Medication recommendation stops at a drug list: RES-MR writes the diagnostic reasoning first, then prescribes inside each patient's own risk boundary

Medication recommendation moves from emitting a drug list to first emitting a step-by-step diagnostic reasoning chain a physician can verify, and then the drug list. Earlier discriminative graph models and generative LLM methods follow a direct-mapping paradigm, and attention or aspect-based alignment only provides post-hoc feature attribution, showing which clinical features contributed but not how the conclusion was deduced. On MIMIC-III, MIMIC-IV and eICU, RES-MR-8B leads all baselines on Jaccard, F1, DDI, RIS and BLEURT, with a 7.7% Jaccard improvement over the state-of-the-art FLAME on MIMIC-III.

The safety constraint moves from a fixed threshold shared by all patients to a risk boundary calibrated to individual clinical status. Most methods enforce safety through static global constraints such as DDI penalties in the loss or molecular-structure-informed representations, without accounting for heterogeneity in patient-specific risk tolerance, so recommendations end up overly conservative for robust patients and potentially risky for vulnerable ones. RES-MR attains the lowest RIS and DDI rates on all three datasets while keeping higher accuracy, and in the ablation, removing RAPO markedly raises the DDI rate, indicating that supervised learning alone does not internalize a patient-specific risk boundary.

The model first learns to explain, then learns to respect the boundary, in two separately trained stages. Earlier work either mixes reasoning generation with safety alignment or applies general-domain preference alignment, lacking explicit modeling of clinical contraindications and individual risk tolerance. Stage one fine-tunes with LoRA on patient-specific evidence retrieved from a knowledge graph and on reasoning trajectories synthesized by GPT-5, and stage two builds patient prototypes from efficacy and risk factors produced by dual-factor disentanglement, then optimizes with adaptive reward shaping and reward clipping.

The method targets medication recommendation over electronic health records and is validated on three critical-care datasets, MIMIC-III, MIMIC-IV and eICU, with structured diagnosis and procedure codes, clinical note text, and knowledge retrieved from PrimeKG as input. Evaluation covers recommendation accuracy, medication safety and explanation quality, with explanation quality judged by GPT-5 and three evaluators with medical or pharmaceutical backgrounds in a blind pairwise comparison on fifty randomly sampled MIMIC-III test instances. The authors propose next to explore agentic reasoning workflows with iterative retrieval and self-verification.

RIS is a proxy metric introduced by the authors to measure how far predicted drugs exceed each patient's risk boundary, because patient-specific risk tolerance is not directly observable from electronic health records. The risk boundary uses historical medications as an empirical proxy for prior risk exposure under clinical supervision, with a parameter controlling strictness. The authors also note that an excessively large efficacy weight diverts the model toward functionally similar rather than exact ground-truth drugs, while extremely high penalties cause an over-conservative policy that refuses prescriptions.