HAMNet: Hierarchical Attention-enhanced Molecular Hypergraph Network for medication recommendation with data imbalance mitigation

AI-driven medication recommendation systems face critical limitations that hinder real-world applicability, including severe data imbalance in Electronic Health Records (EHRs), insufficient modeling of multi-entity clinical relationships, and inadequate incorporation of molecular pharmacology and hierarchical medical knowledge. We propose HAMNet (Hierarchical Attention-enhanced Molecular Hypergraph Network), a framework that introduces four key innovations: (1) hierarchical attention-based pretraining leveraging ICD and ATC taxonomies to transfer knowledge from frequent to rare medical codes, (2) a dual molecular hypergraph representation that captures both functional and topological properties of drug structures, (3) a bilateral hierarchical heterogeneous hypergraph that unifies clinical, molecular, and patient-level attributes, and (4) safety-aware hypergraph convolutions integrating drug–drug interaction constraints. Extensive experiments on MIMIC-III and MIMIC-IV demonstrate that HAMNet surpasses state-of-the-art baselines, achieving 15.2% improvement in F1-score, 12.8% gain in PRAUC, 5.3% improvement in Jaccard index, and a 7.4% reduction in drug–drug interaction rates. Moreover, HAMNet yields over 40% improvement for rare medical codes, effectively addressing clinical data imbalance. These results indicate that HAMNet offers a robust and safety-aware computational foundation for future clinical decision support applications.