文献速递耳科与耳神经小儿耳鼻咽喉

大语言模型用于人工耳蜗教育:ChatGPT、Gemini、Claude与DeepSeek的比较

Large Language Models for Cochlear Implant Education: A Comparison of ChatGPT, Gemini, Claude, and DeepSeek.
PubMed 文献 · Otolaryngol Head Neck Surg · 2026-03-18

本研究选取Google上搜索频率最高的5个人工耳蜗相关问题,分别提问ChatGPT-4、Gemini 2.0、Claude 3.7和DeepSeek v3,由5位耳鼻咽喉科专家按内容准确性和沟通适当性以5分制评分。结果显示模型间在内容与沟通质量上均有统计学显著差异(P < .05),DeepSeek平均得分最高,Claude普遍最低。结论认为部分模型较适合患者教育,但无专家监督时仍存在局限。

要点

  • 四种大语言模型回答人工耳蜗常见问题时内容与沟通质量存在显著差异(P < .05)。
  • DeepSeek在内容准确性和沟通适当性上平均得分最高,Claude普遍最低。
  • 模型用于患者教育时仍需专家监督,不宜单独使用。

核心结论

  • 四种大语言模型在人工耳蜗患者教育中的表现存在统计学显著差异(P < .05),DeepSeek平均得分最高,Claude普遍最低。
  • 尽管部分模型较适合患者教育,但无专家监督时这些系统仍存在局限性。

研究设计与样本

  • 比较分析研究:选取Google上搜索频率最高的5个人工耳蜗问题,分别提问ChatGPT-4、Gemini 2.0、Claude 3.7和DeepSeek v3。
  • 由5位耳鼻咽喉科专家采用5分制量表对回答的内容准确性和沟通适当性进行盲法评估,统计分析采用单因素方差分析及事后检验。

主要发现

  • 模型间在内容准确性和沟通质量上均存在统计学显著差异(P < .05)。
  • DeepSeek在内容与沟通质量两个领域平均得分均最高。
  • Claude平均得分普遍最低;ChatGPT-4表现均衡;Gemini在某些沟通标准上突出。

关键数值与推荐

  • 评估采用5分制量表,基于内容准确性和沟通适当性两个维度。
  • 统计显著性阈值设定为P < .05。

临床怎么用

  • 耳鼻咽喉科智能体医生在向患者提供人工耳蜗教育信息时,可参考模型间表现差异,但不应完全依赖大语言模型输出。
  • 建议在专家监督下使用大语言模型辅助患者教育,尤其需关注内容准确性与沟通适当性。

局限性

  • 原文未明确报告样本量、置信区间等具体统计数值。
  • 评估基于虚拟模拟和盲法专家评分,可能无法完全反映真实临床沟通场景。

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

OBJECTIVE: Artificial intelligence-supported large language models (LLMs) have become increasingly widespread in recent years in the health communication and patient education. Models such as ChatGPT, Claude, Gemini, and DeepSeek are used to provide information on complex medical topics, thanks to their natural language processing capabilities. This study compares the responses of models to 5 frequently asked questions about cochlear implants in terms of content and communication quality. STUDY DESIGN: Comparative analysis of 4 LLMs using expert-evaluated responses to cochlear implant queries. SETTING: Virtual simulation with blinded specialist assessments. METHODS: Five of the most frequently searched cochlear implant questions on Google were selected. Each question was individually posed to ChatGPT-4, Gemini 2.0, Claude 3.7, and DeepSeek v3. The responses from each model were evaluated by 5 otolaryngology specialists using a 5-point scale based on content accuracy and communication appropriateness. One-way ANOVA and post hoc tests were used for statistical analysis. RESULTS: Statistically significant differences were identified among the models in both content and communication quality (P < .05). The DeepSeek model achieved the highest average scores in both areas, while the Claude model generally received the lowest scores. ChatGPT-4 demonstrated a balanced performance, while Gemini stood out in certain communication criteria. CONCLUSION: This study is one of the first comparative analyses evaluating the performance of 4 different large language models in the context of patient education about cochlear implants. Although some models appear more suitable for patient education, the findings indicate that these systems still have limitations when used without expert oversight.

来源:Otolaryngol Head Neck Surg,查看原文。版权归原作者所有。

引用

  1. Birinci M, Kilictas AU, Gül O, et al. Large Language Models for Cochlear Implant Education: A Comparison of ChatGPT, Gemini, Claude, and DeepSeek. Otolaryngol Head Neck Surg. 2026. PMID: 41848221. doi:10.1002/ohn.70192

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。