Components
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
R
2
(a)
Glucose
Lactate
Pyruvate
Compound
0 5 10 15 20 25 30
Number of PLSR Components
0
0.5
1
1.5
2
2.5
R
M
SE
CV
 o
f P
LS
R
(b)
Glucose
Lactate
Pyruvate
Compound
Figure 9.6: A summary of the performance of a PLSR model applied to transmission-FTIR
single-beam spectra of glucose, lactate and pyruvate. The change in R2 with an increasing
number of model components (a) and the change in the RMSECV in predicted concentrations
with an increasing number of model components (b), calculated using leave-one-out cross-
validation, are shown for each compound.
in the number of model components lead to an increase in the RMSECV, most notably
in the transition from 15 to 16 components in the prediction of lactate concentrations.
The performances of PCR and PLSR when applied to transmission-FTIR single-
beam spectra are summarised in Table 9.3. In general, the PLSR model was confirmed
to produce the better results. One notable exception is in the number of components
needed to minimise the RMSECV in the prediction of lactate, which was lower for
the PCR model than the PLSR model due to the aforementioned variation in the
relationship between the RMSECV and the number of model components.
Table 9.3: A summary of the performance of PCR and PLSR statistical models when applied
to transmission-FTIR single-beam spectra.
Compound Parameter PCR PLSR
Glucose
Number of components for max. R2 Value 23 14
Number of components for min. RMSECV 29 26
RMSECV with 20 Components 0.87 0.59
Lactate
Number of components for max. R2 Value 23 21
Number of components for min. RMSECV 20 30
RMSECV with 20 Components 1.23 1.05
Pyruvate
Number of components for max. R2 Value >30 >30
Number of components for min. RMSECV 1 1
RMSECV with 20 Components 0.095 0.097
9.4. TRANSMISSION-FTIR: EFFECT OF SPECTRAL RANGE 159
9.3.3 Comparison of Statistical Models: Conclusions
This section describes a comparison of the performances of PCR and PLSR statistical
models when applied to the same sets of transmission-FTIR ratio and single-beam
spectra. The PLSR model generally produced better results then the PCR model
for both the modelling and prediction of glucose and lactate concentrations, for both
datasets, as observed in the equivalent study of ATR-FTIR spectral data. However, in
contrast to the study of ATR-FTIR spectra, both models performed extremely poorly
when applied to the modelling and prediction of pyruvate concentrations. The reason
for this was unknown, but it was hypothesised that the increased spectral noise in
transmission-FTIR spectra of aqueous solutions, when compared with the spectral
noise seen in equivalent ATR-FTIR spectra, may have masked the contribution of
pyruvate to the spectra.
Both models perfomed less well at both the modelling and prediction of glucose
and lactate concentrations when applied to the single-beam dataset than when applied
to the ratio dataset. The performances of the two models when applied to pyruvate
were similar for both the single-beam and ratio datasets.
These results suggested that predictive statistical modelling could successfully be
used to analyse transmission-FTIR spectra of glucose and lactate. Based on the
results presented in this section, the PLSR model was selected for further refinement
to improve the predictive accuracy of the model when applied to glucose and lactate
and to determine whether its performance with respect to pyruvate could be improved.
9.4 Transmission-FTIR: Effect of Spectral Range
An interval analysis, identical to that carried out during the study of ATR-FTIR
spectra, was performed on both transmission-FTIR datasets to determine the region
of the spectrum to which the PLSR model should be applied in order to minimise
the RMSECV in the prediction of glucose, lactate and pyruvate concentrations. The
analysis was carried out as described in Section 8.6 and was applied to the spectral
range 3000 - 800 cm−1.
9.4.1 Spectral Range: Transmission-FTIR Ratio Spectra
The results of the interval study carried out on transmission-FTIR ratio spectra are
summarised in Figure 9.7.
It is clear from the figure that interval start points between 900 – 1100 cm−1 yielded
the lowest overall RMSECV in the prediction of the concentrations of all compounds.
The lowest RMSECV of 0.202 mM was achieved using an interval range of 1000 – 1500
cm−1 and a 15-component model. Using these parameters, the achievable RMSECV,
in mM, for each individual compound were:
160 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
All Compounds
900 1100 1300 1500 1700 1900
Interval Start / cm -1
100
200
300
400
500
600
700
800
900
1000
In
te
rv
al
 S
ize
 / 
cm
-
1
0.5
1
1.5
2
2.5
3
RMSECV / mM
Figure 9.7: A colourmap showing the effect of spectral interval location and size on the
overall RMSECV of a PLSR model applied to transmission-FTIR ratio spectra.
ˆ Glucose: 0.12
ˆ Lactate: 0.15
ˆ Pyruvate: 0.070
These values of the RMSECV achieved for each compound were lower than those
achieved in the previous section when a 20-component model was applied across the
whole spectral range, indicating that the model had been improved. Notably, the RM-
SECV in the prediction of pyruvate concentrations, which was found in the previous
section to be very poor, had reduced by approximately 30%.
9.4.2 Spectral Range: Transmission-FTIR Single-Beam Spectra
The results of the interval study carried out on transmission-FTIR single-beam spec-
tra are summarised in Figure 9.8. The data indicated that the lowest values of the
RMSECV were achieved using intervals of 300 –500 cm−1 in size, with a start point
between 900 – 1100 cm−1.
The interval which produced the lowest overall RMSECV had a size of 400 cm−1
and was located at the 1000 – 1400 cm−1 region of the spectrum. A 19-component
model was required in order to achieve an overall RMSECV for this interval of 0.224.
9.4. TRANSMISSION-FTIR: EFFECT OF SPECTRAL RANGE 161
All Compounds
900 1100 1300 1500 1700 1900
Interval Start / cm -1
100
200
300
400
500
600
700
800
900
1000
In
te
rv
al
 S
ize
 / 
cm
-
1
0.5
1
1.5
2
2.5
3
RMSECV / mM
Figure 9.8: A colourmap showing the effect of spectral interval location and size on the
overall RMSECV of a PLSR model applied to transmission-FTIR single-beam spectra.
The RMSECV in the prediction of the concentrations, in mM, of each individual
compound over this interval were:
ˆ Glucose: 0.12
ˆ Lactate: 0.17
ˆ Pyruvate: 0.071
Once again, the RMSECV in the prediction of pyruvate concentrations was re-
duced by approximately 30% compared to the value generated when a 20-component
PLSR model was applied across the whole spectral range. The RMSECV for glu-
cose and lactate was also reduced, indicating that the model had successfully been
optimised.
9.4.3 Effect of Spectral Range: Conclusions
This section describes the work undertaken to optimise the predictive performance of
PLSR, as measured by the RMSECV, when applied to both ratio and single-beam
transmission-FTIR spectra. The RMSECV was reduced for the prediction of the con-
centrations of all compounds of interest, compared with the respective values achieved
162 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
by a PLSR model applied over the whole spectral range, indicating that this aim had
been achieved. Significantly, the RMSECV for pyruvate was reduced by approximately
30% for both ratio and single-beam spectra.
The application of the PLSR model to transmission-FTIR ratio spectra was found
to yield slightly lower RMSECV values for lactate and pyruvate than when the model
was applied to transmission-FTIR single-beam spectra. The achieved RMSECV in
the prediction of glucose concentrations was the same for both datasets. This was
a contrast to the results generated by the interval study carried out on ATR-FTIR
spectra, in which the application of PLSR to ATR-FTIR ratio spectra generated
significantly lower values of the RMSECV, for all compounds, than when the model
was applied to ATR-FTIR single-beam spectra.
9.5 Transmission-FTIR: Prediction of LPR
An attempt was made to predict the LPR from both transmission-FTIR datasets,
using the same method as that applied in the equivalent ATR-FTIR study. A test
dataset of 57 spectra, corresponding to approximately 10% of the total dataset, was
chosen in order to maximise the size of the training set. PLSR was applied to the
training dataset, over the spectral region identified by the interval studies as yielding
the lowest overall RMSECV, and calibrated using leave-one-out cross-validation. The
calibrated model was then applied to the spectra in the test dataset to generate
predictions of glucose, lactate and pyruvate concentrations. The predicted LPR was
determined by dividing the predicted lactate concentration by that of pyruvate for
each sample. The actual LPR was calculated by dividing the observed concentrations
of lactate and pyruvate for each sample in the dataset.
9.5.1 Prediction of LPR from Ratio Spectra
Predictions of the LPR from transmission-FTIR ratio spectra were generated by ap-
plying a 15-component PLSR model to the 1500 – 1000 cm−1 region of the datset. A
comparison of the observed and predicted values of the LPR is shown in Figure 9.9.
