🤖 AI

MediRec: Enhancing Chinese Medication Recommendation with Explainable Clinical Reasoning

MediRec is an explainable large language model framework that leverages clinically grounded reasoning-chain distillation and reinforcement learning to enhance the accuracy and interpretability of medication recommendations for Chinese electronic health records.

Original authors: Juntao Li, Haobin Yuan, Ling Luo, Yuanyuan Sun, Jian Wang, Hongfei Lin

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Juntao Li, Haobin Yuan, Ling Luo, Yuanyuan Sun, Jian Wang, Hongfei Lin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where your doctor has a super-smart assistant that can read your entire medical history in seconds and suggest the perfect medicine for you. This isn't science fiction; it's a field called Medication Recommendation, where computers try to help humans make safer, faster decisions about what drugs to prescribe. For years, these computer assistants have been trained mostly on English medical records from the US, learning to guess broad categories of drugs like "antibiotics" or "painkillers." But real life is messy and specific. Doctors don't just need a category; they need to know exactly which pill to give, why, and how it fits with the patient's unique story. The big question researchers are asking is: Can we build an AI that understands Chinese medical records, picks the exact right medicine, and—most importantly—explains its thinking so doctors can trust it?

Enter MediRec, a new project by researchers at Dalian University of Technology who decided to build a smarter, more transparent AI assistant for Chinese hospitals. They realized that simply asking a powerful AI to guess the right medicine wasn't working well enough. The AI was either guessing wildly or giving answers without any logic, like a student who gets the right answer on a test but can't show their work. To fix this, the team created a two-step training camp for their AI.

First, they taught the AI how to think like a doctor. They used a "teacher" AI to read real patient records and drug manuals, then wrote out step-by-step stories explaining why a certain medicine was chosen. They called this "reasoning-chain distillation." It's like having a master chef write down not just the recipe, but the why behind every ingredient choice: "We add salt here because the patient has low blood pressure, not just because the recipe says so." The student AI learned from these detailed stories, practicing how to summarize a patient's condition, analyze their needs, and explain how different drugs work together.

But learning from stories wasn't enough. In the second step, the team put the AI through a rigorous "reality check" using a technique called Reinforcement Learning. Imagine a video game where the AI gets points for two things: getting the medicine list exactly right (accuracy) and keeping its explanation clear and logical (interpretability). If the AI guessed the right drug but gave a confusing reason, it lost points. If it gave a perfect reason but the wrong drug, it also lost points. The AI played this game thousands of times, tweaking its brain to maximize its score.

The results were impressive. On a test set of real Chinese hospital records, MediRec achieved an F1 score of 0.5813 and a Jaccard score of 0.4626. These numbers might look like gibberish, but in the world of medical AI, they mean the system is getting significantly better at picking the right medicines than previous methods. The researchers found that just asking the AI to "think out loud" without training (a method called prompting) didn't work well; the AI needed that structured training camp to learn the ropes. They also discovered that while the AI was great at explaining its logic, it still had some quirks. Sometimes it suggested too many vitamins or missed safety checks like kidney function, showing that while it's getting smarter, it's not quite ready to replace a human doctor just yet.

What makes MediRec special isn't just that it gets the right answers, but that it shows its work. Unlike other systems that just spit out a list of drugs, MediRec generates a transparent clinical story, explaining how a patient's symptoms led to a specific treatment plan. The researchers suggest that this mix of high accuracy and clear reasoning makes it a much more trustworthy tool for real-world hospitals. However, they admit the system still needs to be refined to catch every safety nuance and avoid over-prescribing. For now, MediRec stands as a promising step toward an AI that doesn't just know the answer, but understands the story behind it.

Technical Summary: MediRec – Enhancing Chinese Medication Recommendation with Explainable Clinical Reasoning

Problem Statement
While Large Language Models (LLMs) show promise for clinical decision support, their application to Chinese medication recommendation remains underexplored. Existing approaches suffer from three primary limitations:

Data Bias: Most methods rely on English datasets (e.g., MIMIC), which fail to account for differences in patient demographics, disease distributions, and regional clinical guidelines specific to China.
Granularity: Current models typically predict coarse-grained drug categories rather than specific, fine-grained medications required for discharge.
Lack of Interpretability: Existing systems often function as "black boxes," predicting medication codes without explicit clinical reasoning processes, thereby offering limited support for transparent decision-making.

