Graph-Embedding Enhanced Hierarchical Model for Personalized Rheumatoid Arthritis Drug Recommendations

Rheumatoid arthritis (RA) is an incurable chronic disease with low remission and response rates, making effective drug selection essential. We developed a hierarchical drug recommendation system enhanced by a graph embedding module. By integrating patient feature representations predicted through model probabilities with graph embeddings from EHR and DDI matrices processed by a dual-layer Graph Convolutional Network(GCN), the model predicts drug classes and further refines predictions with subclass information. Using data collected from 352 hospitals across China between 2000 and 2023, we established a cohort of 4975 single-diagnosis RA patients who met the remission criteria. Eight subcategories from four first-line RA drug classes, including glucocorticoids, immunosuppressants, specialized immunosuppressants, and NSAIDs, were selected as recommendation labels. Our model outperforms existing methods, improving F1 and AUC scores by 9% and 19%, respectively, and demonstrates strong potential for personalizing RA medication recommendations.