Privacy Preserving Federated Learning for Myocardial Infarction Detection from ECG Signals

Abstract

Myocardial infarction (MI) detection from 12-lead ECG signals is a critical clinical task, yet two key barriers limit the deployment of deep learning models: data privacy regulations that restrict cross-institutional patient data sharing, and severe class imbalance across heterogeneous hospital populations. This paper proposes a privacy-preserving federated learning framework that addresses both challenges simultaneously. The framework integrates a novel MultimodalMITransformer architecture that encodes 12-lead ECG signals using temporal patch embeddings and a three-layer Transformer encoder, and fuses the resulting global ECG representation with patient clinical metadata (age, sex, height, weight) via late fusion for joint MI classification. Training is coordinated across five simulated hospital clients, each with clinically heterogeneous, non-IID data distributions, using a weighted Federated Averaging (FedAvg) protocol, without sharing any raw patient data. To systematically combat class imbalance under non-IID federated conditions, we propose a quadruple mitigation strategy comprising: (i) per-client positive-class weighting, (ii) balanced mini-batch oversampling, (iii) adaptive focal loss, and (iv) per-client optimal decision threshold calibration derived from the validation precision-recall curve. Experimental evaluation on the PTB-XL dataset demonstrates that the proposed framework achieves an average AUROC of 0.8847, an accuracy of 0.8310, and a recall of 0.7266 among five non-IID clients, representing an absolute improvement in F1 of 21.8% over the baseline of FedAvg and closing 63% of the gap to a fully centralized model, while preserving complete data privacy. The results demonstrate that principled, per-client imbalance correction within a federated setting can substantially recover performance lost to data heterogeneity, offering a clinically viable and privacy-compliant pathway for collaborative multi-site cardiac AI.