ChatGPT-4与医护人员用药审查的比较研究
该回顾性概念验证研究对51例既往由老年内科医生和医院药师完成的用药审查,用ChatGPT-4重新评估,比较两者建议的一致性。ChatGPT与医护人员建议一致率为27.7%,发现医护人员遗漏的19条有效干预(7.6%),但也提出84条错误干预(33.7%)。ChatGPT在结构化患者数据方面表现较好(86.4%正确诊断关联用药),但在结合患者特征确定实验室目标值方面较弱(46.7%)。
要点
- ChatGPT与医护人员用药审查建议一致率仅27.7%。
- ChatGPT发现7.6%医护人员遗漏的有效干预,但33.7%干预错误。
- ChatGPT结构化诊断与用药关联正确率86.4%,实验室目标值仅46.7%。
核心结论
- ChatGPT与医护人员用药审查建议一致性低,但能发现部分遗漏干预,缺乏基于个体患者背景的临床决策能力。
- ChatGPT可作为结构化诊断和用药列表的辅助工具。
研究设计与样本
- 回顾性概念验证研究,纳入51例既往由老年内科医生和医院药师完成的用药审查,使用ChatGPT-4重新评估,由两名药师评分一致性,专家团队审查额外干预。
- ChatGPT接受多重用药指南训练,并获得与医护人员相同的主要数据。
主要发现
- 医护人员提出183条干预,ChatGPT提出202条干预。
- ChatGPT与医护人员干预一致率为27.7%。
- ChatGPT发现19条医护人员遗漏的有效干预(7.6%),同时提出84条错误干预(33.7%)。
- ChatGPT正确诊断关联用药的比例为86.4%,基于患者特征确定合适实验室目标值的比例为46.7%。
临床怎么用
- ChatGPT不宜独立用于老年患者用药审查的临床决策,但可作为结构化诊断和用药列表的辅助工具。
- 使用ChatGPT辅助用药审查时需警惕其提出的错误干预(33.7%),应由医护人员审核。
局限性
- 本研究为回顾性概念验证研究,样本量51例,结果需进一步验证。
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: The increasing prevalence of patients with hyperpolypharmacy (> 10 medications) has made medication reviews increasingly complex. ChatGPT-4-Turbo (ChatGPT), a large language model, has demonstrated potential in healthcare applications and could potentially support medication reviews. OBJECTIVES: This study aimed to evaluate the agreement between medication reviews conducted by ChatGPT compared to healthcare professionals (HCPs) in older people. Secondary objectives included: the validity of additional interventions detected by ChatGPT, its ability to structure diagnoses to medication use, and laboratory target values based on patient characteristics. METHODS: In this retrospective proof-of-concept study, 51 medication reviews previously conducted by a geriatric internist and hospital pharmacist were re-evaluated using ChatGPT. ChatGPT was trained on polypharmacy guidelines, was then provided with the same primary data as HCPs, and was asked to perform medication reviews. Two pharmacists scored the agreement between ChatGPT and HCPs. The structuring of information and additional interventions suggested by ChatGPT were reviewed within an expert team. Descriptive statistics were used. OUTCOMES: The primary outcome was the percentage agreement between the interventions suggested by ChatGPT compared to HCPs. Secondary outcomes included the proportion of valid and incorrect interventions suggested by ChatGPT and its ability to structure patient information. RESULTS: HCPs suggested 183 interventions and ChatGPT 202 interventions. ChatGPT achieved a 27.7% agreement with interventions of HCPs. It identified 19 additional valid interventions which HCPs missed (7.6%), but also proposed 84 incorrect interventions (33.7%). While ChatGPT demonstrated strong capability in structuring patient data (86.4% correct diagnoses linked to medication), it struggled with contextualizing appropriate laboratory target values based on patient characteristics (46.7%). CONCLUSION: ChatGPT had low agreement with HCPs, but found additional interventions that HCPs missed. ChatGPT lacks clinical decision-making capabilities based on individual patient contexts in older people. ChatGPT may, however, serve as a support tool to structure diagnoses and medication lists.
引用
- Ten Hoope SMK, Marongiu S, Siegert CEH, et al. Conducting Medication Reviews: A Comparative Study Between ChatGPT-4 and Healthcare Professionals. J Am Geriatr Soc. 2026. PMID: 41968599. doi:10.1111/jgs.70415
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。