Dietary micronutrients, whole food patterns, and restless legs syndrome in coronary heart disease: A dual dietary profiling machine learning analysis using NHANES data

Abstract

Background

The relationships between dietary patterns and restless legs syndrome (RLS) risk in patients with coronary heart disease (CHD) are poorly understood. This study aimed to identify dietary factors associated with RLS risk in this specific population.

Methods

We analyzed NHANES data (2005–2018) using two complementary dietary representations: micronutrient profiles and food-group profiles derived from the MyPyramid Equivalents Database linked to NHANES dietary recalls. Six machine-learning models were developed using Boruta for feature selection and SMOTE to address class imbalance. Model robustness was further evaluated through temporal validation, and SHAP and LIME were used for model interpretation.

Results

Random forest demonstrated the highest predictive performance in both micronutrient and food-group models. SHAP analysis indicated that Lycopene, Caffeine, Alpha-carotene, Vitamin C, Food folate, and Magnesium showed contributions toward lower predicted RLS probability, whereas Calcium, Moisture, and Vitamin B1 showed contributions toward higher predicted RLS probability. These model-derived directions represent associations rather than causal effects. In the food-group model, total red and orange vegetables, total vegetables, and white potatoes were inversely associated with RLS. Conversely, Milk, Meat, and Added Sugars were positively associated with predicted RLS probability.

Conclusion

Random forest models effectively identify dietary risk factors for RLS in CHD patients. Nutrient-dense dietary patterns, particularly those characterized by higher consumption of specific vegetables and antioxidant-related nutrients, were associated with lower predicted RLS probability, whereas higher consumption of milk, meat, and added sugars was associated with higher predicted RLS probability. These findings suggest that dietary patterns characterized by specific nutrient and food-group profiles may be associated with RLS risk among CHD patients and warrant further investigation in independent prospective studies.

Introduction

Restless legs syndrome (RLS) is a neurological disorder characterized by an uncontrollable urge to move the legs, often accompanied by discomfort that worsens during rest. ¹ These symptoms frequently lead to sleep disturbances and a decline in quality of life among individuals with coronary heart disease (CHD). RLS affects approximately 5–15% of the general population and up to 20–40% of individuals with cardiovascular disease, including CHD, depending on demographic and disease severity factors. ² However, prevalence estimates vary substantially according to the diagnostic approach used. Studies applying symptom-based diagnostic criteria generally report higher prevalence because they capture both treated and untreated cases, whereas medication-based definitions may identify only individuals with clinically recognized and pharmacologically managed RLS. In CHD patients, RLS is associated with hypertension, endothelial dysfunction, autonomic imbalance, and increased cardiovascular morbidity and mortality.^2,3 Longitudinal studies also show that RLS correlates with poor sleep, daytime fatigue, and a higher incidence of major adverse cardiovascular events, underscoring the importance of integrated management in this high-risk population.^4,5 RLS may further aggravate CHD progression through periodic limb movements during sleep, which trigger sympathetic activation and disrupt sleep architecture.^6,7

Dietary micronutrients and daily food patterns influence metabolic regulation and the development of chronic diseases, including cardiovascular and neurological disorders. ⁸ Micronutrients such as iron, magnesium, folate, and vitamins B6, B12, C, D, and E support neurotransmitter synthesis, antioxidant defense, and neuromuscular activity. ⁹ Deficiencies in these nutrients are linked to increased risks of diabetes, hypertension, and neurodegenerative diseases through oxidative and inflammatory pathways. ¹⁰ Diets rich in whole grains, fruits, vegetables, lean proteins, and healthy fats reduce glycemic load and lower risks of obesity, dyslipidemia, and atherosclerosis, whereas high consumption of refined grains, added sugars, and saturated fats increases inflammatory markers and metabolic complications.^11–14 Additionally, foods such as dark-green vegetables, nuts, seeds, and seafood low in n-3 fatty acids provide anti-inflammatory benefits, while excessive intake of processed meat and solid fats promotes oxidative stress and insulin resistance.^15,16

