Medication recommendations are designed to provide physicians and patients with personalized, accurate and safe medication choices to maximize patient outcomes. Although significant progress has been made in related research, three major challenges remain: inadequate modeling of patients' multidimensional and time-series information, insufficient representation of medication substructures, and poor balance between model accuracy and drug-drug interactions. To address these issues , a safe medication recommendation model SDRBT based on patient deep spatio-temporal encoding and medication substructure mapping is proposed in this paper. SDRBT has developed a patient deep temporal and spatial coding module, which combines symptom information, disease diagnosis information, and treatment information from the patient's electronic health record data. It innovatively utilizes the Block Recurrent Transformer to model longitudinal temporal information of patients in different dimensions to obtain the horizontal representation of the patient's current visit. A dual-domain mapping module for medication substructures is designed to perform global and local mapping of medications, fully learning and aggregating medication substructure representations. Finally, a PID LOSS control unit was designed, in which we studied a drug interaction control module based on the similarity calculation between the electronic health map and the drug interaction graph. This module ensures the safety of the recommended medication combination effectively improved the recommendation efficiency and reduced the model training time. Experiments on the public MIMIC-III dataset demonstrate SDRBT's superior accuracy in medication recommendation.
Keywords: Block Recurrent Transformer; Data mining; Electronic health record; Graph neural network; Medication combination recommendation.
Copyright © 2025 Elsevier Inc. All rights reserved.
