基于阶段识别模型的人工智能评估腹腔镜远端胃切除术的手术效率与暴露质量
本研究利用日本多机构2015至2021年的腹腔镜远端胃切除术视频,开发阶段识别模型并构建Total AI评分,发现该评分在高熟练度组显著高于中、低熟练度组,且以67.3为cutoff预测ESSQS认证状态的敏感度为95.7%、特异度为25.0%。
要点
- 阶段识别模型准确率达90.3%,可自动划分手术九阶段。
- Total AI评分在高熟练度组显著高于中、低熟练度组。
- cutoff 67.3预测ESSQS认证敏感度95.7%、特异度25.0%。
核心结论
- 基于自动阶段识别模型的Total AI评分可作为腹腔镜远端胃切除术工作流效率与暴露质量的可重复量化指标。
- 该评分有助于通过风险分层减少专家评审工作量。
研究设计与样本
- 回顾性收集2015至2021年日本多机构腹腔镜远端胃切除术手术视频。
- 256例视频用于开发阶段识别模型,955例用于构建和测试评分系统。
- 根据ESSQS评分将视频分为高、中、低三个熟练度组。
主要发现
- 阶段识别模型准确率为90.3%。
- Total AI评分在高熟练度组显著高于中熟练度组(p=0.003)和低熟练度组(p=0.010)。
- 以67.3为cutoff预测ESSQS认证状态的敏感度为95.7%,特异度为25.0%。
关键数值与推荐
- Total AI评分cutoff 67.3用于预测ESSQS认证状态,敏感度95.7%、特异度25.0%。
- 手术视频按ESSQS评分分为高、中、低三个熟练度组。
临床怎么用
- Total AI评分可用于腹腔镜远端胃切除术的自动化技术评估与风险分层。
- 该评分可能辅助减少专家评审工作量,但特异度较低需结合其他评估。
局限性
- 本研究为概念验证性研究,结论需进一步验证。
- Total AI评分预测ESSQS认证状态的特异度仅为25.0%。
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: Traditional surgical skill assessment tools are inherently subjective, prone to interobserver variability, and place a substantial burden on expert reviewers. This study aimed to evaluate the feasibility of a proof-of-concept artificial intelligence (AI)-based system for automated assessment of selected dimensions of technical performance in laparoscopic distal gastrectomy (LDG) using a phase recognition model and to construct a scoring system for risk-stratification in expert-based surgical performance assessment. METHODS: Surgical videos of LDG procedures performed between 2015 and 2021 were collected from multiple institutions across Japan. Each procedure was divided into nine phases. Based on Endoscopic Surgical Skill Qualification System (ESSQS) scores, surgical videos were classified into three proficiency groups (high, middle, and low). Parameters from the phase recognition model were integrated into a composite score, referred to as the Total AI score, which was compared across groups. A cutoff value was also established for risk-stratification based on ESSQS qualification status. RESULTS: A total of 1399 videos were collected, of which 256 were used to develop the phase recognition model, and 955 were used to construct and test the scoring system after excluding inappropriate cases. The phase recognition model achieved 90.3% accuracy. The Total AI score was significantly higher in the high proficiency group than in the middle (p = 0.003) and low proficiency (p = 0.010) groups. A cutoff of 67.3 identified a sensitivity of 95.7% and specificity of 25.0% for predicting ESSQS qualification status. CONCLUSIONS: The Total AI score, derived from an automated phase recognition model, may serve as a proof-of-concept AI-based quantitative metric for the reproducible assessment of workflow efficiency and exposure quality in LDG and may help reduce expert reviewer workload through risk-stratification.
引用
- Umemiya A, Takenaka S, Kitaguchi D, et al. AI-based assessment of surgical efficiency and exposure quality in laparoscopic distal gastrectomy using a phase recognition model. Surg Endosc. 2026. PMID: 42509496. doi:10.1007/s00464-026-13188-4
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。