Federated Edge Intelligence in Healthcare Wearable Systems

Abstract

ract
Federated edge intelligence represents a paradigm shift in the way continuous health data from wearable systems is collected, processed, and analyzed. Rather than relying on central servers to aggregate raw sensor streams—raising privacy, latency, and bandwidth concerns—federated learning pushes model training to the edge: directly on devices or local edge nodes that possess the data. These models then share only weight updates or gradients with a coordinating server, eliminating the need to transfer sensitive personal health signals over the network. In this study, we design and implement a federated edge intelligence framework tailored for healthcare wearable systems, focusing specifically on arrhythmia detection from ECG streams. Our contributions include: (1) a lightweight convolutional neural network (CNN) architecture optimized for resource-constrained devices; (2) three federated strategies—standard FedAvg, hierarchical aggregation via edge servers, and a transfer-learning–based FedHealth variant; and (3) an end-to-end evaluation on emulated Raspberry Pi nodes using the MIT-BIH Arrhythmia Database. Through systematic experiments, we demonstrate that all federated approaches achieve classification accuracies within 1.6% of a centralized baseline, while reducing communication overhead by over 60% and lowering per-round latency by up to 35%. Moreover, hierarchical aggregation mitigates straggler effects, and transfer learning accelerates convergence under non-IID data distributions. These results underscore the promise of federated edge intelligence in enabling privacy-preserving, low-latency analytics for wearable health monitoring. We discuss practical deployment considerations, including device heterogeneity, intermittent connectivity, and regulatory compliance, and outline avenues for future research in adaptive personalization, secure aggregation, and clinical validation.