Resting-State Electroencephalography for Continuous, Passive Prediction of Coma Recovery After Acute Brain Injury

Abstract

Accurately predicting emergence from disorders of consciousness (DoC) after acute brain injury can profoundly influence mortality, acute management, and rehabilitation planning. While recent advances in functional neuroimaging and stimulus-based EEG offer the potential to enrich shared decision-making, their procedural sophistication and expense limit widespread availability or repeated performance. We investigated continuous EEG (cEEG) within a passive, "resting-state" framework to provide continuously updated predictions of DoC recovery at 24-, 48-, and 72-hour prediction horizons. To develop robust, continuous prediction models from a large population of patients with acute brain injury (ABI), we leveraged a recently described pragmatic approach transforming Glasgow Coma Scale assessment sub-score combinations into frequently assessed DoC diagnoses: coma, vegetative state, minimally conscious state with or without language, and post-injury confusional or recovered states. We retrospectively identified consecutive patients undergoing cEEG following acute traumatic brain injury (TBI), subarachnoid hemorrhage (SAH), or intracerebral hemorrhage (ICH). Models continuously predicting DoC diagnosis for multiple prediction horizons were evaluated utilizing recent clinical assessments with or without cEEG information, which comprised a comprehensive EEG feature set of 288 time, frequency, and time-frequency characteristics computed from consecutive 5-minute EEG epochs, with 6 additional features capturing each EEG features temporal dynamics. Features were fed into a predictive model developed with cross-validation; the ordinal DoC diagnosis was discriminated using an ensemble of XGBoost binary classifiers. For 201 ABI patients (46 TBI, 140 SAH, 15 ICH patients comprising 27,280 cEEG-hours with concomitant clinical assessments), cEEG-augmented models accurately predicted the future DoC diagnosis at 24 hours (one-vs-rest AU-ROC, 92.4%; weighted-F1 84.1%), 48 hours (one-vs-rest AU-ROC=88%, weighted-F1=80%) and 72 hours (one-vs-rest AU-ROC=86.3%, weighted-F1=76.6%). Models were robust to utilizing different ordinal cut-points for the DoC prediction target and evaluating additional models derived from specific sub-populations using a confound-isolating cross-validation framework. The most robust features across evaluation configurations included Petrosian fractal dimension, relative power of high to low (gamma-beta to delta-alpha) EEG frequency spectra, energy within the 12-35 Hz frequency band in the short-time Fourier transform domain, and wavelet entropy. The cEEG-augmented model exceeded the performance of models using preceding clinical assessments, continuously predicting future DoC diagnosis with one-vs-rest AU-ROC in the range of 84.3-92.4% while utilizing approaches to limit overfitting. The proposed continuous, resting-state cEEG prediction method represents a promising tool to predict DoC emergence in ABI patients. Enabling these methods prospectively would represent a new paradigm of continuous prognostic monitoring for predicting coma recovery and assessing treatment response.