Abstract
This study aimed to develop and validate a machine learning model for risk stratification of coronary heart disease (CHD) in patients with carotid atherosclerosis, with CHD presence/absence defined as the target outcome variable. A retrospective analysis was conducted on 442 patients diagnosed with carotid atherosclerosis at a tertiary hospital in China between January 1, 2022, and June 20, 2025. Patients were divided into CHD and non-CHD groups based on clinical outcomes. Data encompassing demographics, laboratory results, and vascular imaging findings were collected. Feature selection involved logistic regression (LR), identifying 5 key predictors: age, diabetes, hyperlipidemia, transient ischemic attack (TIA), and the presence of carotid atherosclerotic plaque. Seven machine learning algorithms (LR, XGBoost, LightGBM, random forest, K-nearest neighbors, support vector machine, and stacking ensemble) were trained and evaluated. Model performance was assessed using 10-fold cross-validation, with metrics including area under the curve, accuracy, sensitivity, specificity, and F1 score. Model interpretability was evaluated using Shapley Additive Explanations, while clinical utility was determined through calibration and decision curve analysis. All models demonstrated satisfactory performance, with the LR model achieving the highest area under the curve of 0.838 on the testing set, indicating balanced sensitivity and specificity. Shapley Additive Explanations analysis identified carotid plaque and TIA as the most influential predictors. Calibration and decision curve analysis curves indicated strong agreement between predicted and observed risks, leading to a significant clinical net benefit. An interpretable LR model incorporating age, diabetes, hyperlipidemia, TIA, and carotid plaque enables reliable CHD risk stratification among patients with carotid atherosclerosis. This model serves as a practical, explainable tool for individualized risk assessment and early clinical decision support in this high-risk population.