ASAF-MedRec: Auxiliary-supervised semantic alignment and fusion for medication recommendation

Medication recommendation aims to predict appropriate medication combinations according to a patient’s clinical status. Existing methods have made important progress by modeling longitudinal EHR trajectories, medication relations, external knowledge graphs, molecular structures, and large language model (LLM) knowledge, but the direct semantic matching between textual patient conditions and candidate medication descriptions remains insufficiently explored. This paper investigates current-visit textual semantic matching and proposes ASAF-MedRec, an auxiliary-supervised semantic alignment and fusion model for medication recommendation. ASAF-MedRec maps diagnosis, procedure, and symptom texts and medication description texts into a shared semantic space for bounded multi-label medication set prediction. Specifically, ASAF-MedRec first adapts BioBERT to task-related clinical and medication texts through task-aware pre-training. It then introduces patient–prescription-set contrastive learning with a medication-overlap-aware negative masking mechanism to align patient states with ground-truth prescription sets while reducing false negatives caused by highly overlapping medication combinations. Furthermore, an auxiliary-supervised adaptive semantic fusion module assigns fusion weights according to the auxiliary prediction capability of diagnosis, procedure, and symptom views, and matches the fused patient representation with candidate medication text representations. Experiments on 13,490 and 126,001 visit records from MIMIC-III and MIMIC-IV show that ASAF-MedRec achieves competitive medication set prediction performance, with Jaccard and F1-score values of 0.5468 and 0.6994 on MIMIC-III, and 0.4988 and 0.6512 on MIMIC-IV. Ablation studies further show that set-level semantic alignment, medication text representations, and auxiliary-supervised fusion all contribute to the reported performance.