Prediction of coronary heart disease in rural Chinese adults: a cross sectional study

Data collection

In this research, 10,691 residents from two townships (Lang Gong Temple township and seven Li Ying township) of Xinxiang county, Henan province, in central China were selected by multistage randomized cluster sampling and investigated with questionnaire survey, physical examination, and lab tests. The participants ranged in age from 18 to 79, with 4,352 males and 6,339 females. All participants received a written informed consent before participating in this study. Questionnaire surveys included demographic information (name, ID, address, contact number, gender, ethnicity, occupation, date of birth, civil state, income, and medical services, etc.), lifestyles (smoking status, alcohol intake, eating habits, tea drinking and athletic activity, etc.), sleep situation, personal and family history of 20 common chronic non-communicable diseases (hypertension, hyperlipidemia, diabetes, CHD, stroke, emphysema, gastrointestinal disease, allergic rhinitis, chronic bronchitis, asthma, kidney disease, gallbladder disease, chronic obstructive pulmonary disease, cancer, liver disease, chronic hepatitis, pancreatic disease, skin disease, gout, and tuberculosis), menstrual, reproductive history and gynecological diseases (for women) and psychologic status. Anthropometric data included height, waistline, hipline, weight, body fat rate (BFR), basal metabolic rate (BMR), visceral fat index (VFI), grip strength, and heart rate (HR). Lab tests included blood routine (24 items), blood lipids, blood glucose (fasting blood glucose (FBG), fasting insulin levels (FINS), and glycated hemoglobin A1c (HbA1c)), hepatic and kidney function, resting blood pressure, urine routine, heavy metal in urine, stool routine, bone mineral density, lung function, 12-lead resting electrocardiography, chest X-ray, and abdominal ultrasonography.

The datasets of this study were derived from a cross-sectional survey of common chronic noncommunicable diseases in rural areas of Henan, China, conducted by Xinxiang Medical University Health Team. All methods were carried out in conformity to the relevant guidelines and regulations approved by the Ethical Committee of Xinxiang Medical University for Human Studies (protocol number XYLL-2016176, approved 9 June 2016) and all participants received written informed consent. Besides, this research meets all the guidelines in the Declaration of Helsinki.

Data preprocessing

This part was to preprocess all the data of 10,691 participants collected above. This dataset had a large number of missing and repeated values, contained variables that were not related to CHD, and the data storage format was inconsistent. Therefore, it was necessary to preprocess the original data to improve the accuracy of subsequent analysis. Process of preparing data for prediction models for CHD comprises the following: (1) Data selection. First, to develop the diagnosis model of CHD in central Chinese rural adult populations, 915 patients with CHD and 2,829 residents without 20 common chronic non-communicable diseases were selected for analysis based on personal disease history. Next, due to the large amount of information contained in the datasets, we only extracted part of the attribute data closely related to CHD for analysis and processing. The variables included ID, gender, age, smoking status, alcohol intake, psychological status, height, hip circumference, weight, BMR, waist circumference, VFI, grip strength, BFR, and HR. Lab tests included blood routine (24 items), blood lipids, blood glucose (FBG, FINS, and HbA1c), hepatic and kidney function, resting blood pressure, and 12-lead resting electrocardiography. (2) Data preprocessing. All variables were combined and deduplicated based on ID. After that, the “impute” function of the R library “Hmsic” was used to handle the missing values. The variable characteristics of the study participants with missing values greater than 30% were deleted, and the variables less than 30% were replaced with the average value. (3) Data transformation. For coding categorical variables into numbers, an integer code was assigned to each category, for example, gender (which had values 0 = male, 1 = female), smoking status (which had values 0 = nonsmoker, 1 = ever smoker, 2 = present smoker), alcohol intake (which had values 0 = nondrinker, 1 = ever drinker, 2 = current drinker), psychological status (which had values 1 = no pressure, 2 = a little pressure, 3 = moderate pressure, 4 = severe pressure, 5 = extreme pressure). Body mass index (BMI) was defined as weight divided by the square of height, expressed in kg/m², and was derived from mass (in kilograms) and height (in meters). WHR was defined as the ratio of waist circumference to hip circumference. The systolic blood pressure (SBP) and diastolic blood pressure (DBP) were the averages of the three measurements.

Statistical analysis

After data preprocessing, 63 variables (Table 1) of 2,311 participants, including 1,458 controls and 853 cases, were selected for statistical analysis. The Shapiro–Wilk W test was applied to detect whether the data came from a normal distribution. Continuous variables that came from a normal distribution were presented as the mean ± standard deviation and Student’s t-test was used to compare the means for these variables in the different groups. For continuous variables not normally distributed, the median (Q1–Q3 quantiles) was displayed and analyzed with the Wilcoxon rank-sum test. For categorical variables, numbers and percentages were presented and prevalence of these variables was compared using Pearson’s Chi-squared test. Due to the small sample groups in some variables, the comparison was using Fisher’s exact test. Results are shown in Table 1. Next, a univariate analysis was performed as a means of identifying the predictor variables related to CHD and screened out the variables with P < 0.05.

