Federated learning with local–global collaboration for predicting acute coronary syndrome

Abstract

Acute Coronary Syndrome (ACS) is a prevalent cardiovascular disease characterized by high incidence and mortality rates. Numerous studies have focused on utilizing artificial intelligence and machine learning algorithms to assess and predict the risk of ACS in patients. However, due to the sensitivity and privacy of medical data, training machine learning models on a centralized server that aggregates ACS data from various institutions poses certain risks. For the first time, this study validates the effectiveness of utilizing federated learning to collaboratively analyze medical data for predicting ACS. A federated learning-based ACS prediction model, i.e., FedLG, which incorporates local–global collaboration for mutual correction, is presented accordingly. On the client side, a regularization term is added to the loss function to reduce deviations caused by heterogeneous data, helping the global model remain accurate and representative. On the server side, gradient normalization is applied to balance contributions from clients with different update frequencies, resulting in a more stable and reliable global model. Comprehensive experiments on the ACS dataset from a tertiary hospital in China show that FedLG consistently outperforms models trained on individual clients, as well as three other federated baselines, across seven evaluation metrics under both IID and non-IID settings. Temporal hold-out validation further indicates that FedLG maintains better generalizability than other baselines. In addition, analysis of feature importance shows that FedLG identifies lipid-related biomarkers, which aligns with clinical knowledge, enhancing the interpretability of the results. The source code of FedLG is freely available at https://github.com/bioinformatics-xu/FedLG.