ChatGPT能否回答患者关于核素负荷试验和18F-氟脱氧葡萄糖PET心肌炎症的问题
研究生成58个关于负荷试验和心脏FDG PET炎症扫描的患者问题,用ChatGPT-3.5和ChatGPT-4o回答,由三位核心脏病学专家按四级评分。ChatGPT-3.5回答中36%正确且完整,4%为有害或误导信息;ChatGPT-4o回答中61%正确且完整,1%为有害或误导信息。结论认为ChatGPT可提供部分准确回答且准确性随时间提高,但在当前发展阶段不适合作为临床诊疗的主要资源。
要点
- 研究纳入58个关于负荷试验和心脏FDG PET炎症扫描的患者问题。
- ChatGPT-4o正确且完整的回答比例为61%,高于ChatGPT-3.5的36%。
- ChatGPT-4o有害或误导信息的回答比例为1%,ChatGPT-3.5为4%。
核心结论
- ChatGPT可以就负荷试验和心脏FDG PET炎症检查的患者问题提供一些准确的回答,其准确性随时间提高;但在当前发展阶段不适合作为临床诊疗的主要资源。
研究设计与样本
- 研究生成58个关于负荷试验和心脏FDG PET炎症扫描的问题,使用OpenAI ChatGPT-3.5和ChatGPT-4o回答,由三位核心脏病学专家按四级分类评分。
主要发现
- ChatGPT-3.5的174个评分中,62/174(36%)为正确且完整,93/174(53%)为部分正确或部分完整,12/174(7%)为不正确且无益,7/174(4%)为不正确且有害或误导。
- ChatGPT-4o的174个评分中,107/174(61%)为正确且完整,62/174(36%)为部分正确或部分完整,3/174(2%)为不正确且无益,2/174(1%)为不正确且有害或误导。
关键数值与推荐
- 评分分类:1=正确且完整,2=部分正确或部分完整,3=不正确且无益,4=不正确且有害或误导信息。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
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
BACKGROUND: Several modalities are used for stress testing and require specific patient preparation. 18F-Flurodeoxyglucose positron emission tomography (FDG PET) is an important tool in the diagnosis and risk stratification of patients suspected of cardiac sarcoidosis and endocarditis [1-3]. There is a need for improved patient access to questions regarding cardiac testing to ensure proper adherence to instructions. We sought to evaluate the effectiveness of ChatGPT in answering questions about stress testing and cardiac FDG PET inflammation scans. METHODS AND RESULTS: We generated fifty-eight questions about stress testing and cardiac FDG PET inflammation scans. OpenAI ChatGPT-3.5 and -4o were used to answer the questions. The answers were graded by three nuclear cardiologists as the following categories: 1 = correct and complete, 2 = somewhat correct/somewhat complete, 3 = incorrect: no benefit, or 4 = incorrect: harmful/misleading information (Table I). Of the 174 grades assigned to responses to the questions from ChatGPT-3.5, 62/174 (36%) were correct and complete, 93/174 (53%) were somewhat correct/somewhat complete, 12/174 (7%) were incorrect: no benefit, 7/174 (4%) were incorrect: harmful/misleading information. Of the grades assigned to responses to questions from ChatGPT-4o, 107/174 (61%) were correct and complete, 62/174 (36%) were somewhat correct/somewhat complete, 3/174 (2%) were incorrect: no benefit, and 2/174 (1%) were incorrect: harmful/misleading information. CONCLUSIONS: ChatGPT can provide some accurate responses to patient questions regarding stress tests and cardiac FDG PET inflammation studies, and its accuracy has improved over time; however, it is not suitable as a primary resource for clinical care at this stage of development.
引用
- Shinn K, Henderson CS, Schenone AL, et al. Can ChatGPT answer patients' questions about nuclear stress tests and 18F-Flurodeoxyglucose PET for myocardial inflammation? J Nucl Cardiol. 2025. PMID: 40194756. doi:10.1016/j.nuclcard.2025.102174
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。