TSAC-DRM: A Drug Recommendation Model Based on Two-Stage Attention Mechanism and Constraint Learning
Shaofu Lin, Ziqian Qiao, Jianhui Chen · 2024
Current drug recommendation systems primarily rely on data-driven approaches, which often overlook the importance of incorporating prior knowledge and capturing temporal dependencies in patient histories. To address these limitations, we propose TSAC-DRM, an innovative drug recommendation model that integrates a two-stage attention mechanism and constraint learning. The two-stage attention mechanism, adapted from the Crossformer, focuses on learning patient historical records and exploring implicit temporal dependencies. This allows the model to better understand the patient's condition and provide more accurate recommendations. Furthermore, TSAC-DRM formalizes drug contraindication relationships as constraints in the loss function, balancing the benefits of drugs for diseases and the risks of contraindications among comorbidities. By integrating prior knowledge through constraint learning, the model ensures the safety and rationality of the recommendations. Experiments conducted on the MIMIC-III dataset demonstrate the superiority of TSAC-DRM compared to traditional recommendation models. The results highlight the advantages of combining the two-stage attention mechanism and constraint learning in drug recommendation tasks, offering new perspectives for developing intelligent decision support systems in healthcare.