Reliability-Aware Adaptive Split Federated Learning with Uncertainty-Guided Personalization for Non-IID ECG Classification in IoMT Networks

The Internet of Medical Things (IoMT) enables distributed electrocardiogram (ECG) monitoring, but statistical and resource heterogeneity complicates collaborative learning. This study proposes Reliability-Aware Adaptive Split Federated Learning (RA-ASFL) for non-independent and non-identically distributed (non-IID) ECG classification. The framework integrates three mechanisms: adaptive split selection based on resource constraints, privacy exposure risk, predictive uncertainty, and split-conditioned reliability; reliability-aware aggregation using client-level learning quality and stability to determine separate contribution weights; and uncertainty-guided global–local personalization. These mechanisms distinguish computation placement, client reliability estimation, and aggregation influence. Experiments used five-class MIT-BIH heartbeat classification and five-superclass multilabel PTB-XL classification under simulated heterogeneous IoMT conditions. On MIT-BIH, RA-ASFL achieved 98.03% accuracy, 97.93% F1-score, and 99.43% ROC-AUC, exceeding resource-only AdaptSFL by 0.55 and 0.58 percentage points in accuracy and F1-score, respectively. On PTB-XL, macro AUROC reached 94.3% and macro F1-score reached 74.6%, compared with 92.8% and 72.3% for AdaptSFL. Ablation results identified reliability-aware aggregation and personalization as the principal contributors to predictive performance, while adaptive splitting provided additional benefits. These findings support coordinated reliability-aware learning for ECG classification under simulated heterogeneous IoMT conditions.