SDRec: Synergistic Drug Recommendation Network Based on Medical Domain Knowledge

Drug recommendation is a prominent topic in the medical field, aiming to suggest a set of safe drug combinations based on patients’ clinical information and medical knowledge. Existing research has primarily focused on utilizing patients’ electronic health records (EHRs) and avoiding adverse drug-drug interaction (ADDI) for drug recommendation. However, they ignore the influence of drug synergy in the recommendation system. To address this limitation, we propose the synergistic drug recommendation network based on medical domain knowledge (SDRec), which introduces the concept of drug synergy into the field for the first time. Specifically, we design a drug synergy module that extracts drug features from the perspective of molecular structures and models the interaction information between different drug pairs to provide safe and effective drug combinations for patients. Additionally, we incorporate a drug synergy graph derived from medical domain knowledge and model it using a graph convolutional network. Considering the safety of the recommended drug, we introduce a contrastive loss function during training to balance the features of drug synergy and adverse drug reactions, thereby minimizing potential side effects. Our experimental findings indicate that SDRec achieves notable performance enhancements compared to several baseline methods on the MIMIC-III and MIMIC-IV datasets.