Scholarlink

Background <p>Carotid artery plaques are a major contributor to ischemic stroke, yet current assessment methods focusing on luminal stenosis often fail to fully capture plaque vulnerability. High-resolution magnetic resonance vessel-wall imaging (HR–MR-VWI) serves as the reference standard for plaque vulnerability assessment but is limited by cost and availability. Computed tomography angiography (CTA) offers a more accessible, non-invasive alternative for plaque evaluation.</p> Objectives <p>This study aimed to develop and validate a plaque-level predictive model for carotid plaque vulnerability using machine learning and CTA features, with HR–MR-VWI serving as the reference standard. The goal was to identify key imaging features that contribute to plaque vulnerability and assess the performance of machine learning models in predicting plaque instability.</p> Methods <p>A retrospective analysis was conducted on patients who underwent both carotid CTA and HR–MR-VWI within one month. Plaque features were extracted from CTA, including plaque composition, vascular lumen geometry, and perivascular adipose tissue (PVAT). Data were randomly divided into training (70%) and validation (30%) sets. Feature selection was performed using LASSO regression, followed by model development with logistic regression, random forest, XGBoost, and support vector machine (SVM). Hyperparameters were optimized using tenfold cross-validation, and model performance was evaluated with AUC, ROC, calibration curves, precision-recall curves, and confusion matrices. SHAP analysis was conducted to assess feature importance.</p> Results <p>Among 144 participants, the cohort of 265 plaques was analyzed to identify key imaging features predictive of carotid plaque vulnerability. Six key imaging features were selected: Perivascular Adipose Tissue (PVAT), Maximum Diameter Stenosis (MDS), Fibrotic Volume to Non-calcified Volume Ratio (FV/NCV), Fibrotic Volume (FV), Lipid-rich Volume to Non-calcified Volume Ratio (LRV/NCV), and Non-calcified Volume (NCV). The Random Forest model achieved the highest AUC of 0.849 in the validation cohort, with sensitivity of 92.2%, specificity of 69.0%, accuracy of 83.8%, PPV of 83.9%, and NPV of 83.3%. SHAP analysis indicated that PVAT was the most influential feature (mean |SHAP value|= 0.174), followed by MDS (mean |SHAP value|= 0.079) and NCV (mean |SHAP value|= 0.078). The calibration curve demonstrated that Random Forest predictions were well-calibrated, with predicted probabilities closely aligning with observed outcomes. Decision Curve Analysis (DCA) revealed that Random Forest provided the highest net benefit, particularly at higher decision thresholds, confirming its superior clinical utility in predicting plaque vulnerability.</p> Conclusion <p>This study demonstrates the potential of machine learning models using CTA to predict carotid plaque vulnerability. The model's reliance on PVAT highlights its promise as a non-invasive biomarker for plaque instability and vulnerability. Future research should focus on multicenter validation and explore additional imaging modalities to further enhance predictive accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Plaque-level machine-learning prediction of carotid plaque vulnerability on computed tomography angiography

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Juan Long,
  • Zhen Wang,
  • He Zhang,
  • Xiaohan Liu,
  • Xiaonan Sun,
  • Shang Jin,
  • Chen Wu,
  • He Zhang,
  • Zhongxiao Liu,
  • Aiyun Sun,
  • Yankai Meng,
  • Kai Xu

摘要

Background

Carotid artery plaques are a major contributor to ischemic stroke, yet current assessment methods focusing on luminal stenosis often fail to fully capture plaque vulnerability. High-resolution magnetic resonance vessel-wall imaging (HR–MR-VWI) serves as the reference standard for plaque vulnerability assessment but is limited by cost and availability. Computed tomography angiography (CTA) offers a more accessible, non-invasive alternative for plaque evaluation.

Objectives

This study aimed to develop and validate a plaque-level predictive model for carotid plaque vulnerability using machine learning and CTA features, with HR–MR-VWI serving as the reference standard. The goal was to identify key imaging features that contribute to plaque vulnerability and assess the performance of machine learning models in predicting plaque instability.

Methods

A retrospective analysis was conducted on patients who underwent both carotid CTA and HR–MR-VWI within one month. Plaque features were extracted from CTA, including plaque composition, vascular lumen geometry, and perivascular adipose tissue (PVAT). Data were randomly divided into training (70%) and validation (30%) sets. Feature selection was performed using LASSO regression, followed by model development with logistic regression, random forest, XGBoost, and support vector machine (SVM). Hyperparameters were optimized using tenfold cross-validation, and model performance was evaluated with AUC, ROC, calibration curves, precision-recall curves, and confusion matrices. SHAP analysis was conducted to assess feature importance.

Results

Among 144 participants, the cohort of 265 plaques was analyzed to identify key imaging features predictive of carotid plaque vulnerability. Six key imaging features were selected: Perivascular Adipose Tissue (PVAT), Maximum Diameter Stenosis (MDS), Fibrotic Volume to Non-calcified Volume Ratio (FV/NCV), Fibrotic Volume (FV), Lipid-rich Volume to Non-calcified Volume Ratio (LRV/NCV), and Non-calcified Volume (NCV). The Random Forest model achieved the highest AUC of 0.849 in the validation cohort, with sensitivity of 92.2%, specificity of 69.0%, accuracy of 83.8%, PPV of 83.9%, and NPV of 83.3%. SHAP analysis indicated that PVAT was the most influential feature (mean |SHAP value|= 0.174), followed by MDS (mean |SHAP value|= 0.079) and NCV (mean |SHAP value|= 0.078). The calibration curve demonstrated that Random Forest predictions were well-calibrated, with predicted probabilities closely aligning with observed outcomes. Decision Curve Analysis (DCA) revealed that Random Forest provided the highest net benefit, particularly at higher decision thresholds, confirming its superior clinical utility in predicting plaque vulnerability.

Conclusion

This study demonstrates the potential of machine learning models using CTA to predict carotid plaque vulnerability. The model's reliance on PVAT highlights its promise as a non-invasive biomarker for plaque instability and vulnerability. Future research should focus on multicenter validation and explore additional imaging modalities to further enhance predictive accuracy.