人工智能在微创妇科手术中的任务特异性表现与临床转化准备度：系统综述
本系统综述检索2015年1月至2025年8月七个数据库的英文研究，纳入8项研究共6577例患者，评估人工智能在微创妇科手术术前、术中、术后阶段的任务特异性表现与临床转化准备度。术前模型对子宫内膜异位症和附件扭转灵敏度高（95%-96%）但特异度有限（11%-45%）；一项随机对照试验显示AI辅助宫腔镜子宫肌瘤切除术规划可显著减少手术时间和术中出血（p<.05）；术后深度学习模型预测宫腔镜粘连分离术后生育结局的区分能力最高（AUC>0.95）。PROBAST-AI评估显示75%研究存在高偏倚风险，证据以观察性为主且外部验证有限，常规应用前需多中心前瞻性验证。
要点
- 术前AI模型对子宫内膜异位症和附件扭转灵敏度95%-96%、特异度11%-45%，适合筛查分诊而非确诊。
- 一项RCT显示AI辅助宫腔镜子宫肌瘤切除术规划可显著减少手术时间和术中出血（p<.05）。
- 术后深度学习预测宫腔镜粘连分离术后生育结局AUC>0.95，但75%研究高偏倚风险，需多中心前瞻性验证。
核心结论
- AI在微创妇科手术中展现有前景的任务特异性表现，但证据以观察性研究为主、高偏倚风险且外部验证有限，制约临床转化。
- 常规采用前需多中心前瞻性验证。
研究设计与样本
- 系统综述遵循PRISMA指南，检索PubMed、Embase、Scopus、Web of Science、Google Scholar、CENTRAL和Cochrane Library，范围2015年1月1日至2025年8月31日英文文献。
- 纳入评估AI或机器学习模型用于微创妇科手术（腹腔镜、宫腔镜或机器人手术）预测或诊断任务的研究，排除动物研究、无临床结局的技术可行性报告和非英文文献。
- 共纳入8项研究、6577例患者，使用PROBAST-AI框架评估偏倚风险。
主要发现
- 术前子宫内膜异位症和附件扭转模型灵敏度95%-96%、特异度11%-45%，提示可作为筛查分诊而非确定性诊断系统。
- 一项随机对照试验显示AI辅助宫腔镜子宫肌瘤切除术手术规划显著减少手术时间和术中出血（p<.05）。
- 术后深度学习模型预测宫腔镜粘连分离术后生育结局的区分能力最高（AUC>0.95），且是唯一纳入前瞻性外部验证的研究。
- PROBAST-AI评估发现75%的研究存在高偏倚风险。
关键数值与推荐
- 术前模型灵敏度95%-96%、特异度11%-45%，对应子宫内膜异位症和附件扭转的预测任务。
- AI辅助宫腔镜子宫肌瘤切除术手术规划显著减少手术时间和术中出血，p<.05，基于一项随机对照试验。
- 术后深度学习预测宫腔镜粘连分离术后生育结局AUC>0.95，为唯一纳入前瞻性外部验证的研究。
- PROBAST-AI评估75%的研究为高偏倚风险。
临床怎么用
- 术前AI模型灵敏度高但特异度低，宜作为筛查分诊工具辅助选择进一步检查，不宜单独用于确诊。
- AI辅助宫腔镜子宫肌瘤切除术手术规划有随机对照试验证据支持可减少手术时间和术中出血，可关注其临床落地。
- 术后生育结局预测模型区分能力高但证据偏倚风险高，临床决策中应谨慎解读，等待多中心前瞻性验证。
局限性
- 纳入研究以观察性为主，PROBAST-AI评估75%存在高偏倚风险。
- 外部验证有限，仅术后生育结局预测模型纳入前瞻性外部验证。
- 仅纳入英文文献，排除动物研究和无临床结局的技术可行性报告。
- 纳入研究数量较少（8项），证据体量有限。
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
OBJECTIVE: To systematically evaluate the task-specific performance and clinical translational readiness of artificial intelligence (AI) applications across the preoperative, intraoperative, and postoperative phases of minimally invasive gynecologic surgery (MIGS). DATA SOURCES: A comprehensive search of PubMed, Embase, Scopus, Web of Science, Google Scholar, CENTRAL, and the Cochrane Library was conducted for English-language studies published between January 1, 2015, and August 31, 2025. METHODS OF STUDY SELECTION: This systematic review followed PRISMA guidelines. Studies evaluating AI or machine learning models for predictive or diagnostic tasks in MIGS (laparoscopy, hysteroscopy, or robotic surgery) were included. Animal studies, technical feasibility reports without clinical outcomes, and non-English publications were excluded. Risk of bias was assessed using a validated tool for AI predicition models (PROBAST-AI) framework. TABULATION, INTEGRATION, AND RESULTS: Eight studies involving 6577 patients met the inclusion criteria. AI applications were categorized according to the perioperative phase. Preoperative models for endometriosis and adnexal torsion demonstrated high sensitivity (95%-96%) but limited specificity (11%-45%), suggesting utility as screening triage rather than definitive diagnostic systems. One randomized controlled trial demonstrated that AI-assisted surgical planning for hysteroscopic myomectomy significantly reduced operative time and intraoperative blood loss (p <.05). Postoperative deep learning models predicting fertility outcomes following hysteroscopic adhesiolysis demonstrated the highest discriminative performance (area under curve >0.95) and were the only studies incorporating prospective external validation. PROBAST-AI assessment identified high risk of bias in 75% of studies. CONCLUSION: AI applications in MIGS demonstrate promising task-specific performance; however, evidence is predominantly observational with high risk of bias and limited external validation, constraining clinical translation. Multicenter prospective validation is required before routine adoption.
引用
- Alnajjar E, Hajeri FA, Alsaad L, et al. Artificial Intelligence in Minimally Invasive Gynecological Surgery: A Systematic Review of Task-Specific Performance and Clinical Translational Readiness. J Minim Invasive Gynecol. 2026. PMID: 42035795. doi:10.1016/j.jmig.2026.04.015
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。