心外膜脂肪CT影像组学预测心房颤动:系统综述
机器学习模型AUC为0.73-0.92,敏感度0.50-0.90,特异度0.53-0.98,但分割与特征提取方法异质性大、外部验证少,尚不能直接用于临床。
要点
- CT心外膜脂肪影像组学模型对房颤预测显示一定潜力,AUC可达0.92。
- 研究间分割方法和特征提取差异大,外部验证罕见,推广性受限。
- 影像组学分析需关注心外膜脂肪分割的可重复性和标准化。
核心结论
- 基于心外膜脂肪影像组学特征的机器学习模型在房颤预测方面显示出前景,但异质性限制其泛化能力,尚不能立即纳入临床实践。
研究设计与样本
- 系统综述遵循PRISMA 2020并在PROSPERO注册,检索Embase、PubMed、MEDLINE、Web of Science和Google Scholar,纳入使用CT心外膜脂肪影像组学预测房颤的原始研究。
- 排除病例报告、动物研究、儿科人群及MRI/超声研究,采用PROBAST评估偏倚风险、影像组学质量评分评估方法学质量。
诊断标准与关键数值
- 预测术后房颤、导管消融后房颤复发、房颤检测及房颤亚型分类的机器学习模型(随机森林、逻辑回归、Cox回归、LASSO)性能为AUC 0.73-0.92、敏感度0.50-0.90、特异度0.53-0.98。
临床应用与阅片要点
- 日常阅片时可关注心外膜脂肪的影像组学特征提取,但需注意分割方法和特征提取的标准化对结果的影响。
- 由于外部验证罕见,影像组学模型结果不应直接用于临床决策,需结合其他临床信息综合判断。
局限性
- 研究间分割方法和特征提取存在异质性,外部验证罕见,限制模型的泛化能力和临床推广。
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
RATIONALE AND OBJECTIVES: Atrial fibrillation (AF) is the most common arrhythmia, affecting 1-2% of the general population. Epicardial adipose tissue (EAT) may contribute to AF through its proximity to the myocardium and coronary vessels, as well as its vasocrine and paracrine functions. Previous studies linked EAT thickness and volume with AF incidence. This review consolidates evidence on the correlation between radiomic analysis of EAT and the incidence of AF.
MATERIAL AND METHODS: This systematic review was conducted in accordance with PRISMA 2020 and registered in the International Prospective Register of Systematic Reviews (PROSPERO). Embase, PubMed, MEDLINE, Web of Science, and Google Scholar were searched for primary studies using CT-based EAT radiomics to predict AF. Case reports, animal studies, pediatric populations, and Magnetic Resonance Imaging (MRI)/ultrasound studies were excluded. Risk of bias was assessed with PROBAST, and methodological quality with the Radiomics Quality Score.
RESULTS: Ten studies, including 2952 patients (53.8-81.4% male; mean age range 42.2-74.1 years), met eligibility. Six were retrospective, and four were prospective. The studies evaluated postoperative AF, AF recurrence after catheter ablation, AF detection, and AF subtype classification. Reported machine learning models included random forest, logistic regression, Cox regression, and Least Absolute Shrinkage and Selection Operator (LASSO). Performance metrics ranged from AUC 0.73-0.92, sensitivity 0.50-0.90 and specificity 0.53-0.98. Segmentation methods and feature extraction varied, and external validation was rare.
CONCLUSION: Machine learning models using EAT radiomic features show promise for AF prediction. However, this heterogeneity limits generalizability and precludes immediate incorporation into clinical practice.
引用
- Gladkiy Y, Hasheminia A, Elsherbini A, et al. CT Radiomics of Epicardial Adipose Tissue for Predicting Atrial Fibrillation: A Systematic Review. Acad Radiol. 2026. PMID: 42827062. doi:10.1016/j.acra.2026.08.071
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。