Dynamic Graph and Multi-view Medication Information for Recommending Medication Combination
Keywords: Medication recommendation, Electronic health record, Drug-drug interaction, Dynamic graph
TL;DR: The DG-MRM Model
Abstract: An accurate and safe medication recommendation system play a vital role to assist medical practitioners in making prescriptions and enhance patient treatment outcomes. Existing approaches often treat electronic health record (EHR) data as sequences in training, failing to capture the complex dependencies between different medical events. Additionally, these systems encounter challenges in mitigating drug-drug interactions (DDIs). In response to these limitations, we propose a novel medication recommendation model named DG-MRM, which constructs dynamic graphs with multi-source EHR data and utilizes multi-view medication information from external knowledge database. Specifically, we leverage dynamic graph neural networks to investigate the temporal and structural relationships between different treatment data in patients’ medical history, aiming to generate comprehensive patient representations with several longitudinal visits. Moreover, we capture internal-view medication molecule structure and functional-view interactions between medication molecules to generate safe medication combinations. Finally, our proposed DG-MRM is extensively evaluated on a benchmark dataset and the results reveal superior performance in terms of efficacy and safety. Comparative analysis with state-of-the-art drug recommendation models, DG-MRM significantly improves the accuracy of drug recommendations while maintains a low risk of DDIs.
Submission Number: 34
Loading