Multi-Visit Interactive Recalibration Network for Drug Recommendation with a Triple Graph Encoder

Electronic health records (EHRs) comprehensively describe the health status of many patients during their visits. Combining the records in EHRs with the patient's current medical treatment status can generate personalized medication combinations. However, the increasing number of drugs poses significant challenges for clinical experts to recommend combination drugs. In recent years, deep learning models have been widely studied and applied in drug recommendation task. Currently, the existing models either lack sufficient mining of patients' health data or ignore the modelling of patients' longitudinal medical information. Therefore, we propose a multi-visit interactive recalibration network (MIRNet) for drug recommendation with a triple graph encoder. Specifically, we design a medical recalibration module to capture the feature representations in patient diagnosis and procedure information through cascaded convolutions. To achieve friendly interaction of medical codes between relatively necessary historical visits and the current visit in the drug recommendation process, we propose a multi-visit filter module. Furthermore, we design a triple graph encoder to fuse molecular, EHR, and Drug-Drug Interation (DDI) graphs, which aims to extract implicit drug feature representations from different medical knowledge. We perform experiments on the real-world MIMIC-III dataset, and the experimental results reveal that the model MIRNet outperforms other competitive baselines regarding major indicators.