Despite growing attention to nutrient-disease interactions, evidence regarding dietary micronutrients, food intake patterns, and RLS risk in CHD remains limited. Existing studies rarely assess causal or dose-response relationships. ¹⁷ Potential mechanisms involve inflammation, where micronutrient deficiencies impair cytokine regulation and amplify inflammatory responses relevant to both RLS and CHD. ¹⁸ Oxidative stress may also contribute, as inadequate antioxidant intake fails to neutralize reactive oxygen species, worsening dopaminergic dysfunction and vascular injury. ¹⁹ Iron dysregulation further influences RLS development, as reduced ferritin levels common in CHD may impair brain iron metabolism and intensify symptoms through endothelial damage.^20,21 Other pathways include mitochondrial dysfunction and neuronal hyperexcitability, particularly when low polyphenol or omega-3 intake is insufficient to counter lipid peroxidation.^22,23

To address these knowledge gaps, we analyzed NHANES data and incorporated the MyPyramid Equivalents Database as a complementary food-group-based dietary representation to build machine-learning models focusing on whole-food dietary patterns. By combining multiple algorithms and interpretability techniques, this study aimed to identify dietary components associated with RLS in CHD patients and provide evidence-based guidance for nutritional intervention strategies.

Materials and methods

Study population

The National Health and Nutrition Examination Survey (NHANES), conducted by the National Center for Health Statistics (NCHS), is a nationally representative program designed to evaluate the health and nutritional status of non-institutionalized U.S. civilians. In this study, NHANES 2005–2018 data were analyzed using two complementary dietary representations: dietary micronutrient profiles directly obtained from NHANES dietary recalls and food-group profiles generated from the MyPyramid Equivalents Database linked to the same NHANES dietary recall information through USDA food codes. After applying exclusion criteria, 68,446 individuals were removed due to: (1) missing CHD diagnostic information or no confirmed CHD diagnosis (n=67,254); (2) lack of RLS diagnostic information (n=201); (3) incomplete dietary micronutrient records (n=632); and (4) missing data on education, PIR, BMI, hypertension, diabetes, smoking, or alcohol consumption (n=359). Ultimately, 1,744 participants were included in the NHANES analytical cohort (Supplementary Figure 1).

Using the MyPyramid Equivalents Database, 68,186 individuals were excluded based on similar criteria: (1) no CHD diagnosis information or absence of CHD (n=67,254); (2) missing RLS diagnostic information (n=201); (3) unavailable daily food consumption records, including meat intake (n=325); and (4) incomplete sociodemographic or lifestyle information (n=406). As a result, 2,004 participants were retained for final analysis (Supplementary Figure 1).

Assessment of dietary micronutrients

Micronutrient data, including vitamins and carbohydrate-related variables, were obtained using two 24-hour dietary recall interviews conducted on non-consecutive days. The first recall took place during the mobile examination center (MEC) visit, and the second was completed by telephone several days later. Both interviews were administered by trained personnel using the Automated Multiple-Pass Method (AMPM) to enhance accuracy and ensure complete reporting of dietary intake.

Assessment of daily food intake

Daily food consumption was assessed using the MyPyramid Equivalents Database, which converts dietary recall information into standardized servings for each food group. This conversion enables quantitative evaluation of diet quality and supports nutrition-related surveillance. The MyPyramid data were linked to NHANES using USDA food codes to generate daily intake estimates across food categories.

Diagnosis of restless legs syndrome and CHD

Medication use within the previous 30 days was obtained from the prescription medication questionnaire. Participants were classified as having RLS if they reported the use of medications specifically indicated for “Restless legs syndrome” (ICD-10-CM G25.81) within the previous 30 days according to the NHANES Prescription Medication Questionnaire. Because NHANES does not provide a standardized symptom-based assessment or clinician-confirmed diagnosis of RLS across the 2005–2018 cycles, this medication-indication approach was used as the available proxy for clinically recognized and treated RLS. However, this definition may not capture untreated or undiagnosed RLS cases and should therefore be interpreted as identifying pharmacologically treated RLS rather than the complete spectrum of disease.

