HiNet3D: Integrating patient hierarchical aggregation network and molecular 3D-aware for medication recommendation.

Accurate medication recommendation from electronic health records (EHR) is crucial for clinical decision support. However, prior methods often treat EHR as a flat visit sequence and encode drugs with 2D molecular graphs, which limits capturing inadequate patient representation and drug molecular semantics. Therefore, we propose a framework named HiNet3D for the recommendation task. Specifically, HiNet3D builds a patient history graph by representing each visit as a node and performs cross-visit pairwise aggregation to form a layer of virtual nodes, enabling rich cross-visit context modeling to improve downstream medication recommendation. Then, HiNet3D integrates 3D molecular priors and uses a geometric vector perceptron (GVP) to encode atomic coordinates and bonding geometry, producing 3D-aware drug embeddings and building rich 3D structural context through multi-hop aggregation at the molecular level. Finally, the fused patient-drug representations are used to predict the medication set. Experiments on the real-world clinical dataset show that HiNet3D significantly outperforms strong baselines on Jaccard, PR-AUC, F1, and ROC-AUC. HiNet3D delivers scalable and clinically impactful EHR-based medication recommendations by unifying hierarchical patient modeling with 3D molecular structure analysis.