As seen in Figure 9.9(a), there was very little correlation between the observed and
predicted values of the LPR, as confirmed by a calculated R2 value of 0.00. However,
when the plot shown in Figure 9.9(a) was scaled to show only the region where the
majority of the datapoints were located (Figure 9.9(b)) there did appear to be a
linear relationship in the data, suggesting that the overall R2 value was affected by
outlying data points with very large errors. In addition, one value of the LPR was
predicted to be negative, indicating that the PLSR model had calculated a negative
value for the concentration of either lactate or pyruvate. The magnitude of the error
in LPR prediction is explained by the fact that the individual errors in the prediction
9.5. TRANSMISSION-FTIR: PREDICTION OF LPR 163
0 200 400 600 800
Observed LPR
-500
0
500
1000
1500
2000
2500
3000
3500
4000
4500
Pr
ed
ict
ed
 L
PR
(a)
0 20 40 60 80 100 120
Observed LPR
0
20
40
60
80
100
120
Pr
ed
ict
ed
 L
PR
(b)
Figure 9.9: Scatter plots showing the observed vs. predicted LPR. Predicted values of the
LPR were derived from the division of the concentrations of lactate and pyruvate predicted
from transmission-FTIR ratio spectra using a PLSR model. Plot (a) shows all results, whilst
(b) shows only the region containing the most data points.
of lactate and pyruvate were compounded by the division of the two values to generate
the predicted LPR.
9.5.2 Prediction of LPR from Single-Beam Spectra
The PLSR model used to predict the LPR from transmission-FTIR single-beam spec-
tra used 19 components and was applied over the spectral range identified as yielding
the lowest overall RMSECV (1400 - 1000 cm−1). A comparison of the observed and
predicted values of the LPR is shown in Figure 9.10(a). Once again, there was a
significant error in the prediction of the LPR, as confirmed by an R2 value for a lin-
ear relationship of the data of 0.01, but a more qualitatively linear relationship was
observed when the plot was scaled to show only the region where the majority of the
datapoints were located (Figure 9.10(b)). A small number of predicted LPR values
were negative and therefore it was concluded that the PLSR model had, once again,
predicted negative concentrations of either lactate or pyruvate.
9.5.3 Prediction of LPR: Conclusions
This section describes an attempt to predict the LPR from both ratio and single-beam
transmission-FTIR spectra, via a straightforward division of the concentrations of lac-
tate and pyruvate predicted using a PLSR model. It has been shown that the RMSEP
of the model, particularly with respect to the prediction of pyruvate concentrations,
was too high to allow accurate prediction of the LPR using this method. A particular
164 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
0 200 400 600 800 1000 1200
Observed LPR
-100
-50
0
50
100
150
200
250
Pr
ed
ict
ed
 L
PR
(a)
0 20 40 60 80
Observed LPR
0
10
20
30
40
50
60
70
80
90
Pr
ed
ict
ed
 L
PR
(b)
Figure 9.10: Scatter plots showing the observed vs. predicted LPR. Predicted values of the
LPR were derived from the division of the concentrations of lactate and pyruvate predicted
from transmission-FTIR single-beam spectra using a PLSR model. Plot (a) shows all results,
whilst (b) shows only the region containing the most data points.
concern was the prediction of invalid negative values of the concentrations of lactate
and pyruvate. These results correlate with those obtained during the equivalent ATR-
FTIR study, although it was noted that fewer negative concentration values of lactate
and pyruvate were predicted during this transmission-FTIR study.
9.6 Transmission-FTIR: Analysis of Clinical Samples
The optimised PLSR model was predictively applied to the transmission-FTIR spectra
of the six clinical microdialysate samples analysed in Section 6.3, in order to gain an
insight into how applicable the results generated in this chapter might be to clinical
samples. As explained in Section 6.3, this study could not be performed with ATR-
FTIR spectroscopy due to the small volumes of the clinical samples.
Both ratio and single-beam transmission-FTIR spectra of the microdialysate sam-
ples were acquired and used as two separate test datasets. For each type of spectrum,
the equivalent transmission-FTIR dataset of 575 spectra generated using synthetic
samples was used as a training dataset. The training dataset was converted to ab-
sorbance data and mean-centred before being used to train a PLSR model using
leave-one-out cross-validation. The test datasets were also converted to absorbance
values and mean-centred before the model was applied.
Initially, a 20-component PLSR model was trained and applied to the test dataset
over the full spectral range of 3000 – 800 cm−1. The model was then trained and
applied a second time, using the spectral interval and number of components identi-
9.6. TRANSMISSION-FTIR: ANALYSIS OF CLINICAL SAMPLES 165
fied in Section 9.4 as yielding the lowest overall RMSECV. This process provided an
insight into whether the model optimisation achieved in this study would be directly
applicable to clinical samples.
9.6.1 Analysis of Clinical Samples: Ratio Spectra
This section describes the prediction of glucose, lactate and pyruvate concentrations
from transmission-FTIR ratio spectra of clinical samples.
The predictions generated by a 20-component PLSR model, applied over the full
spectral range (3000 – 800 cm−1), are compared with the observed values of the
analyte concentrations in Figure 9.11.
0 0.5 1 1.5 2
Observed Concentration / mM
0
0.5
1
1.5
2
2.5
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Glucose
1 2 3 4 5
Observed Concentration / mM
0
1
2
3
4
5
6
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
 Lactate
0 0.1 0.2 0.3
Observed Concentration / mM
0
0.1
0.2
0.3
0.4
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Pyruvate
Figure 9.11: A comparison of the concentrations of glucose, lactate and pyruvate predicted
from transmission-FTIR ratio spectra of clinical samples, calculated using a 20-component
PLSR model applied over the spectral range 3000 – 800 cm−1, with the observed concentrations
measured using the ISCUS Flex Microdialysis Analyser. The solid blue line represents a perfect
correlation of the observed and predicted concentrations, whilst the dashed black lines indicate
10% error intervals.
The results in Figure 9.11 show that the predicted glucose concentrations did not
fall within 10% of the observed concentration for any sample. This was also the case for
the prediction of pyruvate concentrations. The concentration of lactate was predicted
to within 10% of the observed concentration for two samples, yet the prediction of
lactate concentrations within the remaining samples was significantly outside the 10%
boundary.
Application of a 15-component PLSR model over the spectral range 1500 - 1000
cm−1 was found, during the interval study described in Section 9.4, to produce the
lowest RSMECV of the large transmission-FTIR ratio spectra dataset. This model
was therefore applied to the transmission-FTIR ratio spectra of the clinical micro-
dialysate samples to determine whether it would improve the predictive accuracy.
The concentrations of glucose, lactate and pyruvate predicted by this model, and
their comparison with the observed concentrations, are shown in Figure 9.12.
The data shown in Figure 9.12 indicated that the accuracy in prediction of glu-
166 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
0 0.5 1 1.5 2
Observed Concentration / mM
0
0.5
1
1.5
2
2.5
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Glucose
1 2 3 4 5
Observed Concentration / mM
0
1
2
3
4
5
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
 Lactate
0 0.1 0.2 0.3
Observed Concentration / mM
-0.2
-0.1
0
0.1
0.2
0.3
0.4
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Pyruvate
Figure 9.12: A comparison of the concentrations of glucose, lactate and pyruvate predicted
from transmission-FTIR ratio spectra of clinical samples, calculated using a 15-component
PLSR model applied over the spectral range 1500 – 1000 cm−1, with the observed concentra-
tions measured using the ISCUS Flex Microdialysis Analyser. The solid blue line represents
an exact fit of the observed and predicted concentrations, whilst the dashed black lines indicate
10% error intervals.
cose concentrations changed very little with the optimised model, as indicated by a
qualitative examination of the correlation between observed and predicted values. The
accuracy in prediction of lactate concentrations appeared to have improved, whilst the
prediction of pyruvate concentrations appeared to have become less accurate. These
results contradicted the findings of the interval study and suggested that a developed
sensor would need to carry out a similar study to find the optimum range over which
to apply a PLSR model to minimise the overall error in prediction of clinical samples.
9.6.2 Analysis of Clinical Samples: Single-Beam Spectra
This section describes the prediction of glucose, lactate and pyruvate concentrations
from transmission-FTIR single-beam spectra of clinical samples.
The predictions generated by a 20-component PLSR model applied over the full
spectral range (3000 – 800 cm−1) are shown in Figure 9.13. It is immediately clear
from this figure that the prediction of glucose, lactate and pyruvate concentrations
from transmission-FTIR single-beam spectra was less accurate than a prediction from
transmission-FTIR ratio spectra. This conclusion was in agreement with the results
generated using synthetic samples.
