EEG and Deep Learning for Predicting the Outcome of Comatose Patients After Cardiac Arrest

Most patients who experience cardiac arrest remain comatose after restoration of blood circulation due to post-anoxic brain injury. Accurate outcome prediction is essential for guiding clinical management and communicating with relatives. Among the various neuroprognostication modalities, resting-state electroencephalogram (EEG) analysis is the most widely used, as it directly reflects brain activity. Currently, EEG-based clinical evaluation relies on visual interpretation, which is prone to intra- and inter-rater variability. To overcome these limitations, computer-based methods, particularly convolutional neural networks (CNNs), have recently emerged as promising alternatives. The aim of this work is to investigate the potential of a CNN to predict post-cardiac arrest outcome from EEG data, based on Cerebral Performance Category (CPC). Outcomes were classified as favourable (FO) for CPC scores of 1 or 2 and unfavourable (UO) otherwise. The dataset included EEG signals from 483 patients (39% FO) from the public I-CARE database, recorded within 12–24 hours after cardiac arrest. Model optimization was performed via repeated 5-fold cross-validation, combining preprocessing strategies evaluation and Bayesian hyperparameter tuning. The best model achieved a mean validation area under the receiver operating characteristic curve (AUC) of 0.844 ± 0.050 and, on an independent test set, an AUC of 0.838 with 76% balanced accuracy. Optimization led to only marginal improvements, suggesting model robustness and limited sensitivity to preprocessing or parameter variations. Similar validation and test results indicate effective generalization to unseen data. To gain insights into the model’s decision-making process, gradient-weighted class activation mapping (Grad-CAM) was used. This method highlighted the EEG segments that most contributed to the model’s predictions. The analysis revealed that the network’s decisions aligned with expert knowledge, indicating that it had effectively learned clinically relevant EEG patterns. This study confirms the strong potential of CNNs to provide stable and objective outcome predictions, supporting increasing trust in artificial intelligence and encouraging its integration into routine critical care practice.