Federated Learning Framework with Prototype Clustering for Personalized Healthcare in IoT Wearables

With the growing adoption of IoT-enabled wearables, the continuous monitoring of physiological signals has become a vital component of personalized healthcare. However, existing frameworks still face significant challenges, including non-IID data distribution across patients and high communication costs in resource-constrained wearables. To address these limitations, we propose a drift-aware prototype Clustering (DAPC)-enabled Federated Learning Framework (FL) for personalized healthcare diagnostics. Initially, data collected from the Medical Information Mart for Intensive Care III (MIMIC-III) raw wearable signals were denoised and normalized, followed by feature extraction from physiological and activity data. Each device then generates a lightweight prototype vector to summarize its data distribution. At the edge server, the prototypes are clustered using cosine similarity to form user groups with similar health patterns. Federated aggregation is performed within each cluster, and drift detection dynamically reassigns the devices during user behavior. Finally, a personalized adapter layer was fine-tuned on each device, and the optimized TinyML model was deployed for real-time lowlatency diagnostics. Experimental results show that the proposed DAPC-FL attains higher results than the existing model FL-Synthetic Minority Oversampling Technique (SMOTE) in terms of accuracy (98 %), precision (98 %), recall (97 %), and F1-score (97 %).