Baseline data collection was divided into two phases. In the first stage, a baseline survey was conducted from 2010 to 2012. Through stratified cluster random sampling, the Uyghur population in Jiangbazi Township, Jiashi County, Kashi Prefecture, and southern Xinjiang, and the Kazakhs in Nalati Township, Xinyuan County, Ili Prefecture, and northern Xinjiang were selected. In the second stage, a baseline survey was conducted from September to December 2016, and the Uyghur population of the 51st Regiment of the Third Division of the Xinjiang Corps was selected as the research cohort through stratified cluster random sampling. A total of 19,549 people who were aged ≥18 years and lived in the local area for >6 months were included in the study. The exclusion criteria included CVD at baseline, those lost to follow-up, and those with incomplete blood information. Follow-up continued until December 2017 for the first stage (median: 6.07 years) and until August 2021 for the second stage (median: 4.94 years). According to the inclusion and exclusion criteria, 5,335 and 7,357 people were included in the first and second stages, respectively, for a total of 12,692 individuals (Supplementary Figures 1.1, 1.2). Then do data analysis (Supplementary Figure 1.3). All participants provided written informed consent. This study was approved by the Ethics Committee of the First Affiliated Hospital of Shihezi University School of Medicine (NO. SHZ2010LL01).

Data were collected via questionnaire, physical examination, and laboratory examination. Questionnaires were completed face-to-face. Anthropometric measurements such as height, weight, waist circumference (WC), hip circumference (HC), and blood pressure were obtained by trained professionals. Blood pressure was measured three times for each participant using a mercury sphygmomanometer after 5-min seated rest, and the average value was calculated. Hypertension was defined as systolic blood pressure (SBP) of ≥140 mmHg or diastolic blood pressure (DBP) of ≥90 mmHg. Prehypertension was defined as 140 > SBP ≥ 120 mmHg or 90 > DBP ≥ 80 mmHg (25). Synthetic indices were calculated based on anthropometric measurements: BMI [weight (kg)/height² (m)]; BAI (HC/height^1.5-18); pulse pressure (SBP–DBP); and waist-to-hip ratio [WHR; WC (cm)/HC (cm)]. A family history of diabetes was defined as a history of diabetes in at least one parent or sibling; the same criteria were used for a family history of stroke and coronary heart disease (CHD). Current smokers were defined as participants who had been smoking for >6 months (26). Drinking was defined as consuming alcoholic beverages (beer, red wine, and white wine) ≥2 times a month (27). A 5 ml fasting blood sample was collected from each subject and levels of the fasting blood glucose (FBG), triglycerides (TGs), high-density lipoprotein cholesterol (HDL-C), total cholesterol (TC), low-density lipoprotein cholesterol (LDL-C), and other indicators were obtained using an automatic biochemical analyser (Olympus AU 2700; Olympus Diagnostics, Hamburg, Germany) at the First Affiliated Hospital of Shihezi University School of Medicine. In this study, individuals with diabetes (28) were defined as having FBG level of ≥7.0 mmol/L and 2-h postprandial blood glucose level of ≥11.1 mmol/L, a previous diabetes diagnosis, and use of blood sugar control drugs. We also calculated other synthetic indices, including TyG, (TG [mg/dl]^*FBG [mg/dl]), (LAP) (men: [WC-65]^*TC [mmol/L]; women: [WC-58]^*TG [mmol/L]); lipoprotein combine index (LCI) (TC^*TG [mmol/L]^*LDL-C/HDL-C); atherogenic index (AI) (TC [mmol/L]-HDL-C)/HDL-C); atherogenic index of plasma (AIP) (Log[TG/HDL]); LpH (LDL-C/HDL-C ratio); and bilirubin comprehensive index (THT) (TC [mmol/L]/[HDL-C+TBIL (μmol/mL)]).

There were some missing values in the database, and direct deletion of missing values resulted in the loss of sample information. Since there were a few variables with missing values in this study, continuous variables were filled using the mean, while categorical variables were filled using the mode. By standardizing continuous variables, categorical variables were processed by one-hot encoding to reduce the influence of different variable units and quantity levels on the analysis. For the description of missing variables in this study, see Supplementary Table 1.

The diagnostic criteria for CVD (29) pertained to the detection of ischaemic heart disease, cerebrovascular disease, and related diseases [International Classification of Diseases (ICD)-9: code 390–495]; hospitalization; or death due to CVD (ICD-10) during the follow-up period. Data regarding patient questionnaire answers, medical records, and the diagnosis of CVD during the follow-up period were obtained and recorded. If the same type of CVD event occurred more than once in a patient, the first occurrence of CVD was the final event. The time of onset was recorded. Self-reported patients needed to provide proof of their clinical diagnosis.

