A Medication Recommendation Method Based on Multi-View Medication Representation and a Time-Aware Copy Mechanism

Abstract

Electronic health records provide an important data foundation for intelligent medication assistance. Although existing medication recommendation methods have achieved notable progress in recommendation accuracy, they still suffer from insufficient medication representation learning and coarse-grained utilization of historical prescriptions. To address these limitations, this paper proposes a medication recommendation method based on multi-view medication representation and a time-aware copy mechanism. At the medication representation level, information from the EHR co-occurrence graph, drug--drug interaction graph, and drug molecular structure graph is integrated to construct a multi-view medication representation, thereby enhancing the structural and semantic expressiveness of medication embeddings. At the level of historical information utilization, a time-aware copy mechanism is introduced, which jointly considers temporal decay and diagnosis similarity during historical prescription reuse, so as to improve the specificity and reliability of historical prescription selection. Experiments conducted on the public MIMIC-III dataset demonstrate that the proposed method outperforms multiple baseline models in terms of Jaccard, F1, and PR-AUC, while also achieving better safety performance on DDI Rate. Furthermore, ablation studies verify the effectiveness of both the multi-view medication representation module and the time-aware copy mechanism. The results indicate that the proposed method can improve recommendation accuracy while simultaneously accounting for medication safety, thus providing an effective solution for intelligent medication assistance.