The interval study described in Chapter 9.4 found that the overall RMSECV
in the prediction of glucose, lactate and pyruvate concentrations from transmission-
FTIR single-beam spectra was minimised using a 19-component PLSR model over
the spectral range 1400 – 1000 cm−1. This model was predictively applied to the
transmission-FTIR single-beam spectra of the clinical samples, to determine whether
it would improve the accuracy in prediction of glucose, lactate and pyruvate concen-
trations. The results of this study, shown in Figure 9.14, indicated that the accuracy
9.6. TRANSMISSION-FTIR: ANALYSIS OF CLINICAL SAMPLES 167
0 0.5 1 1.5 2
Observed Concentration / mM
-1
0
1
2
3
4
5
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Glucose
1 2 3 4 5
Observed Concentration / mM
-5
0
5
10
15
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
 Lactate
0 0.1 0.2 0.3
Observed Concentration / mM
0
0.1
0.2
0.3
0.4
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Pyruvate
Figure 9.13: A comparison of the concentrations of glucose, lactate and pyruvate pre-
dicted from transmission-FTIR single-beam spectra of clinical samples, calculated using a 20-
component PLSR model applied over the spectral range 3000 – 800 cm−1, with the observed
concentrations measured using the ISCUS Flex Microdialysis Analyser. The solid blue line
represents an exact fit of the observed and predicted concentrations, whilst the dashed black
lines indicate 10% error intervals.
in prediction of glucose and lactate concentrations had increased from that of the
PLSR model applied over the full spectral range, whilst the accuracy in prediction of
pyruvate concentrations had decreased.
0 0.5 1 1.5 2
Observed Concentration / mM
-1
0
1
2
3
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Glucose
1 2 3 4 5
Observed Concentration / mM
0
2
4
6
8
10
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
 Lactate
0 0.1 0.2 0.3
Observed Concentration / mM
-0.4
-0.2
0
0.2
0.4
Pr
ed
ict
ed
 C
on
ce
nt
ra
tio
n 
/ m
M
Pyruvate
Figure 9.14: A comparison of the concentrations of glucose, lactate and pyruvate pre-
dicted from transmission-FTIR single-beam spectra of clinical samples, calculated using a 19-
component PLSR model applied over the spectral range 1400 – 1000 cm−1, with the observed
concentrations measured using the ISCUS Flex Microdialysis Analyser. The solid blue line
represents an exact fit of the observed and predicted concentrations, whilst the dashed black
lines indicate 10% error intervals.
9.6.3 Analysis of Clinical Samples: Conclusions
This section has described the application of a PLSR model, trained on a large dataset
of transmission-FTIR spectra of synthetic solutions of glucose, lactate and pyruvate,
to the prediction of glucose, lactate and pyruvate concentrations in clinical micro-
dialysate samples. The errors in prediction for each spectrum type and model used
168 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
are summarised in Table 9.4.
Table 9.4: A summary of the performance of PLSR models when predictively applied to both
ratio and single-beam transmission-FTIR spectra of clinical microdialysate samples.
Spectrum
Type
Model Range
No. of
Components
Compound RMSEP / mM
Glucose Lactate Pyruvate
Ratio
3000 – 800 20 0.40 0.52 0.034
1500 - 1000 15 0.44 0.30 0.078
Single-
Beam
3000 – 800 20 1.02 2.35 0.04
1400 – 1000 19 0.62 1.47 0.11
The errors in prediction, for all compounds and both types of spectrum, are larger
than would be acceptable for a developed sensor. This may be due to overfitting of
the model to the training data, as all tested models used a relatively high number
of components. Alternatively, the errors in prediction may be due to the fact that
the model used to predict the glucose, lactate and pyruvate concentrations had been
calibrated using spectra of synthetic solutions. The clinical solutions were likely to
have contained a number of molecules in addition to glucose, lactate and pyruvate, as
discussed in Section 6.2, and these were not reflected in the training data. It is likely
that both of these factors contributed to the generation of predictive errors.
The trend observed in the analysis of synthetic samples, in which a predictive
PLSR model generally performed better in the analysis of ratio spectra than single-
beam spectra, was also found to be the case in this study. However, the models
that had been optimised for use with synthetic samples, through the identification of
the spectral range and number of model components that yielded the lowest overall
RMSECV, did not always improve the predictive performance of the model when
applied to clinical samples. It can be seen from Table 9.4 that the prediction of
glucose and pyruvate concentrations from transmission-FTIR ratio spectra became
less accurate when the optimised model was applied.
Overall, these results suggest that transmission-FTIR spectroscopy could be com-
bined with a predictive PLSR model to analyse clinical microdialysis samples, al-
though it is apparent that the model would need to be optimised using data from
clinical samples.
9.7 Conclusions
This chapter has described a systematic study of the potential of transmission-FTIR
to be used as the basis of a sensor for the analysis of glucose, lactate and pyruvate
concentrations. The experiments carried out during this study were the same as those
carried out for the systematic ATR-FTIR study described in Chapter 8, in order to
allow a comparison of the generated data.
9.7. CONCLUSIONS 169
Many of the general trends observed in the ATR-FTIR study were also seen in
the results generated by this study. PLSR generally perfomed better than PCR in
the modelling and prediction of both ratio and single-beam spectra. Further studies
carried out using PLSR showed that the prediction of glucose, lactate and pyruvate
concentrations from ratio spectra was generally more accurate than prediction from
single-beam transmission-FTIR spectra.
The RMSECV for the prediction of glucose, lactate and pyruvate concentrations
from transmission-FTIR ratio spectra, acquired following a systematic study to iden-
tify the spectral range and number of model components that minimised the RM-
SECV, are summarised in Table 9.5. The table compares these errors with the min-
imum concentration of each compound that a developed sensor for the metabolic
profiliing of TBI patients would be required to detect.
Table 9.5: A comparison of the RMSECV of the optimised PLSR model, as
applied to the detection of glucose, lactate and pyruvate from transmission-FTIR
ratio spectra, with the lowest concentration of each compound that a developed
sensor would be required to detect.
Compound RMSECV / mM Minimum Concentration / mM
Glucose 0.12 0.1
Lactate 0.15 0.1
Pyruvate 0.07 0.01
The data in Table 9.5 indicated that the RMSECV for all compounds would need
to be lowered further before the model could successfully be used to detect clinically-
relevant concentrations of glucose, lactate and pyruvate with the desired accuracy.
The RMSECV for pyruvate requires the most significant improvement, as was found
in the ATR-FTIR study. In order to allow a direct comparison of both studies, the
table summarising the ATR-FTIR results is reproduced in Table 9.6.
Table 9.6: A comparison of the RMSECV of the optimised PLSR model, as
applied to the detection of glucose, lactate and pyruvate from ATR-FTIR ratio
spectra, with the lowest concentration of each compound that a developed sensor
would be required to detect.
Compound RMSECV / mM Minimum Concentration / mM
Glucose 0.2 0.1
Lactate 0.08 0.1
Pyruvate 0.068 0.01
A comparison of the data in Tables 9.5 and 9.6 shows that there is not a clear dis-
tinction in the predictive performance of PLSR applied to ATR-FTIR and transmission-
FTIR ratio spectra. The optimised prediction of glucose concentrations yielded an
RMSECV that was approximately 50% lower for transmission-FTIR spectra than for
ATR-FTIR spectra. Conversely, the optimised prediction of lactate concentrations
170 CHAPTER 9. MIXTURE ANALYSIS: TRANSMISSION-FTIR SPECTRA
yielded an RMSECV that was approximately 50% lower for ATR-FTIR spectra than
for transmission-FTIR spectra. The error in pyruvate prediction was similar for both
ATR-FTIR and transmission-FTIR spectra. The optimised spectral intervals over
which the PLSR models were applied in order to achieve these RMSECV values were
similar for the transmission-FTIR and ATR-FTIR ratio spectra (1000 – 1500 cm−1 and
900 – 1500 cm−1, respectively). However, the number of model components required
was lower for transmission-FTIR spectra than for ATR-FTIR spectra (15 components
vs. 20 components, respectively), indicating that the model of the transmission-FTIR
spectral data was less complex.
Prediction of the LPR via a straightforward division of the lactate and pyruvate
concentrations predicted from transmission-FTIR spectra, using the optimised PLSR
models, generated very large errors for both ratio and single-beam spectra. Some
negative LPR values were predicted, for both ratio and single-beam spectra, show-
ing that the PLSR models had calculated invalid negative values of the lactate and
pyruvate concentrations. This was also the case when applying the same method to
ATR-FTIR spectra.
