The NHANES collects comprehensive background information through household screenings, interviews, and physical examinations. It provides data on the general health and nutrition of the American population, employing advanced multi-stage probability sampling methods. For this analysis, data from NHANES cycles covering the years 2007–2018 were utilized. The inclusion criteria consisted of: (1) participants with a previous diagnosis of depression; (2) individuals aged 20 years and older; and (3) complete information on all relevant variables. The exclusion criteria included: (1) participants without a previous diagnosis of depression; (2) individuals younger than 20 years; and (3) cases with missing values for any variable. Initially, 59,744 participants contributed data for the survey. Following the application of the inclusion and exclusion criteria, a final cohort of 2,085 individuals aged 20 years and older was selected for our study. The study protocol for NHANES was approved by the Institutional Review Board at the Centers for Disease Control and Prevention, with informed consent obtained from all participants. Figure 1 illustrates the screening process for the subject population.

This study utilized the PHQ-9 scale, a nine-question tool designed to evaluate depression. Responses are rated on a four-point Likert scale, ranging from 0 (not at all) to 3 (nearly every day), resulting in a total score that can vary from 0–27. Generally, a score of 10 or above indicates a likelihood of depression.

In the NHANES survey, participants were inquired whether a doctor or other healthcare provider had ever informed them of a diagnosis of coronary heart disease. Those who responded affirmatively were categorized as having coronary heart disease. The covariates included demographic information [age, gender, race, marital status, education level, and poverty-income ratio (PIR)], lifestyle factors (alcohol use, smoking habits, sedentary time, and sleep duration on workdays), chronic health conditions [hypertension (No/Yes), myocardial infarction (No/Yes), diabetes, chest pain (No/Yes)], “Yes” represents participants with the corresponding disease and “No” represents participants without the disease. screening data [body mass index [BMI], waist circumference [WC]], and laboratory measurements (uric acid [UA], total cholesterol [TCHOL], creatinine [CR], albumin [ALB], blood urea nitrogen [BUN], high-density lipoprotein [HDL], HbA1c, triglycerides [TG], alanine aminotransferase [ALT], and aspartate aminotransferase [AST]). Gender was categorized as male or female, race was divided into Mexican American, non-Hispanic white people, non-Hispanic black people, Hispanic people, and other races. Marital status was classified as unmarried, married or cohabitating, and married but living alone (separated, divorced, or widowed). Education levels were grouped into below 9th grade, 9th–11th grade, high school graduate, some college, or associate degree and above.

Household or individual income was modified according to the survey year and the poverty threshold specific to each state in order to determine the poverty-to-income ratio. Participants provided information on their alcohol use and smoking habits. Smoking was categorized into three distinct groups: nonsmokers, former smokers, and current smokers. Alcohol consumption was divided into five classifications: never drinkers, former drinkers, heavy drinkers (three or more drinks per day for women and four or more for men), moderate drinkers (up to two drinks per day for women and three for men), and light drinkers (not included in the other categories). Sleep duration (in hours) and sedentary time (in minutes) were assessed through a questionnaire. Medical professionals measured waist circumference and body mass index at mobile examination centers. Additional questionnaires collected data on participants' chronic conditions, including hypertension, diabetes mellitus, myocardial infarction, and chest pain. Laboratory tests provided values for UA, TCHOL, ALB, CR, HDL, BUN, TG, HbA1c, ALT, and AST.

In this study, the dataset was randomly split into a training set and a validation set in an 8:2 ratio. Univariate and multivariate logistic regression analyses were employed to identify predictor variables. Using the training data, eight machine learning models were developed: LR, RF, GBM, SVM, XGBoost, CART, KNN, and NNET. The validation set was employed to evaluate the predictive accuracy of the models. Discrimination between models was assessed using ROC curves and PR curves, while calibration was determined through calibration curves and the Brier score, comparing predicted outcomes to actual results. The clinical applicability of the models was analyzed via DCA. To mitigate overfitting, internal validation was conducted using the Bootstrap technique. Additionally, a nomogram and web calculator derived from logistic regression was developed to visually illustrate the predictive model. Lastly, the significance of variables in the top-performing model was ranked utilizing SHAP (Shapley Additive Explanations) plots.

Given the complex sampling design of NHANES, data were weighted during the analysis of baseline information and the logistic regression. Continuous variables were reported as means with standard errors, whereas categorical variables were presented in terms of counts and percentages. T-tests were utilized to assess continuous variables between the two groups, while chi-square tests or Fisher's exact tests were employed for comparing categorical variables. All statistical analyses were conducted using R software (version 4.4.1), with statistical significance defined as P < 0.05 for all analyses.

