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

OBJECTIVES

Continuous assessment of physiology after traumatic brain injury is essential to prevent secondary brain insults. The present work aims at the development of a method for detecting physiologic states associated with the outcome from time-series physiologic measurements using a hidden Markov model.

DESIGN

Unsupervised clustering of hourly values of intracranial pressure/cerebral perfusion pressure, the compensatory reserve index, and autoregulation status was attempted using a hidden Markov model. A ternary state variable was learned to classify the patient's physiologic state at any point in time into three categories ("good," "intermediate," or "poor") and determined the physiologic parameters associated with each state.

SETTING

The proposed hidden Markov model was trained and applied on a large dataset (28,939 hr of data) using a stratified 20-fold cross-validation.

PATIENTS

The data were collected from 379 traumatic brain injury patients admitted to Addenbrooke's Hospital, Cambridge between 2002 and 2016.

INTERVENTIONS

Retrospective observational analysis.

MEASUREMENTS AND MAIN RESULTS

Unsupervised training of the hidden Markov model yielded states characterized by intracranial pressure, cerebral perfusion pressure, compensatory reserve index, and autoregulation status that were physiologically plausible. The resulting classifier retained a dose-dependent prognostic ability. Dynamic analysis suggested that the hidden Markov model was stable over short periods of time consistent with typical timescales for traumatic brain injury pathogenesis.

CONCLUSIONS

To our knowledge, this is the first application of unsupervised learning to multidimensional time-series traumatic brain injury physiology. We demonstrated that clustering using a hidden Markov model can reduce a complex set of physiologic variables to a simple sequence of clinically plausible time-sensitive physiologic states while retaining prognostic information in a dose-dependent manner. Such states may provide a more natural and parsimonious basis for triggering intervention decisions.