Disease Spectrum-aware and Time-evolving Dependency Learning for Medication Recommendation

Medication recommendation is a core component of clinical decision support, aiming to tailor effective drug combinations for patients based on longitudinal Electronic Health Records (EHRs). Despite the success of deep learning in this domain, existing methods typically suffer from two fundamental limitations: (1) they treat diagnosis codes as orthogonal labels, thereby fragmenting shared pharmacological treatment patterns among related diagnoses within latent therapeutic disease spectra, and (2) they infer disease evolution solely based on a coarse-grained holistic hidden state, thereby obscuring the fine-grained temporal dependencies across visit sequences. To address these gaps, we propose SpecTD-MR, a novel medication recommendation framework for Disease Spectrum-aware and Time-evolving Dependency Learning. Specifically, we design a Disease Spectrum-aware Hypergraph Learning module that constructs a hypergraph initialized with multi-source clinical knowledge. By employing task-guided clustering with contrastive alignment, this module unifies disparate diagnoses into cohesive, spectrum-aware representations. Furthermore, we introduce a Time-evolving Dependency Modeling module that explicitly quantifies the dependency strength of current diseases on historical contexts. By incorporating a Mixture-of-Experts (MoE) mechanism, this module synergizes disease spectrum with time-evolving dependency to adaptively regulate patient health state transitions. Extensive experiments on real-world datasets demonstrate that SpecTD-MR surpasses state-of-the-art baselines in accuracy, while providing visualizable structural associations to support insights into dynamic disease evolution. 1