Abstract：
OBJECTIVE: This study explores the potential of a generative artificial intelligence tool (ChatGPT) as clinical support for nurses. Specifically, we aim to assess whether ChatGPT can demonstrate clinical decision-making equivalent to that of expert nurses and novice nursing students. This will be evaluated by comparing ChatGPT responses to clinical scenarios to those of nurses on different levels of experience.
METHODS: This is a cross-sectional study.
METHODS: Emergency room registered nurses (i.e. experts; n = 30) and nursing students (i.e. novices; n = 38) were recruited during March-April 2023. Clinical decision-making was measured using three validated clinical scenarios involving an initial assessment and reevaluation. Clinical decision-making aspects assessed were the accuracy of initial assessments, the appropriateness of recommended tests and resource use and the capacity to reevaluate decisions. Performance was also compared by timing response generations and word counts. Expert nurses and novice students completed online questionnaires (via Qualtrics), while ChatGPT responses were obtained from OpenAI.
RESULTS: Concerning aspects of clinical decision-making and compared to novices and experts: (1) ChatGPT exhibited indecisiveness in initial assessments; (2) ChatGPT tended to suggest unnecessary diagnostic tests; (3) When new information required re-evaluation, ChatGPT responses demonstrated inaccurate understanding and inappropriate modifications. In terms of performance, the mean number of words utilized in ChatGPT answers was 27-41 times greater than that utilized by both experts and novices; and responses were provided approximately 4 times faster than those of novices and twice faster than expert nurses. ChatGPT responses maintained logical structure and clarity.
CONCLUSIONS: A generative AI tool demonstrated indecisiveness and a tendency towards over-triage compared to human clinicians.
CONCLUSIONS: The study shows that it is important to approach the implementation of ChatGPT as a nurse\'s digital assistant with caution. More study is needed to optimize the model\'s training and algorithms to provide accurate healthcare support that aids clinical decision-making.
UNASSIGNED: This study adhered to relevant EQUATOR guidelines for reporting observational studies.
UNASSIGNED: Patients were not directly involved in the conduct of this study.
METHODS: This is a cross-sectional study.
METHODS: Emergency room registered nurses (i.e. experts; n = 30) and nursing students (i.e. novices; n = 38) were recruited during March-April 2023. Clinical decision-making was measured using three validated clinical scenarios involving an initial assessment and reevaluation. Clinical decision-making aspects assessed were the accuracy of initial assessments, the appropriateness of recommended tests and resource use and the capacity to reevaluate decisions. Performance was also compared by timing response generations and word counts. Expert nurses and novice students completed online questionnaires (via Qualtrics), while ChatGPT responses were obtained from OpenAI.
RESULTS: Concerning aspects of clinical decision-making and compared to novices and experts: (1) ChatGPT exhibited indecisiveness in initial assessments; (2) ChatGPT tended to suggest unnecessary diagnostic tests; (3) When new information required re-evaluation, ChatGPT responses demonstrated inaccurate understanding and inappropriate modifications. In terms of performance, the mean number of words utilized in ChatGPT answers was 27-41 times greater than that utilized by both experts and novices; and responses were provided approximately 4 times faster than those of novices and twice faster than expert nurses. ChatGPT responses maintained logical structure and clarity.
CONCLUSIONS: A generative AI tool demonstrated indecisiveness and a tendency towards over-triage compared to human clinicians.
CONCLUSIONS: The study shows that it is important to approach the implementation of ChatGPT as a nurse\'s digital assistant with caution. More study is needed to optimize the model\'s training and algorithms to provide accurate healthcare support that aids clinical decision-making.
UNASSIGNED: This study adhered to relevant EQUATOR guidelines for reporting observational studies.
UNASSIGNED: Patients were not directly involved in the conduct of this study.
摘要：
目的：本研究探讨了生成人工智能工具（ChatGPT）作为护士临床支持的潜力。具体来说,我们的目的是评估ChatGPT是否能够证明临床决策与专业护士和新手护生相同.这将通过比较ChatGPT对临床情景的反应与不同经验水平的护士的反应来进行评估。
方法：这是一项横断面研究。
方法：在2023年3月至4月期间招募急诊室注册护士（即专家；n=30）和护理专业学生（即新手；n=38）。使用涉及初始评估和重新评估的三种经过验证的临床方案来衡量临床决策。评估的临床决策方面是初始评估的准确性，推荐测试和资源使用的适当性以及重新评估决策的能力。还通过时序响应世代和字数比较了性能。专家护士和新手学生完成在线问卷调查（通过Qualtrics），而ChatGPT应答是从OpenAI获得的。
结果：关于临床决策的各个方面以及与新手和专家的比较：（1）ChatGPT在初始评估中表现出优柔寡断；（2）ChatGPT倾向于建议不必要的诊断测试；（3）当需要重新评估新信息时，ChatGPT反应显示不准确的理解和不适当的修改。在性能方面,ChatGPT答案中使用的平均字数比专家和新手使用的字数大27-41倍;提供的回答比新手快约4倍,比专家护士快2倍.ChatGPT响应保持逻辑结构和清晰度。
结论：与人类临床医生相比，一种生成AI工具表现出优柔寡断和过度分类的趋势。
结论：该研究表明，谨慎地将ChatGPT作为护士的数字助理来实施是很重要的。需要更多的研究来优化模型的训练和算法，以提供准确的医疗保健支持，帮助临床决策。
■本研究遵循相关的EQUATOR指南报告观察性研究。
■患者未直接参与本研究的进行。
方法：这是一项横断面研究。
方法：在2023年3月至4月期间招募急诊室注册护士（即专家；n=30）和护理专业学生（即新手；n=38）。使用涉及初始评估和重新评估的三种经过验证的临床方案来衡量临床决策。评估的临床决策方面是初始评估的准确性，推荐测试和资源使用的适当性以及重新评估决策的能力。还通过时序响应世代和字数比较了性能。专家护士和新手学生完成在线问卷调查（通过Qualtrics），而ChatGPT应答是从OpenAI获得的。
结果：关于临床决策的各个方面以及与新手和专家的比较：（1）ChatGPT在初始评估中表现出优柔寡断；（2）ChatGPT倾向于建议不必要的诊断测试；（3）当需要重新评估新信息时，ChatGPT反应显示不准确的理解和不适当的修改。在性能方面,ChatGPT答案中使用的平均字数比专家和新手使用的字数大27-41倍;提供的回答比新手快约4倍,比专家护士快2倍.ChatGPT响应保持逻辑结构和清晰度。
结论：与人类临床医生相比，一种生成AI工具表现出优柔寡断和过度分类的趋势。
结论：该研究表明，谨慎地将ChatGPT作为护士的数字助理来实施是很重要的。需要更多的研究来优化模型的训练和算法，以提供准确的医疗保健支持，帮助临床决策。
■本研究遵循相关的EQUATOR指南报告观察性研究。
■患者未直接参与本研究的进行。
