Blockchain-coordinated Federated Learning with Label-Aware Proximal Aggregation for Robust ECG Classification
Selma Aouamria, Djalila Boughareb, Mohamed Nemissi, Hamid SeridiIntroduction/Objective:
Privacy preservation and trustworthy coordination remain major challenges in distributed healthcare analytics, particularly for ECG arrhythmia classification. Although Federated Learning (FL) enables collaborative model training without sharing raw patient data, its performance degrades under non-IID data distributions and lacks transparent coordination mechanisms. This study proposes a trustworthy, heterogeneity-aware federated framework for collaborative ECG classification that enhances training stability while enabling verifiable inter-institutional coordination.
Methods:
We introduce FedLA-Prox, an aggregation strategy that integrates label-aware aggregation with proximal regularization to jointly mitigate class imbalance and client drift. The framework combines a 1D-CNN classifier with blockchain-based smart contract orchestration and InterPlanetary File System (IPFS) for off-chain storage to enable verifiable, auditable, and coordinated model exchange. Experiments were conducted on the MIT-BIH Arrhythmia dataset under Dirichlet-based non-IID settings with varying levels of heterogeneity.
Results:
FedLA-Prox showed its clearest advantage under heterogeneous settings, recording the highest F1-scores at α = 0.1, α = 0.3, and α = 0.5, namely 0.8476, 0.8831, and 0.9298, respectively. Its strongest result was obtained at α = 0.5, where it reached 94.83% accuracy and a 92.98% F1-score at Round 30.
Discussion:
In near-IID regimes, namely α = 0.7 and α = 10, the performance gap narrowed substantially, and FedAvg remained the strongest baseline. However, FedLA-Prox continued to show competitive performance, suggesting that the proposed strategy is most beneficial when client data are heterogeneous and class distributions are imbalanced.
Conclusion:
By jointly addressing statistical heterogeneity and class imbalance within a blockchain-coordinated FL framework, FedLA-Prox improves training stability and classsensitive performance, particularly under heterogeneous client data distributions. The proposed system supports collaborative ECG analytics while reducing exposure of raw data and enabling verifiable coordination in healthcare-oriented federated settings.