Finally, two PLSR models, trained on the transmission-FTIR ratio and single-
beam spectra of synthetic solutions generated during this study, were predictively
applied to the transmission-FTIR ratio and single-beam spectra of six clinical mi-
crodialysate samples. Once again, the predictive performance of the model applied
to ratio spectra was more accurate than the performance of that applied to single-
beam spectra. The modelling of both spectrum types produced a RMSEP for each
compound that was higher than those achieved for the prediction of concentrations
from synthetic samples. This was thought to be due to a combination of two fac-
tors: the fact that the model was trained on spectra of synthetic solutions, whilst
clinical solutions contain a number of competing molecules that were not present in
the synthetic samples, and overfitting of the model to the training data. However,
the results are encouraging and suggest that, with further work, it may be possible
to reach the required RMSEP for each compound in clinical microdialysate samples.
This is discussed further in Chaper 10.
Chapter 10
Discussion and Future Outlook
This Chapter provides a brief overview of the results presented in this thesis. The
conclusions arising from this work, with respect to the stated aim of this thesis, are
subsequently discussed and suggestions for further work are made.
10.1 Summary of Work
The stated aim of the thesis was to determine whether vibrational spectroscopy could
be used as the basis of an on-line sensor for metabolic profiling of TBI patients.
This aim was first addressed by comparing the ability of two forms of vibrational
spectroscopy, FTIR spectroscopy and Raman spectroscopy, to detect glucose, lactate
and pyruvate in solution. All results were assessed against the specifications required
of the developed on-line sensor, as listed in Section 2.6.
The initial investigation into the ability of FTIR spectroscopy to detect glucose,
lactate and pyruvate in solution was carried out using two techniques: ATR-FTIR
spectroscopy and transmission-FTIR spectroscopy. ATR-FTIR spectroscopy was cho-
sen due to its relative insensitivity to water, which was expected to be a significant
advantage of the technique, given the highly dilute nature of clinical microdialysis
samples. Transmission-FTIR, on the other hand, is highly sensitive to water but
was selected for investigation because it is arguably the most simple of the available
FTIR spectroscopic techniques and was thus expected to be the most amenable to
miniaturisation into a bedside device.
ATR-FTIR spectroscopy and transmission-FTIR spectroscopy were used to anal-
yse solutions of glucose, lactate and pyruvate in perfusion fluid. The compounds were
analysed both as pure solutions and as a three-component mixture and peaks unique
to each compound were successfully identified. Consideration was given to the nature
and likely concentration of other compounds which may be present in microdialysate
samples from TBI patients. Two of these compounds, glycerol and glutamate, were
studied via ATR-FTIR and transmission-FTIR spectroscopy both as single-component
171
172 CHAPTER 10. DISCUSSION AND FUTURE OUTLOOK
solutions and as five-component solutions of glucose, lactate, pyruvate, glycerol and
glutamate. This study indicated that peaks unique to the three analytes of interest
would still be visible against a background of glycerol and glutamate when all com-
pounds were present at clinically-relevant concentrations. Finally, transmission-FTIR
spectroscopy was used to acquire FTIR spectra of six clinical microdialysate samples.
Characterisation of these spectra revealed that peaks unique to glucose and lactate
could be identified but the intensity of pyruvate absorption peaks was too low to allow
their identification.
Following these studies, FTIR spectroscopy was considered a promising technique
for inclusion in an on-line sensor for metabolic profiling of TBI patients. It was con-
cluded that in-depth studies should be carried out into the ability to accurately identify
glucose, lactate and pyruvate concentrations from both ATR-FTIR and transmission-
FTIR spectra. Systematic studies were therefore carried out to determine which ATR-
FTIR crystal and transmission-FTIR path length should be used for these in-depth
investigations.
Raman spectroscopy was investigated as a possible technique upon which to base
an on-line sensor for metabolic profiling of TBI patients due to its insensitivity to wa-
ter, potential for miniaturisation into a bedside device and previous reported use in
medical applications. It was quickly established that Raman spectroscopy of glucose,
lactate and pyruvate was too insensitive to enable the detection of these compounds
at clinically-relevant concentrations. A SERS-based sensor using silver colloids sus-
pended in solution was proposed as an alternative and its feasibility was assessed
using cysteine as a model analyte. However, an assessment of the difficulties involved
in implementing this device, combined with the promising results generated by the
initial FTIR spectroscopy study, led to the decision not to proceed further with SERS.
The in-depth investigations of ATR-FTIR and transmission-FTIR spectroscopy
were based on the analysis of synthetic solutions containing glucose, lactate and pyru-
vate at clinically-relevant concentrations. Both investigations followed the same work-
flow in order to allow the results obtained using each technique to be compared.
The first stage in each investigation was to compare the relative performances of
PCR and PLSR in the modelling and prediction of glucose, lactate and pyruvate con-
centrations from both ratio and single-beam spectra. PLSR was the best-performing
model for all types of spectra and was selected for further optimisation. The model
was optimised to determine the spectral interval and number of components which
yielded the lowest overall RMSECV. These model parameters were used to generate
a prediction of the LPR via a division of the predicted concentrations of lactate and
pyruvate.
Finally, two PLSR models trained on transmission-FTIR ratio and single-beam
spectra were used to predict the concentrations of glucose, lactate and pyruvate from
10.2. CONCLUSIONS 173
the transmission-FTIR ratio and single-beam spectra of clinical samples. The errors
in prediction were higher than those obtained for models applied to the spectra of
synthetic samples, but were considered promising.
10.2 Conclusions
Following the work described in this thesis, it is recommended that an on-line sensor
for continuous metabolic profiling of TBI patients is not based on Raman spectroscopy
due to its inherent insensitivity, particularly with respect to the detection of molecules,
such as glucose, lactate and pyruvate, which exhibit a low Raman scattering intensity.
The use of SERS in such a sensor is also not recommended due to inherent character-
istics of the SERS technique which would complicate the operation of the device and
affect its reliability.
The results generated in this work suggest that it may be feasible to produce a
sensor for on-line metabolic profiling of TBI patients that is based upon a combination
of FTIR spectroscopy and statistical modelling, although further reduction of the
RMSECV would be required. These results strongly suggest that PLSR would be a
more appropriate statistical model than PCR for the analysis of spectra. In addition,
the results indicate that a sensor based on predictive PLSR modelling of FTIR spectra
should be used to analyse ratio spectra, rather than single-beam spectra, in order to
minimise the error in prediction.
Based upon the results generated in this work, it is concluded that there is little
overall difference in the predictive accuracy of a PLSR model applied to ratio spectra
acquired using ATR-FTIR or transmission-FTIR. It is therefore concluded that the
differences between ATR-FTIR and transmission-FTIR spectra, as observed in Section
6.1, do not appear to have had an impact on the predictive accuracy of PLSR. This is
thought to be due to the fact that the spectral intervals identified as yielding the lowest
RMSECV, for both ATR-FTIR and transmission-FTIR spectra, did not incorporate
regions in which the spectra were significantly different. Regions of significant water-
induced noise (3700 – 3000 cm−1, 1700 – 1600 cm−1 and 900 – 450 cm−1) were excluded
from the optimised intervals, as was the region in which peaks are absent from the
ATR-FTIR spectra of lactate and pyruvate due to orientation effects (1500 – 1600
cm−1). A final decision regarding which form of FTIR spectroscopy would be most
suitable for integration into an on-line sensor for metabolic profiling of TBI patients
could therefore be based on practical aspects of device design, such as which would be
easier to miniaturise. In this instance, it is likely that transmission-FTIR would be
the recommended technique as the miniaturisaton of the crystals used in ATR-FTIR
would be difficult.
The prediction of the LPR via a division of the predicted concentrations of lactate
and pyruvate was determined not to be feasible with the current error in prediction
174 CHAPTER 10. DISCUSSION AND FUTURE OUTLOOK
of the two compounds. An alternative method may be to generate a statistical model
for the prediction of the LPR directly from FTIR spectra; however, this would be
extremely complex. It is therefore recommended that attempts are made to reduce
the error in prediction of the two compounds, and the resulting improvement in the
LPR assessed, before the construction of such a model is attempted.
It should be noted that the FTIR spectra acquired, and subsequently analysed,
in this work required approximately 2.5 minutes to generate. This length of time
is compatible with the temporal resolution required of a sensor for on-line metabolic
profiling of TBI patients, which necessitates the taking of a measurement at least once
every five minutes.
10.3 Considerations for Future Device Development
An on-line sensor for metabolic profiling of TBI patients that is based upon a combina-
tion of FTIR spectroscopy and statistical modelling would require a higher predictive
accuracy than has been achieved in this work, particularly with respect to pyruvate.
