Federated Learning with Differential Privacy for Secure Multi-Institutional Healthcare Data Sharing

Abstract

Sharing patient data across hospitals is essential for training robust clinical prediction models, yet privacy regulations and institutional barriers prevent centralized data aggregation. We present FedHealth-DP, a federated learning framework that combines secure aggregation with calibrated (ε, δ)-differential privacy to enable collaborative model training on electronic health records (EHR) from 12 hospitals spanning three countries. FedHealth-DP employs adaptive gradient clipping, per-client privacy budget allocation based on data sensitivity, and a novel hospital-specific layer normalization scheme that mitigates non-IID distribution effects. On MIMIC-IV mortality prediction and eICU sepsis detection tasks, FedHealth-DP achieves AUC of 0.891 and 0.876 respectively — within 1.8% of centralized training — while guaranteeing ε = 3.2 differential privacy. Privacy audit simulations confirm zero successful membership inference attacks across 10,000 adversarial queries.