A Secure and Interpretable Federated Learning Framework for Diabetes Prediction with Blockchain-Enabled Security

The rapid advancement of privacy-preserving artificial intelligence in healthcare highlights the need for secure, interpretable, and scalable learning frameworks. To address the limitations of conventional Federated Learning (FL) in heterogeneous clinical environments, we propose the Secure Inter-pretable Federated Learning (SIFL) framework. SIFL integrates advanced FL optimization, blockchain-based verification, and post hoc interpretability to enable trustworthy and decentralized diabetes prediction. We evaluate five FL algorithms: FedAvg, FedProx, FedOpt, Q-FedAvg, and FedNova on two real-world diabetes datasets, the Diabetes Prediction Dataset (DPD) and the Diabetes Health Indicators Dataset (DHID), across client populations ranging from$K = 10$to 100 and under both 50/50 and 90/10 client-server splits. Experimental results show that FedNova within SIFL achieves the highest global accuracy (92.24% on DPD and 91.73% on DHID for the 50/50 split), the fastest execution time per round (