Privacy-Preserving Federated AI for Early Disease Risk Prediction Across Distributed U.S. Healthcare Systems

Privacy law, institutional policy, and re-identification risk make it difficult to pool hospital records for early disease risk prediction. This paper presents a framework in which hospitals train a shared risk-prediction model without exchanging raw records. It combines federated averaging, record-level differential privacy (DP) enforced at each hospital, secure aggregation, and site-specific model calibration, and it is evaluated on four dimensions: predictive accuracy, privacy leakage, performance across hospital populations, and communication overhead. Experiments use a synthetic network of ten hospitals (eight training, two held out) with heterogeneous populations in a simulated three-year type 2 diabetes risk task; the results demonstrate the method and are not findings about real patients. Federated training reached a mean hospital AUC of 0.801, compared with 0.766 for local-only models and 0.811 for centralized pooling, and raised AUC in the smallest hospital from 0.703 to 0.810. With DP, AUC was 0.791 at ε = 4 per hospital (δ = 10⁻⁵), fell to 0.768 at ε = 2, and dropped below local-only performance at ε = 1. DP degraded calibration (expected calibration error 0.082 versus 0.040 without DP), but local recalibration restored it to 0.038, including at two hospitals that never took part in training. A membership-inference attack performed no better than chance against the federated models, whereas local-only models in the smallest hospitals leaked measurably (advantage up to 13.4% against a 6.2% null). Federated traffic was 4.5% of the volume of the raw training data. The results indicate a practical budget near ε = 4 for this task.