Aims: Prediction of sudden cardiac death (SCD) from routine 12-lead ECG is clinically important, yet published models vary in architecture, input representation, loss function, and evaluation protocol, making cross-study comparison unreliable. We present a controlled comparison using existing implementations of four model architectures and two training objectives using the same patient cohort and evaluation framework.
Methods: We analyzed 460,154 ECGs from 22,482 patients, where 1045 patients had SCD endpoint. Four key architectures were evaluated: XGBoost and LightGBM operating on standard tabulated algorithmically derived ECG features (GE 12SL), a multilayer perceptron (MLP) also operating on 12SL features, and a one-dimensional convolutional neural network (CNN) operating on raw ECG waveforms. Each architecture was trained under both a survival objective (accelerated failure time for XGBoost; Cox proportional hazards for LightGBM; cause-specific hazard for MLP and CNN) and a binary classification objective, all targeting event occurrence within a two-year horizon. Hyperparameter optimization was conducted via nested cross-validation and Bayesian search.
Results: For the SCD endpoint under binary classification, the CNN achieved an AUROC of 76.42%, compared with 79.55% for XGBoost, 78.85% for LightGBM, and 79.20% for the MLP, with survival objective showing similar results. MLP ablation revealed that gains came primarily from the initial embedding layers, while model architectural or training optimization choices had limited impact.
Conclusion: The comparison isolates the contributions of input representation (engineered features vs. raw waveforms), model capacity (gradient-boosted trees vs. neural networks), and training objective (survival vs. binary) to predictive performance. Engineered 12SL features outperformed the raw waveform input for two-year SCD prediction, and model architecture choice contributed less than input representation to observed performance differences.