The NHANES, conducted by the National Center for Health Statistics (NCHS), evaluates the health and nutritional status of the non-institutionalized U. S. population. This study utilized NHANES data from 2005 to 2018, initially comprising 70,190 participants. After applying exclusion criteria, 67,515 individuals were removed for the following reasons: (1) age under 50 years (n = 50,495); (2) missing data needed for calculating phenotypic age and PhenoAgeAccel (n = 15,696); (3) absence of CHD information (n = 40); (4) missing dietary micronutrient data (n = 805); (5) missing education information (n = 4); and (6) missing data on PIR, BMI, smoking, hypertension, or alcohol use (n = 475). The final analytical sample included 2,675 participants (Figure 1).

Micronutrient intake, including vitamins and carbohydrates, was assessed using 24-h dietary recall data collected during the first and second interview days of NHANES. The first interview was conducted in person at the Mobile Examination Center (MEC), while the second was administered by telephone several days later. Both interviews were conducted by trained professionals using the Automated Multiple-Pass Method (AMPM), which enables comprehensive and standardized documentation of food and beverage consumption.

CHD status was determined based on participants’ self-reported physician diagnoses, including “coronary heart disease,” “angina pectoris,” and “heart attack.” Data were collected through in-person interviews conducted by trained personnel using the Computer-Assisted Personal Interviewing (CAPI) system. Relevant items were derived from the Medical Conditions Questionnaire (MCQ). Participants were classified as having CHD if they responded “yes” to any of the diagnoses of “coronary heart disease,” “angina pectoris,” or “heart attack.”

Phenotypic age was calculated using the algorithm developed by Levine et al., which integrates chronological age with nine biomarkers: albumin, creatinine, glucose, log-transformed C-reactive protein (CRP), lymphocyte percentage, mean cell volume, red cell distribution width, alkaline phosphatase, and white blood cell count. The calculation was performed using a Cox proportional hazards elastic net model with 10-fold cross-validation. Phenotypic age acceleration was defined as the residual from a linear regression of phenotypic age on chronological age. Negative values indicated a biologically younger state, while positive values indicated a biologically older state. The complete formula used in the calculation is provided below:

Where

Sociodemographic and lifestyle covariates included age, sex, race/ethnicity (Mexican American, Other Hispanic, Non-Hispanic White, Non-Hispanic Black, Other Race), education level (<9th grade, 9–11th grade, high school diploma/GED, some college/AA degree, ≥college graduate), family income-to-poverty ratio (PIR), BMI, smoking status, alcohol use, hypertension, and diabetes. Hypertension was defined as a self-reported physician diagnosis and current use of antihypertensive medication. Diabetes was defined as physician-diagnosed diabetes, a 2-h OGTT glucose ≥11.1 mmol/L, or fasting glucose ≥7.0 mmol/L. Prediabetes was defined as a prior diagnosis or intermediate glucose levels (2-h glucose 7.8–11.1 mmol/L or fasting glucose 6.1–6.9 mmol/L). Smoking status was categorized as never/long-term former (never smoked or quit >1 year ago) or current smoker (smoked within the past 30 days, smoked upon waking, or smoked >2 cigarettes/day after quitting). Drinking status was classified as lifetime abstainer (<12 drinks in lifetime) or current drinker (≥12 drinks/year or drinking on >6 occasions in the past 12 months). BMI was calculated as weight in kilograms divided by height in meters squared (kg/m²).

A total of 56 features were considered, including 49 continuous and 7 categorical variables. Features with variance inflation factors (VIF) > 3, adjusted for degrees of freedom, were excluded to reduce multicollinearity. To address class imbalance and improve recognition of the minority class, the Synthetic Minority Over-sampling Technique (SMOTE) was applied. SMOTE generates synthetic data points by interpolating between existing samples and their k-nearest neighbors, thereby enhancing dataset balance (Supplementary Figure 1). All variables were standardized to minimize disproportionate influence due to differing scales.

Feature selection was performed using the Boruta algorithm, a random forest-based method that evaluates feature importance over 500 iterations by comparing real features with randomized shadow features. Only features classified as “confirmed” were retained for model development.

All statistical procedures followed NHANES analytical guidelines. Continuous variables were presented as means ± standard deviations (SD), and categorical variables as frequencies and percentages. Group comparisons used chi-square tests for categorical variables and Student’s t-tests for continuous variables. To examine the joint effect of dietary micronutrient mixtures on CHD in older adults with accelerated aging, Weighted Quantile Sum (WQS) regression was employed. Weights for each nutrient component were estimated using 1,000 bootstrap iterations. Data were randomly split into training (60%) and testing (40%) sets to enhance model reliability.