Methodology: The MediRec Framework
The authors propose MediRec, a two-stage framework designed to improve both recommendation accuracy and interpretability for Chinese Electronic Health Records (EHRs).

Stage 1: Clinical Reasoning Distillation and Supervised Fine-Tuning (SFT)
- Data Sources: The framework utilizes de-identified EHRs from the CHIP 2025 benchmark and drug package inserts (indications, contraindications, pharmacological effects) from the DingXiangYuan website.
- Reasoning Distillation: A powerful teacher LLM (Qwen3-32B) generates interpretable reasoning chains from patient records and drug instructions. These chains follow a four-step structure:
  1. Patient Condition Summary.
  2. Disease and Treatment Needs Analysis.
  3. Individual Medication Analysis (linking drugs to conditions and mechanisms).
  4. Overall Treatment Plan Synthesis.
- Cold-Start Training: The distilled reasoning trajectories and drug instruction QA pairs are converted into standard SFT samples. The model is initialized using LoRA-based parameter-efficient fine-tuning to jointly learn patient-specific clinical reasoning and drug-specific medical knowledge.
Stage 2: Reinforcement Learning via Policy Optimization (GRPO)
- Optimization: Following SFT initialization, the model is further optimized using Group Relative Policy Optimization (GRPO). This stage focuses exclusively on patient-specific medication recommendation (excluding drug instruction QA pairs).
- Hybrid Reward Function: To balance structural validity and prediction accuracy, a hybrid reward function is employed:
  - Format Reward: A rule-based metric ensuring the output adheres to the four-stage reasoning structure.
  - Outcome Reward: A performance-driven metric measuring alignment with ground-truth medications using F1 and Jaccard scores.
- Final Reward Calculation: $R = (1.5 \times F1 + 1.5 \times Jaccard + R_{format}) / 4$ .

Key Contributions

Systematic Benchmarking: The authors evaluated representative LLMs (Llama, GLM, Qwen, Baichuan) under various prompting and fine-tuning strategies, demonstrating that general-purpose LLMs struggle with fine-grained Chinese medication tasks without domain-specific adaptation.
Novel Framework: MediRec is the first framework to explicitly combine clinical reasoning distillation with reinforcement learning for Chinese medication recommendation, addressing both accuracy and interpretability.
Explainable Output: Unlike methods that output only medication lists, MediRec generates transparent clinical rationales, making it more viable for practical clinical decision support.

Experimental Results
Experiments were conducted on the CHIP 2025 discharge medication recommendation benchmark (3,190 patients, 5,894 visit records).

Performance: MediRec (based on Qwen3) achieved an F1 score of 0.5813 and a Jaccard score of 0.4626.
Comparison:
- Prompting vs. Fine-tuning: Supervised Fine-Tuning (SFT) drastically outperformed zero-shot, one-shot, and Chain-of-Thought (CoT) prompting, indicating that intrinsic LLM reasoning is insufficient for this task without domain training.
- Impact of RL: While adding distilled reasoning chains (CoT_SFT) slightly decreased direct prediction metrics compared to standard SFT (due to the trade-off of learning structured explanations), subsequent GRPO optimization significantly improved performance. The Qwen3-CoT_SFT-GRPO configuration achieved the best results.
- State-of-the-Art: MediRec's performance is highly competitive with top-performing approaches in the CHIP 2025 challenge (e.g., Zhu et al., DP-EMR), despite relying on a single model without complex ensemble methods or full-parameter fine-tuning.
Ablation Study: Removing any component of the hybrid reward function (Format, Jaccard, or F1) resulted in a decrease in the F1 score, confirming the necessity of the combined optimization strategy.

Significance and Claims
The paper claims that MediRec provides a transparent alternative to traditional black-box medication prediction systems. By generating clinically plausible recommendations accompanied by step-by-step reasoning, the framework addresses the critical need for interpretability in Chinese healthcare settings.

The authors modestly acknowledge limitations: the model may occasionally over-recommend certain drugs (e.g., Vitamin D) and may rely on broad clinical assumptions rather than strict evidence for all explanations. Safety factors like renal function and bleeding risk are not always explicitly discussed. Consequently, the authors position MediRec as a significant step forward that requires further refinement—such as integrating medical knowledge graphs and drug-drug interaction data—to fully enhance precision and safety awareness for clinical deployment.

Technical Summary: MediRec – Enhancing Chinese Medication Recommendation with Explainable Clinical Reasoning

More like this