Federated Learning for Privacy-Preserving Healthcare Data Analytics

Abstract

The proliferation of electronic health records (EHR) and medical imaging data offers unprecedented opportunities for developing predictive healthcare models using machine learning. However, stringent data privacy regulations such as HIPAA and GDPR prohibit the centralized aggregation of sensitive patient data across healthcare institutions. Federated learning (FL) addresses this challenge by enabling collaborative model training without sharing raw data. This paper proposes FedHealth, a federated learning framework enhanced with differential privacy and adaptive aggregation for privacy-preserving healthcare analytics. The framework incorporates a novel client-adaptive weighting scheme that accounts for non-IID data distributions typical in multi-hospital settings and a gradient compression mechanism that reduces communication overhead by 42% compared to standard FedAvg. Experiments on two healthcare tasks—EHR-based disease prediction and chest X-ray classification—demonstrate that FedHealth achieves 93.2% and 91.0% accuracy respectively under a privacy budget of ε = 5.0, within 2% of centralized training performance. The framework converges in 28 communication rounds, 38% faster than FedAvg, while providing formal differential privacy guarantees. These results establish FedHealth as a viable approach for multi-institutional healthcare AI collaboration that respects patient privacy.