Ultra-low communication overhead federated learning strategy for distributed wearable ECG networks

Abstract

Abstract Federated learning (FL) over wearable electrocardiogram (ECG) networks enables privacy-preserving cardiac monitoring at scale, yet standard FL protocols are incompatible with the severe bandwidth and energy constraints of heterogeneous LoRa IoT deployments. This paper proposes ECG-FedCS, a hardware-aware FL framework that dynamically modulates gradient sparsification intensity based on real-time ECG signal quality, QRS complex morphology, and heart-rate variability indices extracted on-device. Nodes operating in high-quality clinical environments transmit a larger gradient fraction, preserving feature diversity for accurate multi-class cardiac classification; bandwidth-constrained LoRa nodes receive protocol-aware maximum compression via a hardware-aware factor, jointly optimizing diagnostic accuracy and energy efficiency. ECG-FedCS is evaluated on a federated simulation of ten heterogeneous nodes spanning BLE and LoRa wireless links, using patient records from the PhysioNet CinC 2020 dataset across nine cardiac conditions. The proposed framework achieves a classification accuracy of 91.24%, converging faster than all baselines while achieving an 80.8 × per-parameter compression ratio. Per-round energy consumption is substantially reduced, extending LoRa node communication-related battery lifetime by 2.4 × under baseline (idealized) channel conditions; a sensitivity analysis under realistic packet-delivery degradation (Section VI-H) shows this relative multiplier is preserved, and increases further for the most distance-constrained rural nodes as channel conditions degrade. ECG-FedCS simultaneously outperforms all four baselines — FedAvg, FedProx, QSGD, and TopK — in accuracy, convergence speed, and compression ratio, establishing the first FL framework to achieve this Pareto improvement in a multi-protocol wearable cardiac IoT setting. These results demonstrate that ECG-signal guidance enables privacy-preserving, communication-efficient federated cardiac monitoring on resource-constrained remote nodes, advancing arrhythmia detection in low-connectivity clinical environments.