LASSO model

In this section, 2,311 participants with 63 variables (Table 1) were used to variable selection and parameter estimation. LASSO regression algorithm is useful for fitting a wide variety of models with high prediction accuracy and interpretability. LASSO shrinks the absolute sum of the coefficients (L1 regularization) to be less than a fixed value (λ) and assigns zero weights to some irrelevant features to perform variable selection and parameter estimation. λ is chosen with an automated ten-fold cross-validation approach. LASSO can minimize collinearity and overfitting when multicollinearity is present among input variables (Muthukrishnan & Rohini, 2016; Tibshirani, 2011). In this research, LASSO was applied for variable selection and parameter estimation of the influencing factors for CHD. Variables with a regression coefficient of zero after shrinkage were eliminated from the model. Then the remaining variables were entered into the regression model implemented with the R package (Friedman, Hastie & Tibshirani, 2010).

At last, 16 predictors (BFR, WHR, TG, HbA1c, LDL_C, TC, PSQI, VFI, age, SBP, FBG, ALT, creatinine, ECGNOTE, smoking, and drinking) of 2,311 participants were screened out for developing the CHD model with the LASSO, SVM and RF.

SVM model

SVM is a supervised learning algorithm. Its learning strategy is to seek the optimal segmentation hyperplane using support vectors and margins to classify the data. SVM can be used for classification and regression analysis. As a training algorithm, SVM has high accuracy and less prone to overfitting (Noble, 2006). In this research, the “tune.svm” function with an automated ten-fold cross-validation was applied for parameter tuning and the “svm” function in the “e1071” R package was used for the construction of CHD model (Meyer et al., 2021).

RF model

RF is a classifier containing multiple decision trees, and its output classification result is determined by the highest votes of all the tree. It is an ensemble learning method can be used for classification and regression. RF can solve the overfitting problem that often occurs in decision trees, and get more accurate prediction models (Breiman, 2001; Cutler, Cutler & Stevens, 2012). RF, compiled in the R package “randomForest” was chosen as a classifier to predict future possibilities of CHD with GHC data (Liaw & Wiener, 2002). A ten-fold cross-validation procedure was employed to optimize the parameters (ntree and mtry).

Effectiveness of risk prediction models

The performance of risk prediction models can be evaluated in various ways. In this study, the effectiveness of the three CHD models generated by LASSO, SVM, and RF classifiers were all assessed using classification accuracy, sensitivity, specificity, precision, F1-score, Matthews correlation coefficient (MCC), and AUC measures of ROC curves. Besides, the NRI and IDI were also calculated using the “reclassification” functions in the “PredictABEL” package in R (Kundu et al., 2011). NRI and IDI were used to assess model discrimination with the cutoff = 0.5. AUC, NRI, and IDI are three metrics that are increasingly used together in the assessment of binary outcomes models (Martens, Tonk & Janssens, 2019). In addition, DCA was used to evaluate the net benefit across a range of threshold probabilities among the three models developed by LASSO, SVM, and RF classifiers, and identified the optimal model (Vickers & Elkin, 2006).

General characteristics of the participants

After data preprocessing, 2,311 rural residents including 779 males and 1,532 females with 63 variables of Xinxiang county, Henan province in China were screened out for analysis. Results of descriptive statistics for the selected 63 variables in groups with and without CHD for the 2,311 individuals are shown in Table 1. Compared with the controls, patients with CHD were more likely to be males, older, ever-smokers, and ever-drinkers, and had higher PSQI, BFR, BMR, VFI, SBP, DBP, WHR, and BMI (P < 0.001). Besides, blood test results (24 items) showed that patients with CHD had a higher WBC count including five subtypes NEUT count, LYMPH count, MONO count, BASO count, EOS count, and the percentage of EOS than the control group (P < 0.001). The levels of RBC, HGB, RDW-CV, RDW-SD, HCT, MCV, and MCH were also elevated in patients with CHD (P < 0.001). In contrast, PCT and PLT counts were significantly lower in CHD patients compared to the normal group (P < 0.001 and P = 0.002). Additionally, the level of blood glucose (FBG, HbA1c, and FINS) and blood lipid (LDL_C, TG, and TC) of patients with CHD was significantly higher than that without CHD (P < 0.001). However, compared with the control group, the level of HDL_C was decreased in patients with CHD. All indices of hepatic and renal function (ALT, creatinine, IBIL, ALP, urea, uric_acid, AST, TBIL, and THB) were all significantly elevated in patients with CHD except DBIL. Prolonged PR and QT intervals, QRS duration, QTc, and SV1 + RV5 on resting 12 lead resting ECGs were observed in the CHD patients (P < 0.001). All of these variables except for RDW_CV and PR intervals that differed between the CHD group and the normal group were also statistically significant in univariate logistic regression (P < 0.01) (Table S1). For example, the prevalence of CHD increased with age (OR = 1.15, 95% CI = 1.14–1.16) and was higher in men than in women (OR = 0.62, 95% CI = 0.52–0.74). In addition, CHD prevalence also increased with high MONO counts (OR = 5.40, 95% CI = 2.47–11.80) and EOS counts (OR = 5.82, 95% CI = 2.72–12.43), elevated LDL_C (OR = 5.99, 95% CI = 4.98–7.20), TG (OR = 20.16, 95% CI = 15.24–26.68) and TC (OR = 5.23, 95% CI = 4.43–6.16) levels, high FBG (OR = 3.39, 95% CI = 2.91–3.95) and HbA1c (OR = 8.01, 95% CI = 6.47–9.91) levels with all P < 0.001.

