Medication mapping and diagnosis enhancement for fine-grained medication recommendation

Medication recommendation aims to assist physicians by recommending personalized drugs based on electronic health records (EHRs). However, existing methods typically recommend a single drug set for multiple diseases without establishing fine-grained correspondences between drugs and individual diseases, which limits interpretability and clinical trust. Moreover, most approaches ignore the severity of diagnosed diseases, preventing medication recommender systems from accurately assessing a patient’s health condition. To address these critical gaps, we propose a novel fine-grained medication recommendation method via medication mapping and diagnosis enhancement, called FineMed. To address the first gap, FineMed decomposes the drug recommendation for a set of diseases into multiple sub-recommendations, each focusing on a single disease, thus enabling the respective recommendation of appropriate drugs for each individual disease. To address the second gap, we design a diagnosis enhancement mechanism that establishes explicit links among various medical elements to reflect disease severity. In extensive retrospective experiments on MIMIC-III, MIMIC-IV, and SMH, FineMed achieves Jaccard scores of 0.3868, 0.3621, and 0.4588, respectively, with an average relative improvement of 3.06% over the strongest competing baselines across Jaccard, F1-score, and PRAUC. FineMed provides fine-grained drug-disease correspondences that offer more explicit justifications for its medication recommendations. It should be noted that the current validation is retrospective and partially dependent on LLM-generated annotations with limited human evaluation, rather than prospective clinical validation. Our code and datasets are available at: https://github.com/liyifo/FineMed .