# MedRAG: Enhancing Retrieval-augmented Generation with Knowledge Graph-Elicited Reasoning for Healthcare Copilot

**Authors:** Xuejiao Zhao, Siyan Liu, Su-Yin Yang, Chunyan Miao  
**arXiv:** [2502.04413](https://arxiv.org/abs/2502.04413)  
**Date:** 2025-02-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2502.04413v1/

## Key points

- Authors Xuejiao Zhao, Siyan Liu, Su-Yin Yang, and Chunyan Miao introduce MedRAG, a model enhanced by knowledge graph (KG)-elicited reasoning for improving diagnostic accuracy in the medical domain.
- MedRAG retrieves diagnosis and treatment recommendations based on manifestations from Electronic Health Records (EHR) using a comprehensive four-tier hierarchical diagnostic KG.
- The model leverages reasoning within a large language model to provide more accurate decision support and can generate follow-up questions to enhance personalized medical decision-making.
- Experimental results show that MedRAG outperforms existing RAG methods in reducing misdiagnosis rates on public dataset DDXPlus and a private chronic pain diagnostic dataset (CPDD).
- MedRAG offers specific diagnostic insights for healthcare practitioners and patients, making it a valuable tool for improving healthcare outcomes.

## Summary

, , , , 


In their paper titled "MedRAG: Enhancing Retrieval-augmented Generation with Knowledge Graph-Elicited Reasoning for Healthcare Copilot," authors Xuejiao Zhao, Siyan Liu, Su-Yin Yang, and Chunyan Miao introduce a novel approach to improving diagnostic accuracy in the medical domain. They address the limitations of existing heuristic-based Retrieval-augmented generation (RAG) models by proposing MedRAG, a model enhanced by knowledge graph (KG)-elicited reasoning. <break>
 
<break>

<break>

<break>
 
MedRAG is designed to retrieve diagnosis and treatment recommendations based on manifestations from Electronic Health Records (EHR). The model systematically constructs a comprehensive four-tier hierarchical diagnostic KG that encompasses critical diagnostic differences of various diseases. By dynamically integrating these differences with similar EHRs retrieved from a database, MedRAG leverages reasoning within a large language model to provide more accurate and specific decision support. One key feature of MedRAG is its ability to proactively generate follow-up questions, enhancing personalized medical decision-making. <break>
 
The model is evaluated on both public dataset DDXPlus and a private chronic pain diagnostic dataset (CPDD) collected from Tan Tock Seng Hospital. Experimental results demonstrate that MedRAG outperforms existing RAG methods in reducing misdiagnosis rates, thanks to its information integration and relational abilities provided by the KG. Overall, MedRAG offers more specific diagnostic insights for healthcare practitioners and patients, making it a valuable tool for improving healthcare outcomes. <break>

The authors have made their code available at https://github.com/SNOWTEAM2023/MedRAG for further exploration and implementation in the medical field.