HyperEnsemble Learning from Multimodal Biosignals to Robustly Predict Functional Outcome after Cardiac Arrest

Aligned with the George B. Moody PhysioNet Challenge 2023, we aim to create an open-source algorithm to predict coma recovery 3–6 months following return of spontaneous circulation after cardiac arrest. Our algorithm employs multimodal biosignals from a dataset spanning five U.S. and European hospitals. The preprocessing pipeline enhances signal quality through bandpass and notch filtering, resampling, and z-score normalization. EEG channels are organized into 21 bipolar montages and standardized ECG lead selection. We extract 362 EEG expert-based features, encompassing temporal, spectral, time-frequency, and connectivity metrics. ECG features include 16 heart rate variability indices, features indicating atrial tachyarrhythmia, and shockable rhythms. Our classification approach involves initial positional encoding, followed by an ensemble of CatBoost combined with stacked ensembles with diverse base learners in the first layer. To improve robustness and reduce bias, our ensemble strategy employs various crossvalidation committees, including confounder-isolating cross-validation. Our proposed predictive framework demonstrates an average challenge score mechanism of 0.651 ± 0.077 through 5-fold cross-validation. Our “AIrhythm” scored 0.792 on the hidden test set and officially ranked 1^st in the PhysioNet Challenge 2023.