RES-MR: Risk-Aware Reasoning for Explainable and Safe Medication Recommendation

This is the official implementation of RES-MR: Risk-Aware Reasoning for Explainable and Safe Medication Recommendation. This repository provides the complete research pipeline, including data processing for MIMIC-III, MIMIC-IV, and eICU , prompt templates for clinical reasoning distillation , and the training code for our two-stage paradigm (SFT & RL).

RES-MR addresses two critical challenges in medication recommendation: the lack of explainability in "black-box" models, which often results in clinician trust issues due to unverifiable diagnostic logic , and the neglect of personalized safety, where reliance on static, patient-agnostic constraints can lead to adverse outcomes by failing to account for individual risk tolerance. By integrating retrieval-augmented reasoning with adaptive safety boundaries, our framework ensures that recommendations are both clinically transparent and tailored to the unique physiological conditions of each patient.

• We propose RES-MR, a novel framework for explainable and safe medication recommendation using LLMs, which integrates reasoning generation and safety alignment into a two-stage paradigm.

• In Stage 1, we propose explainable clinical reasoning distillation to ensure transparent decision-making by integrating a knowledge graph retrieval module with a reasoning distillation strategy.

• In Stage 2, we devise risk-aware policy optimization to achieve personalized risk-benefit trade-offs through a dual-factor disentanglement module and patient prototypes that quantify personalized risk boundaries.

• Extensive experiments on benchmark datasets demonstrate that RES-MR consistently outperforms state-of-the-art baselines across recommendation accuracy, safety, and explainability metrics.

Clone the repository and install dependencies:

# Clone the repository
cd RES-MR
# Install dependencies
pip install -r requirements.txt

1. You must have obtained access to MIMIC-III, MIMIC-IV and eICU databases before running the code.

2. Download the MIMIC-III, MIMIC-IV and eICU datasets, then unzip and put them in the data/input/ directory.

3. Download the drugbank_drugs_info.csv, drugbank2name.csv, ndc2db_all.csv, and ddi_data_all.csv files from previous works, and place them in the data/input/ directory.

4. Run the data processing scripts:

   # Process the dataset
   python data/mimic3.py

5. Reasoning Data Synthesis:

   # Due to anonymity principles and repository size constraints, our synthesized reasoning data 
   # has been uploaded to Google Drive and will be made public in the next step.
   
   python data/LLM/ehrsummary.py
   python data/LLM/reasoning_synthesis.py

6. GraphRAG Knowledge Graph Index Construction:

   # Ensure Neo4j is installed and the medical knowledge graph PrimeKG is downloaded.
   # Configure your connection settings and run the following script to build the GraphRAG index:
   python data/GraphRAG/build_kg.py
   
   # Run Hybrid Retrieval
   python data/GraphRAG/hybrid_retriever.py

7. Build Pre-training Data for Dual-Factor Disentanglement:

   python data/Disent/build_dataset.py
   python data/Disent/build_embeddings.py

1. Dual-Factor Disentanglement (DFD) Pre-training:

   cd src/DFD
   bash train.sh

2. Stage 1: Explainable Clinical Reasoning Distillation (ECRD) Fine-tuning:

   cd src/ECRD
   bash train.sh

3. Stage 2: Risk-Aware Policy Optimization (RAPO) Model RL Training:

   cd src/RAPO
   # Build drug risk cache
   python build_cache.py
   # Start training
   bash train.sh

4. Inference:

   cd src/RAPO
   bash infer_vllm.sh

If you find this work useful for your research, please cite our paper:

@inproceedings{wang2026resmr,
  author    = {Wang, Cong and Li, Jin and Wang, Shoujin and Li, Yishuo and Gu, Huilin and Lu, Wenpeng},
  title     = {{RES-MR}: {Risk-Aware Reasoning for Explainable and Safe Medication Recommendation}},
  booktitle = {Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '26)},
  year      = {2026},
  month     = {July 20-24},
  address   = {Melbourne, VIC, Australia},
  publisher = {ACM},
  pages     = {12},
  doi       = {10.1145/3805712.3809604}
}