Performance of federated learning models in health services research: A systematic review and meta-analysis

This study aimed to assess the performance of federated learning (FL) models and compare their performance with local and centralized models.

We conducted a systematic search of Ovid MEDLINE and PubMed from inception to June 10, 2025, to identify studies using patient data to train or validate FL algorithms and reporting at least one model performance outcome. Two reviewers independently screened articles and extracted data on study characteristics, FL frameworks and model training methodologies, and reported performance metrics. We summarized model performance using medians and interquartile ranges for federated, local, and centralized models. Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Diagnostic Test Accuracy Studies was followed.

Thirteen studies involving 247 sites and 158,435 samples were included, with eight studies contributing to this meta-analysis. Most FL models (85%) used the Federated Averaging (FedAvg) algorithm for parameter aggregation across sites. FL model performance metrics were compared with local and centralized models within each machine learning task in each study. FL showed notable gains over local models, improving the AUC by 8.2%, F1 score by 7.9%, sensitivity by 4.1%, PPV by 26.1%, and PRAUC by 1.6%. Compared with centralized models, FL showed modest losses of 4.1% in AUC, 9.9% in F1 score, 9.3% in sensitivity, and 2.8% in PRAUC.

FL outperformed local models and demonstrated comparable performance to centralized models while preserving data privacy. Standardized reporting and improved methodological transparency are needed to support its broader application in health research.