MED-COPILOT: A Medical Assistant Powered by GraphRAG and Similar Patient Case Retrieval
New medical AI combines WHO guidelines with similar-patient retrieval to reduce hallucinations by 40%.
A research team led by Shuheng Chen has introduced MED-COPILOT, an interactive clinical decision-support system designed to address the critical limitations of current large language models (LLMs) in medical settings. The system tackles LLMs' propensity for hallucinations and their difficulty integrating long, structured documents by combining two advanced retrieval methods: guideline-grounded GraphRAG and hybrid semantic-keyword similar-patient retrieval. This architecture allows the AI to synthesize heterogeneous evidence from patient histories, clinical guidelines, and trajectories of comparable cases, providing a transparent and evidence-aware foundation for clinical reasoning. The tool is aimed at clinicians and medical trainees and is already available as a full interactive system online.
Technically, MED-COPILOT constructs a structured knowledge graph from authoritative sources like WHO and NICE guidelines, applying community-level summarization for efficiency. Its core innovation is a hybrid retrieval engine that queries a vast database of 36,000 similar patient cases, derived from SOAP-normalized MIMIC-IV clinical notes and Synthea-generated synthetic records. In evaluations for clinical note completion and medical question answering, the system consistently outperformed parametric LLM baselines and standard RAG (retrieval-augmented generation) pipelines, demonstrating measurable improvements in both generation fidelity and clinical reasoning accuracy. The live demo enables users to inspect retrieved evidence, visualize token-level similarity contributions, and conduct guided follow-up analysis, marking a significant step toward practical, interpretable AI assistants in healthcare.
- Combines GraphRAG on WHO/NICE guidelines with retrieval from a 36,000-case patient database (MIMIC-IV & Synthea)
- Outperforms standard LLMs and basic RAG in clinical note completion and medical QA, improving reasoning accuracy
- Provides full interpretability with evidence inspection, similarity visualizations, and a live demo for user testing
Why It Matters
Offers clinicians a more accurate, evidence-transparent AI tool that reduces dangerous hallucinations by grounding decisions in real guidelines and past cases.