机器学习预测腹主动脉瘤生长与破裂的系统评价
生长预测AUC为0.79至0.93,均方根误差0.94至2.4 mm;破裂预测AUC为0.75至0.91。内部性能较好,但外部验证少见,报告规范性与公平性不足。
要点
- 机器学习模型对腹主动脉瘤生长与破裂的预测内部性能较好,但外部验证少见。
- 生长预测AUC 0.79至0.93,均方根误差0.94至2.4 mm;破裂预测AUC 0.75至0.91。
- 临床落地需统一特征定义、稳健外部验证及前瞻性评估对决策的影响。
核心结论
- 机器学习模型在改善腹主动脉瘤风险预测方面展现潜力,但临床实施仍需标准化特征定义、稳健外部验证及前瞻性评估对决策的影响。
研究设计与样本
- 系统评价,遵循PRISMA,检索PubMed/MEDLINE、Embase与Web of Science(截至2025年3月),纳入开发术前机器学习模型预测腹主动脉瘤生长或破裂的研究。
- 数据提取采用CHARMS清单,偏倚与适用性采用PROBAST+AI评估,报告质量采用TRIPOD+AI评估。
主要发现
- 生长模型常用支持向量机、卷积神经网络与梯度提升,纳入直径、腔内血栓厚度、迂曲度、壁面剪切应力、影像组学与临床变量。
- 生长预测的曲线下面积(AUC)为0.79至0.93,均方根误差为0.94至2.4 mm。
- 破裂模型使用多种分类器与相似的多模态输入,AUC为0.75至0.91。
- 内部性能较好,但外部验证不常见;PROBAST+AI显示61%的研究偏倚风险低,但适用性问题常见。
- TRIPOD+AI显示报告不一致,尤其在透明度与公平性方面。
临床怎么用
- 机器学习模型有望辅助腹主动脉瘤风险分层,但当前不宜直接替代基于直径的监测决策。
- 临床使用前需关注模型的外部验证情况、特征定义标准化与报告透明度。
局限性
- 外部验证不常见,多数研究仅报告内部性能。
- 适用性问题常见,透明度与公平性报告不一致。
- 纳入研究样本量有限,生长模型合计745例、破裂模型合计1394例。
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: Abdominal aortic aneurysm (AAA) rupture remains a major cause of mortality, and diameter-based surveillance is an imperfect predictor of risk. Some aneurysms rupture below operative thresholds, whereas others remain stable despite exceeding them. Machine learning (ML) may improve risk stratification by integrating geometric, hemodynamic, radiomic, and clinical data. We performed a systematic review to evaluate ML models predicting AAA growth and rupture, characterize their performance, and assess readiness for clinical translation. METHODS: A Preferred Reporting Items for Systematic Reviews and Meta-Analyses-compliant search of PubMed/MEDLINE, Embase, and Web of Science (through March 2025) identified studies developing preoperative ML models for AAA growth or rupture. Data extraction followed the CHecklist for Critical Appraisal and data extraction for systematic Reviews of prediction Modelling Studies checklist. Risk of bias and applicability were assessed using the Prediction model Risk of Bias ASsessment Artificial Intelligence Tool, and reporting quality was assessed using Transparent Reporting of a multivariable prediction model for Individual Prognosis Artificial Intelligence. RESULTS: Eighteen studies met inclusion criteria: 13 addressed growth (n = 745 patients) and five rupture (n = 1,394). Growth models commonly employed support vector machines, convolutional neural networks, and gradient boosting, incorporating diameter, intraluminal thrombus thickness, tortuosity, wall shear stress, radiomics, and clinical variables. The area under the curve for growth prediction ranged from 0.79 to 0.93, with root mean square errorof 0.94-2.4 mm. Rupture models used diverse classifiers with similar multimodal inputs and area under the curve from 0.75 to 0.91. Although internal performance was promising, external validation was uncommon. PROBAST + AI demonstrated low risk of bias in 61% of studies but frequent applicability concerns. Transparent Reporting of a multivariable prediction model for Individual Prognosis Artificial Intelligence revealed inconsistent reporting, particularly regarding transparency and fairness. CONCLUSION: ML models show promise for improving AAA risk prediction. Clinical implementation will require standardized feature definitions, robust external validation, and prospective evaluation of impact on decision-making.
引用
- Aljobeh A, Parthasarathy P, Liao J, et al. Machine Learning Models Used to Predict Abdominal Aortic Aneurysm Growth and Rupture: A Systematic Review and Critical Appraisal. Ann Vasc Surg. 2026. PMID: 41895595. doi:10.1016/j.avsg.2026.03.026
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。