FedMed: Federated Learning-Based Personalized and Safe Medication Recommendation

Medication recommendation for patients with multimorbidity is an essential task for AI in healthcare. Existing works rely on centralized storage of all patient clinical records (e.g., diagnoses and procedures) to develop a centrally managed medication recommendation model. However, this setting exposes sensitive patient data to privacy risks and potential leakage. To address this issue, we propose FedMed, a privacy-preserving, personalized, and safe medication recommender based on federated learning. In FedMed, each patient client develops a prediction module using its private clinical records for medication recommendation, and a self-supervised strategy is further introduced to alleviate record data sparsity. On the server side, the global server performs a personalized graph-guided aggregation mechanism for updating and synchronizing client public parameters without accessing patient private information, thereby preserving privacy. In addition, our FedMed mitigates adverse interactions between medications, known as drug-drug interaction (DDI), by incorporating DDI regularization at the client and server sides to promote predicted medication safety. The results on the two Electronic Health Records (EHR) datasets show the effectiveness of our federated medication recommender FedMed.