To avoid overfitting, a 6:4 training-validation split was maintained throughout model construction. Six machine learning models were developed using the MLR3 framework: Random Forest, LightGBM, K-Nearest Neighbors (KNN), Naive Bayes, Support Vector Machine (SVM), and XGBoost. Random Forest: Aggregates multiple decision trees to deliver robust predictions and is inherently resistant to overfitting. LightGBM: An efficient gradient-boosted decision tree model optimized for speed, memory usage, and parallel computation. K-NN: Classifies samples based on proximity to neighbors; performs well on small or non-linear datasets. Naive Bayes: A fast, probabilistic classifier based on Bayes’ theorem, effective even with missing values. SVM: Identifies optimal hyperplanes to separate classes, particularly effective in high-dimensional data. XGBoost: A highly efficient gradient boosting framework that balances accuracy and computational efficiency.

Model performance was evaluated using standardized datasets and the following six metrics: accuracy, F beta score, area under the ROC curve (AUC-ROC), sensitivity, specificity, and area under the precision-recall curve (AUC-PR). AUC-ROC was used as the primary evaluation metric. Ten-fold cross-validation was applied to enhance generalizability. ANOVA and Kruskal-Wallis H tests were used to compare model performance metrics.

To enhance model interpretability, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) were used. SHAP, grounded in cooperative game theory, quantifies each feature’s contribution by considering all feature combinations, thus offering transparent and consistent interpretations. LIME creates interpretable local approximations (e.g., linear models) to explain predictions made by complex models.

All statistical analyses were conducted using IBM SPSS Statistics (version 24.0) and R (version 4.5.0). A two-tailed p-value < 0.05 was considered statistically significant.

Table 1 presents baseline characteristics of older adults aged 50 years and above with accelerated aging, stratified by CHD status. A total of 2,675 participants from NHANES 2005–2018 were included. The mean age was 65.62 years (SD = 9.38), comprising 902 females (33.72%) and 1,773 males (66.28%). Among them, 565 individuals were diagnosed with CHD, with a higher mean age of 69.53 years (SD = 8.52).

Compared with participants without CHD, those with CHD had significantly lower intakes of several dietary components, including energy (1,831.58 vs. 2,008.67 kcal, p < 0.001), protein (71.01 vs. 76.52 g, p = 0.002), carbohydrates (217.45 vs. 240.12 g, p < 0.001), total sugar (94.39 vs. 105.44 g, p < 0.001), dietary fiber (14.57 vs. 15.71 g, p = 0.022), total fat (73.43 vs. 78.21 g, p = 0.033), polyunsaturated fatty acids (16.38 vs. 17.68 g, p = 0.006), alpha-carotene (319.46 vs. 352.58 μg, p = 0.039), niacin (21.94 vs. 23.56 mg, p = 0.022), vitamin B6 (1.74 vs. 1.89 mg, p = 0.030), total folate (364.50 vs. 339.76 μg, p = 0.037), food folate (179.48 vs. 203.16 μg, p = 0.003), phosphorus (1,180.50 vs. 1,261.34 mg, p = 0.005), magnesium (254.89 vs. 275.97 mg, p = 0.001), copper (1.13 vs. 1.22 mg, p = 0.019), potassium (2,392.68 vs. 2,539.28 mg, p = 0.032), selenium (100.98 vs. 106.97 μg, p = 0.042), and alcohol (5.67 vs. 8.94 g, p = 0.005).

In contrast, individuals with CHD had higher intakes of alpha-tocopherol (0.72 vs. 0.48 mg, p < 0.033) and added vitamin B12 (0.75 vs. 0.72 μg, p = 0.039).

Table 2 presents results from the WQS regression, which assessed associations between dietary nutrient mixtures and CHD risk in older adults with accelerated aging. After adjustment for potential confounders (age, sex, race/ethnicity, education, income-to-poverty ratio, BMI, smoking, and alcohol use), an inverse association was observed in the unconstrained model (adjusted OR = 0.90; 95% CI: 0.81–0.99; p = 0.048). The primary contributors were alcohol (weight = 0.281), selenium (0.108), protein (0.100), and cholesterol (0.063) (Figure 2; Supplementary Table 1). A similar trend was observed in the positively constrained model (adjusted OR = 0.84; 95% CI: 0.70–1.01; p = 0.064), although not statistically significant. Major contributors included added vitamin B12 (0.430), alcohol (0.116), caffeine (0.103), and vitamin A (0.055) (Figure 3; Supplementary Table 2).

Figure 4 displays VIFs used to detect multicollinearity. Variables with adjusted VIFs exceeding 3 were excluded, resulting in the removal of 20 dietary components. These included alcohol, alpha-carotene, beta-carotene, beta-cryptoxanthin, carbohydrates, energy, folate (DFE), folic acid, food folate, magnesium, monounsaturated fatty acids, phosphorus, polyunsaturated fatty acids, protein, retinol, saturated fatty acids, total choline, total fat, total folate, and vitamin A.

