大语言模型在耳科与神经耳科中的临床效用：系统综述与荟萃分析
Clinical utility of LLMs in otology and neurotology: a systematic review and meta-analysis.
文献信息
| PMID | 42931836 |
|---|---|
| 原文 | 在 PubMed 查看原文 ↗ |
| 发表日期 | 2026 |
| 作者 | Mikhael Makhoul |
| 作者单位 | Department of Otorhinolaryngology, Hôpital Pitié-Salpêtrière, AP-HP, Sorbonne Université, Paris, France. |
| 期刊 | European archives of oto-rhino-laryngology : official journal of the European Federation of Oto-Rhino-Laryngological Societies (EUFOS) : affiliated with the German Society for Oto-Rhino-Laryngology - Head and Neck Surgery |
| SCI 分区 | Q1 |
| IF | 2.3 |
| 研究类型 | 综述 Meta · 临床 |
| 所属专科 | 耳科 |
中文摘要
背景： 大语言模型（LLMs）正日益被探索用于耳鼻咽喉科的患者教育和临床决策支持。然而，其可靠性在不同临床任务间存在差异，且在耳科与神经耳科——该领域决策常需整合复杂的听力学、前庭和影像学数据——中的表现仍不明确。
方法： 按照PRISMA指南开展系统综述与荟萃分析。纳入标准采用人群、干预、对照、结局、时间与场景（PICOTS）框架定义。主要结局为LLM总体准确率。次要结局包括与专家的一致性、幻觉率、安全性、可读性、可重复性及任务特异性表现。采用随机效应荟萃分析估计合并准确率。
结果： 共纳入41项研究，其中31项纳入荟萃分析。基于GPT的LLM合并准确率为73.5%（95% CI，64.5%-80.8%），显著优于非GPT LLM（56.5%；95% CI，43.2%-68.9%；p = 0.029）。表现因任务类型而异，患者教育准确率最高（83.5%），诊断（54.6%）和治疗建议（57.9%）最低。尽管观察到显著异质性，正式评估未提供明确的小样本效应证据，但这些分析应谨慎解读。
结论： 基于GPT的LLM在耳科与神经耳科的患者教育和知识型任务中显示出前景。然而，其在诊断和治疗决策方面表现有限，支持将其作为临床医生判断的辅助工具而非替代品。
英文摘要
BACKGROUND: Large language models (LLMs) are increasingly being explored for patient education and clinical decision support in otolaryngology. However, their reliability varies across clinical tasks, and their performance in otology and neurotology, where decision-making often requires integration of complex audiological, vestibular, and imaging data, remains unclear.
METHODS: A systematic review and meta-analysis were conducted according to PRISMA guidelines. Eligibility criteria were defined using the Population, Intervention, Comparator, Outcomes, Timing, and Setting (PICOTS) framework. The primary outcome was overall LLM accuracy. Secondary outcomes included agreement with experts, hallucination rate, safety, readability, reproducibility, and task-specific performance. Random-effects meta-analyses were performed to estimate pooled accuracy.
RESULTS: Forty-one studies were included, with 31 contributing to the meta-analysis. GPT-based LLMs achieved a pooled accuracy of 73.5% (95% CI, 64.5%-80.8%), significantly outperforming non-GPT LLMs (56.5%; 95% CI, 43.2%-68.9%; p = 0.029). Performance varied by task type, with the highest accuracy for patient education (83.5%) and the lowest for diagnosis (54.6%) and treatment recommendations (57.9%). Although substantial heterogeneity was observed, formal assessment did not provide clear evidence of small-study effects, although these analyses should be interpreted cautiously.
CONCLUSION: GPT-based LLMs show promise for patient education and knowledge-based tasks in otology and neurotology. However, their limited performance in diagnostic and therapeutic decision-making supports their use as adjuncts rather than replacements for clinician judgment.