Logistic regression belongs to probabilistic nonlinear regression and is one of the most widely used classification models. Logistic regression usually uses regularization to optimize the model. The adjustable parameters include inverse regularization parameters and methods (30). By adding a regularization coefficient to Logistic regression, the parameters of the variable are sparse, so that the weight of most of the feature vectors is 0, thereby reducing the dimension of the variable. SVM is currently one of the most common ML algorithms that can effectively solve the classification problem of small samples and nonlinear and high-dimensional data. It classifies samples by finding a set of hyperplanes in a high-dimensional space, and the samples closest to the hyperplane are called support vectors. When the training data are inseparable, this problem can be solved using the kernel trick (31).That is, the original features of the samples are mapped to a higher dimensional space that makes the samples linearly separable through the mapping function. The RF algorithm is an ensemble learning algorithm based on the decision tree algorithm. The basic idea is to integrate weak classifiers into a more robust model (32). AdaBoost (33) is an ensemble learning algorithm based on boosting. The algorithm first builds a weak learner based on the training data and then according to AdaBoost, increases the weight of the samples that were misclassified by weak learning in the previous round. Then, it reduces the weight of the correctly classified samples, loops this process until the weak learner reaches the specified value, and then linearly combines all weak learners to obtain the final strong classifier by weighted majority voting. In this study, both random forest and Adaboost are ensemble learning algorithms based on decision trees. The decision tree algorithm selects variables by evaluating the characteristics and depth of dividing nodes, reducing the dimension of variables. The integrated model has better generalization error and can effectively reduce the overfitting combination phenomenon.

The datasets were randomly divided into training datasets (927CVD/10153) and test datasets (249CVD/2539). The KS test was performed on the training and test datasets, and the P-values were both >0.05. The ratio of the training and test datasets was 8:2. We considered four variable selection methods: forward partial likelihood estimation (FLR) with logistic regression (LR), lasso regularization with logistic regression (Lasso-LR), permutation-based selection with random forest (RF), and characteristic importance with RF. Variables were established using a subset of algorithms, such as L1-LR, RF, SVM, and AdaBoost. A prediction model of each algorithm was then established. The optimal prediction model of the same algorithm was then selected by discrimination and calibration, and the most suitable prediction model for the population was obtained by comparing the discrimination, calibration, and clinical effectiveness of the optimal prediction models of different algorithms.

The discrimination of the model was determined by comparing the area under the receiver operating curve (AUC), Net Reclassification Index (cNRI), and Integrated Discrimination Improvement Index (IDI) (34) between models, and the calibration degree was compared by calculating the Brier Score (BS) and Homser–Lemeshow χ² (35, 36). This study evaluated the clinical validity of the model using decision curve analysis (DCA) (37). The horizontal axis of the decision curve represents the threshold probability and vertical axis represents the net benefit obtained after subtracting the harm from the benefit under the threshold probability. Using DCA to determine the net benefit that can be obtained using the model to screen high-risk groups compared with assuming that all participants are high-risk groups of CVD and implanting undifferentiated interventions, followed by calculating the net benefit without increasing the number of positive results, can reduce unnecessary interventions.

To avoid over-fitting the problem of the model in the process of model selection and hyper-parameter tuning, we used a 10-fold cross-validation to optimize the parameters of the training set and subsequently selected the optimal model. This method divided the training data in 10 equal, non-repeated parts, nine of which were used for model training, and the remaining one was used for model verification. This process was repeated 10 times, and combination of Bayesian optimisation and grid search was used to select the optimal hyperparameters. The AUC was used as the model selection criterion to determine the hyperparameter value that optimized the model predictive performance. Afterwards, we used the optimal hyperparameter value. We built the model on all training data sets. Finally, the independent test data set was used to make a final evaluation of model performance.

Since machine learning algorithms, such as SVM output, predicted CVD occurrence by default, they did not directly predict CVD probability. We used the Platt scaling method (38) to calibrate the predicted probabilities output using the four models for more accurate prediction of CVD risk and identification of high-risk groups. The data used in this study were unbalanced to enable the use of the threshold probability movement method. The default 0.5 of the model was not used as the standard for dividing the incidence of CVD. However, the optimal threshold probability of each model was determined according to the Youden Index, which was the basis for dividing the high-risk population of CVD. All statistical analyses were performed using the Python 3.7 or R version 4.0. A two-sided test with a P-value of <0.05 was considered statistically significant.

A total of 12,692 people (6,264 men, 6,398 women; average age 41.24 years) were included in this study. A total of 1,176 CVD events were observed during a median follow-up of 4.94 years. The cumulative incidence was 9.26%. Compared with people without CVD events, those with CVD showed a higher trend in study indicators, such as age, BMI, TC, alkaline phosphatase (ALP), WC, and HC. Moreover, subjects with high blood pressure and type 2 diabetes were also at a higher risk of CVD development. The comparison of different characteristics between participants with CVD and those without training and test datasets listed is shown in Supplementary Tables 2.1,2.2.