Cardiovascular disease diagnoses were derived from self-reported medical conditions using the computerized personal interview system. Participants reporting diagnoses of coronary heart disease, angina pectoris, or myocardial infarction were categorized as having CHD.

Covariates

Covariates included demographic and lifestyle factors: 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/GED, some college/associate degree, ≥college graduate), income-to-poverty ratio, BMI, smoking status, drinking status, hypertension, and diabetes. Hypertension was defined as a self-reported history of hypertension with current antihypertensive medication use. Diabetes was defined by a physician diagnosis or laboratory criteria including 2-hour OGTT glucose ≥11.1 mmol/L, HbA1c ≥6.5%, or fasting glucose ≥7.0 mmol/L. Prediabetes was defined as physician-reported prediabetes or laboratory indices consistent with OGTT glucose 7.8–11.1 mmol/L, fasting glucose 6.1–6.9 mmol/L, or HbA1c 5.7–6.4%. Smoking status was classified as non-smokers (never smokers or quit ≥1 year) and current smokers (smoked within the last 30 days or ≥2 cigarettes/day after relapse). Drinking status was categorized as never drinkers (<12 lifetime drinks) and current drinkers (≥12 drinks/year or >6 episodes within 12 months). BMI was calculated as weight (kg) divided by height squared (m²).

Feature preprocessing and variable selection

A total of 93 variables were initially examined, including 86 continuous dietary features (46 micronutrient variables from NHANES and 37 food intake variables from MyPyramid) and 7 categorical covariates. To prevent information leakage, the dataset was first randomly divided into training and validation sets at a ratio of 70:30 before any preprocessing procedures were performed. All preprocessing steps, including standardisation, correlation-based feature filtering, Boruta feature selection, and SMOTE-based class balancing, were conducted exclusively within the training dataset. Specifically, features with correlation coefficients >0.9 were identified and removed using the training dataset only. Subsequently, the Boruta algorithm was applied to the training dataset based on random forest classification, and features consistently identified as important after 500 iterations and classified as “Confirmed” were retained for model development, whereas Tentative and Rejected features were excluded. Finally, SMOTE was applied only to the training dataset to address class imbalance, while the validation dataset remained completely independent and was used exclusively for model evaluation.

Secondary feature selection was conducted using the Boruta algorithm, which generates shadow features from a random forest model. After 500 iterations, only features consistently identified as important were retained for model construction.

Statistical analyses

All analyses adhered to NHANES analytical guidelines. Continuous variables were reported as mean ± standard deviation (SD), and categorical variables were presented as frequencies and percentages. Dietary intake variables derived from 24-hour recalls may demonstrate skewed distributions due to variability in individual consumption patterns; therefore, descriptive statistics were interpreted considering their distributional characteristics. Group differences were evaluated using chi-square tests for categorical variables and Student’s t-tests for continuous variables.

