公开多模态大语言模型在眼表感染分诊、诊断与治疗中的基准测试:与角膜专科医生的对比
该研究首次将公开多模态大语言模型(包括 GPT-5)与角膜专科医生进行头对头基准测试,使用真实裂隙灯图像和临床案例,评估其在感染性角膜炎和结膜炎分诊、诊断和治疗中的表现。结果显示,当同时提供图像和文本输入时,模型达到了接近人类的诊断和分诊准确率。多模态大语言模型有望在高流量或资源有限的眼科环境中辅助早期分诊和决策。
要点
- 公开多模态大语言模型在图像联合文本输入下,诊断和分诊准确率接近角膜专科医生。
- 研究使用真实裂隙灯图像和临床案例,首次将 GPT-5 等模型与人类专家头对头比较。
- 模型可辅助高流量或资源有限环境下的早期分诊,并与微生物诊断整合以缩短诊断延迟。
核心结论
- 公开多模态大语言模型在同时接收图像和文本输入时,对眼表感染的分诊和诊断准确率接近角膜专科医生。
研究设计与样本
- 研究采用头对头基准测试设计,使用真实裂隙灯图像和临床案例,将公开多模态大语言模型(包括 GPT-5)与角膜专科医生进行对比。
主要发现
- 当提供图像和文本输入时,模型达到了接近人类的诊断和分诊准确率。
临床怎么用
- 多模态大语言模型可辅助高流量或资源有限眼科环境中的早期分诊和决策。
- 模型与微生物学或宏基因组诊断整合可能有助于简化感染管理并减少诊断延迟。
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
WHAT IS ALREADY KNOWN ON THIS TOPIC Diagnosing infectious keratitis and conjunctivitis remains challenging due to limited sensitivity and delayed turnaround of culture-based tests. Artificial intelligence tools have shown promise in image analysis but lack validation against human specialists in real-world ophthalmic infections. WHAT THIS STUDY ADDS This study provides the first head-to-head benchmarking of publicly available multimodal large language models, including GPT-5, against corneal specialists using real slit-lamp images and clinical vignettes. When provided with both image and text inputs, models achieved near-human diagnostic and triage accuracy. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY Multimodal large language models could complement specialist care by supporting early triage and decision-making in high-volume or resource-limited ophthalmic settings. Their integration with microbiological or metagenomic diagnostics may help streamline infection management and reduce diagnostic delays. Introduction Infectious keratitis (IK) is a severe...
引用
- British Journal of Ophthalmology:Publicly available multimodal large language models for ocular surface infections: benchmarking against corneal specialists in triage, diagnosis and treatment
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。