A Federated Learning Approach for Predicting Heart Disease Risk

Deep learning has demonstrated significant potential in utilizing electrocardiogram (ECG) signals for arrhythmia prediction. However, the sensitivity of ECG data often results in its storage as isolated 'data islands' across different medical institutions. Federated learning techniques have been employed for cross-institutional collaborative learning; however, practical challenges such as imbalanced ECG data categories and the unequal allocation of client sample weights during server aggregation can lead to biased calculations and suboptimal model performance. To address these challenges and improve the identification of high-risk patients, we propose a novel heart disease risk prediction algorithm based on federated learning. This algorithm incorporates patient-centered adjustments on the client side through cost-sensitive learning, which mitigates internal class bias. On the server side, it optimizes federated aggregation weights by considering both sample size and quality. The algorithm was calibrated and evaluated according to the AAMI EC57 standard using the MIT-BIH arrhythmia database, and the results show improved accuracy in handling imbalanced data, facilitating efficient cross-institutional learning, and enhancing patient-centered heart disease risk identification.