Knowledge enhanced representation learning network for drug recommendation

Drug recommendation systems have attracted considerable attention within medical healthcare, which aim to deliver personalized and efficacious drug combinations tailored to patients’ clinical records. Through extensive investigation, we identify two key issues with existing methods: (1) class imbalance distribution, where common diseases occur more frequently than rare ones, resulting in biased and insufficient patient representations; and (2) inadequate modelling of historical medications, where the historical drugs may contain drug information that is valuable for current medical treatment is often overlooked. In this paper, we propose a Knowledge Enhanced Representation Learning (KERL) network for drug recommendation. To address the first issue, we introduce external medical knowledge, using disease entities of different granularities to enhance patient representation. Meanwhile, we construct multiple medical knowledge graphs based on extracted entities and design a graph knowledge enhancement mechanism to integrate global clinical information, alleviating the imbalanced distribution of medical entities. To address the second issue, we design a dual-path drug representation network to model longitudinal historical information from both visit-level and drug-level perspectives. Extensive experiments on two real-world datasets MIMIC-III and MIMIC-IV demonstrate the effectiveness of the proposed KERL in drug recommendation task. Specifically, our KERL achieves improvements of 2.15%, 2.09%, 1.92% and 2.09%, 2.74%, 1.96% over current state-of-the-art methods in terms of F1-score, PRAUC, and Jaccard, respectively.