Data were randomly divided into training and validation sets at a ratio of 70:30 before model development. To ensure reproducibility, a fixed random seed (seed = 1234) was applied for dataset partitioning and all subsequent stochastic procedures, including cross-validation, hyperparameter tuning, and SMOTE-based oversampling. Six machine-learning models were constructed using the MLR3 framework: Random Forest, LightGBM, K-NN, Naive Bayes, SVM, and XGBoost. The learner backends were implemented using the following R packages: Random Forest using ranger, XGBoost using xgboost, LightGBM using lightgbm, Support Vector Machine using kernlab, K-Nearest Neighbors using class, and Naive Bayes using naivebayes. Model implementation and workflow management were conducted using the mlr3, mlr3learners, and related mlr3 extension packages in R version 4.3.0. The validation dataset was not involved in preprocessing or model training and was used only for final performance evaluation. Hyperparameter optimization was performed using grid-search tuning within the training dataset rather than relying on default model parameters. Ten-fold cross-validation was used as the resampling strategy during hyperparameter optimization. The predefined search spaces and final selected hyperparameters for each machine-learning algorithm are provided in Supplementary Table 7. Briefly, Random Forest parameters included the number of trees (100–500), maximum tree depth (5–None), minimum samples required for splitting,^2–10 minimum leaf size,^1–4 and bootstrap strategy. XGBoost parameters included learning rate, number of estimators, maximum depth, subsampling ratio, feature sampling ratio, and gamma. LightGBM tuning included learning rate, number of estimators, maximum depth, number of leaves, and sampling parameters. SVM optimization included the penalty parameter (C), kernel function, and gamma. Naive Bayes tuning focused on variance smoothing, while K-NN optimization included neighborhood size, weighting strategy, distance metric, and Minkowski parameter. SMOTE-based class balancing was implemented exclusively within the training dataset using the smotefamily package, with synthetic samples generated based on the default k-nearest-neighbor strategy (k = 5). SMOTE was applied after feature preprocessing and before model training, while the independent validation dataset remained untouched. During ten-fold cross-validation, all preprocessing procedures, including standardisation, feature selection, hyperparameter tuning, and SMOTE-based oversampling, were independently performed within each training fold, whereas the corresponding validation folds were used only for performance evaluation. ANOVA or Kruskal–Wallis tests were used to assess performance differences across models. Given the relatively limited number of RLS events, model complexity and feature interpretation were considered cautiously. Following recommendations for prediction model development, including sample size considerations described by Riley et al., the results were interpreted as exploratory and hypothesis-generating rather than as a definitive clinical prediction tool. ²⁴

Feature interpretability was assessed using SHAP and LIME. SHAP values quantified global feature contributions, whereas LIME provided local interpretability through linear approximation of complex classification boundaries.

All analyses were performed using IBM SPSS Statistics 24.0 and R version 4.3.0. Statistical significance was defined as a two-sided p-value <0.05. This study was reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis using Artificial Intelligence (TRIPOD+AI) guideline.

Results

Characteristics of participants according to the presence of RLS among patients with CHD

Supplementary Table 1 presents baseline characteristics of CHD patients with and without RLS in the NHANES dataset. A total of 1,744 participants from NHANES 2005–2018 were included, with a mean age of 67.40 years (SD = 11.30). Of these, 620 (35.55%) were female and 1,124 (64.45%) were male. Fifty-three participants were identified with comorbid RLS, whose mean age was 64.77 years (SD = 12.36). Compared with patients without RLS, those with RLS had lower intakes of Energy (1,803.35 vs 1,841.54 kcal), Protein (71.24 vs 72.73 g), Total Fat (70.26 vs 72.15 g), Beta-cryptoxanthin (2,022.47 vs 2,102.45 μg), and several other micronutrients. Similarly, 2,004 participants from the MyPyramid Equivalents Database were analyzed, with an average age of 67.61 years (SD = 11.50); 714 (35.63%) were female and 1,290 (64.37%) were male. Sixty-three CHD patients had RLS with a mean age of 66.00 years (SD = 13.00). Compared with those without RLS, participants with RLS reported lower consumption of Citrus, Melons, and Berries (0.20 vs 0.22 servings/day), Dark Green Vegetables (0.06 vs 0.12 servings/day), Legumes as Vegetables (0.09 vs 0.11 servings/day), and Refined Grains (4.44 vs 4.58 servings/day) (Supplementary Table 2).