Parameter evaluation and feature selection

Parameters tuning for different models were performed by ten-fold cross-validation that gave the best hyper parameters with the lowest cross-validation average error. The optimal parameters combination for svm model were 0.01 for gamma and 100 for cost. The best mtry and ntree for RF model were 150 and eight. To choose the most useful explanatory variables from 63 candidate variables, first, univariate logistic regression analysis was used to preliminarily screen the variables significantly associated with CHD. A total of ten variables including PLCR%, RDW_CV, LYMPH%, HCHC, MPV, PDW, NEUT%, DBIL, HR, and RR not statistically significant associated with CHD (P > 0.05) were excluded (Table S1). Next, automated ten-fold cross-validation of the LASSO regression (cv.glmnet) was used for selecting the most parsimonious model with the largest λ (lambda.lse = 0.02) whose error was no more than one standard error above the error of the best model. Sixteen variables (BFR, WHR, TG, HbA1c, LDL_C, TC, PSQI, VFI, age, SBP, FBG, ALT, creatinine, ECGNOTE, smoking, and drinking) with non-zero coefficient were screened out. Among the 16 variables, WHR, TG, and ECGNOTE had the largest coefficients, namely 1.5355, 1.004, and 0.8523 respectively. In contrast, BFR, ALT, and creatinine had the smallest coefficients, which were 0.0035, 0.0031, and 0.0002 respectively. Finally, a total of 2,311 participants (1,458 controls and 853 CHD patients) including 16 variable features were selected to construct the CHD model with the LASSO, SVM, and RF methods. A training dataset (70% of the patients with CHD and 70% of the controls without CHD) was for developing the risk prediction models, and a test dataset (30% of the patients with CHD and 30% of the controls without CHD) to test the performance of the model.

Comparison of predictive model performance

Performance assessment of the three models is presented in Table 2. Overall, the three classifiers (LASSO, SVM, and RF) performed well in categorizing CHD patients and non-CHD individuals. The three models all achieved significantly higher AUC values (>0.95). Among the three classifiers, RF attained the best performance on risk prediction of CHD, with highest accuracy (1.00 and 0.9389), sensitivity (1.00 and 0.8725), specificity (1.00 and 0.9771), precision (1.00 and 0.9563), F1-score (1.00 and 0.9125), and MCC (1.00 and 0.8678) on both the training set and test set, followed by SVM model, while the LASSO model performed worst. Details are shown in Table 2.

The ROC curve results of the three CHD models are shown in Fig. 1. The CHD models constructed by the RF algorithm can provide a consistently high accuracy for predicting CHD outcome, with the highest AUC both on the training set and test set (1.00 and 0.9711), followed by SVM (0.9860 and 0.9589) and LASSO (0.9733 and 0.9587). Besides, the IDI and categorical NRI were also calculated. Results showed that the RF model had significantly improvement in CHD risk prediction both on the training set (NRI = 8.52%, 95% CI = 6.32%~10.72%, P < 0.001 and IDI = 9.54%, 95% CI = 8.15%~10.94%, P < 0.001) and the test set (NRI = 6.39%, 95% CI = 2.59%~10.19%, P = 0.00099 and IDI = 4.21%, 95% CI = 1.68%~6.74%, P = 0.00112) compared with the SVM model. The RF model was also better than the LASSO model both on the training set (NRI = 17.92%, 95% CI = 14.92%~20.92%, P < 0.001 and IDI = 19.50%, 95% CI = 17.71%~21.29%, P < 0.001) and the test set (NRI = 8.85%, 95% CI = 5.07%~12.63%, P < 0.001 and IDI = 8.78%, 95% CI = 6.45%~11.11%, P < 0.001). The SVM model was superior to the LASSO model both on the training set (NRI = 9.40%, 95% CI = 6.88%~11.92%, P < 0.001 and IDI = 9.95%, 95% CI = 8.57%~11.34%, P < 0.001) and the test set (NRI = 2.46%, 95% CI = −1.30%~6.23%, P = 0.19969 and IDI = 4.57%, 95% CI = 2.58%~6.57%, P = 10⁻⁵), but not statistically significant of NRI on the test set. Besides, results of DCA also indicated the CHD risk prediction model developed by the RF classifier had a higher net benefit than the models constructed by SVM and Lasso methods both on the training set and test set (Fig. 2).