文献速递创伤与烧伤

人工智能辅助的创伤聚焦腹部超声检查质量评估工具

A quality assessment tool for focused abdominal sonography for trauma examinations using artificial intelligence.
PubMed 文献 · J Trauma Acute Care Surg · 2024-12-14
一句话结论AI模型以0.015决策阈值筛查FAST图像,每复核5段即可发现1段不合格,较人工每25段发现1段的效率提升五倍。

本研究开发并验证了一种基于人工智能(AI)的创伤聚焦腹部超声(FAST)图像质量自动评分工具。五名编码者将FAST片段按质量1–5分标注,≥3分为合格。最终数据集含1514段合格片段(1399段合格、115段不合格),两两编码者合格/不合格判定一致率94%,Krippendorff α为66%。模型决策阈值可调,以0.015为阈值时,复核者每检查约5段即可发现1段不合格,同时正确识别85%的合格片段,而未用AI时平均每25段才发现1段不合格。作者认为AI整合有望提高FAST图像评估准确性并减轻专家医师工作负担。

要点

  • 最终数据集1514段FAST片段中1399段合格、115段不合格,两两编码者合格/不合格判定一致率94%,Krippendorff α为66%。
  • 质量评分1(最低)至5(最高)分,≥3分为合格;AI预测分数四舍五入至最近质量等级以区分合格与不合格。
  • 不用AI模型时,复核者平均每检查约25段才发现1段不合格;模型决策阈值0.015时,每检查约5段即可发现1段不合格,复核片段数减少五倍,同时正确识别85%的合格片段。
  • 决策阈值可调,以在精确度与敏感度之间取得所需权衡。
  • 作者认为AI整合有望提高FAST图像评估准确性并减轻专家医师工作负担。

适用人群与场景

  • 创伤聚焦腹部超声(FAST)检查图像的质量评估场景

诊断标准与关键阈值

  • FAST片段质量评分1(最低质量)至5(最高质量)分,≥3分为合格(passing),<3分为不合格(failing)。
  • AI预测分数四舍五入至最近质量等级,以区分合格与不合格片段。

研究设计与关键结果

  • 研究类型:诊断试验/标准研究,开发AI模型对FAST片段进行质量评分;五名编码者标注FAST片段,每人负责不同子集,其中10%片段三重标注以评估编码者间信度;片段分为训练集、验证集和测试集。
  • 最终数据集共1514段合格片段,其中1399段合格、115段不合格。
  • 两两编码者合格/不合格判定一致率为94%,Krippendorff α为66%。
  • 决策阈值可调,以在精确度与敏感度之间取得所需权衡。
  • 不用AI模型时,复核者平均每检查约25段才发现1段不合格;使用模型决策阈值0.015时,复核者每检查约5段即可发现1段不合格,复核片段数减少五倍,同时正确识别85%的合格片段。

证据类型、等级与局限

  • 证据等级:诊断试验/标准研究,Level II。
  • 局限:原文未报告局限性。

对床旁实践的改变

  • 作者认为AI整合有望提高FAST图像评估准确性,同时减轻专家医师工作负担。

提炼说明

  • 本条依据论文摘要整理;摘要未涉及「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,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

BACKGROUND: Current tools to review focused abdominal sonography for trauma (FAST) images for quality have poorly defined grading criteria or are developed to grade the skills of the sonographer and not the examination. The purpose of this study is to establish a grading system with substantial agreement among coders, thereby enabling the development of an automated assessment tool for FAST examinations using artificial intelligence (AI).

METHODS: Five coders labeled a set of FAST clips. Each coder was responsible for a different subset of clips (10% of the clips were labeled in triplicate to evaluate intercoder reliability). The clips were labeled with a quality score from 1 (lowest quality) to 5 (highest quality). Clips of 3 or greater were considered passing. An AI training model was developed to score the quality of the FAST examination. The clips were split into a training set, a validation set, and a test set. The predicted scores were rounded to the nearest quality level to distinguish passing from failing clips.

RESULTS: A total of 1,514 qualified clips (1,399 passing and 115 failing clips) were evaluated in the final data set. This final data set had a 94% agreement between pairs of coders on the pass/fail prediction, and the set had a Krippendorff α of 66%. The decision threshold can be tuned to achieve the desired tradeoff between precision and sensitivity. Without using the AI model, a reviewer would, on average, examine roughly 25 clips for every 1 failing clip identified. In contrast, using our model with a decision threshold of 0.015, a reviewer would examine roughly five clips for every one failing clip - a fivefold reduction in clips reviewed while still correctly identifying 85% of passing clips.

CONCLUSION: Integration of AI holds significant promise in improving the accurate evaluation of FAST images while simultaneously alleviating the workload burden on expert physicians.

LEVEL OF EVIDENCE: Diagnostic Test/Criteria; Level II.

来源:J Trauma Acute Care Surg,查看原文。版权归原作者所有。

参考来源

  1. Cull J, Morrow D, Manasco C, et al. A quality assessment tool for focused abdominal sonography for trauma examinations using artificial intelligence. J Trauma Acute Care Surg. 2024. PMID: 39327643. doi:10.1097/TA.0000000000004425

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。