Resting-state EEG for continuous prognostic monitoring and prediction of coma recovery after acute brain injury.

OBJECTIVE

We developed a continuous prognostic monitoring tool to predict recovery from disorders of consciousness (DoC) following acute brain injury (ABI), utilizing resting-state EEG recorded during routine clinical care.

METHODS

Predictive models updating every 5 min were developed using serial neurologic assessments and continuous resting-state EEG to predict future consciousness level at 24-, 48-, and 72-hour time horizons. An ensemble of CatBoost classifiers was utilized for multi-class DoC grade prediction, leveraging a comprehensive set of 242 computed EEG features encoding time, frequency, and time-frequency characteristics. Conventional and confound-isolating cross-validation mitigated biases and increased robustness. Performance was compared across multiple ordinal DoC grade cut-points and time horizons.

RESULTS

201 patients met inclusion criteria. Models incorporating EEG outperformed behavioral assessments alone, achieving a mean one-vs-rest AUROC of 0.88-0.89 (EEG + GCS) across 24-72-hour horizons and various trichotomized cut-points, with 95% bootstrap confidence intervals. The most robust features included global field power, theta-band (4-8 Hz) bandpower, beta-band (13-30 Hz) phase-locking value, and spectral entropy.

CONCLUSIONS

EEG-augmented models enabled continuous prediction of future DoC grade after ABI.

SIGNIFICANCE: Combining EEG with other serial measures during routine clinical care creates a novel paradigm for improved shared decision-making through continuous prognostic monitoring and serial assessment.