Intelligent Medication Recommendation via Dynamic Prescription Modeling and Molecular Substructure Learning

Medication recommendation system is a critical application of artificial intelligence in healthcare, supporting clinicians in prescribing effective and safe drug combinations. To achieve more comprehensive recommendation, further research is required in three key areas: (1) Identifying more detailed and relevant drug substructure information to optimize drug combinations, as specific drug substructures impact therapeutic effects and side effects. (2) Capturing historical prescription dynamics to reflect patient condition progression over time. (3) Integrating correlations among medical codes to provide a comprehensive representation of patient conditions, given that certain diagnoses often co-occur. This paper proposes dynamic medication recommendation system DynMedRec, a computational model for medication recommendation that prioritizes prescription dynamics and structured drug molecule information. DynMedRec captures fine-grained drug information by modeling the interactions between substructures and employing a clinical-context query mechanism to generate adaptive molecular representations. Additionally, it integrates structural correlations among medical codes to enhance visit-level patient representations. To preserve long-term longitudinal dependencies, an enhanced recurrent neural network is proposed to model dynamic changes in historical prescriptions. DynMedRec was trained on the MIMIC-III dataset using a visit-by-visit training approach. Extensive experiments and in-depth analyses further demonstrate that DynMedRec outperforms existing state-of-the-art methods across diverse scenarios.