Federated Machine Learning for Privacy-Preserving Healthcare Risk Prediction

This preprint presents an Explainable Federated Machine Learning framework, called XFML, for privacy-preserving diabetes risk prediction across distributed healthcare institutions. The framework enables multiple hospital nodes to collaboratively develop a predictive model without transferring or centralizing sensitive patient-level data. XFML combines locally trained XGBoost and Random Forest models with a federated stacking meta-learner, adaptive variance-penalized feature importance aggregation, Gaussian differential privacy, and dual-layer explainability using SHAP and LIME. The framework is evaluated using 445,132 records from the 2022 CDC Behavioral Risk Factor Surveillance System, partitioned across five geographically distributed, non-IID hospital nodes. The proposed model achieves an AUC-ROC of 0.8247 and sensitivity of 0.8249, with a minimal performance difference of −0.0005 AUC compared with the centralized baseline. It also provides improved probability calibration, achieving a Brier score of 0.1077 compared with 0.1748 for the centralized model. Explainability analysis identifies general health, age, and body mass index as the most influential diabetes-risk factors, while stability experiments demonstrate highly consistent federated SHAP explanations. The findings demonstrate that privacy preservation, predictive performance, probability calibration, and model interpretability can be jointly achieved in federated healthcare machine learning. This manuscript is deposited as a preprint and has not yet completed formal peer review.