Feature selection for machine learning models

Supplementary Figure 2 illustrates correlation analyses for micronutrient variables (NHANES) and food intake variables (MyPyramid). Features with correlation coefficients >0.9 were considered collinear and removed. In NHANES, Total Fat and Total Folate were excluded (Supplementary Figure 2A–B). In MyPyramid, Total Starchy Vegetables, Legumes as Vegetables, Total Grains, and Total Protein Foods were removed (Supplementary Figure 2C–D). All feature selection procedures were performed exclusively within the training dataset to avoid information leakage. The validation dataset was not involved in correlation filtering, Boruta selection, or model training and was reserved solely for evaluating the generalization performance of the developed models.

Secondary feature screening was conducted using the Boruta algorithm. In NHANES, 15 features were classified as Confirmed, 19 as Tentative, and 20 as Rejected (Figure 1(A)). In MyPyramid, 13 features were Confirmed, 13 were Tentative, and 17 were Rejected (Figure 1(B)). Rejected features were removed, and Confirmed features were retained for model development. Based on the Boruta selection results, only confirmed features were retained for model construction. Therefore, 15 micronutrient features from NHANES and 13 food-group features from the MyPyramid Equivalents Database were included in the final machine-learning models.

Figure 1.

Ridge plots presenting the results of variable selection using the BORUTA algorithm. (A) NHANES; (B) MyPyramid Equivalents Database.

Model development and evaluation

Six machine-learning models—Random Forest, LightGBM, K-NN, Naive Bayes, SVM, and XGBoost—were trained using the processed training dataset and subsequently evaluated using the independent validation dataset. All preprocessing procedures were restricted to the training data to ensure unbiased performance estimation. Performance metrics included AUC-ROC (Figure 2, Supplementary Figure 5B, Supplementary Figure 6B), AUC-PR (Supplementary Figures 4, 5F, 6F), Accuracy (Supplementary Figures 5A, 6A), F-Beta (Supplementary Figures 5C, 6C), Sensitivity (Supplementary Figures 5D, 6D), Specificity (Supplementary Figures 5E, 6E), Brier Score (Supplementary Figure 12), and calibration plots for the best-performing Random Forest models (Supplementary Figure 13). NHANES results are shown in Supplementary Tables 3–4, and MyPyramid results in Supplementary Tables 5–6. NHANES micronutrient models: During training, Random Forest achieved the best performance with the highest AUC-ROC (0.874), Accuracy (0.759), F-Beta (0.667), Specificity (0.875), AUC-PR (0.859), and the lowest Brier Score (0.174). LightGBM and XGBoost ranked second and third (AUC-ROC = 0.808 and 0.798). K-NN and SVM showed moderate predictive ability, while Naive Bayes had the weakest performance (AUC-ROC = 0.538; AUC-PR = 0.514) (Supplementary Table 3; Figures 4A, 5A, 6A).

Figure 2.

ROC curves of the six machine-learning models in the training and validation sets. (A) NHANES training set; (B) NHANES validation set; (C) MyPyramid Equivalents Database training set; (D) MyPyramid Equivalents Database validation set.

In validation, Random Forest remained superior (AUC-ROC = 0.849; Accuracy = 0.763; F-Beta = 0.792; Sensitivity = 0.858; Specificity = 0.666; AUC-PR = 0.879; Brier Score = 0.178) (Supplementary Table 4; Supplementary Figure 3B, Figure 5B, Supplementary Figure 4B). Calibration analysis further demonstrated good agreement between predicted probabilities and observed RLS event rates across risk deciles in the validation dataset (Supplementary Figure 13A). The calibration curve showed that the predicted probabilities generated by the Random Forest model generally followed the observed event rates, indicating acceptable probabilistic calibration.

