Federated Learning for Early Cardiac Anomaly Prediction in Cross-Silo IoMT Environments

Abstract

Early detection of cardiovascular anomalies remains critical for proactive patient care, especially within the growing ecosystem of Internet of Medical Things (IoMT) devices. This study explores the application of Federated Learning (FL) to predict early cardiac events using electrocardiogram (ECG) signals across heterogeneous IoMT silos without centralized data sharing. We focus on Premature Ventricular Contraction (PVC) as an example of early event prediction. Using three realworld ECG datasets (PTB-XL, Chapman-Shaoxing, and MITBIH), we simulate cross-silo environments where local models are trained independently and aggregated through FL. Our experiments demonstrate that local models can already achieve high classification performance, but global models obtained via FL lead to consistent improvements in macro precision, recall, and F1-scores across datasets. Visual analysis of early ECG segments further highlights inter-dataset variability, emphasizing the importance of silo-specific characteristics. The results validate that FL is a promising strategy to enable scalable, privacypreserving, and accurate early cardiovascular event prediction in IoMT systems, bridging clinical silos while safeguarding sensitive patient data.