The ideal way to improve the accuracy in prediction would be to improve the signal-
to-noise ratio of the spectral data acquired. An obvious way in which this could be
achieved would be to improve the instrumental setup used to acquire the data. A
developed sensor would not be based on a commercially-available spectrometer, such
as that used in this work, but would comprise a bespoke combination of commercially-
available components. As a result, there would be numerous opportunities to improve
the hardware of the instrument compared with the hardware of the spectrometer used
in this work. For example, the use of QCLs in FTIR spectroscopy is reported to
produce spectra with a higher signal-to-noise ratio than that achieved with SiC glo-
bar light sources, such as that used in this work. The integration of a QCL into the
final device would impose a limit on the region of the spectrum to be analysed, as
an individual QCL will emit only a narrow range of wavelengths of electromagnetic
radiation. However, the interval data presented in Chapters 8 and 9 suggest that
the use of small spectral intervals can improve the error in prediction. One obvious
disadvantage of the use of a QCL would be a significant increase in the cost of the
device.
It may also be possible to improve the signal-to-noise ratio of spectra using a
technique known as surface-enhanced infrared absorption spectroscopy, which is based
upon the same principles as SERS. However, this is not recommended as it would
introduce the same complications that were encountered using SERS.
If the quality of spectral data could not be improved sufficiently to achieve the
desired accuracy in the prediction of glucose, lactate and pyruvate concentrations,
further development of the statistical model used to analyse the data would be neces-
sary. It is possible that, at very low concentrations, the data are not linear. If this is
10.4. CONCLUDING REMARKS 175
the case, it may be necessary to apply a non-linear statistical model to the data, such
as nonlinear regression.
The analysis of clinical samples via a PLSR model trained on the transmission-
FTIR spectra of the synthetic solutions generated in this work has highlighted the
importance of training a final model on spectra of clinical samples. However, it is
recommended that this work is carried out using the hardware and sensing chip applied
to the final device, in order to ensure that the effects of these components can be
accounted for by the model.
Consideration would also need to be given to the calibration of the device to ensure
that the measured concentrations of glucose, lactate and pyruvate remain accurate.
The frequency with which this would need to be carried out would depend upon the
stability of the device. For example, the enzyme-based rsMD device described in
Chapter 4 is reportedly re-calibrated every three hours.112
A final consideration for the development of a sensor relates to the generation of
background spectra. It has been shown in this work that the application of both PLSR
and PCR to FTIR ratio spectra yields a lower RMSECV than that achieved when the
models are applied to FTIR single-beam spectra. The generation of FTIR ratio spectra
requires the acquisition of background spectra, which would need to be generated on
a regular basis in order to ensure the background continued to reflect the environment
of the sensor. This would require a sensing chip with an additional channel for the
introduction of a suitable reference sample and may require an automated process to
ensure that device operation is not labour-intensive.
10.4 Concluding Remarks
The work presented in this thesis has shown that it may be possible to improve upon
the current method for metabolic profiling of TBI patients, using a sensor based upon
FTIR spectroscopy. It is hoped that the results presented here, and the conclusions
drawn from them, will be useful in guiding the next phase of sensor development.

Appendices
177

Appendix A
Combinations of Glucose,
Lactate and Pyruvate Analysed
in Mixture Studies
A.1 Three-Component Mixtures
The concentrations of glucose, lactate and pyruvate in each of the three-component
mixtures studied via ATR-FTIR spectroscopy and transmission-FTIR spectroscopy
in their respective mixture studies are shown in Table A.1.
Table A.1: The concentrations of glucose, lactate and pyruvate in the three-component mix-
ture analysed as part of the ATR-FTIR spectroscopy and transmission-FTIR spectroscopy
mixture studies.
Glucose / mM Lactate / mM Pyruvate / mM
0.05 0.05 0.005
0.05 2.00 0.005
0.10 0.10 0.005
0.10 0.10 0.010
0.10 0.10 0.050
0.10 0.10 0.100
0.10 0.10 0.150
0.10 0.10 0.200
0.10 2.00 0.005
0.10 2.00 0.010
0.10 2.00 0.050
0.10 2.00 0.100
Continued on next page
179
180 APPENDIX A. GLUCOSE, LACTATE & PYRUVATE COMBINATIONS
Table A.1 – Continued from previous page
Glucose / mM Lactate/ mM Pyruvate / mM
0.10 2.00 0.150
0.10 2.00 0.200
0.10 4.00 0.005
0.10 4.00 0.010
0.10 4.00 0.050
0.10 4.00 0.100
0.10 4.00 0.150
0.10 4.00 0.200
0.10 6.00 0.005
0.10 6.00 0.010
0.10 6.00 0.050
0.10 6.00 0.100
0.10 6.00 0.150
0.10 6.00 0.200
0.70 0.10 0.005
0.70 0.10 0.010
0.70 0.10 0.050
0.70 0.10 0.100
0.70 0.10 0.150
0.70 0.10 0.200
0.70 2.00 0.005
0.70 2.00 0.010
0.70 2.00 0.050
0.70 2.00 0.100
0.70 2.00 0.150
0.70 2.00 0.200
0.70 4.00 0.005
0.70 4.00 0.010
0.70 4.00 0.050
0.70 4.00 0.100
0.70 4.00 0.150
0.70 4.00 0.200
0.70 6.00 0.005
0.70 6.00 0.010
0.70 6.00 0.050
0.70 6.00 0.100
0.70 6.00 0.150
Continued on next page
A.1. THREE-COMPONENT MIXTURES 181
Table A.1 – Continued from previous page
Glucose / mM Lactate/ mM Pyruvate / mM
0.70 6.00 0.200
1.00 0.05 0.005
1.00 0.10 0.005
1.00 0.10 0.010
1.00 0.10 0.050
1.00 0.10 0.100
1.00 0.10 0.150
1.00 0.10 0.200
1.00 2.00 0.005
1.00 2.00 0.010
1.00 2.00 0.050
1.00 2.00 0.100
1.00 2.00 0.150
1.00 2.00 0.200
1.00 4.00 0.005
1.00 4.00 0.010
1.00 4.00 0.050
1.00 4.00 0.100
1.00 4.00 0.150
1.00 4.00 0.200
1.00 6.00 0.005
1.00 6.00 0.010
1.00 6.00 0.050
1.00 6.00 0.100
1.00 6.00 0.150
1.00 6.00 0.200
3.00 0.10 0.005
3.00 0.10 0.010
3.00 0.10 0.050
3.00 0.10 0.100
3.00 0.10 0.150
3.00 0.10 0.200
3.00 2.00 0.005
3.00 2.00 0.010
3.00 2.00 0.050
3.00 2.00 0.100
3.00 2.00 0.150
Continued on next page
182 APPENDIX A. GLUCOSE, LACTATE & PYRUVATE COMBINATIONS
Table A.1 – Continued from previous page
Glucose / mM Lactate/ mM Pyruvate / mM
3.00 2.00 0.200
3.00 4.00 0.005
3.00 4.00 0.010
3.00 4.00 0.050
3.00 4.00 0.100
3.00 4.00 0.150
3.00 4.00 0.200
3.00 6.00 0.005
3.00 6.00 0.010
3.00 6.00 0.050
3.00 6.00 0.100
3.00 6.00 0.150
3.00 6.00 0.200
A.2 Two-Component Mixtures
The concentrations of glucose, lactate and pyruvate in each of the two-component
mixtures studied via ATR-FTIR spectroscopy and transmission-FTIR spectroscopy
in their respective mixture studies are shown in Table A.2.
Table A.2: The concentrations of glucose, lactate and pyruvate in the two-component mix-
tures analysed as part of the ATR-FTIR spectroscopy and transmission-FTIR spectroscopy
mixture studies.