MyPyramid food-intake models: Random Forest also demonstrated the best performance. In training, it achieved an Accuracy of 0.766, F-Beta 0.755, AUC-ROC 0.841, Sensitivity 0.750, Specificity 0.781, AUC-PR 0.826, and the lowest Brier Score (0.178). Validation results remained consistent (Accuracy = 0.781; F-Beta = 0.810; AUC-ROC = 0.834; Sensitivity = 0.859; Specificity = 0.686; AUC-PR = 0.845). Calibration analysis of the MyPyramid-based Random Forest model also showed acceptable agreement between predicted probabilities and observed RLS event rates across risk deciles in the validation dataset (Supplementary Figure 13B). Performance differences among models were statistically significant (P < 0.001) (Supplementary Figures 5–6, Supplementary Tables 3–6).

SHAP and LIME interpretation of feature importance

SHAP analysis identified the top 25 predictors for RLS risk (Figure 3). Because the number of RLS events was relatively limited, SHAP-derived rankings were interpreted as relative contributions within the fitted models rather than definitive estimates of population-level importance. These findings should therefore be considered exploratory and require confirmation in larger independent cohorts with standardized RLS assessment and dietary measurements.

Figure 3.

SHAP beeswarm plots showing feature importance and SHAP values for the best-performing model. (A) NHANES; (B) MyPyramid Equivalents Database.

In NHANES micronutrient models, Lycopene (0.0301), Caffeine (0.0265), Alpha-carotene (0.0159), Vitamin C (0.0138), Food Folate (0.0135), and Magnesium (0.0123) showed inverse associations with predicted RLS probability within the model, whereas Calcium (0.0181), Moisture (0.0138), and Vitamin B1 (0.0106) showed positive associations with predicted RLS probability. These SHAP-derived directions represent model-based feature contributions rather than causal effects or statistically significant differences between RLS and non-RLS groups.

In MyPyramid models, Total Red and Orange Vegetables (0.0340), Cheese (0.0255), White Potatoes (0.0162), and Total Vegetables (0.0140) showed inverse associations with predicted RLS probability, whereas Milk (0.0228), Meat (0.0154), and Added Sugars (0.0133) showed positive associations with predicted RLS probability. These findings reflect model-derived patterns and should not be interpreted as evidence of causal dietary effects. These model-derived directions should be interpreted cautiously because several dietary variables did not differ significantly between RLS and non-RLS groups in baseline comparisons. Waterfall and force plots illustrated feature contributions (Figure 4). In NHANES, baseline RLS-free probability increased from 0.485 to 0.786 after feature aggregation (Figure 4(A)–(B)). A similar increase was observed in MyPyramid (0.545→0.717; Figure 4(C)–(D)). Interaction plots of top predictors are shown in Figure 11.

Figure 4.

Force plots and waterfall plots illustrating individual prediction results of the best-performing model. (A) NHANES; (B) MyPyramid Equivalents Database.

LIME analysis was conducted to provide local explanations for individual model predictions. Unlike SHAP, which summarizes average feature contributions across the entire dataset, LIME estimates feature effects within localized regions of the prediction space. Therefore, LIME-derived contributions may vary between individuals and should not be interpreted as population-level directional associations. In NHANES, LIME illustrated heterogeneous local contributions of Vitamin B6, Iron, Dietary Fiber, Alpha-carotene, and Food Folate (Supplementary Figure 8A). Similarly, in the MyPyramid model, local contribution patterns were observed for Poultry, Nuts and Seeds, Total Vegetables, and Meat (Supplementary Figure 8B). Because some LIME contribution intervals crossed zero, these features were not classified as consistently protective or harmful. The overall direction of dietary associations was therefore interpreted primarily based on SHAP analysis.

Sensitivity analysis

Temporal validation within the NHANES framework was conducted using participants from 2005–2012 for model development and participants from 2013–2018 for temporal robustness assessment. In NHANES, Random Forest showed the strongest stability, with training AUC-ROC 0.783, Accuracy 0.809, F-Beta 0.886, and AUC-PR 0.911 (Figure 13(A),(C),(E)). In validation, performance remained high (Accuracy 0.730; F-Beta 0.805; Sensitivity 0.964; AUC-PR 0.809), and although AUC-ROC (0.765) ranked third, overall robustness supported Random Forest as the preferred model (Supplementary Figure 9B,D,F).

