project
MedRAG - A medical diagnostic model developed by the Nanyang Technological University team
MedRAG is a medical diagnostic model proposed by a research team at Nanyang Technological University. It enhances the diagnostic capabilities of Large Language Models (LLMs) by incorporating knowledge graph reasoning. The model constructs a four-layer fine-grained diagnostic knowledge graph, enabling precise classification of...
What is MedRAG?
MedRAG, a medical diagnostic model proposed by a research team at Nanyang Technological University, enhances the diagnostic capabilities of Large Language Models (LLMs) by incorporating knowledge graph reasoning. The model constructs a four-layer fine-grained diagnostic knowledge graph, accurately classifying different disease manifestations and filling in gaps in patient information through an active question-filling mechanism. MedRAG improves diagnostic accuracy by 11.32% on real clinical datasets, demonstrating good generalization ability and applicability to various LLM base models. MedRAG supports multimodal input and can analyze symptoms in real time to generate accurate diagnostic suggestions.
Main functions of MedRAG
- Precision diagnostic supportMedRAG constructs a four-layer fine-grained diagnostic knowledge graph, enabling precise diagnosis based on key differences between disease manifestations. Through the diagnostic difference knowledge graph search module, patient symptoms are matched with diagnostic features in the knowledge graph, accurately locating the most similar symptom nodes and identifying the most important manifestations for disease differentiation, providing strong support for accurate diagnosis and personalized treatment plans.
- Intelligent supplementary questionsMedRAG features a proactive diagnostic questioning mechanism that automatically generates efficient and accurate supplementary questions to help doctors quickly fill in information gaps and improve the accuracy and reliability of diagnoses. When the information provided by the patient is insufficient to distinguish certain diseases, the model will be prompted to generate targeted follow-up questions to refine the symptom description.
- Highly efficient patient information analysisIn terms of UI/UX design, MedRAG supports multimodal input, including non-intrusive voice monitoring during consultations, text input, and uploading of electronic health records, ensuring doctors can quickly enter patient information. The system analyzes symptoms in real time, searches for similar cases in the local case database, and uses knowledge graph reasoning to generate accurate diagnostic suggestions.
MedRAG's technical principles
- Construction of a four-layer fine-grained diagnostic knowledge graphThis research addresses the shortcomings of existing medical knowledge bases, such as insufficient granularity and a lack of comparative information on specific disease symptoms. Researchers constructed a four-layer knowledge graph encompassing disease categories, subcategories, specific disease names, and disease characteristics using techniques including disease clustering, hierarchical aggregation, semantic embedding, medical expert knowledge, and large language model enhancement.
- Diagnostic Differences Knowledge Graph SearchThis tool is used to match patient symptoms with diagnostic features in a knowledge graph. The core process includes: clinical feature decomposition, breaking down the patient description into independent symptom representations; clinical symptom matching, calculating the similarity between patient symptoms and knowledge graph features to locate the most similar symptom nodes; upward traversal, finding the most relevant disease category in the knowledge graph; and extraction of key diagnostic features, identifying the most important representations for disease differentiation.
- Knowledge Graph-Guided LLM ReasoningTraditional RAG relies solely on retrieved cases for diagnosis generation, while MedRAG enhances the reasoning capabilities of LLM through knowledge graphs. First, it constructs an efficient index based on FAISS to accurately locate clinically similar cases; then, it matches key diagnostic features to extract the most discriminative representational information for each patient; finally, it uses LLM to fuse retrieved case information, diagnostic difference knowledge graphs, and patient information for joint reasoning, generating accurate diagnostic suggestions.
MedRAG project address
- Github repository:https://github.com/SNOWTEAM2023/MedRAG
- arXiv technical paper:https://arxiv.org/pdf/2502.04413
Application scenarios of MedRAG
- Smart Health AssistantMedRAG uses knowledge graph reasoning to generate intelligent supplementary questions, helping doctors quickly obtain key diagnostic information.
- Emergency MedicineIn emergency medicine, MedRAG can quickly analyze a patient's symptoms and provide preliminary diagnostic suggestions, shortening the treatment time.
- Chronic disease managementMedRAG can develop personalized health management plans based on a patient's medical history and lifestyle.
- Medical researchMedical researchers can use MedRAG to quickly access the latest medical research data.
- Medical EducationEducational institutions can use MedRAG as a teaching tool to help students better understand and master medical knowledge.
- Hospital online consultation serviceHospitals can use the MedRAG model to provide patients with 24/7 online medical consultation services.