Scholarlink

Background <p>Dental trauma represents a common emergency condition requiring rapid and accurate guidance to prevent permanent damage. In urgent scenarios where access to dental professionals may be delayed, large language models (LLMs) have the potential to provide patients with timely and relevant information. The aim of this study was to comparatively evaluate the dentist rated performance of five widely used LLMs in answering frequently asked questions related to dental trauma emergencies across three predefined components: accuracy, comprehensiveness, and clinical applicability.</p> Methods <p>Based on the guidelines of the International Association of Dental Traumatology (IADT), the ToothSOS application, and frequently asked patient questions, 27 open-ended questions in Turkish were prepared and divided into five clinical subcategories: avulsion, post-replantation care, luxation, fractures, and other traumas. The questions were posed to the models ChatGPT-4o, Claude 3.5, DeepSeek, Microsoft Copilot, and Gemini 2.0 Flash. Twenty experienced dentists evaluated the responses using a 5-point Likert scale, with the evaluators jointly considering three predefined components: accuracy, comprehensiveness, and clinical applicability. A total of 2,700 individual ratings were analyzed using the Friedman test, Bonferroni-corrected Wilcoxon test, and Intraclass Correlation Coefficient (ICC).</p> Results <p>A significant difference in overall performance was observed among the models (<i>p</i> &lt; 0.001). ChatGPT-4o achieved the highest mean score (4.63 ± 0.57), whereas Gemini received the lowest score (3.87 ± 0.91). Claude, DeepSeek, and Copilot demonstrated similar, moderate performance, ranging approximately between 4.1 and 4.3. Median values were 5 (IQR 4–5) for ChatGPT-4o and Claude and 4 (IQR 3–5) for the other models, indicating that the responses of ChatGPT-4o and Claude were rated more consistently high. Within each model, no statistically significant differences were observed in mean Likert scores across clinical subcategories (<i>p</i> &gt; 0.05).</p> Conclusions <p>Although LLMs may not be able to replace professional clinical examination, they may serve as a rapid supportive source of patient-oriented information in dental trauma emergencies where access to a dentist is limited. However, model dependent differences, highlight the need for regular verification before patient facing use.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative performance of large language models for patient-oriented support in dental trauma emergencies

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月7日

查看全文
  • Emre Sözen,
  • Hasan Akpınar,
  • Sevda Yaman

摘要

Background

Dental trauma represents a common emergency condition requiring rapid and accurate guidance to prevent permanent damage. In urgent scenarios where access to dental professionals may be delayed, large language models (LLMs) have the potential to provide patients with timely and relevant information. The aim of this study was to comparatively evaluate the dentist rated performance of five widely used LLMs in answering frequently asked questions related to dental trauma emergencies across three predefined components: accuracy, comprehensiveness, and clinical applicability.

Methods

Based on the guidelines of the International Association of Dental Traumatology (IADT), the ToothSOS application, and frequently asked patient questions, 27 open-ended questions in Turkish were prepared and divided into five clinical subcategories: avulsion, post-replantation care, luxation, fractures, and other traumas. The questions were posed to the models ChatGPT-4o, Claude 3.5, DeepSeek, Microsoft Copilot, and Gemini 2.0 Flash. Twenty experienced dentists evaluated the responses using a 5-point Likert scale, with the evaluators jointly considering three predefined components: accuracy, comprehensiveness, and clinical applicability. A total of 2,700 individual ratings were analyzed using the Friedman test, Bonferroni-corrected Wilcoxon test, and Intraclass Correlation Coefficient (ICC).

Results

A significant difference in overall performance was observed among the models (p < 0.001). ChatGPT-4o achieved the highest mean score (4.63 ± 0.57), whereas Gemini received the lowest score (3.87 ± 0.91). Claude, DeepSeek, and Copilot demonstrated similar, moderate performance, ranging approximately between 4.1 and 4.3. Median values were 5 (IQR 4–5) for ChatGPT-4o and Claude and 4 (IQR 3–5) for the other models, indicating that the responses of ChatGPT-4o and Claude were rated more consistently high. Within each model, no statistically significant differences were observed in mean Likert scores across clinical subcategories (p > 0.05).

Conclusions

Although LLMs may not be able to replace professional clinical examination, they may serve as a rapid supportive source of patient-oriented information in dental trauma emergencies where access to a dentist is limited. However, model dependent differences, highlight the need for regular verification before patient facing use.