Plaque-Level Machine Learning Prediction of Intraplaque Hemorrhage in Carotid Arteries Using Computed Tomography
Juan Long1,2, Xiaohan Liu1,2, He Zhang3
1Department of Radiology, the Affiliated Hospital of Xuzhou Medical University, Xuzhou, China.
Insights
A machine learning model using computed tomography angiography (CTA) shows promise in predicting carotid plaque vulnerability, identifying intraplaque hemorrhage (IPH) with high sensitivity. Key predictors include perivascular adipose tissue (PVAT), maximum diameter stenosis (MDS), and fibrotic volume (FV).
Area of Science:
- Vascular imaging and AI
- Cardiovascular disease research
- Stroke prevention strategies
Background:
- Carotid artery plaques, particularly with intraplaque hemorrhage (IPH), are major causes of ischemic stroke.
- High-resolution magnetic resonance vessel-wall imaging (HR-MR-VWI) is the gold standard but is limited by availability and cost.
- Computed tomography angiography (CTA) is more accessible but its utility in detecting IPH and plaque instability is underexplored.
Purpose of the Study:
- To develop and validate a machine learning model using CTA to predict IPH in carotid plaques.
- To integrate plaque composition, vascular lumen geometry, and perivascular adipose tissue (PVAT) features.
- To assess the model's potential for non-invasive plaque vulnerability prediction in settings without MRI.
Main Methods:
- Retrospective analysis of patients with both carotid CTA and HR-MR-VWI.
- Extraction of plaque features (composition, geometry, PVAT) from CTA.
- Machine learning model development (logistic regression, random forest, XGBoost, SVM) with LASSO feature selection and 10-fold cross-validation.
Main Results:
- The Random Forest model achieved an AUC of 0.679 with 86.0% sensitivity and 81.2% negative predictive value.
- Key predictors identified were PVAT attenuation, maximum diameter stenosis (MDS), and fibrotic volume (FV).
- SHAP analysis confirmed MDS, FV, and PVAT as most influential features; the model showed good calibration.
Conclusions:
- A CTA-based machine learning model shows potential for predicting carotid plaque vulnerability, with PVAT, MDS, and FV as key predictors.
- The model's high sensitivity and NPV position it as a non-invasive screening tool for identifying patients needing further HR-MR-VWI.
- Further refinement with additional clinical and morphological data, along with multi-center validation, is recommended to improve accuracy and clinical applicability.
Background:
Carotid artery plaques, especially those with intraplaque hemorrhage (IPH), are significant contributors to ischemic stroke. Although high-resolution magnetic resonance vessel-wall imaging (HR-MR-VWI) is the gold standard for assessing plaque vulnerability, its limited availability and high cost pose challenges. Computed tomography angiography (CTA) offers a more accessible, non-invasive alternative, but its ability to detect IPH and assess plaque instability remains underexplored. Machine learning techniques have shown promise in improving the prediction of carotid plaque vulnerability using CTA.
Objectives:
This study aimed to develop and validate a machine learning model using CTA to predict IPH in carotid plaques. The model integrated key imaging features, including plaque composition, vascular lumen geometry, and perivascular adipose tissue (PVAT), with HR-MR-VWI serving as the reference standard. The goal was to evaluate the model's potential for non-invasive plaque vulnerability prediction, particularly in clinical settings where MRI is not available.
Methods:
A retrospective analysis was conducted on patients who underwent both carotid CTA and HR-MR-VWI within one month. Key plaque features, including composition, vascular lumen geometry, and PVAT, were extracted from CTA. The dataset was split 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). Hyperparameter tuning was performed using 10-fold cross-validation. Model performance was assessed using AUC, ROC curves, calibration curves, precision-recall curves, and confusion matrices. SHAP analysis was employed to evaluate the importance of each feature.
Results:
In the validation cohort, the Random Forest model achieved an AUC of 0.679, sensitivity of 86.0%, specificity of 45.6%, accuracy of 63.0%, positive predictive value (PPV) of 54.4%, and negative predictive value (NPV) of 81.2%. Feature selection using LASSO regression identified perivascular adipose tissue (PVAT) attenuation, maximum diameter stenosis (MDS), and fibrotic volume (FV) as the most important predictors. SHAP analysis confirmed MDS as the most influential feature, followed by FV and PVAT. The model demonstrated good calibration, with predicted probabilities aligning closely with observed outcomes. Decision Curve Analysis (DCA) showed that the Random Forest model provided the highest net benefit at higher decision thresholds, supporting its clinical potential for predicting plaque vulnerability.
Conclusion:
This study demonstrates the potential of a CTA-based machine learning model for predicting carotid plaque vulnerability, with PVAT, MDS, and FV as key predictors. While the model shows good sensitivity, its moderate specificity suggests the need for further refinement, particularly by incorporating additional clinical and morphological data. Given its high sensitivity and negative predictive value, this model is positioned as a non-invasive screening and triage tool to identify patients who may benefit from subsequent HR-MR-VWI, rather than a standalone definitive diagnostic classifier. Future research should focus on multi-center validation and the integration of clinical data to improve accuracy and clinical applicability.
