DKGM: A diffusion knowledge graph model for medication recommendation leveraging medical and drug information

Medication recommendation systems aim to assist clinicians by providing effective and personalized treatment options through advanced computational methods. Despite significant progress, existing medical recommendation systems face challenges such as incomplete domain knowledge and noisy data, which limit their reliability and effectiveness. To address these issues, we propose a novel Diffusion Knowledge Graph on Medication (DKGM) recommendation that leverages knowledge graphs and drug-drug interactions. Specifically, our proposed DKGM approach first reconstructs a patient diagnosis-medication knowledge graphs and encodes the entities’ representation from their neighboring nodes. The aggregated encoding is then fed into a diffusion graph model to predict probabilistic patient’s diagnosis-medication connections while managing potential drug-drug interactions. To further refine and personalize recommendations, a collaborative filtering module is incorporated into the probabilistic output of the diffusion module. Consequently, the final recommendation is reliable since only low-risk medical combinations are extracted from the augmented knowledge graphs. Experimental results in the MIMIC-3 dataset demonstrate that the proposed DKGM method outperforms existing graph-based approaches in multiple evaluation metrics, validating its effectiveness and reliability in clinical applications.