High density EEG and deep learning outcome prediction on the first day of coma after cardiac arrest

We assessed the coma outcome prediction using a deep learning analysis applied to resting EEG on the first and second day after cardiac arrest, and its complementarity to clinical prognosis. We recorded 62-channel resting-state EEG in comatose patients across three Swiss hospitals during the first (N = 165) and second (N = 100) day of coma. Patient outcome was classified as favorable if the best Cerebral Performance Category was 1-2. A convolutional neural network provided a predicted probability for favorable outcome for each patient and recording day. Predictive performance was additionally evaluated on an external 19-channel dataset collected outside Switzerland (N = 60). The deep learning prediction was compared to EEG-based clinical markers, brainstem reflexes and motor responses. On the first day, for 62 channels, sensitivity and specificity for favorable outcome were 0.98±0.01 and 0.88±0.05 when maximizing both metrics. A sensitivity of 0.98±0.01 and a specificity of 0.64±0.14 was achieved when maximizing the sensitivity for favorable outcome and a sensitivity of 0.41±0.11 and a specificity of 0.99±0.01 when maximizing unfavorable outcome specificity. On the first day, using 19 channels, we obtained marginally lower values for sensitivity at 0.95±0.02 and specificity at 0.84±0.05 for favorable outcome. On the external dataset, sensitivity and specificity for favorable outcome were 0.83 and 0.92. The second day was less predictive with 0.80±0.09 sensitivity and 0.63±0.04 specificity for favorable outcome. The outcome prediction was consistent with clinical markers, except brainstem reflexes. On the first day of coma, a deep learning analysis of resting-state EEG provides accurate outcome prediction, complementing clinical markers.