CuraView: Multi-Agent Framework for Medical Hallucination Detection
CuraView is an innovative multi-agent system aimed at identifying faithfulness hallucinations in discharge summaries produced by large language models (LLMs). It develops a knowledge graph based on GraphRAG from patient-specific electronic health records (EHRs) and features a closed-loop pipeline for generation and detection, utilizing sentence-level evidence retrieval and classification across four evidence tiers (E1-E4), ranging from strong support to outright contradiction. This framework offers clear and structured evidence chains to pinpoint statements that conflict with original records, thereby mitigating potential risks to patient safety. The evaluation of CuraView was conducted on a sample of 250 patients from the Di dataset.
Key facts
- CuraView is a multi-agent framework for detecting faithfulness hallucinations in LLM-generated discharge summaries.
- It uses GraphRAG to build a knowledge graph from patient-level EHRs.
- The framework employs a closed-loop generation-detection pipeline.
- Evidence retrieval and classification span four grades: E1 (strong support) to E4 (direct contradiction).
- It produces structured and interpretable evidence chains.
- Faithfulness hallucinations are statements that contradict source records.
- LLMs can improve generation efficiency but are prone to hallucinations.
- The framework was evaluated on 250 patients from the Di dataset.
Entities
—