Glucose / mM Lactate / mM Pyruvate / mM
0.00 0.05 0.005
0.00 0.05 0.010
0.00 0.05 0.050
0.00 0.05 0.150
0.00 0.05 0.200
0.00 0.10 0.005
0.00 0.10 0.010
0.00 0.10 0.050
0.00 0.10 0.150
0.00 0.10 0.200
Continued on next page
A.2. TWO-COMPONENT MIXTURES 183
Table A.2 – Continued from previous page
Glucose / mM Lactate/ mM Pyruvate / mM
0.00 0.50 0.005
0.00 0.50 0.010
0.00 0.50 0.050
0.00 0.50 0.150
0.00 0.50 0.200
0.00 4.00 0.005
0.00 4.00 0.010
0.00 4.00 0.050
0.00 4.00 0.150
0.00 4.00 0.200
0.05 0.00 0.005
0.05 0.00 0.010
0.05 0.00 0.050
0.05 0.00 0.150
0.05 0.00 0.200
0.05 0.05 0.000
0.05 0.10 0.000
0.05 0.50 0.000
0.05 4.00 0.000
0.10 0.00 0.005
0.10 0.00 0.010
0.10 0.00 0.050
0.10 0.00 0.150
0.10 0.00 0.200
0.10 0.05 0.000
0.10 0.10 0.000
0.10 0.50 0.000
0.10 4.00 0.000
0.70 0.00 0.005
0.70 0.00 0.010
0.70 0.00 0.050
0.70 0.00 0.150
0.70 0.00 0.200
0.70 0.05 0.000
0.70 0.10 0.000
0.70 0.50 0.000
0.70 4.00 0.000
Continued on next page
184 APPENDIX A. GLUCOSE, LACTATE & PYRUVATE COMBINATIONS
Table A.2 – Continued from previous page
Glucose / mM Lactate/ mM Pyruvate / mM
1.00 0.00 0.005
1.00 0.00 0.010
1.00 0.00 0.050
1.00 0.00 0.150
1.00 0.00 0.200
1.00 0.05 0.000
1.00 0.10 0.000
1.00 0.50 0.000
1.00 4.00 0.000
Bibliography
[1] A. I. R. Maas, N. Stocchetti and R. Bullock, Lancet Neurol., 2008, 7, 728–741.
[2] B. Roozenbeek, A. I. Maas and D. K. Menon, Nat. Rev. Neurol., 2013, 9, 231–
236.
[3] Head Injury: Triage, Assessment, Investigation and Early Management of Head
Injury in Children, Young People and Adults, National Institute for Health and
Care Excellence Costing Report, 2014.
[4] K. E. Saatman, A. C. Duhaime, R. Bullock, A. I. R. Maas, A. Valadka and
G. T. Manley, J. Neurotrauma, 2008, 25, 719–738.
[5] J. Ghajar, The Lancet, 2000, 356, 923–929.
[6] I. Jalloh, K. L. H. Carpenter, A. Helmy, T. A. Carpenter, D. K. Menon and
P. J. Hutchinson, Metab. Brain Dis., 2015, 30, 615–632.
[7] P. J. Hutchinson, I. Jalloh, A.Helmy, K. L. H. Carpenter, E. Rostami, B. M.
Bellander, M. G. Boutelle, J. W. Chen, J. Claassen, C. Dahyot-Fizelier, P. En-
blad, C. N. Gallagher, R. Helbok, L. Hillered, P. D. L. Roux, S. Magnoni,
a. D. K. M. H. S. Mangat, C. H. Nordstro¨m, K. H. O’Phelan, M. Oddo, J. P.
Barcena, C. Robertson, E. Ronne-Engstro¨m, J. Sahuquillo, M. Smith, N. Stoc-
chetti, A. Belli, A. T. Carpenter, J. P. Coles, M. Czosnyka, N. Dizdar, J. C.
Goodman, A. K. Gupta, T. H. Nielsen, N. Marklund, A. Montcriol, M. T.
O’Connell, M. A. Poca, A. Sarrafzadeh, R. J. Shannon, J. Skjøth-Rasmussen,
P. Smielewski, J. F. Stover, I. Timofeev, P. Vespa, E. Zavala and U. Ungerstedt,
Intensive Care Med., 2015, 41, 1517–1528.
[8] A. Helmy, K. L. H. Carpenter, D. K. Menon, J. D. Pickard and P. J. A. Hutchin-
son, J. Cereb. Blood Flow Metab., 2011, 31, 658–670.
[9] K. L. H. Carpenter, I. Timofeev, P. G. Al-Rawi, D. K. Menon, J. D. Pickard
and P. J. Hutchinson, Acta Neurochir. Suppl., 2008, 102, 207–213.
[10] CMA/Microdialysis, ISCUS Flex Microdialysis Analyser Technical Manual,
2008.
185
186 BIBLIOGRAPHY
[11] Y. Tholance, G. Barcelos, I. Quadrio, B. Renaud, F. Dailler and A. Perret-
Liaudet, Clin. Chim. Acta, 2011, 412, 647–654.
[12] P. Mergenthaler, U. Lindauer, G. A. Dienel and A. Meisel, Trends Neurosci.,
2013, 36, 587–597.
[13] P. Reinstrup, N. St˚ahl, P. Mellerg˚ard, T. Uski, U. Ungerstedt and C. H. Nord-
stro¨m, Neurosurgery, 2000, 47, 701–710.
[14] S. Sahu, D. S. Nag, A. Swain and D. P. Samaddar, World J. Biol. Chem., 2017,
8, 21–31.
[15] I. Jalloh, K. L. H. Carpenter, P. Grice, D. J. Howe, A. Mason, C. N. Gallagher,
A. Helmy, M. P. Murphy, D. K. Menon, T. A. Carpenter, J. D. Pickard and
P. J. Hutchinson, J. Cereb. Blood Flow Metab., 2015, 35, 111–120.
[16] I. Timofeev, K. L. H. Carpenter, J. Nortjie, P. G. Al-Rawi, M. T. O’Connell,
M. Czosnyka, P. Smielewski, J. D. Pickard, D. K. Menon, P. J. Kirkpatrick,
A. K. Gupta and P. J. Hutchinson, Brain, 2011, 134, 484–494.
[17] A. Helmy, K. L. H. Carpenter and P. J. Hutchinson, Curr. Med. Chem., 2007,
14, 1525–1537.
[18] T. Paraforou, K. Paterakis, K. Fountas, G. Paraforos, A. Chovas, A. Tasiou,
M. Mpakopoulou, D. Papadopoulos, A. Karavellis and A. Komnos, BMC Res.
Notes, 2011, 4, 1–6.
[19] C. N. Gallagher, K. L. H. Carpenter, P. Grice, D. J. Howe, A. Mason, I. Timo-
feev, D. K. Menon, P. J. Kirkpatrick, J. D. Pickard, G. R. Sutherland and P. J.
Hutchinson, Brain, 2009, 132, 2839–2849.
[20] K. L. H. Carpenter, I. Jalloh, C. N. Gallagher, P. Grice, D. J. Howe, A. Mason,
I. Timofeev, A. Helmy, M. P. Murphy, D. K. Menon, P. J. Kirkpatrick, T. A.
Carpenter, G. R. Sutherland, J. D. Pickard and P. J. Hutchinson, Eur. J. Pharm.
Sci., 2014, 57, 87–97.
[21] H. J. Butler, L. Ashton, B. Bird, G. Cinque, K. Curtis, J. Dorney, K. Esmond-
White, N. Fullwood, B. Gardner, P. L. Martin-Hirsch, M. J. Walsh, M. R.
McAinsh, N. Stone and F. L. Martin, Nat. Protoc., 2016, 11, 664–687.
[22] C. L. Haynes, A. D. McFarland and R. P. V. Duyne, Anal. Chem., 2005, 77,
338A–346A.
[23] M. L. Juan, M. Righini and R. Quidant, Nat. Photonics, 2011, 5, 349–356.
[24] A. P. F. Turner, Chem. Soc. Rev., 2013, 42, 3184–3196.
BIBLIOGRAPHY 187
[25] H. Yao, A. J. Shum, M. Cowan, I. La¨hdesma¨ki and B. A. Parviz, Biosens.
Bioelectron., 2011, 26, 3290–3296.
[26] H. Lee, T. K. Choi, Y. B. Lee, H. R. Cho, R. Ghaffari, L. Wang, H. J. Choi,
T. D. Chung, N. Lu, T. Hyeon, S. H. Choi and D. H. Kim, Nat. Nanotechnol.,
2016, 11, 566–572.
[27] J. R. Anusha, C. J. Raj, B. B. Cho, A. T. Fleming, K. H. Yu and B. C. Kim,
Sens. Actuators, B, 2015, 215, 536–543.
[28] S. K. Krishnan, E. Prokhorov, D. Bahena, R. Esparza and M. Meyyappan, ACS
Omega, 2017, 2, 1896–1904.
[29] M. Mazaheri, H. Aashuri and A. Simchi, Sens. Actuators, B, 2017, 251, 462–471.
[30] Y. Tang, Q. Liu, Z. Jiang, X. Yang, M. Wei and M. Zhang, Sens. Actuators, B,
2017, 251, 1096–1103.
[31] M. Jarosz, R. P. Socha, P. Jo´z´wik and G. D. Sulka, Appl. Surf. Sci., 2017, 408,
96–102.
[32] J. Zhang, C. Xu, R. Zhang, X. Guo, J. Wang, X. Zhang, D. Zhang and B. Yuan,
Mater. Lett., 2018, 210, 291–294.
[33] S. J. Cho, H. B. Noh, M. S. Won, C. H. Cho, K. B. Kim and Y. B. Shim,
Biosens. Bioelectron., 2018, 99, 471–478.
