Feasibility of Hidden Markov Models for the Description of Time-Varying Physiologic State After Severe Traumatic Brain Injury.

Abstract

Objective: Continuous assessment of physiology after traumatic brain injury (TBI) is essential to prevent secondary brain insults. The present work aims at the development a method for detecting physiological states associated with outcome from time-series physiological measurements using a Hidden Markov Model (HMM). Design: Unsupervised clustering of hourly values of intracranial/cerebral perfusion pressure (ICP/CPP), the compensatory reserve index (RAP), and autoregulation status (PRx) was attempted using a HMM. A ternary state variable was learned to classify the patient’s physiological state at any point in time into three categories (‘good’, ‘intermediate’ or ‘poor) and determined the physiological parameters associated with each state. Setting: The proposed HMM was trained and applied on a large dataset (28,939 hours of data) using a stratified 20-fold cross-validation. Patients: The data was collected from 379 TBI patients admitted to Addenbrooke’s Hospital, Cambridge between 2002 and 2011. Intervention: Retrospective observational analysis. Measurements and Main Results: Unsupervised training of the HMM yielded states characterized by ICP, CPP, RAP and PRx that were physiologically plausible. The resulting classifier retained a dose-dependent prognostic ability. Dynamic analysis suggested that the HMM was stable over short periods of time consistent with typical timescales for TBI pathogenesis. Conclusions: To our knowledge, this is the first application of unsupervised learning to multidimensional time-series TBI physiology. We demonstrated that clustering using a HMM can reduce a complex set of physiological variables to a simple sequence of clinically plausible time-sensitive physiological states while retaining prognostic information in a dose-dependent manner. Such states may provide a more natural and parsimonious basis for triggering intervention decisions.