Stud Health Technol Inform. 2026 May 7;335:236-241. doi: 10.3233/SHTI260090.
ABSTRACT
BACKGROUND: Machine learning models for predicting major adverse cardiovascular events (MACE) often generalize poorly across populations, and multinational development is limited by data-sharing constraints.
OBJECTIVES: We investigate whether federated learning (FL) can reduce the generalization gap of MACE prediction models across international clinical cohorts while preserving data privacy.
METHODS: Using harmonized electronic medical record (EMR) data from Austria, Brazil, and the USA, we train federated and local XGBoost and multilayer perceptron (MLP) models and evaluate performance using AUROC.
RESULTS: Our preliminary results show that the performance of local models degrades substantially on external cohorts, particularly when trained on smaller datasets. FL reduces this gap, with the greatest gains observed when compared to models trained on smaller cohorts and evaluated on the largest cohort. Local models performed best in-country, and XGBoost consistently outperformed MLPs.
CONCLUSION: Federated learning improves cross-site generalizability of MACE prediction models, with trade-offs between global robustness and local performance.
PMID:42119126 | DOI:10.3233/SHTI260090
