| 摘要: |
| ［摘要］ 大语言模型（LLMs）在医疗领域应用广泛，但幻觉问题严重制约其在临床的安全部署。该文系统梳理了医学场景下LLMs幻觉的定义、分类、机制及治理对策。首先，构建了涵盖事实性与忠实性的医学幻觉分类，剖析了从数据、架构、算法到交互的偏差传导机制。其次，结合临床问答、病历生成及多模态场景，分析了事实矛盾、证据伪造等幻觉表现。在此基础上，总结了数据治理、检索增强生成、事实对齐及不确定性量化等主流检测与缓解策略。最后，提出人机协同的治理机制，为推动医疗LLMs向安全、可信方向部署提供参考。 |
| 关键词: 大语言模型 医学幻觉 检索增强生成 安全可信 |
| DOI：10.3969/j.issn.1674-3806.2026.07.02 |
| 分类号:R 319 |
| 基金项目:浙江省卫生健康行业科技计划委省共建项目（编号：WKJ-ZJ-26039） |
|
| A study on the hallucination phenomena of large language models in medical scenarios and their countermeasures |
|
Shen Feng1,2, Qi Xinglun3, Yang Dagan3
|
|
1.School of Medical Technology and Information Engineering, Zhejiang Chinese Medical University, Hangzhou 310053, China; 2.Qiaosi Branch, the First People′s Hospital of Linping District, Hangzhou 311100, China; 3.Department of Laboratory Medicine, the First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou 310003, China
|
| Abstract: |
| ［Abstract］ Large language models(LLMs) are widely used in the medical field, but the problem of hallucinations seriously restricts their safe deployment in clinical practice.This paper systematically reviews the definition, classification, underlying mechanisms and mitigation strategies of the hallucinations of LLMs under medical scenarios. First, a classification system encompassing “factuality” and “faithfulness” is constructed, and a bias transmission mechanism spanning data, architecture, algorithms and interaction is analyzed. Second, clinical scenarios such as medical question-answering, medical record generation and multimodal tasks are integrated to analyze specific manifestations of hallucinations, including factual contradiction and evidence fabrication. On this basis, the mainstream detection and mitigation strategies such as data governance, retrieval-augmented generation, factual alignment and uncertainty quantification are summarized. Finally, a human-AI collaborative governance mechanism is proposed, aiming to provide a reference for advancing the safe and trustworthy deployment of LLMs in clinical practice. |
| Key words: large language models(LLMs) medical hallucinations retrieval-augmented generation safety and trustworthiness |
