Federated learning approach to optimize coronary revascularization decisions with privacy-preserving data integration from multiple hospitals

Abstract Background Coronary revascularization has a crucial role in the treatment of acute coronary syndrome and significantly affects patient mortality. We aimed to develop a federated learning model to predict the receipt of coronary revascularization within 1 week after an emergency room visit among patients who obtained ECG and cardiac biomarker testing (troponin and CK-MB), while preserving privacy using multicenter data. We conducted a retrospective analysis using de-identified data from three hospitals in South Korea (Korea University Anam Hospital, Ansan Hospital, and Guro Hospital). Methods This study included patients who underwent electrocardiography, troponin, and CK-MB tests, and excluded those who were discharged, deceased, transferred, or had trauma. The data were divided into training (January 2017 to July 2020) and test (August 2020 to July 2021) sets for model development and evaluation. We developed models using local, centralized, and federated learning approaches. To enhance privacy, we added Gaussian noise (SD = 0.01) to each site’s parameters before server aggregation. Model performance was assessed using the area under the receiver operating characteristic curve, the area under the precision-recall curve, precision, recall, and accuracy. Clinical usefulness was evaluated using Cohen’s kappa, and the model predictions were compared with physicians’ decisions regarding coronary angiography. Results This study included 26,906 cases from Anam, 27,194 cases from Ansan, and 21,397 cases from Guro, with revascularization rates ranging from 2.6% to 3.4%. Both the federated learning and centralized learning models outperformed the local learning model, achieving an AUROC of 0.95 and an AUPRC of 0.64. Furthermore, the federated learning model demonstrated robust performance even with added noise and showed moderate-to-substantial agreement with physicians’ decisions to perform coronary angiography. Conclusions Federated learning using multi-center emergency room improved discrimination for predicting receipt of coronary revascularization among patients undergoing ECG and cardiac biomarker testing, while avoiding direct data sharing. This approach may enable scalable decision support that reduces unnecessary coronary angiography while maintaining identification of patients likely to require intervention.