Integrated Radiomics of Plaque Wall Morphology and Hemodynamics for Risk-Stratifying Symptomatic Progression in Asymptomatic Carotid Disease

Abstract

Carotid plaque causes approximately 15% of strokes, yet objective criteria for accurately stratifying high-risk asymptomatic patients are lacking. This study aimed to develop a model to predict the symptomatic progression of asymptomatic plaques using multi-modal MR-based radiomics. Radiomics features quantifying plaque vulnerability were extracted from 3D-TSE images (reflecting wall morphology) and 3D-TOF MRA images (indicating hemodynamics). We selected predictive features using mutual information and compared four machine learning models: Logistic Regression, SVM, LightGBM, and Random Forest. The models were evaluated on a dataset of 18 asymptomatic patients (7 who developed symptoms and 11 who did not). The Random Forest model, with features selected via mutual information, achieved the highest discriminative performance with an Area Under the Curve (AUC) of 0.922. Furthermore, SHAP analysis was used to ensure the model's clinical interpretability by visualizing its prediction basis. The novelty of this work is in demonstrating that fusing plaque wall and hemodynamic radiomics effectively stratifies high-risk patients, complementing conventional risk assessment based on stenosis severity.