Consistent results were obtained in MyPyramid models. Random Forest achieved optimal performance in both training (Accuracy 0.756; F-Beta 0.846; AUC-ROC 0.704; AUC-PR 0.832) and validation (Accuracy 0.713; F-Beta 0.804; AUC-ROC 0.697; Sensitivity 0.932; AUC-PR 0.820) (Supplementary Figure 10A–F). Sensitivity analyses confirmed strong temporal stability, reinforcing Random Forest as the optimal modeling approach.

Discussion

In this study, NHANES data were analyzed using complementary micronutrient-based and food-group-based dietary profiles derived from the same survey framework to develop machine-learning models examining dietary associations with RLS risk in patients with CHD. Random Forest demonstrated the strongest predictive performance in both the micronutrient-based model derived from NHANES and the food-group-based model derived from the MyPyramid Equivalents Database. SHAP analysis identified lycopene, caffeine, alpha-carotene, vitamin C, food folate, and magnesium as the most influential features associated with lower predicted RLS probability in the NHANES model, whereas calcium, moisture, and vitamin B1 were associated with higher predicted RLS probability among CHD patients. In the food-group model, SHAP analysis identified total red and orange vegetables, cheese, white potatoes, and total vegetables as negatively associated with predicted RLS risk, whereas milk, meat, and added sugars showed positive associations. These global associations were interpreted alongside LIME results, which provided complementary individual-level explanations rather than population-level effect directions. Together, these findings suggest that nutrient-dense dietary patterns may be associated with lower predicted RLS probability in this population, whereas certain dietary components may be associated with higher predicted RLS probability. These findings highlight the potential value of whole-food–based approaches for identifying dietary patterns related to RLS in CHD patients.^8,25

The inverse associations between RLS risk and micronutrients such as magnesium, vitamin C, and folate are consistent with previous research reporting similar trends across diverse populations. These nutrients may provide benefits by enhancing antioxidant capacity and reducing neuroinflammation. Magnesium, for example, has consistently been linked to reduced RLS severity in both hemodialysis patients and the general population. Lower serum magnesium correlates with more severe symptoms, possibly due to its role in stabilizing neuronal membranes and modulating NMDA receptors to reduce dopaminergic hyperexcitability.^26,27 Vitamin C supplementation has also been shown to alleviate RLS symptoms in patients with kidney disease through antioxidant mechanisms that reduce peripheral nerve oxidative injury, which may similarly apply to CHD patients with vascular impairment.^10,28 Folate may exert protective effects by regulating homocysteine metabolism and preventing endothelial dysfunction, thereby reducing RLS risk through improved vascular integrity and suppression of inflammatory mediators such as C-reactive protein.^29,30 The negative association between vegetable intake—particularly red and orange vegetables—and RLS risk is consistent with dietary pattern studies showing lower RLS prevalence in individuals consuming more vegetables. Carotenoids and vitamins in these vegetables may activate Nrf2-related antioxidant pathways and inhibit pro-inflammatory cytokines such as IL-6.^15,31 White potatoes, a source of resistant starch and micronutrients, have been linked to improved sleep indices in observational studies, potentially through gut microbiome regulation and reductions in systemic inflammation. ³² Collectively, these findings strengthen the growing evidence that nutrient-rich dietary patterns can mitigate RLS symptoms, especially in CHD, where oxidative stress exacerbates vascular and neural dysfunction.^33,34

