MoTRANet: Safety-Aware Drug Recommendation via Molecular Representation Learning and Dynamic Patient Modeling

With the rapid development of artificial intelligence in healthcare, medication recommendation systems have become important tools for achieving personalized treatment. However, existing methods still have limitations in modeling drug molecular functions and perceiving dynamic changes in patient conditions. To address these issues, this paper proposes MoTRANet, a safety-aware combinatorial medication recommendation model. The model designs a dynamic patient modeling module that integrates GRU, Transformer, and diagnosis-procedure cross attention to effectively capture the dynamic evolution of patient health status. On the drug side, an ECFP-enhanced graph attention network and LightGCN-based motif propagation are employed to construct drug representations with molecular functional information. Furthermore, a DDI-aware encoder and a curriculum learning strategy are introduced to dynamically balance recommendation accuracy and medication safety during training. Experimental results on two public datasets, MIMIC III and MIMIC-IV, demonstrate that MoTRANet outperforms existing mainstream baseline methods in terms of Jaccard, F1, and PRAUC, while achieving a relatively low DDI rate, validating the model's accuracy and safety. Future work will focus on multi center real-world data validation, dose sensitivity analysis, and model interpretability.