The study cohort was derived from the routine health check-up system of the First Affiliated Hospital of Shandong First Medical University in Jinan, China. All the participants were free of CAS at the first check-up and underwent three health checks during the follow-up. Individuals who had been diagnosed with coronary heart disease, previous coronary heart disease, cerebral ischaemia, cerebral infarction, cerebral artery stenosis, cerebral artery spasm, coronary artery stenosis, coronary atherosclerotic heart disease, and those with missing information were excluded. CAS was diagnosed by carotid B-mode ultrasonography as a carotid intima-media thickness of 1.0 mm or greater or plaque formation. The study was approved by the Ethics Committee of the First Affiliated Hospital of Shandong First Medical University, and informed consent was obtained from all eligible participants.

The study variables consisted of three sets of data: demographic data, laboratory indicators, and clinical history. All the individuals in this study cohort underwent anthropometric and laboratory tests. The height and weight of the participants were measured while they were wearing light clothing and no shoes. Peripheral blood samples were collected from the subjects after an overnight fast, and the variables included blood urea nitrogen (BUN), lymphocyte percentage (LYM), aspartate aminotransferase (AST), red cell volume distribution width standard deviation (RDW-SD), red blood cell count (RBC), mean corpuscular haemoglobin concentration (MCHC), mean platelets (MPV), fasting blood glucose (GLU), platelet count (PLT), eosinophil percent (PEOS), white blood cell count (WBC), and carcinoembryonic antigen (CEA). Disease history was also collected, such as carotid stenosis (CS), diabetes mellitus (DM), and hypertension. All the measurements were collected following the same standard procedures.

To address the issue of data imbalance, we applied the synthetic minority oversampling technique (SMOTE) in our study. Since the number of individuals without CAS was larger than those with CAS, SMOTE was employed to generate synthetic samples of the minority class (22). The new synthetic records were generated using the existing samples of the minority class via linear interpolation. After we obtain new minority sample data, a balanced dataset can be obtained by merging with majority samples.

Variable selection is an important step in the application of machine learning to ensure that the most relevant predictors are used. Both filtering and embedded feature selection methods were used to select the predictors. The variables were first selected using univariate logical regression with a threshold P value of 0.1. Second, we applied three tree-based machine learning methods-random forest (RF), eXtreme Gradient Boost (XGBoost), and gradient boosting decision tree (GBDT))-to assess the importance of each variable. These methods are well-suited for identifying important variables due to their ability to capture complex interactions and non-linear relationships. Variables were ranked based on their importance scores from these models. Finally, the correlation coefficients of the continuous variables were calculated to address multicollinearity, which can distort the model’s estimates and reduce interpretability. Features with low importance among the highly related variables were eliminated for determining the predictors.

The baseline characteristics were assessed for CAS and non-CAS patients during the follow-up. Continuous variables were described by the mean and standard deviation (SD), and categorical features were described as proportions; we compared the baseline features using the t test and the chi-square test. To predict the probability of CAS, we employed five machine learning models: logistic regression (LR), support vector machine (SVM), RF, XGBoost, and GBDT (11, 23–26). We then used a stacking ensemble model, specifically the super learner, which combines these individual models by assigning them different weights to optimize predictive performance (27). The final predicted value is a weighted sum of the individual model predictions, where the weights are determined to minimize the cross-validation risk (28). Figure 1 shows the framework of the super learner ensemble model. We compared the predictive performance of the super learner against the individual models (LR, RF, SVM, XGBoost, and GBDT).

We optimized the hyperparameters using a random search method, which is a traditional and efficient technique for tuning in classification methods (29).This process was conducted within a 5-fold cross-validation framework. Specifically, each model configuration was trained on four folds and validated on the remaining fold. This cycle was repeated five times, with each fold serving as the validation set once, to ensure a comprehensive evaluation across the entire dataset. We performed the random search 1,000 times, selecting the hyperparameter combination with the highest average areas under the receiver operating characteristic curve (AUC).

The performance of the prediction model was validated using both the testing and hold-out validation set. Several metrics were used to evaluate the performance of the prediction models: accuracy, precision, recall, F1 score, and AUC. Compared to commonly used performance metrics, the AUC better reflects model performance in unbalanced datasets. Hence, the AUC was the main metric, while the others were considered of secondary priority.

To solve the “black box” problem in machine learning, we report the feature importance ranking of each predictor based on SHAP values (22). SHAP values are useful for explaining the prediction of a machine learning model by computing the contribution of each feature to the prediction. Kernel-based SHAP values were used to rank the variables in terms of their ability to predict the CAS, which is an additive feature attribution method using kernel functions enabling consistent explanation of feature importance (30).

