Abstract / Summary
Carotid plaque vulnerability is a key predictor of ischemic stroke (IS). We aimed to evaluate the performance of machine learning models combining multimodal ultrasound and clinical features to assess IS risk in patients with carotid plaques. This prospective study enrolled 231 inpatients with ultrasound-verified carotid plaques (December 2022–December 2024), partitioned into IS and non-IS groups based on recent neuroimaging. The dataset was randomly split into training ( n = 162) and testing ( n = 69) cohorts. Candidate variables included multimodal ultrasound (shear wave elastography [SWE], contrast-enhanced ultrasound [CEUS], Doppler, grayscale) and clinical parameters. Feature selection utilized the intersection of univariate logistic regression (LR) and 10-fold cross-validated LASSO regression, followed by stepwise LR minimizing the Akaike Information Criterion. Five machine learning models—LR, k-nearest neighbors, support vector machine, decision tree, and random forest (RF)—were constructed. Performance was evaluated via receiver operating characteristic (ROC) curves and decision curve analysis (DCA). The SHapley Additive exPlanations (SHAP) approach interpreted the optimal model. Five predictive features were selected: plaque thickness, plaque SWE near-shoulder (NS), intraplaque neovascularization, triglycerides, and smoking history. The RF model exhibited the optimal predictive performance, yielding an area under the curve (AUC) of 0.866 in the testing set, with accuracy, sensitivity, and specificity of 0.81, 0.80, and 0.83, respectively. SHAP analysis identified Plaque SWE_NS as the primary contributor to model output, where lower SWE values indicated higher IS risk. Machine learning models integrating multimodal ultrasound and clinical factors demonstrate robust predictive capability for IS events. The optimal RF model facilitates accurate identification and risk stratification, providing valuable adjunctive information for individualized clinical assessment.