影像组学提升股腘动脉慢性完全闭塞血管内治疗成功预测:一项概念验证研究
本研究回顾性纳入76例有症状的外周动脉疾病患者,均为股浅动脉和/或腘动脉慢性完全闭塞,评估将影像组学特征与临床数据整合能否提升AI模型预测血管内开通成功的性能。在CatBoost、随机森林、XGBoost和TabPFN四种架构中,XGBoost在联合数据集上表现最佳,MCC为0.71,特异度0.9,AUC为0.84(95%CI 0.7-0.9)。SHAP分析显示模型输出受影像组学和临床特征共同影响。
要点
- 联合影像组学与临床数据的XGBoost模型预测股腘动脉CTO血管内开通成功,AUC达0.84(95%CI 0.7-0.9)。
- 该模型特异度为0.9,MCC为0.71,适用于类别不平衡数据的评估。
- SHAP分析提示影像组学特征与临床特征共同影响模型决策。
核心结论
- 整合影像组学与临床数据在提升AI模型预测介入放射学手术结局方面显示出显著潜力,XGBoost在联合数据集上表现最优,且模型能正确识别临床重要参数。
研究设计与样本
- 回顾性收集76例有症状的外周动脉疾病患者,均为股浅动脉和/或腘动脉慢性完全闭塞,评估CatBoost、随机森林、XGBoost和TabPFN四种机器学习架构在临床数据、影像组学特征及联合数据集上的表现。
诊断标准与关键数值
- 最佳模型为联合数据集训练的XGBoost,MCC为0.71,特异度0.9,AUC为0.84(95%CI 0.7-0.9);因类别不平衡采用MCC评估,并报告AUC和F1分数作比较。
局限性
- 样本量为76例,作者提出需扩大样本量。
- 作者提出需处理数据不平衡问题。
- 作者提出需在多样化患者人群中进一步验证这些发现。
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)。
英文原文(摘要)· 点击展开
OBJECTIVE: This study aims to investigate if integrating radiomics features with clinical data can enhance the performance of an AI model that predicts surgical outcomes among patients undergoing interventional radiology procedures, such as endovascular crossing for chronic total occlusions (CTOs) in peripheral arteries. The proposed model and all code are open-access.
METHODS: The dataset included 76 retrospectively collected patients with symptomatic peripheral arterial disease, specifically those with CTOs in the superficial femoral artery (SFA) and/or popliteal artery (POPA). Four machine learning architectures: CatBoost, Random Forest, XGBoost, and TabPFN were evalauted on three datasets: clinical data (A), radiomics features (B) and combined (C). Model performance was evaluated using Matthews Correlation Coefficient (MCC) due to class imbalance, with additional metrics like AUC and F1-score reported for comparison. SHAP values were used to interpret the best-performing model's decisions.
RESULTS: The best performing model was the XGBoost trained on the combined dataset, achieving 0.71 MCC, 0.9 specificity and 0.84 (95% CI 0.7-0.9) AUC. The SHAP analysis revealed that the model's output is strongly influenced by both radiomics and clinical features.
CONCLUSION: The integration of radiomics and clinical data has shown significant potential in improving the performance of AI models that predict outcomes of interventional radiology procedures. Among the tested architectures, XGBoost demonstrated superior performance when trained on a combined dataset featuring both radiomics and clinical data. Importantly, the AI model also managed to correctly identify the most clinically important parameters from the data. Moving forward, expanding the sample size and addressing data imbalance could further validate these findings in diverse patient populations.
引用
- Vrettos K, Kasioumi K, Galanakis N, et al. Radiomics enhance the prediction of endovascular treatment success for femoropopliteal chronic total occlusions: a proof-of-concept study. Eur J Radiol. 2025. PMID: 41166916. doi:10.1016/j.ejrad.2025.112496
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。