Large language models (LLMs) have significantly transformed medical education, yet the issue of information hallucination has emerged as a critical concern, introducing potential risks to medical education. This study systematically maps the manifestations of information hallucination within medical education contexts, identifying three distinct types, including fact-conflict hallucinations in knowledge transmission scenarios, logic-disruption hallucinations in learning support settings, and context-misleading hallucinations in clinical simulation environments. The study further analyzes the underlying causes across three interrelated dimensions including limitations inherent in data-driven training paradigms, architectural and algorithmic constraints in model design and training mechanisms, and the inherently dynamic and complex nature of medical education environments. On this basis, the study proposes constructing a multi-dimensional collaborative inhibition strategy system encompassing technical optimization, educational adaptation, ethical calibration, and human-machine collaboration. Key components include the development of a multimodal knowledge fusion system with dynamic validation protocols, enhancement of reasoning traceability through interpretable inference chains and causal modeling, integration of diverse clinical contexts with value-sensitive calibration frameworks, and the establishment of a human-machine collaborative meta-cognitive learning architecture. In conclusion, this multidimensional strategy system is conducive to preserving the transformative potential of LLMs while systematically minimizing the propagation of hallucinated content, enhancing the reliability and instructional suitability of model outputs, thereby providing a practical pathway for the safe and reliable advancement of intelligent medical education.
大语言模型(Large Language Model, LLM)是一种基于大规模语料库进行预训练的超大型深度学习模型[1],以ChatGPT、DeepSeek-R1、Med-PaLM等为代表,凭借其强大的自然语言处理能力和知识生成效率,正在深刻重塑着医学教育生态。据统计,全球超过67%的顶尖医学院校已在其课程体系中引入AI辅助教学组件,用于基础医学知识传授、临床思维训练及考核评估等场景[2]。这种技术融合不仅能够模拟临床对话、生成教学案例、解答医学疑问,而且还能参与临床决策支持。为医学教育提供了前所未有的数字化辅助,显著提升教学设计的个性化、教学资源的可及性和教学评价的科学性。然而,随着大语言模型应用的深入,信息幻觉(Hallucination)问题在医学教育领域逐渐凸显。模型在运用过程中极易生成与医学事实不符或缺乏可靠依据的误导性内容,给医学教育带来了潜在风险。因此,厘清医学教育场景中的信息幻觉表现形式,对其深层根源进行系统剖析,建立具有针对性的风险防控策略,助力AI技术安全融入医学教育体系。
第三,基于SNOMED-CT的语义验证管道。医学术语系统命名法-临床术语(systematized nomenclature of medicine - clinical terms,SNOMED-CT)作为世界上最全面的国际标准医学术语系统,其中包含了30多万个医学概念和超过130万个关系,具备全面稳固的语义基础和综合完善的结构化术语表达形式,在国际医疗与健康领域和临床医学数据分析的研究中都有着广泛应用[19-20]。利用SNOMED-CT的标准化术语和语义关系,对大语言模型生成或处理的医学文本进行语义验证。例如,通过对医学数据进行应用前清晰分析和处理、应用中概念对照、文本关系投射分析。通过这一语义验证管道降低医学术语知识整合的认知负荷,使医学文本中医学概念和逻辑更为畅通,提高医学文本处理的准确性和可靠性,为医疗领域的应用提供更有力的支持。
BUESSL, KEICHERM, NAVABN,et al .From large language models to multimodal AI:A scoping review on the potential of generative AI in medicine[J].Biomedical Engineering Letters,2025,15(5):845-863.
JIS X, PANS R, CAMBRIAE,et al .A survey on knowledge graphs:Representation,acquisition,and applications[J].IEEE Transactions on Neural Networks and Learning Systems,2022,33(2):494-514.
MCKENNAN, LIT Y, CHENGL,et al .Sources of hallucination by large language models on inference tasks[EB/OL].(2023-10-22)[2025-09-17].
[18]
TONMOYS M T I, MEHEDI ZAMANS M, JAINV,et al .A comprehensive survey of hallucination mitigation techniques in large language models[EB/OL].(2024-01-08)[2025-09-17].
[19]
PUQ F, GUPTAS, GOLLAKOTAS,et al .Whole-home gesture recognition using wireless signals[C]//Proceedings of the 19th Annual International Conference on Mobile Computing & Networking.ACM,2013:27-38.
[20]
YANGJ, GEY, XIONGH,et al .Performing joint learning for passive intrusion detection in pervasive wireless environments[C]//2010 Proceedings IEEE INFOCOM.March 14-19,2010,San Diego,CA,USA.IEEE,2010:1-9.