Electrocardiogram-Based Deep Learning for Time-Resolved Prediction of Heart Failure With Reduced Ejection Fraction: A Multinational Study

Heart failure with reduced ejection fraction (HFrEF) is associated with substantial morbidity and mortality worldwide, and early identification of individuals at risk remains a major clinical challenge. Although existing risk models and biomarkers can provide prognostic information, their application is often limited by the need for laboratory testing or imaging, as well as variability in performance across populations.

In this study, we develop a deep learning-based survival model using raw 12-lead electrocardiogram (ECG) data to predict the future onset of HFrEF. The model is designed to generate individualized, time-to-event risk estimates over a 5-year follow-up period, allowing for dynamic assessment of risk trajectories rather than static classification.

The model is trained on data from Zhongshan Hospital and externally validated in independent cohorts from Shanghai Tenth People's Hospital and Beth Israel Deaconess Medical Center. These cohorts include a broad spectrum of patients, ranging from individuals without known cardiovascular disease to those with diverse clinical conditions, thereby enabling evaluation of model generalizability across different healthcare systems and demographic subgroups.

Model performance is comprehensively assessed using multiple metrics, including the concordance index, time-dependent area under the receiver operating characteristic curve, area under the precision-recall curve, Brier score, calibration analysis, and decision curve analysis. Risk stratification capability is evaluated using Kaplan-Meier survival analysis.

To enhance interpretability, complementary representation-based and attention-based methods are applied. These include variational autoencoder-derived latent feature analysis, correlation with conventional ECG parameters, and gradient-based visualization techniques to identify waveform regions contributing to model predictions. These approaches aim to ensure that the model captures physiologically meaningful signals associated with myocardial remodeling and cardiac dysfunction.