DFHD: dual-granularity fusion network using historical drugs for drug recommendation

Abstract

Drug recommendation is a task in clinical medicine aimed at suggesting a set of safe and effective medications based on a patient’s electronic health records. Current approaches either rely on diagnoses and procedures documented in electronic health records to recommend drug combinations or focus on enhancing drug recommendation safety by considering drug-drug interactions. However, these approaches often overlook the significance of historical medication information in drug recommendation despite its strong correlation with current diagnostic and prescription recommendation. Therefore, we propose a Dual-granularity Fusion Network using Historical Drugs. Specifically, at the time-series modeling level, recurrent neural networks are used to extract time-series features from historical drug data to construct coarse-grained drug characterizations. At the molecular structure modeling level, a graph neural network is used to build a relationship map between drug molecular structures and drug substructures to capture the fine-grained interactions within drug molecules. In addition, we designed a historical drug molecule awareness module to capture historical drug information during drug molecule modeling so as to identify the drugs that really help to cure patients. To effectively integrate dual-granularity information, we design a dual-granularity fusion module to realize the synergistic learning of temporal and structural features. To ensure drug safety, we introduce the DDI loss function to adaptively adjust the loss weights based on the drug interaction risk results, taking into account the optimization goals of efficacy and safety. Our source code is available at https://github.com/AK-321/DFHD.