An encrypted dual-attention federated learning framework for personalized electrocardiogram classification

Electrocardiogram (ECG) analysis is fundamental to the diagnosis and monitoring of cardiovascular diseases. In multi-center healthcare settings, collaborative model development is challenged by heterogeneous patient populations and stringent privacy requirements that restrict the sharing of sensitive data and model parameters. This study presents privacy-preserving dual-attention federated learning (PriDA-FL), a framework that integrates personalized federated learning with encrypted parameter aggregation for multi-center ECG classification. The proposed approach employs a transformer adaptor attached to a frozen, pretrained ECG foundation model, with attention heads partitioned into globally shared and client-specific components. This design simultaneously enables local adaptation, controls parameter transmission, and protects privacy by ensuring that client-specific parameters are excluded from server-visible communication channels. The framework is evaluated on the federated cardiovascular disease (FedCVD) benchmark, comprising four clinical ECG datasets under a 20-class multi-label classification setting. PriDA-FL is compared with federated representation learning (FedRep), federated tuning and personalization (FedTP), and low-rank adaptation (LoRA) under a matched eight-million-parameter configuration. PriDA-FL achieves an average per-client Micro-F1 score of 85.80 ± 0.27 % , remaining within 1% of FedRep ( 86.27 ± 0.10 % ) and FedTP ( 84.91 ± 0.21 % ) while outperforming personalization-free baselines by 13%–26%. Adjusting the attention-head ratio yields global Micro-F1 scores ranging from 69.67 ± 1.23 % to 71.96 ± 1.07 % , comparable to FedTP within the seed variance. A 10,000-iteration gradient inversion attack fails to reconstruct ECG waveforms from plaintext global updates. Furthermore, restricting Cheon-Kim-Kim-Song (CKKS) homomorphic encryption to globally shared parameters reduces the encrypted communication payload by 71.5% while increasing per-round computational overhead by only 1.06 times. These results demonstrate that a construction-time partition of transformer attention heads can simultaneously support personalized learning and privacy-preserving parameter exchange, providing an efficient and secure framework for collaborative ECG analytics and clinical decision support across distributed healthcare institutions.