Digital phenotyping through multimodal, unobtrusive sensing
University of Cambridge
Doctor of Philosophy (PhD)
MetadataShow full item record
Perez Pozuelo, I. (2021). Digital phenotyping through multimodal, unobtrusive sensing (Doctoral thesis). https://doi.org/10.17863/CAM.68489
The growing adoption of multimodal wearable and mobile devices, such as smartphones and wrist-worn watches has generated an increase in the collection of physiological and behavioural data at scale. This digital phenotyping data enables researchers to make inferences regarding users’ physical and mental health at scale, for the first time. However, translating this data into actionable insights requires computational approaches that turn unlabelled, multimodal time-series sensor data into validated measures that can be interpreted at scale. This thesis describes the derivation of novel computational methods that leverage digital phenotyping data from wearable devices in large-scale populations to infer physical behaviours. These methods combine insights from signal processing, data mining and machine learning alongside domain knowledge in physical activity and sleep epidemiology. First, the inference of sleeping windows in free-living conditions through a heart rate sensing approach is explored. This algorithm is particularly valuable in the absence of ground truth or sleep diaries given its simplicity, adaptability and capacity for personalization. I then explore multistage sleep classification through combined movement and cardiac wearable sensing and machine learning. Further, I demonstrate that postural changes detected through wrist accelerometers can inform habitual behaviours and are valuable complements to traditional, intensity-based physical activity metrics. I then leverage the concomitant responses of heart rate to physical activity that can be captured through multimodal wearable sensors through a self-supervised training task. The resulting embeddings from this task are shown to be useful for the downstream classification of demographic factors, BMI, energy expenditure and cardiorespiratory fitness. Finally, I describe a deep learning model for the adaptive inference of cardiorespiratory fitness (VO2max) using wearable data in free living conditions. I demonstrate the robustness of the model in a large UK population and show the models’ adaptability by evaluating its performance in a subset of the population with repeated measures ~6 years after the original recordings. Together, this work increases the potential of multimodal wearable and mobile sensors for physical activity and behavioural inferences in population studies. In particular, this thesis showcases the potential of using wearable devices to make valuable physical activity, sleep and fitness inferences in large cohort studies. Given the nature of the data collected and the fact that most of this data is currently generated by commercial providers and not research institutes, laying the foundations for responsible data governance and ethical use of these technologies will be critical to building trust and enabling the development of the field of digital phenotyping.
Digital Phenotyping, Wearables, Health Data Science, Applied Machine Learning, Sleep, Epidemiology, Self-supervised learning, Fitness
I was funded by GlaxoSmithKline and the Engineering and Physical Sciences Research Council. I was also supported by the Alan Turing Institute through their Enrichment Scheme.
This record's DOI: https://doi.org/10.17863/CAM.68489
All rights reserved, All rights reserved