Personalized Federated Learning Improves ECG Classification Accuracy

Federated learning (FL) allows multiple institutions to collaboratively train AI models without directly sharing sensitive data, which is crucial in healthcare. However, the inherent variability in medical data, such as electrocardiogram (ECG) readings across different providers, poses significant challenges for model robustness and accuracy. A new personalized federated learning method, FedDualAtt, addresses this by modifying transformer attention heads. It splits these heads into two branches: global heads, which are aggregated across all participating clients to capture universal patterns, and local heads, which remain specific to each client to adapt to unique institutional characteristics and data variations. Evaluated on the FedCVD benchmark for cardiovascular disease detection, FedDualAtt demonstrated superior performance in ECG classification compared to both standard and personalized federated learning approaches. The study also revealed that the optimal balance between global and local attention varies, indicating that different clients benefit from tailored levels of architectural personalization.