The research database included demographic characteristics, physical examination findings, and serology results. There were 62 variables in total. After removing the missing ratio of ≥50% and 11 variables unrelated to the research, a total of 51 variables were included. The following methods were used to filter and establish a subset of variables: FLR-LR (22 variables) and Lasso-LR (34 variables). The top 35 variables were selected according to the built-in random forest importance. The top 30 variables were subsequently selected as the screening subset according to permutation feature importance of RF. The variable subsets formed by the selected variables using the four methods are shown in Supplementary Tables 3–6.

To further explore the predictive performance of different variable subsets on different algorithms, we used the above variable subsets and the full variable set to build predictive models using different algorithms to find the algorithm based on the optimal model. Through Bayesian optimization and grid search, the hyperparameter values with the best prediction performance of each model were selected (Supplementary Tables 7.1–7.4). The AUC values of different algorithms in the training and test datasets are shown in Supplementary Table 8. There was no risk of overfitting and, to comprehensively consider the results of discrimination and calibration, this study concluded that the optimal models based on the four algorithms were Lasso-AdaBoost, FLR-L1-LR, FLR-RF, and FLR-SVM (Supplementary Tables 9.1–9.4).

The predictive performance indicators of the optimal models for each algorithm are listed in Table 1. All models have a moderate or higher (AUC value between 0.798 and 0.817) distinguishing ability. The AUC of FLR-L1-LR, FLR-SVM, FLR-RF, and Lasso-AdaBoost was 0.817 (95% CI, 0.801–0.832), 0.814 (95% CI, 0.798–0.829), 0.804 (95% CI, 0.788–0.820), and 0.798 (95% CI, 0.782–0.81), respectively. The receiver operating characteristic (ROC) curve of the prediction model is shown in Figure 1.

Compared with other optimal models, the FLR-L1-LR model performed better in terms of Youden index, specificity, and PPV when the optimal threshold was 0.11. BS and Homser–Lemeshow χ² also demonstrated that the FLR-L1-LR model was better than others. In the FLR-L1-LR model, 26.7% of the participants were identified as high risk for CVD development (Table 1). The results of the calibration curve showed that FLR-L1-LR, FLR-SVM, Lasso-AdaBoost, and FLR-RF predicted the number of patients with CVD to be 234.12, 234.05, 230.55, and 223.93, respectively. The corresponding predicted CVD events/objective CVD events (P/O) values were 94.02, 94.00, 92.59, and 89.93, respectively (Figure 2).

To further select a prediction model suitable for this population, we compared the differences between the AUC value, IDI, and cNRI of the optimal models. We found that the AUC values of FLR-L1-LR and FLR-SVM were similar (P > 0.05), and both were higher than the AUC values of Lasso-AdaBoost and FLR-RF (P < 0.05). The reclassification capabilities of each model were compared with that of the FLR-L1-L model. The cNRI values of FLR-SVM and Lasso-AdaBoost values were 0.278 and 0.208, respectively. Compared with the FLR-L1-LR model, the Lasso-AdaBoost and the FLR-SVM models had a correct classification rate of 21 and 28%, respectively. Similarly, FLR-SVM was compared with Lasso-AdaBoost in terms of the proportion of correct classification. The FLR-SVM had a 17% increased proportion of correct classification compared with that of the Lasso-AdaBoost. The difference between the reclassification capabilities of the remaining models was not statistically significant. The results of the comprehensive discrimination ability of each model, from best to worst, were FLR-L1-LR > FLR-SVM > FLR-RF > Lasso-AdaBoost. This is described in Table 2.

The clinical effectiveness of FLR-L1-LR, FLR-SVM, FLR-RF, and Lasso-AdaBoost based on the results of the decision curve are shown in Figure 3. It is evident that the clinical application value of the FLR-L1-LR model is higher than that of FLR-SVM, Lasso-AdaBoost, and FLR-RF (Figure 3, Table 3). Under the optimal threshold, we assumed that all participants were in a high-risk group for CVD. We then administered undifferentiated interventions for primary and secondary prevention. The net benefit of using the FLR-L1-LR model was 0.061. This showed that without increasing the positive results, 49 out of every 1,000 people could avoid unnecessary interventions.

Previous studies indicated that compared with FRS and PCE, the ML algorithm could better determine the nonlinear and complex relationships between variables and outcomes. Furthermore, the ML algorithm identified potential risk factors more effectively (39–41). We further analyzed the relative relationship among the importance rankings of the algorithm variables using the coefficients of variables that could not be obtained based on the Gaussian kernel function. Therefore, this study only highlights the importance of the optimal model variables established by the AdaBoost, RF, and L1-LR algorithms to compare the ability of each variable to predict the incidence of CVD (Figure 4). This study found that the risk factors for CVD included factors that reflected the degree and type of body obesity, such as age, sex, ethnicity, DBP, HDL-C level, TC level, BAI, and BMI. Risk factors also included those that reflected glucose and lipid metabolism, such as TyG, LpH level, AI, and occupation type. The indicators were also risk factors for CVD and could predict CVD risk.