Mixture of experts based medication recommendation for multimorbidity

Medication combination recommendation aims to assist clinicians in providing rational treatment decisions based on patient clinical information, which remains challenging and error-prone, especially in multimorbidity scenarios. However, existing approaches often fail to generate comprehensive recommendations that integrate insights from multiple expert perspectives. To address this limitation, we propose Rx-Expert, a Mixture-of-Experts-based intelligent recommendation system designed to deliver accurate and interpretable medication combination suggestions. Rx-Expert consists of two complementary models: a coarse-grained model that simulates the real-world medical consultation process by training specialized expert networks for distinct recommendation tasks, and a fine-grained model that incorporates molecular representations and known drug-drug interaction information to support drug-level recommendation. Moreover, a Cross Fusion Module is introduced to integrate global, substructure, and textual molecular features, ensuring a holistic understanding of medication properties. Experimental results on the MIMIC-III, MIMIC-IV and eICU datasets demonstrate that our dual-track framework consistently outperforms existing baselines across all evaluation metrics. This work presents a practical and interpretable expert system that enhances clinical decision-making, particularly for patients with complex multimorbidity, by combining patient-specific routing, molecular drug representations, and known interaction constraints.