Federated Learning for Privacy-Preserving Healthcare Analytics: Addressing Data Security, Model Accuracy, and Interoperability

Abstract

Healthcare data is abundant but siloed: privacy regulation and institutional boundaries prevent the pooling that machine learning needs, so models are trained on narrow single-site data and generalise poorly. Federated learning (FL) offers a route around this by moving the model to the data rather than the data to the model. This study developed and evaluated a federated analytics framework across five simulated hospital nodes for a clinical-outcome prediction task, comparing it against isolated single-site models and a (privacy-violating) centralised baseline, and assessing the accuracy cost of adding differential privacy and the burden of cross-site interoperability harmonisation. The federated model reached an AUROC of 0.911, recovering 98.3% of centralised performance while sharing no raw data, and improved mean single-site AUROC by 9.4 percentage points. Differential privacy at ε = 4.0 cost 1.8 points of AUROC. Interoperability mapping resolved 92.7% of feature-schema mismatches. Federated learning delivers near-centralised accuracy with strong privacy, provided interoperability is actively managed.