文献速递牙体牙髓病

两种大语言模型与牙髓专科医生对需初次根管治疗牙齿诊断能力的比较:多中心回顾性研究

Comparing Diagnostic Ability of Two Off-the-Shelf Large Language Models with That of Endodontists for Teeth Requiring Initial Root Canal Treatment: A Multicenter Retrospective Investigation.
PubMed 文献 · J Endod · 2026-10-13

这项多中心回顾性研究比较了ChatGPT-5.2和Gemini 3.0与经过校准的牙髓专科医生对645颗需初次根管治疗牙齿的诊断准确性。结论支持将LLM作为牙髓诊断和治疗的辅助工具,但需专家监督。

要点

  • 两个LLM在单个终点准确率尚可(74%-88%),但在综合评估和低患病率类别上性能明显下降。
  • LLM可用于牙髓诊断和治疗计划作为辅助工具,但不能脱离专家监督。

核心结论

  • 两个模型在单个终点和常见表现上表现可接受,但性能因终点而异,在综合评估和低患病率类别上显著降低。
  • 研究结果支持将LLM作为牙髓诊断和治疗计划的辅助工具,但需专家监督。

研究设计与样本

  • 多中心回顾性研究,纳入645颗需初次根管治疗的牙齿。
  • 比较ChatGPT-5.2和Gemini 3.0与经过校准的牙髓专科医生的诊断准确性。
  • 评估四个终点:牙髓诊断、根尖周诊断、治疗计划和病例难度。
  • 使用标准化临床摘要和X线片,采用固定单次通过方案提交给专科医生和两个LLM。

主要发现

  • 各终点准确率范围为74%至88%。
  • 宏平均F1在根尖周诊断最高(ChatGPT-5.2为0.724;Gemini 3.0为0.719),在治疗计划最低(0.385;0.550)。
  • 两个模型在少数类别中观察性能较低,但作者因亚组样本量小将其解释为提示性。
  • 三个临床终点的复合准确率分别为61.1%(95%CI 57.4%-64.8%)和59.4%(95%CI 55.3%-63.3%)。
  • 四个终点的准确率分别为47.3%(95%CI 43.3%-51.2%)和46.2%(95%CI 42.3%-49.8%),表明约半数病例存在至少一个临床相关错误。

临床怎么用

  • LLM可用于牙髓诊断和治疗计划作为辅助工具,但需专家监督。
  • 在综合评估和低患病率类别中应谨慎依赖LLM输出。

局限性

  • 少数类别亚组样本量小,相关观察性能解释为提示性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

INTRODUCTION: Despite growing interest in large language models (LLMs) for endodontic diagnosis, prior studies have used simulated cases, examination formats, or small selected series. This multicenter retrospective study compared the diagnostic accuracy of ChatGPT-5.2 and Gemini 3.0 with that of calibrated endodontists across 645 teeth requiring initial root canal treatment. METHODS: Cases were evaluated across four endpoints: pulpal diagnosis, periapical diagnosis, treatment planning, and case difficulty. Standardized clinical summaries and radiographs were submitted to endodontists and both LLMs using a fixed single-pass protocol. Exact-match accuracy, Cohen's κ, Gwet's AC1, macro-averaged F1 score, McNemar tests, and generalized estimating equations were used for statistical comparison. RESULTS: Endpoint accuracies ranged from 74% to 88%. Macro F1 was highest for periapical diagnosis (ChatGPT-5.2, 0.724; Gemini 3.0, 0.719) and lowest for treatment planning (0.385; 0.550), where both models showed lower observed performance in minority categories, interpreted as suggestive given small subgroup sizes. Composite accuracy across the three clinical endpoints was 61.1% (95% CI, 57.4%-64.8%) and 59.4% (95% CI, 55.3%-63.3%), and four-endpoint accuracy was 47.3% (95% CI, 43.3%-51.2%) and 46.2% (95% CI, 42.3%-49.8%), indicating at least one clinically relevant error in about half of cases. Treatment planning was the only endpoint differing between models after Holm-Bonferroni correction, favoring Gemini 3.0 (OR, 1.457; adjusted p = .003). CONCLUSIONS: Both models performed acceptably on individual endpoints and common presentations, but performance was endpoint-specific and substantially reduced on integrated assessment and low-prevalence categories. These findings support use of LLMs in endodontic diagnosis and treatment planning as adjunctive tools, but not without expert supervision.

来源:J Endod,查看原文。版权归原作者所有。

引用

  1. Torabinejad M, Tordik PA, Alenezi M, et al. Comparing Diagnostic Ability of Two Off-the-Shelf Large Language Models with That of Endodontists for Teeth Requiring Initial Root Canal Treatment: A Multicenter Retrospective Investigation. J Endod. 2026. PMID: 42869378. doi:10.1016/j.joen.2026.10.003

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。