ChatGPT在听力图解读中的版本特异性表现：ChatGPT-4.0与ChatGPT-5.1的比较

文献信息

中文摘要

目的： 大型语言模型（LLMs）在临床决策支持中的应用日益受到探索，但同一模型的连续版本在诊断任务上的比较仍不明确。本研究旨在通过在标准化纯音听力图解读上直接比较ChatGPT-4.0与ChatGPT-5.1，量化诊断表现的版本特异性变化。
研究设计： 比较性诊断表现研究。
设置： 使用模拟软件生成听力图的对照实验研究。
方法： 使用AudSim Flex®软件生成100份听力图，其中20份包含故意设置的掩蔽错误，并通过基于ASHA标准的标准化7步诊断提示提交给两个模型版本。两名具有听力学专长的委员会认证耳鼻咽喉科医师独立采用5点Likert量表对每个回答评分，每个模型获得1400次评估。采用二次加权Cohen's kappa评估评分者间一致性，并采用Wilcoxon符号秩检验进行模型比较。
结果： 总体评分者间信度可接受（κ = 0.74-0.78），但ChatGPT-5.1在听力损失程度分类上的一致性较低（κ = 0.44）。ChatGPT-5.1在7个问题中的6个上评分显著高于ChatGPT-4.0（P < .001），效应量为0.53至0.94。最大改善出现在听力损失程度分类（中位差异，+1.0分）。骨导阈值识别是唯一无显著版本间差异的项目（P = .121）。两个模型在包含掩蔽错误的听力图上表现均较差。
结论： ChatGPT-5.1在听力图解读方面较ChatGPT-4.0表现出有意义的改善，尤其是在高阶诊断推理任务中。然而，两个版本在临床有缺陷的输入数据上均表现困难。这些发现强调，在将LLMs整合到听力学实践时，需要进行版本特异性基准测试和专家监督。

英文摘要

OBJECTIVE: Large language models (LLMs) are increasingly explored for clinical decision support, yet how successive versions of the same model compare on diagnostic tasks remains unclear. This study aimed to quantify version-specific changes in diagnostic performance by directly comparing ChatGPT-4.0 and ChatGPT-5.1 on standardized pure-tone audiogram interpretation.
STUDY DESIGN: Comparative diagnostic performance study.
SETTING: Controlled experimental study using audiograms generated with simulation software.
METHODS: One hundred audiograms, including 20 with intentional masking errors, were generated using AudSim Flex® software and submitted to both model versions through a standardized 7-step diagnostic prompt based on ASHA criteria. Two board-certified otolaryngologists with audiological expertise independently scored each response on a 5-point Likert scale, yielding 1400 evaluations per model. Inter-rater agreement was assessed using quadratic weighted Cohen's kappa, and model comparisons were performed with the Wilcoxon signed-rank test.
RESULTS: Inter-rater reliability was acceptable overall (κ = 0.74-0.78), though agreement was lower for degree-of-loss classification with ChatGPT-5.1 (κ = 0.44). ChatGPT-5.1 scored significantly higher than ChatGPT-4.0 on 6 of 7 questions (P < .001), with effect sizes ranging from 0.53 to 0.94. The largest improvement appeared in degree-of-hearing-loss classification (median difference, +1.0 point). Bone conduction threshold identification was the only item without a significant between-version difference (P = .121). Both models performed poorly on audiograms containing masking errors.
CONCLUSION: ChatGPT-5.1 demonstrated meaningful improvements over ChatGPT-4.0 in audiogram interpretation, particularly in higher-order diagnostic reasoning tasks. However, both versions struggled with clinically flawed input data. These findings underscore the need for version-specific benchmarking and expert oversight when integrating LLMs into audiological practice.