CORE-Coma: Deep Learning Framework for Coma Prognosis from Auditory Event-Related Potentials

Accurate prognosis of coma emergence is difficult because bedside behavioral scales can fail to detect residual consciousness. Auditory oddball event-related potentials (ERPs) offer a physiological readout, but single-component markers (e.g., MMN or P3) have limited sensitivity and generalizability. We present CORE-Coma, a deep learning framework for full-waveform ERP analysis, trained exclusively on healthy controls and evaluated zero-shot in coma patients. We analyzed ERPs from 39 healthy controls and 8 coma patients in the intensive care unit (ICU), segmenting EEG recordings into ~5-minute sub-blocks to capture temporal fluctuations. We define two complementary, model-derived metrics: a time-resolved ERP Separability Score (ESS) and a subject-level Global ERP Separability Index (GESI). Controls showed near-ceiling standard–deviant separability (ROC AUC=0.99), while separability was reduced in coma (ROC AUC=0.68). CORE-Coma identified all patients who emerged from coma (3/3; sensitivity 100%) and 4/5 patients who did not emerge (specificity 80%), yielding accuracy=87.5% (7/8). ESS revealed temporal fluctuations (waxing–waning) of responsiveness in coma at ~5-minute resolution, absent in controls. SHAP explanations localized influential features, including frontocentral electrodes and time windows consistent with canonical oddball components: 100–150 ms (N1/MMN) and 270–370 ms (P3a/P3b). By combining bedside-feasible scalp EEG with time-resolved and subject-level metrics, CORE-Coma offers an etiology-agnostic approach to coma prognosis. Prospective multicenter studies are needed to validate performance and support clinical deployment.