文献速递重症综合管理与镇静镇痛

多模态大语言模型整合进床旁超声记录:一项先导研究

Integration of multimodal large language models into point-of-care ultrasound documentation: A pilot study
RSS 订阅 · Am J Emerg Med · 2026-11-05
一句话结论多模态大语言模型可接入床旁超声记录流程,但心脏切面识别与异常判读和专家一致性差,目前不应进入临床实践。

该回顾性先导研究将具备视频能力的多模态大语言模型(MLLM)应用于QPathE中存储的心脏床旁超声(POCUS)检查,目标纳入350例(100例正常、250例有病变),实际应用370例、失败20例,以专家图像回顾为金标准。切面识别与专家一致的仅151例(43.1%);心尖四腔心一致性为一般,胸骨旁长轴、胸骨旁短轴、剑突下及下腔静脉为中等。正常检查判读一致86/100,但心包积液、射血分数、主动脉流出道、心腔大小、下腔静脉的一致率分别为8%、4%、70%、0%、12%。作者认为该厂商无关流程证明可整合MLLM,但在当前状态下不应纳入临床实践。

要点

  • 回顾性先导研究:MLLM用于370例心脏POCUS检查,20例失败;以专家图像回顾为金标准。
  • 切面识别与专家一致仅151例(43.1%);心尖四腔心一致性一般,胸骨旁长轴、胸骨旁短轴、剑突下、下腔静脉为中等。
  • 正常检查判读一致86/100;心包积液一致8%、射血分数4%、主动脉流出道70%、心腔大小0%、下腔静脉12%,各病变间差异大。
  • 作者认为该厂商无关流程可把MLLM整合进图像判读,但当前状态不应纳入临床实践,未来模型改进或可实现自动化POCUS判读与记录。

适用人群与场景

  • 研究纳入QPathE中存储的心脏POCUS检查,计划纳入350例(100例正常、250例有病变),实际应用370例、失败20例;为回顾性先导研究场景。

研究设计与关键结果

  • 研究类型:回顾性先导研究;样本:MLLM应用于370例检查、失败20例;以专家图像回顾为金标准,MLLM切面识别与判读结果经Qualtrics记录。
  • 切面识别与专家一致151例(43.1%);Cohen kappa显示心尖四腔心一致性为一般,胸骨旁长轴、胸骨旁短轴、剑突下及下腔静脉为中等。
  • 判读与专家一致:正常检查86/100;按病变为心包积液8%、射血分数4%、主动脉流出道70%、心腔大小0%、下腔静脉12%。

证据类型、等级与局限

  • 证据类型:回顾性先导研究(单臂,无对照干预)。
  • 局限:原文未报告局限性。

对床旁实践的改变

  • 作者认为该厂商无关流程证明可将MLLM整合进POCUS图像判读与记录,但在当前状态下不应纳入临床实践,未来模型改进或可实现自动化POCUS判读与记录。

指南或共识出处

  • Telexy Health开发的MLLM系统部署于POCUS管理器QpathE;原文无指南或共识出处。

提炼说明

  • 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「处置流程与药物剂量」「红旗征与不能漏的情况」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开

INTRODUCTION: There are no studies evaluating the potential to deploy artificial intelligence in the interpretation of cardiac point-of-care ultrasound (POCUS) documentation. Telexy Health developed an image interpretation and documentation system using multimodal model extensions with video capabilities (MLLM) and deployed it on the POCUS manager QpathE. This study evaluated the MLLM's agreement to experts in identifying cardiac POCUS views and interpreting normal and abnormal examinations.

METHODS: This was a retrospective pilot study using cardiac POCUS studies stored in QPathE. The MLLM was applied to POCUS examinations until 350 examinations were reached, 100 normal, and 250 with pathology. MLLM view identification and interpretation were recorded using Qualtrics, with expert image review serving as the gold standard.

RESULTS: The MLLM was applied to 370 examinations, with 20 failures. View identification agreed with expert review in 151 examinations (43.1%). Cohen's kappa showed fair agreement for apical four-chamber views between expert and MLLM, with moderate agreement for parasternal long axis, parasternal short axis, subxiphoid, and inferior vena cava views. The MLLM agreed with expert interpretation 86 of 100 normal examinations; the agreement varied by pathology with pericardial effusion, ejection fraction, aortic outflow tract, chamber size, and inferior vena cava having an agreement of 8%, 4%, 70%, 0%, and 12% respectively.

CONCLUSION: While agreement in image interpretation was poor, this vendor agnostic workflow demonstrated the ability to integrate an MLLM model into image interpretation. While it should not be incorporated into clinical practice in this state, future model improvement may enable automated POCUS interpretation and documentation.

来源:Am J Emerg Med,查看原文。版权归原作者所有。

参考来源

  1. Am J Emerg Med:Integration of multimodal large language models into point-of-care ultrasound documentation: A pilot study
  2. Boivin Z, Jaffa E, Agaraj I, et al. Integration of multimodal large language models into point-of-care ultrasound documentation: A pilot study. Am J Emerg Med. 2026. PMID: 43002824. doi:10.1016/j.ajem.2026.11.001

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。