Quantifying energy and accuracy trade-offs of federated learning on wearable health devices.

The rapid development of wearable health tools has made it possible to continuously monitor physiological conditions for preventive care. However, stringent privacy laws, including HIPAA and GDPR, require decentralized methods such as federated learning (FL) to safeguard personal patient information. Nonetheless, empirical profiling in this paper finds that typical FL implementations are plagued by a serious performance trilemma; a naive federated model attains a 35.3 percent energy savings (3.84 vs. 5.93 kJ in the centralized models), but at the cost of a disastrous performance penalty of 13.87 percentage points (84.94 vs. 98.81 percent in centralized models). The failure in research is largely due to the on-device computational load of 4.24 MFLOPs per training sample, resulting in a "straggler" bottleneck that increases the total training duration to 1,066.26 s, almost 70 times longer than centralized training. As a result, the introduction of the hybrid hierarchical federated split learning (H-FedSL) architecture helps in strategically splitting the neural network at a cut layer to divide the workload between wearable and nearby edge servers. The methodology provides a new framework that offloads the heavy and deep-layer computations to the edge server, leaving the shallow feature extraction to the point of operation, and sends only privacy-sensitive abstractions of the smashed data, rather than raw signals. The integration of asynchronous protocols will help manage device heterogeneity and resource-aware client selection, thereby achieving the aim of H-FedSL to restore the gold-standard accuracy of 98.81% with the state-of-the-art 35.3% energy efficiency of the federated model. Thus, a technically and economically feasible pathway will be provided for deploying medical-grade AI on resource-constrained Internet of Medical Things (IoMT) devices.