大语言模型在口腔医学的应用

Application of Large Language Models in Stomatology

摘要: 目的：系统梳理大语言模型(Large language models, LLMs)在口腔医学领域的应用现状、技术进展与未来发展方向，为口腔医学的智能化转型提供参考。方法：通过文献综述与分析，从临床诊断与治疗规划、患者教育与沟通、口腔医学教育三个核心应用领域，系统归纳LLMs的应用效果与性能表现，并深入探讨其在口腔医学应用中面临的挑战与未来发展方向。结果：LLMs在口腔医学多个领域展现出巨大潜力。在临床诊断方面，通过微调可显著提升诊断准确性(如多生牙检测准确率从63%提升至91%)，在病理诊断中与专家一致性最高可达68.6%；在治疗规划中，模型在遵循临床指南方面准确率可达80%，但整体性能仍不稳定且存在异质性。在患者教育与沟通方面，LLMs回答问题的综合准确率为81.87%，并能生成科学可靠的科普信息与知情同意材料。在口腔医学教育中，LLMs作为辅助工具，在多项目牙科考试中的表现已达到甚至超过考生平均水平。然而，当前应用面临准确性波动、“幻觉”现象、数据隐私风险、模型“黑箱”问题以及缺乏标准化评估框架等主要挑战。结论：LLMs在口腔医学中具有广阔的应用前景，其定位应是辅助而非替代临床决策。未来，通过发展多模态模型、建立统一评估基准、增强模型可解释性以及开展前瞻性临床验证，LLMs有望成为可靠的医生助手，助力口腔医学服务体系向更高效、高质量的方向发展。

Abstract: Objective: This paper aims to systematically review the current state, technological advancements, and future directions of large language models (LLMs) in the field of stomatology, providing insights for the intelligent transformation of oral healthcare. Methods: A comprehensive literature review was conducted to synthesize applications and performance of LLMs across three core domains: clinical diagnosis and treatment planning, patient education and communication, and dental education. The analysis also focused on identifying the primary challenges and outlining future development pathways for integrating LLMs into stomatology. Results: LLMs have demonstrated significant potential across multiple facets of stomatology. In clinical diagnosis, domain-specific fine-tuning substantially improved performance, for instance, increasing the accuracy of supernumerary tooth detection from 63% to 91%. In pathological diagnosis, the highest level of agreement with experts reached 68.6%. For treatment planning, LLMs showed an 80% accuracy in adhering to clinical guidelines; however, overall performance remained inconsistent and was assessed using heterogeneous methods. In patient education and communication, LLMs achieved a pooled accuracy of 81.87% in answering patient inquiries and generated scientifically reliable information and informed consent materials. Within dental education, LLMs served as powerful adjunct tools, performing at or above the average level of human examinees on various dental board examinations. Despite these promising results, several critical challenges persist, including performance variability, the propensity to generate inaccurate “hallucinations”, data privacy concerns, the inherent “black box” nature of the models, and the absence of a standardized evaluation framework. Conclusion: LLMs hold immense promise for stomatology, positioned as assistive tools to augment rather than replace clinical expertise. Future progress hinges on the development of multimodal models, the establishment of unified and reliable benchmarks, advancements in model interpretability, and rigorous prospective clinical validation. With these advances, LLMs are poised to evolve into reliable assistants for clinicians, ultimately facilitating a more efficient and high-quality oral healthcare delivery system.