Neurological Outcome Prediction After Cardiac Arrest: A Multi-Level Deep Learning Approach with Feature and Decision Fusion
Cardiac arrest leads to complex neurological outcomes, demanding accurate predictions to guide post-arrest care. Using the International Cardiac Arrest Research Consortium (I-CARE) dataset, we developed models to discern between 'good' and 'poor' neurological outcomes post-cardiac arrest. We concatenated clinically relevant, manually extracted EEG features with autoencoder-derived, automatically extracted features to train transformer and Bi-LSTM models. Additionally, we ensembled the predicted probabilities between these deep learning models with a statistical model trained on non-EEG clinical variables. This ensemble approach demonstrated that the transformer excel at capturing long-term temporal dependencies, and the fusion of features and prognosis decisions led to improved model performance in terms of AUROC in predicting neurological outcomes post-cardiac arrest.
