环境人工智能记录员的使用与急诊医师客观文档工作量的关系
该研究在20个急诊科(3个学术中心、17个社区中心)的275名主治医师中,通过电子健康记录(EHR)直接识别AI记录员使用情况(以病历中的患者同意声明为标志),排除了教学和见证记录。采用医师内固定效应模型比较同一医师使用与未使用AI记录员的病历。结果显示:AI记录病历的活跃时间中位数10.0 vs 12.5分钟(调整后减少2.4分钟),开始书写时间延迟约15分钟(31 vs 17分钟),病历长度增加约14%(5,220 vs 4,468字符,医师内+629字符),病历完成未延迟(-6分钟),72小时急诊复诊率未增加(-0.8个百分点,95%CI -1.5至-0.1)。
要点
- 在275名急诊主治医师中,31%的接诊使用了AI记录员;AI记录病历的活跃书写时间中位数10.0分钟,未使用者12.5分钟,调整后减少2.4分钟。
- AI记录病历的开始书写时间中位数31分钟(首次看患者后),未使用者17分钟,延迟约15分钟。
- AI记录病历长度中位数5,220字符,未使用者4,468字符,医师内增加629字符(约14%)。
- 病历完成时间未延迟(-6分钟),72小时急诊复诊率未增加(-0.8个百分点,95%CI -1.5至-0.1)。
- 该研究为多中心观察性设计,暴露与结局均直接来自EHR,但因果关系需谨慎解读。
适用人群与场景
- 20个急诊科(3个学术中心、17个社区中心)的主治急诊医师,排除教学接诊和见证记录。
诊断标准与关键阈值
- AI记录员使用以急诊病历中的患者同意声明标记;分析限于成熟采用窗口(2025年12月起,Epic审计日志数据保留期间)。
处置流程与药物剂量
- 原文未提及
红旗征与不能漏的情况
- 原文未提及
研究设计与关键结果
- 多中心观察性研究,20个急诊科、275名主治医师,31%接诊为AI记录;主要终点活跃书写时间中位数10.0 vs 12.5分钟(调整后-2.4分钟),次要终点包括开始书写延迟约15分钟(31 vs 17分钟)、病历长度增加约14%(5,220 vs 4,468字符,医师内+629字符)、完成时间-6分钟、72小时急诊复诊-0.8个百分点(95%CI -1.5至-0.1)。
证据类型、等级与局限
- 观察性多中心研究,采用医师内固定效应模型;局限包括非随机暴露、同意声明可能不完全代表实际使用、仅限成熟采用窗口。
对床旁实践的改变
- AI记录员可减少急诊主治医师病历活跃书写时间,但开始书写时间延迟、病历更长;不增加72小时复诊,完成时间未延迟;适用于主治医师非教学接诊,教学接诊和见证记录未纳入。
指南或共识出处
- Ann Emerg Med 2026 Oct 27:S0196-0644(26)01292-8
以上提炼依据论文摘要(全文未开放获取,或许可不允许翻译转载);全文请点「阅读原文」。
英文原文(摘要)· 点击展开
Ann Emerg Med . 2026 Oct 27:S0196-0644(26)01292-8. doi: 10.1016/j.annemergmed.2026.08.737. Online ahead of print. ABSTRACT STUDY OBJECTIVE: Ambient artificial intelligence (AI) scribes ("AI scribes") are purported to reduce emergency department (ED) documentation burden, but ED evidence is largely single-site, uncontrolled, or self-reported, with exposure inferred rather than measured. We evaluated AI scribe use among attending emergency physicians across a multisite network, with exposure and outcomes read directly from the electronic health record. METHODS: Across 20 EDs (3 academic and 17 community), AI scribe use was flagged per note from the patient consent statement within the ED note. Analyses were restricted to attending physicians and, to isolate the physician's own documentation, excluded teaching encounters and attestation notes identified from standardized attestation text. Endpoints, computed over the mature-adoption window (December 2025 onward, when Epic audit-log data are retained), were active time in the note (primary), note length, note-start time, completion lag, and 72-hour ED return. Within-physician models (physician and calendar-month fixed effects, clustered by physician) compared each adopter's AI-scribed notes with their own unassisted notes, adjusting only for pre-exposure case mix. RESULTS: Among 275 attending physicians, 31% of encounters were AI-scribed. Physicians spent less active time in AI-scribed notes (median 10.0 versus 12.5 min-2.4 minutes after adjustment), started them about 15 minutes later (median 31 versus 17 minutes after first seeing the patient, and produced notes about 14% longer [median 5,220 versus 4,468 characters; within-physician +629 characters]). Note completion was not delayed (-6 minutes), and 72-hour ED return was not increased (-0.8 percentage point; 95% confidence interval -1.5 to -0.1). CONCLUSION: Among emergency physicians, AI scribe use was associated with less active time in the note, deferred note-start, and longer notes, without delaying completion or increasing short-term returns, with effect sizes consistent with ambulatory randomized data. PMID: 42949694 | DOI: 10.1016/j.annemergmed.2026.08.737
参考来源
- PubMed · Ann Emerg Med:Ambient Artificial Intelligence Scribe Use and Objective Documentation Effort Among Emergency Physicians
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。