[34] S. Soylemez, B. Yoon, L. Toppare and T. M. Swager, ACS Sens., 2017, 2, 1123–
1127.
[35] C. Jiang, J. Zhu, Z. Li, J. Luo, J. Wang and Y. Sun, RSC Adv., 2017, 7, 44463–
44469.
[36] Y. M. Wang, J. W. Liu, J. H. Jiang and W. Zhong, Anal. Bioanal. Chem., 2017,
409, 4225–4232.
[37] W. Dong, L. Yang and Y. Huang, Talanta, 2017, 167, 359–366.
[38] L. Jin, Z. Meng, Y. Zhang, S. Cai, Z. Zhang, C. Li, L. Shang and Y. Shen, ACS
Appl. Mater. Interfaces, 2017, 9, 10027–10033.
[39] Q. Li, G. Tang, X. Xiong, Y. Cao, L. Chen, F. Xu and H. Tan, Sens. Actuators,
B, 2015, 215, 86–92.
[40] Z. Huang, J. Yang, L. Zhang, X. Geng, J. Ge, Y. Hu and Z. Li, Anal. Methods,
2017, 9, 4275–4281.
188 BIBLIOGRAPHY
[41] L. Wang, X. Kang and D. Pan, Inorg. Chem., 2017, 56, 6122–6130.
[42] L. Chen, L. Chen, M. Dotzert, C. W. J. Melling and J. Zhang, J. Nanobiotech-
nol., 2017, 15, 1–9.
[43] S. O. Alijanpour, R. Akhoondi and M. J. Chaichi, J. Anal. Chem., 2017, 72,
120–127.
[44] J. Ghosh, R. Ghosh and P. K. Giri, Sens. Actuators, B, 2018, 254, 681–689.
[45] C. Wang, Y. Li and Y. Wei, Sens. Actuators, B, 2017, 247, 595–601.
[46] H. Li, L. Zhu, X. Zhu and H. Zhang, Sens. Actuators, B, 2017, 246, 819–825.
[47] M. Liang, Y. Ren, H. Zhang, Y. Ma, X. Niu and X. Chen, Luminescence, 2017,
32, 1031–1038.
[48] F. Molaabasi, S. Hosseinkhani, A. A. Moosavi-Movahedi and M. Shamsipur,
RSC Adv., 2015, 5, 33123–33135.
[49] C. Li, X. Chen, F. Zhang, X. He, G. Fang, J. Liu and S. Wang, Anal. Chem.,
2017, 89, 10431–10438.
[50] C. Liu, Y. Sheng, Y. Sun, J. Feng, S. Wang, J. Zhang, J. Xu and D. Jiang,
Biosens. Bioelectron., 2015, 70, 455–461.
[51] J. S. Hansen, T. Hoeg-Jensen and J. B. Christensen, Tetrahedron, 2017, 73,
3010–3013.
[52] Z. Zhang, R. T. K. Kwok, Y. Yu, B. Z. Tang and K. M. Ng, ACS Appl. Mater.
Interfaces, 2017, 9, 38153–38158.
[53] A. L. Hu, Y. H. Liu, H. H. Deng, G. L. Hong, A. L. Liu, X. H. Lin, X. H. Xia
and W. Chen, Biosens. Bioelectron., 2014, 61, 374–378.
[54] D. Calabria, C. Caliceti, M. Zangheri, M. Mirasoli, P. Simoni and A. Roda,
Biosens. Bioelectron., 2017, 94, 124–130.
[55] Q. Chen, T. Sun, X. Song, Q. Ran, C. Yu, J. Yang, H. Feng, L. Yu and D. Wei,
Nanotechnology, 2017, 28, 315501–1–315501–9.
[56] S. RoyChoudhury, P. Manickam, Y. Umasankar and S. Bhansali, ECS Trans.,
2015, 69, 7–15.
[57] J. Kim, G. Valde´s-Ramı´rez, A. J. Bandodkar, W. Jia, A. G. Martinez,
J. Ramı´rez, P. Mercier and J. Wang, Analyst, 2014, 139, 1632–1636.
BIBLIOGRAPHY 189
[58] W. Jia, A. J. Bandodkar, G. Valde´s-Ramı´rez, J. R. Windmiller, Z. Yang,
J. Ramı´rez, G. Chan and J. Wang, Anal. Chem., 2013, 85, 6553–6560.
[59] N. V. Zaryanov, V. N. Nikitina, E. V. Karpova, E. E. Karyakina and A. A.
Karyakin, Anal. Chem., 2017, 89, 11198–11202.
[60] A. S. Mart´ın, S. Ceballo, F. Baeza-Lehnert, R. Lerchundi, R. Valdebenito,
Y. Contreras-Baeza, K. Alegr´ıa and L. F. Barros, PLoS One, 2014, 9, 1–9.
[61] N. Gajovic, G. Binyamin, A. Warsinke, F. W. Scheller and A. Heller, Anal.
Chem., 2000, 72, 2963–2968.
[62] M. Azhar, J. Mandon, A. H. Neerincx, Z. Liu, J. Mink, P. J. F. M. Merkus,
S. M. Cristescu and F. J. M. Harren, Appl. Phys. B, 2017, 123, 268–1–7.
[63] V. Lirtsman, M. Golosovsky and D. Davidov, Rev. Sci. Instrum., 2017, 88,
103105–1–7.
[64] K. Ataka, S. T. Stripp and J. Heberle, Biochim. Biophys. Acta, 2013, 1828,
2283–2293.
[65] R. Ebrahimifard, S. van den Driesche, H. Breiteneder, C. Hafner and M. J.
Vellekoop, Sens. Actuators, B, 2017, 247, 981–991.
[66] R. R´ıos-Reina, R. M. Callejo´n, C. Oliver-Pozo, J. M. Amigo and D. L. Garc´ıa-
Gonza´lez, Food Control, 2017, 78, 230–237.
[67] J. Zhang, P. Huang, Z. Wang and H. Dong, Biosci. Rep., 2017, 37, 1–10.
[68] O. J. Old, L. M. Fullwood, R. Scott, G. R. Lloyd, L. M. Almond, N. A. Shepherd,
N. Stone, H. Barr and C. Kendall, Anal. Methods, 2014, 6, 3901–3917.
[69] J. Nallala, G. R. Lloyd, N. Shepherd and N. Stone, Analyst, 2016, 141, 630–639.
[70] J. Nallala, G. R. Lloyd, M. Hermes, N. Shepherd and N. Stone, Vib. Spectrosc.,
2017, 91, 83–91.
[71] F. Palombo, M. Madami, D. Fioretto, J. Nallala, H. Barr, A. David and
N. Stone, J. Biophotonics, 2016, 9, 694–700.
[72] O. J. Old, G. R. Lloyd, J. Nallala, M. Isabelle, L. M. Almond, N. A. Shepherd,
C. A. Kendall, A. C. Shore, H. Barr and N. Stone, Analyst, 2017, 142, 1227–
1234.
[73] C. Petibois, A. M. Melin, A. Perromat, G. Cazorla and G. De´le´ris, J. Lab. Clin.
Med., 2000, 135, 210–215.
190 BIBLIOGRAPHY
[74] D. Lafrance, L. C. Lands and D. H. Burns, Talanta, 2003, 60, 635–641.
[75] M. Meinke, H. Albrecht, C. Antoniou and H. Richter, J. Biomed. Opt., 2008,
13, 014021–1–6.
[76] S. Liakat, K. A. Bors, L. Xu, C. M. Woods, J. Doyle and C. F. Gmachi, Biomed.
Opt. Express, 2014, 5, 2397–2404.
[77] S. Liakat, K. A. Bors, T. Y. Huang, A. P. M. Michel, E. Zanghi and C. F.
Gmachl, Biomed. Opt. Express, 2013, 4, 1083–1090.
[78] J. Kuligowski, G. Quinta´s, M. de la Guardia and B. Lendl, Anal. Chim. Acta,
2010, 679, 31–42.
[79] A. V. Ewing, G. S. Clarke and S. G. Kazarian, Biomicrofluidics, 2016, 10,
024125–1–13.
[80] G. R. Lloyd, J. Wood, C. Kendall, T. Cook, N. Shepherd and N. Stone, Vib.
Spectrosc., 2012, 60, 43–49.
[81] P. Crow, B. Barrass, C. Kendall, M. Hart-Prieto, M. Wright, R. Persad and
N. Stone, Br. J. Cancer, 2005, 92, 2166–2170.
[82] Y. Yue, J. J. Wang, P. A. M. Basheer, J. J. Boland and Y. Bai, Constr. Build.
Mater., 2017, 135, 369–376.