A total of 1669 participants were included in this study, including 1426 (85.4%) males and 243 (14.6%) females. A total of 441 participants were diagnosed with CAS during the follow-up, including 395 men and 46 women. A total of 1228 participants were not diagnosed with CAS. The SMOTE method was used to address the sample imbalance problem. Figure 2 shows the roadmap of the data processing. All the samples were first divided into a training set and a hold-out validation set. Specifically, 5% of the subjects were randomly selected in advance as the validation set, and the remaining 95% were used for model construction; 419 patients with CAS and 1167 non-CAS patients were included. Of the remaining data, 70% of the data were used as the training set where SMOTE resampling was applied to address class imbalance. The remaining 30% was used as the testing set. In the original dataset, the ratio of non-CAS to CAS was approximately 2.78 to 1, which was adjusted in the training set to approximately 1 to 1.

A total of 28 variables with a P value < 0.1 were retained in the univariate logistic regression and subsequently included in the three machine learning models. Supplementary Figure S1 shows the variable importance rankings for RF, GBDT, and XGBoost. A total of 22 of these variables were common among all three machine learning algorithms (please see the Supplementary File ). The correlations between the selected continuous variables are shown in Supplementary Figure S2 in the Supplementary File . Variables with high correlations and the lowest importance were removed; thus, five variables were excluded. Finally, 17 variables were selected, including age, sex, BUN, LYM, AST, RDW-SD, RBC, MCHC, MPV, GLU, PLT, PEOS, WBC, CEA, CS, hypertension and DM.

Table 1 summarizes the baseline characteristics of the incident CAS status. Overall, individuals who developed CAS were more likely to be male CS, DM, hypertension, older age, MCHC, BUN, RDW-SD, MPV, GLU, PEOS, WBC, and CEA and lower PLT, LYM, AST, and RBC at baseline; these variables were significantly different at a level of 0.1.

The super learner algorithm creates an optimal weighted average of the five models (LR, RF, SVM, XGBoost, and GBDT). Supplementary Table S1 in the Supplementary File depicts the weight coefficients of the super learner model. Supplementary Table S2 shows the hyperparameters used in the models. The weight coefficients of LR and SVM are 0, indicating that they were not used in the super learner model, while the coefficient of RF is 0.820, which is much greater than that of the other four models, indicating that RF contributes most in the prediction model.

Table 2 shows the predictive performance of the six models on the testing set and the hold-out validation set. It can be seen that the predictive performance varies across the five models. SVM has the highest precision, while its performance in the validation set is inferior. Logistic regression had the lowest performance in the testing set. These results also indicate the reason that the two methods are not selected in the super learner model. Combined with the advantages of RF, XGBoost and GBDT, the performance of the super learner was improved. The super learner model had the optimal performance measures in the validation set, with considerable performance in the testing set. The ROC curves of the different machine learning models are shown in Supplementary Figure S3 . The super learner has the largest ROC curve area in the validation set, and its AUC is 0.861. Overall, the predictive performance of the super learner model is superior to that of the other five models, especially regarding the overfitting problem.

In this paper, the SHAP value was used to quantify the impact of each variable on the prediction of CAS, and the results are shown in Figure 3 . Figure 3A shows the contribution of all the features to the prediction, which was sorted according to the average SHAP values. CS and age are the two most important predictors with the largest SHAP values, followed by RDW-SD, GLU and hypertension.

To further explain how each variable affects the occurrence of CAS, we illustrate two sample cases. Figures 3B, C depict the SHAP value of each variable for individuals I and II. The blue bars on the left (SHAP value less than 0) indicate variables that reduce the probability of the individual being predicted as CAS; the orange bars on the right (SHAP value greater than 0) indicate variables that increase the probability of the individual being predicted as CAS. Larger areas indicate greater impacts of that factor. For individual I, diabetes and an increase in glucose are the main reasons for the increased risk of CAS. Due to the absence of CS, relatively young age and other variables with negative impacts, the predicted probability of CAS for individuals is low. In contrast, for individual II, CS is the main reason for the increased risk of CAS, and most of the variables have a positive impact in predicting CAS. The probability of CAS for individual II is only slightly reduced by the absence of hypertension, diabetes and young age; thus, this individual is more likely to develop CAS in the future. Therefore, through the SHAP framework, we can directly determine the main causes for the increased individual probability of CAS; thus, corresponding interventions could be taken to reduce the risk.