Machine Learning for Prognosis of Acute Brain-Injured Patients in the ICU Using EEG Complexity Analysis and Naturalistic Narrative Stimuli
Abstract
Assessing the residual cognition of behaviourally unresponsive patients with critical brain injuries continues to be a challenge. There is a need for tools that can predict patient outcomes to better inform decisions being made in the intensive care unit. In this study, we examine brain complexity through various measures that quantify the intricate patterns and dynamics of electrical signals recorded through high-density electroencephalography (EEG). We compare responses to a naturalistic auditory stimulation task with those from the scrambled versions of the same stimuli to determine if these differences can predict the survival outcomes of patients with acute brain injury. I assess 52 acutely brain-injured patients and 18 healthy controls. Results are presented in three parts. Part One showed that while complexity measures could distinguish between conditions in healthy controls, they were less successful in predicting patient outcomes. However, Part Two found a significant difference between patients with favourable and unfavourable outcomes when complexity was examined independent of Task differences. In Part Three, I found that various complexity measures can be used to predict outcomes at an individual level with machine learning. These findings suggest that EEG complexity measures have the potential as prognostic tools for individual patients, with the quality of complex brain activity being more informative for prognosis than task-based differences in acute brain-injury patients.