The BORUTA algorithm then identified 36 variables with significant contributions to the comorbidity of diabetes and hypertension. These included 10 demographic variables (age, hypertension, sex, diabetes, race/ethnicity, income-to-poverty ratio, smoking, education, alcohol use, BMI) and 26 dietary factors (caffeine, added vitamin B12, theobromine, potassium, calcium, sodium, moisture, vitamins D, B2, K, niacin, cholesterol, B12, B1, selenium, iron, copper, B6, total sugar, zinc, dietary fiber, alpha-tocopherol, lutein + zeaxanthin, lycopene, vitamin E, and vitamin C) (Figure 5). Supplementary Figure 2 shows Z-score trends across iterations for feature selection.

Figures 6 and 7 present heatmaps for six machine learning models: Random Forest, LightGBM, K-KNN, Naive Bayes, SVM, and XGBoost. These models were trained and validated using demographic and dietary variables. Performance was evaluated using AUC-ROC (Figures 8–10), AUC-PR (Figures 11–13), accuracy (Figure 14), F beta score (Figure 15), sensitivity (Supplementary Figure 1), and specificity (Supplementary Figure 2).

In the training set, Random Forest achieved the best performance across all metrics: accuracy (0.813), F beta (0.852), AUC-ROC (0.881), sensitivity (0.934), specificity (0.647), and AUC-PR (0.894). XGBoost and LightGBM followed closely, with AUC-ROC values of 0.872 and 0.869, AUC-PR of 0.882 and 0.873, accuracy of 0.792 and 0.799, and F-beta scores of 0.828 and 0.837. K-KNN and SVM showed moderate performance, while Naive Bayes had the lowest metrics, particularly AUC-ROC (0.675) and AUC-PR (0.708) (Table 3). In the validation set, Random Forest again outperformed other models, with accuracy (0.823), F beta (0.862), AUC-ROC (0.890), sensitivity (0.940), specificity (0.656), and AUC-PR (0.908) (Table 4), confirming its superior generalizability.

When only dietary variables were used, Random Forest remained the best-performing model. In the training set, it achieved an accuracy of 0.725, F beta of 0.786, AUC-ROC of 0.770, sensitivity of 0.875, specificity of 0.519, and AUC-PR of 0.799. XGBoost and LightGBM showed comparable results: AUC-ROC of 0.761 and 0.757, AUC-PR of 0.796 and 0.791, accuracy of 0.698, and F beta scores of 0.748 and 0.750 (Table 5). In the validation set, Random Forest again led all metrics: accuracy (0.747), F beta (0.805), AUC-ROC (0.811), sensitivity (0.883), specificity (0.550), and AUC-PR (0.828) (Table 6).

Across all analyses, Random Forest consistently demonstrated superior performance regardless of input variable type. Statistically significant differences in model performance were observed in all comparisons (p < 0.001) (Tables 3–6).

The SHAP algorithm was employed to interpret feature contributions to CHD risk prediction in individuals over age 50 with accelerated aging. Two scenarios were considered: one including both demographic and dietary variables, and one using dietary variables alone. Supplementary Figure 3 illustrates the top 25 features under each scenario using the Random Forest model, with SHAP values quantifying each feature’s importance.

When both demographic and dietary variables were included, Age (SHAP = 0.0509) and Hypertension (0.0472) were the strongest positive contributors. Negative contributors included vitamin B12 (0.0107), lycopene (0.0101), potassium (0.0097), total sugar (0.0081), and lutein + zeaxanthin (0.0075). In the dietary-only model, vitamin B12 (0.0365), lycopene (0.0256), theobromine (0.0193), total sugar (0.0187), and lutein + zeaxanthin (0.0160) had the strongest negative contributions, while caffeine (0.0231) and cholesterol (0.0113) contributed positively.

Force plots and waterfall plots (Supplementary Figures 4, 5) were used to visualize individual-level predictions. In the combined model, the baseline CHD prediction was 0.584, increasing to 0.889 after accounting for feature contributions. In the dietary-only model, the baseline was 0.593 and rose to 0.737 after feature inclusion.

SHAP interaction dependency plots further illustrated the nonlinear relationships between key dietary nutrients and CHD risk (Supplementary Figures 6, 7). In the fully adjusted model, higher intakes of caffeine, lycopene, potassium, total sugar, vitamin B12, and lutein + zeaxanthin were associated with lower SHAP values, indicating protective effects, with notable interactions by age, sex, and hypertension status (Supplementary Figure 6). Similar trends were observed in the unadjusted model using dietary variables alone, though effect magnitudes were slightly attenuated (Supplementary Figure 7). LIME explanations for individual predictions (Supplementary Figures 8, 9) consistently identified vitamin B6, dietary fiber, zinc, vitamin B12, and lutein + zeaxanthin as key negative contributors to CHD risk, reinforcing the robustness of these findings across adjusted and unadjusted models.