[83] M. Jermyn, K. Mok, J. Mercier, J. Desroches, J. Pichette, K. Saint-Arnaud,
L. Bernstein, M. C. Guiot, K. Petrecca and F. Leblond, Sci. Transl. Med., 2015,
7, 1–11.
[84] I. E. I. Petterson, J. C. C. Day, L. M. Fullwood, B. Gardner and N. Stone, Anal.
Bioanal. Chem., 2015, 407, 8311–8320.
[85] G. R. Lloyd, L. E. Orr, J. Christie-Brown, K. McCarthy, S. Rose, M. Thomas
and N. Stone, Analyst, 2013, 138, 3900–3908.
[86] A. Rajput, S. Kumar and J. P. Singh, Analyst, 2017, 142, 3959–3966.
[87] G. Barbillon, V. E. Sandana, C. Humbert, B. Be´lier, D. J. Rogers, F. H.
Teherani, P. Bove, R. McClintock and M. Razeghi, J. Mater. Chem., 2017,
5, 3528–3535.
[88] F. Xu, S. Xie, H. Xu, X. Chen, H. Yu and L. Wang, Appl. Surf. Sci., 2017, 410,
566–573.
[89] C. Y. Zhang, R. Hao, B. Zhao, Y. Z. Fu, Y. W. Hao and Y. Q. Liu, J. Mater.
Sci., 2017, 52, 11391–11401.
BIBLIOGRAPHY 191
[90] J. Li, W. Zhang, H. Lei and B. Li, Nano Res., 2017, 1–9.
[91] Y. Liu, C. Xu, Z. Zhu, Q. Zhu, A. G. Manohari and Z. Shi, Appl. Surf. Sci.,
2018, 427, 830–836.
[92] Y. Zhao, X. Li, Y. Liu, L. Zhang, F. Wang and Y. Lu, Sens. Actuators, B, 2017,
247, 850–857.
[93] S. Yuan, F. Ge, M. Zhou, Z. Cai and S. Guang, Nanoscale Res. Lett., 2017, 12,
1–5.
[94] Y. Kang, L. Zhang, H. Zhang, T. Wu and Y. Du, J. Appl. Spectrosc., 2017, 84,
225–230.
[95] X. Yang, Y. He, X. Wang and R. Yuan, Appl. Surf. Sci., 2017, 416, 581–586.
[96] Y. Wu, T. Jiang, Z. Wu and R. Yu, Biosens. Bioelectron., 2018, 99, 646–652.
[97] Y. Zhang, P. Yang, M. Abubaker, H. Muhammed, S. K. Alsaiari, B. Moosa,
A. Almalik, A. Kumar, E. Ringe and N. M. Khashab, ACS Appl. Mater. Inter-
faces, 2017, 9, 37597–37605.
[98] H. Shi, N. Chen, Y. Su, H. Wang and Y. He, Anal. Chem., 2017, 89, 10279–
10285.
[99] W. Zhou, Y. F. Tian, B. C. Yin and B. C. Ye, Anal. Chem., 2017, 89, 6120–6128.
[100] A. Zengin, U. Tamer and T. Caykara, J. Raman Spectrosc., 2017, 48, 668–672.
[101] K. E. Shafer-Peltier, C. L. Haynes, M. R. Glucksberg and R. P. V. Duyne, J.
Am. Chem. Soc., 2003, 125, 588–593.
[102] N. C. Shah, O. Lyandres, J. T. W. Jr., M. R. Glucksberg and R. P. V. Duyne,
Anal. Chem., 2007, 79, 6927–6932.
[103] K. Ma, J. M. Yuen, N. C. Shah, J. T. Walsh, M. R. Glucksberg and R. P. V.
Duyne, Anal. Chem., 2011, 83, 9146–9152.
[104] B. Sharma, P. Bugga, L. R. Madison, A. I. Henry, M. G. Blaber, N. G.
Greeneltch, N. Chiang, M. Mrksich, G. C. Schatz and R. P. V. Duyne, J. Am.
Chem. Soc., 2016, 138, 13952–13959.
[105] A. M. K. Enejder, T. G. Scecina, J. Oh, M. Hunter, W. C. Shih, S. Sasic, G. L.
Horowitz and M. S. Feld, J. Biomed. Opt., 2005, 10, 031114–1–9.
[106] A. Tycova, J. Prikryl and F. Foret, Electrophoresis, 2017, 38, 1977–1987.
192 BIBLIOGRAPHY
[107] O. M. Buja, O. D. Gordan, N. Leopold, A. Morschhauser, J. Nestler and D. R. T.
Zahn, Beilstein J. Nanotechnol., 2017, 8, 237–243.
[108] D. A. Jones, M. C. Parkin, H. Langemann, H. Landolt, S. E. Hopwood, A. J.
Strong and M. G. Boutelle, J. Electroanal. Chem., 2002, 538–539, 243–252.
[109] M. C. Parkin, S. E. Hopwood, A. J. Strong and M. G. Boutelle, Trends Anal.
Chem., 2003, 22, 487–497.
[110] D. Feuerstein, A. Manning, P. Hashemi, R. Bhatia, M. Fabricius, C. Tolias,
C. Pahl, M. Ervine, A. J. Strong and M. G. Boutelle, J. Cereb. Blood Flow
Metab., 2010, 30, 1343–1355.
[111] M. L. Rogers, P. A. Brennan, C. L. Leong, S. A. N. Gowers, T. Aldridge, T. K.
Mellor and M. G. Boutelle, Anal. Bioanal. Chem., 2013, 405, 3881–3888.
[112] M. L. Rogers, D. Feuerstein, C. L. Leong, M. Takagaki, X. Niu, R. Graf and
M. G. Boutelle, ACS Chem. Neurosci., 2013, 4, 799–807.
[113] M. L. Rogers, C. L. Leong, S. A. N. Gowers, I. C. Samper, S. J. Jewell, A. Khan,
L. McCarthy, C. Pahl, C. M. Tolias, D. C. Walsh, A. J. Strong and M. G.
Boutelle, J. Cereb. Blood Flow Metab., 2017, 37, 1883–1895.
[114] R. Kurita, K. Hayashi, X. Fan, K. Yamamoto, T. Kato and O. Niwa, Sens.
Actuators, B, 2002, 87, 296–303.
[115] C. K. Su, S. C. Yen, T. W. Li and Y. C. Sun, Anal. Chem., 2016, 88, 6265–6273.
[116] J. Zhao, Unpublished.
[117] G. Socrates, Infrared Characteristic Group Frequencies, Wiley & Sons Ltd, 2nd
edn., 1994.
[118] Contact and Orientation Effects in FT-IR ATR Spectra, 2011.
[119] Orientation Effects in ATR Spectra, Perkin Elmer technical report, 2012.
[120] T. Clausen, O. L. Alves, M. Reinert, E. Doppenberg, A. Zauner and R. Bullock,
J. Neurosurg., 2005, 103, 233–238.
[121] D. W. Nelson, B. Thornquist, R. M. MacCallum, H. Nystro¨m, A. Holst,
A. Rudehill, M. Wanecek, B. M. Bellander and E. Weitzberg, BMC Med., 2011,
9, 1–17.
[122] ATR Accessories: An Overview, Perkin Elmer technical report, 2004.
[123] How to Select an Infrared Transmission Window, Pike Technologies Application
Note 0602.
BIBLIOGRAPHY 193
[124] J. Shao, M. Lin, Y. Li, X. Li, J. Liu, J. Liang and H. Yao, PLoS One, 2012, 7,
e48127–1–e48127–6.
[125] C. H. Munro, W. E. Smith, M. Garner, J. Clarkson and P. C. White, Langmuir,
1995, 11, 3712–3720.
[126] S. Nie and S. R. Emory, Science, 1997, 275, 1102–1106.
[127] K. S. Lee and M. A. El-Sayed, J. Phys. Chem. B, 2006, 110, 19220–19225.
[128] S. Kruszewski and M. Cyrankiewicz, Acta Phys. Pol., A, 2012, 121, A68–A74.
[129] J. T. Edsall, J. W. Otvos and A. Rich, J. Am. Chem. Soc., 1950, 72, 474–477.
[130] G. D. Fleming, J. J. Finnerty, M. Campos-Vallette, F. Ce´lis, A. E. Aliaga,
C. Fredes and R. Koch, J. Raman Spectrosc., 2009, 40, 632–638.
[131] S. F. Parker, Chem. Phys., 2013, 424, 75–79.
[132] M. Kaminski, A. Kudelski and M. Pecul, Journal of Physical Chemistry B, 2012,
116, 4976–4990.
[133] C. Jing and Y. Fang, Chem. Phys., 2007, 332, 27–32.
[134] Raman Data and Analysis, HORIBA Jobin Yvon application note.