An explainable CT-based machine learning model integrating carotid plaque and perivascular adipose tissue for predicting symptomatic plaques.

Rationale and objectives: Accurate identification of symptomatic carotid plaques remains a clinical challenge, as conventional imaging focuses mainly on luminal stenosis and lacks sensitivity to plaque vulnerability and perivascular inflammation. This study aimed to develop and validate an explainable machine learning model integrating CT-based radiomics features from carotid plaque and perivascular adipose tissue (PVAT) to identify symptomatic carotid plaques. Materials and methods: 324 patients with extracranial carotid atherosclerosis and stenosis who had undergone head and neck computed tomography angiography (CTA) were retrospectively included. Three-dimensional radiomics features were extracted from segmented carotid plaque, PVAT and combined carotid plaque and PVAT (CP-PVAT) regions. Independent clinical factors were identified using univariate and multivariate logistic regression analyses. A combined model integrating the radiomics signature with selected clinical factors was developed. Models were developed and underwent internally validated using five-fold cross-validation to enhance robustness and minimize overfitting. Model interpretability was assessed using Shapley Additive Explanations (SHAP). Results: The combined model, which integrated CP-PVAT features and clinical factors, achieved excellent discriminative performance, with mean AUCs of 0.903 and 0.904 in the training and testing sets, respectively. It significantly outperformed models based solely on carotid plaque, PVAT, CP-PVAT or clinical factors ( Conclusion: This explainable radiomics-based model, combining CP-PVAT features with clinical risk factors, may serve as a promising tool for identifying symptomatic carotid plaques and supporting individualized cerebrovascular risk assessment.