A continuous convolutional neural network: very early EEG most predictive for poor neurological outcome in postanoxic coma

Objective: Our first objective was to characterize the predictive power of electroencephalography (EEG) over time after cardiac arrest (CA) using a continuous, convolutional neural network (CNN). Our second objective was to investigate the performance difference between a 9 and a 4-electrode-model in predicting poor neurological outcome. Methods: We trained a CNN model using hourly 5-min EEG epochs from 366 postanoxic coma patients from the start of recording up to 72 h after CA. Primary outcome was best Cerebral Performance Category (CPC) within 6 months after CA, classified as good (CPC score 1–2) or poor (CPC score 3–5). Prediction was based on the average of predictions over all available hours, both independently and cumulatively. We report (1) model discrimination of poor neurological outcome up to 72 h after CA and (2) model performance difference between a 9 and 4-electrode configuration. Results: The 9-electrode cumulative CNN model reached an area under the receiver operating characteristic curve (AUC) of 82 % and a sensitivity of 71 % at 0 % false-positive rate (FPR) over all available hours. Highest predictive performance was observed before 14 h after CA, peaking at 8 h (AUC = 0.93, sensitivity at 0 % FPR = 0.81). Performance differences over hours were negligible between the 9 and 4-electrode model. Conclusion: The CNN seems valuable for poor neurological outcome prediction in postanoxic coma patients, especially for very early (6–14 h after CA) EEG. A reduced electrode configuration is equally sufficient as the full set. Future studies could explore capturing nuanced temporal dependencies in EEG signals for prognostication.