This study included 2,085 individuals with a prior diagnosis of depression, who had an average age of 46.12 years. Among these participants, 36.69% were male, 63.31% were female, 44.32% identified as non-Hispanic White persons, 20.29% as non-Hispanic Black persons, 14.58% as Mexican American, 13.29% as Hispanic American, and 7.53% as other races. Among the 2,085 participants, 124 were diagnosed with coronary heart disease, while 1,961 had no prior diagnosis of coronary heart disease.

In the training set, participants with depression were divided into two groups according to whether they had coronary heart disease. Significant statistical differences were identified between the two groups in terms of waist circumference, age, marital status, sedentary behavior, history of myocardial infarction, chest pain, hypertension, diabetes, alcohol consumption, CR, UA, ALB, BUN, TG and HbA1c (P < 0.05). In the validation set, significant statistical differences were found between the two groups with respect to age, sedentary time, BUN, ALT levels, marital status, history of myocardial infarction, chest pain, and diabetes (P < 0.05). The results are shown in Table 1.

To identify independent risk factors for coronary heart disease in individuals with depression, subsequent univariate and multivariate logistic regression analyses were conducted. The univariate logistic regression analysis revealed that factors such as WC, age, sedentary behavior, race, marital status, educational level, alcohol consumption, CR, UA, ALB, BUN, TG, ALT, HbA1c, diabetes, myocardial infarction, chest pain, and hypertension were significantly linked to the risk of coronary heart disease. Following this, multivariate logistic regression analysis indicated that age, education level, TG, history of myocardial infarction, and presence of chest pain emerged as independent predictive factors for coronary heart disease risk in individuals with depression (P < 0.05). The findings are summarized in Table 2.

To assess and validate the eight models, ROC curves were generated for all nine machine learning algorithms. In the training dataset (Figure 2A), the RF model achieved the highest AUC at 0.987, followed by the XGBoost with an AUC of 0.915, the GBM at 0.910, the NNET also at 0.902, SVM at 0.901, LR at 0.896, KNN at 0.892, and CART at 0.826. In the validation dataset (Figure 2B), the RF model maintained its leading position among the eight algorithms, with an impressive AUC of 0.996, indicating a robust discriminative capability. Additionally, in both the training set and validation set, the random forest model recorded PR curve AUC values of 0.848 and 0.960, respectively, highlighting its superior discriminative performance relative to the other models (Figure 3A,B).

Moreover, calibration curves from both datasets were employed to evaluate the predictive accuracy of the nine models against the actual incidence rates. The findings demonstrated a strong alignment between the actual and predicted values for the random forest model in both the training dataset (Figure 4A) and the validation dataset (Figure 4B). To further examine the model's discriminative power, Brier scores were calculated for both sets (Table 3). The random forest algorithm achieved the best Brier score of 0.026 in the training set, outperforming all other models. Likewise, in the validation set, it recorded the lowest Brier score of 0.021, further confirming its superior discriminative performance.

DCA was also conducted for the training set and validation set to evaluate the clinical utility of the models. The random forest model provided a significant net benefit in predicting coronary heart disease among depressed populations, further demonstrating its substantial clinical utility (Figure 5A,B).

To prevent model overfitting, the Bootstrap method was employed for internal validation, yielding an AUC of 0.864, indicating good performance. As a result, the random forest model was ultimately chosen as the predictive model for this research.

Given the strong performance of the traditional logistic regression model in previous analyses, a nomogram was subsequently developed based on eight identified risk factors. By incorporating these eight risk factors, the nomogram enables a more precise estimation of the likelihood of specific outcomes (Figure 6). In addition, a web calculator was constructed based on the nomogram for clinicians to predict the risk of coronary heart disease in depressed patients (https://xwzxwang.shinyapps.io/DynNomapp/) (Figure 7).

Figure 8 illustrates the SHAP value interpretation for a single instance when using the random forest model to predict coronary heart disease. This figure employs a horizontal bar plot to represent the contribution of each feature to the model's prediction (SHAP value). The length and direction of each bar indicate the magnitude and direction of the feature's contribution to the prediction. Red bars represent a positive contribution towards predicting coronary heart disease, while blue bars indicate a negative contribution, suggesting a non-CHD outcome. It is evident from the figure that the feature representing MI history is the most influential, as it has the largest absolute SHAP value, showing symmetry around zero. This indicates that different values of MI introduce significant uncertainty in the model's prediction. Specifically, the SHAP value for MI is ±0.027, suggesting that variations in this feature have a substantial impact on both CHD and non-CHD predictions. The age feature follows, with a SHAP value of ±0.009, indicating a notable influence on the prediction outcome as well. In contrast, other features, such as TG, education, and chest pain, have relatively smaller SHAP values, all less than 0.005, implying their limited contribution to the model's prediction for this particular instance.