人工智能预测外周动脉疾病截肢及相关结局的系统综述与荟萃分析
汇总AUC为0.88(95%CI 0.83-0.94),但异质性高(I2=95.4%),GRADE证据确定性极低,且无模型经独立外部验证。
要点
- 机器学习模型对PAD截肢相关结局判别性能较高,汇总AUC为0.88
- 证据确定性极低、偏倚风险高且缺乏外部验证,限制临床落地
- 注册库研究汇总AUC高于非注册库研究(0.92 vs 0.80;p=0.004)
核心结论
- 机器学习模型对PAD截肢相关结局具有较高判别性能,但证据确定性极低,外部验证缺失与高偏倚风险限制其临床实施。
研究设计与样本
- 遵循PRISMA/PRISMA-S指南,检索PubMed、Embase、Cochrane Library和Web of Science,检索时间自建库至2026年3月15日。
- 采用随机效应荟萃分析(DerSimonian-Laird法联合Hartung-Knapp-Sidik-Jonkman校正)与HSROC分析,并用GRADE评级证据确定性。
主要发现
- 截肢相关结局的汇总AUC为0.88(95%CI 0.83-0.94;预测区间0.79-0.98),异质性显著(I2=95.4%)。
- HSROC分析得出汇总灵敏度0.855、特异度0.836。
- 注册库研究的汇总AUC高于非注册库研究(0.92 vs 0.80;p=0.004)。
- 所有模型在分析域均为高或不明确偏倚风险,且无一经独立外部验证;GRADE证据确定性为极低。
临床怎么用
- 机器学习模型目前不宜直接用于PAD截肢风险的临床决策,需等待前瞻性多中心外部验证。
- 临床应用前应遵循人工智能报告规范,并关注模型偏倚风险与适用人群。
局限性
- 所有模型分析域偏倚风险高或不明确,且无独立外部验证,GRADE证据确定性极低。
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: Peripheral artery disease (PAD) affects over 236 million individuals globally and carries substantial risks of amputation, major adverse limb events (MALE), and mortality. Existing clinical scoring systems focus on short-term surgical outcomes and lack generalizability. Machine learning (ML) has emerged as a promising tool for PAD outcome prediction. OBJECTIVE: To conduct the first quantitative meta-analysis evaluating the discriminative performance, methodological rigor, and certainty of evidence of ML-based prognostic models for adverse limb outcomes in PAD. METHODS: Following PRISMA/PRISMA-S guidelines, PubMed, Embase, Cochrane Library, and Web of Science were searched from inception through March 15, 2026. Two reviewers independently performed study selection, data extraction, and PROBAST risk-of-bias assessment. Random-effects meta-analysis (DerSimonian-Laird with Hartung-Knapp-Sidik-Jonkman correction) and hierarchical summary ROC (HSROC) analysis were performed; certainty was rated with GRADE. RESULTS: Ten studies were included. The pooled AUC for amputation-related outcomes was 0.88 (95% CI, 0.83-0.94; prediction interval, 0.79-0.98) with substantial heterogeneity (I2 = 95.4%). HSROC yielded pooled sensitivity 0.855 and specificity 0.836. Registry-based studies showed higher pooled AUC than non-registry studies (0.92 vs 0.80; p = 0.004). All models had high or unclear risk of bias in the analysis domain; none underwent independent external validation. GRADE certainty was very low. CONCLUSIONS: ML models demonstrate high discriminative performance for predicting amputation-related outcomes in PAD, but very low certainty of evidence, absence of external validation, and high risk of bias currently limit clinical implementation. Prospective multicenter validation and adherence to AI reporting guidelines are essential before clinical translation.
引用
- Hsu H, Wang MJ, Chu YY, et al. Artificial Intelligence for Predicting Amputation and Related Outcomes in Peripheral Artery Disease: A Systematic Review and Meta-Analysis. Ann Vasc Surg. 2026. PMID: 42874860. doi:10.1016/j.avsg.2026.10.003
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。