Cardiovascular disease (CVD) is the leading cause of morbidity and mortality worldwide. Carotid plaque (CP), as an important marker of atherosclerosis, can be used to predict cardiovascular events. However, conventional ultrasound screening is limited in large-scale populations. Exploring machine learning-based prediction models may facilitate the early identification of CP and the prevention and control of CVD.

To explore the construction of a machine learning prediction model for CP risk among high-risk CVD populations.

Data from 5 666 participants in Nanjing, collected between September 2023 and August 2025 as part of the "Jiangsu Province Early Screening and Comprehensive Intervention Program for High-Risk Cardiovascular Populations" were analyzed. The dataset was randomly divided into training and testing sets at a 7∶3 ratio. Variables selected via Least Absolute Shrinkage and Selection Operator (LASSO) regression were used as independent variables, with CP as the outcome. Logistic Regression (LR), Categorical Boosting (CatBoost), Extreme Gradient Boosting (XGBoost), Elastic Net (EN), and Support Vector Machine (SVM) models were constructed. Model performance was comprehensively evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score on the test set. The Shapley Additive Explanations (SHAP) method was employed to interpret the best-performing model.

A total of 5 666 participants were included, comprising 3 005 males (53.0%) and 2 661 females (47.0%). There were 2 027 (35.8%) in the normal group and 3 639 (64.2%) in the CP group. LASSO regression identified 21 key variables with non-zero coefficients, including sex, region, electrocardiogram results, age, serum uric acid, estimated glomerular filtration rate, glycated hemoglobin A_1c (HbA_1c), homocysteine (Hcy), non-high-density lipoprotein cholesterol, apolipoprotein B (ApoB), creatine kinase-MB (CK-MB), diastolic blood pressure, pulse pressure, urine pH, urine specific gravity, education level, current occupation, smoking status, alcohol consumption, physical activity, and sleep quality. Among the models, XGBoost demonstrated the best overall predictive performance in the training set (AUC=0.747, accuracy=0.693, sensitivity=0.732, specificity=0.623, F1 score=0.754). SHAP analysis ranked the feature importance as follows: age, Hcy, CK-MB, region, HbA_1c, ApoB, and sex.

The CP risk prediction model based on the XGBoost machine learning algorithm showed the best performance among high-risk CVD populations. This model may provide a simple and low-cost tool for primary healthcare institutions to facilitate early identification of individuals with CP, support timely intervention at the early stage of atherosclerosis, and help reduce the risk of adverse CVD outcomes.