In contrast, several positive associations—such as those involving calcium, vitamin B1, and food groups including milk and meat—differ from some prior studies. These discrepancies may reflect population-specific characteristics, methodological variation, or unmeasured confounding inherent to cross-sectional research. Although calcium has shown variable associations with RLS in previous studies, its positive contribution in our model should be interpreted cautiously. Because calcium intake did not differ significantly between RLS and non-RLS participants at baseline, this finding may reflect complex interactions with other dietary factors, medication use, or unmeasured characteristics of CHD patients rather than a direct adverse effect of calcium intake.^35,36 The positive contribution of vitamin B1 in our model also differs from previous evidence suggesting potential neurological benefits or neutral effects. Given the lack of significant baseline differences and the exploratory nature of SHAP interpretation, this finding should not be interpreted as evidence that higher vitamin B1 intake increases RLS risk. ³⁷ Future studies are needed to determine whether this association reflects dietary patterns, underlying metabolic status, or other confounding factors. Milk and cheese showed divergent contributions in the MyPyramid model, although both belong to dairy-related food categories.^38,39 Because the mechanisms underlying this difference cannot be determined from NHANES dietary recall data, we avoid speculative explanations regarding fermentation, probiotics, or specific dairy components. The observed discrepancy may reflect differences in dietary patterns, preparation methods, accompanying foods, residual confounding, or model-derived nonlinear interactions. Additional studies incorporating detailed dairy composition, fermentation status, and longitudinal dietary assessment are needed to clarify these associations. Meat consumption showed a positive association with predicted RLS risk in the SHAP-based global interpretation, although individual-level LIME explanations demonstrated heterogeneous contributions among participants. This discrepancy likely reflects nonlinear interactions within machine-learning models and highlights the importance of distinguishing global feature importance from local prediction explanations. This inconsistency may stem from the effects of heme-iron-rich processed or red meats, as excess heme iron can promote lipid peroxidation and inflammatory cascades that intensify cardiovascular stress. ⁴⁰ These discrepancies may also arise from dietary recall bias or variability in RLS diagnostic criteria across studies, highlighting the need for standardized assessment methods.^4,41

Our findings emphasize the interconnected roles of dietary factors, inflammation, and oxidative stress in CHD, offering new insights into RLS pathophysiology and extending existing dopaminergic and iron-metabolism models. Although traditional theories attribute RLS primarily to brain iron deficiency and dopaminergic dysfunction, the observed associations with antioxidants such as vitamin C and alpha-carotene support a broader framework that incorporates peripheral oxidative stress. These nutrients may mitigate endothelial dysfunction and protect dopaminergic pathways by limiting reactive oxygen species production.^42,43 The inverse contribution of caffeine observed in our model differs from common clinical recommendations that suggest caffeine restriction in some patients with RLS. ⁴⁴ Therefore, this finding should not be interpreted as evidence that caffeine consumption improves RLS symptoms. Instead, it may reflect complex nonlinear relationships, residual confounding, differences in individual caffeine sensitivity, or dietary patterns correlated with caffeine intake in this CHD population. Further studies with detailed information on caffeine dose, timing, and symptom severity are required to clarify this association. Conversely, the positive associations involving calcium and added sugars challenge purely anti-inflammatory models, suggesting that these components may disrupt calcium-dependent signaling or contribute to persistent low-grade inflammation in CHD. ⁴⁵ Overall, the results complement current mechanistic understanding by suggesting that plant-derived carotenoids may be associated with RLS-related pathways under ischemic cardiovascular conditions. However, these potential mechanisms require confirmation in experimental and longitudinal studies. These findings support hybrid mechanistic frameworks that integrate nutritional, vascular, and neurological factors. ⁴⁶

From a clinical perspective, these findings may provide preliminary insights into dietary patterns that could be considered during nutritional counseling for CHD patients with comorbid RLS. The observed associations suggest that dietary patterns characterized by higher consumption of vegetables and antioxidant-related nutrients may be relevant components of overall dietary quality, whereas limiting excessive intake of added sugars and certain processed dietary components may be consistent with general cardiometabolic dietary recommendations. However, these findings should not be interpreted as evidence that individual nutrients or food groups can prevent or treat RLS, given the cross-sectional design and exploratory nature of machine-learning-based interpretation. In clinical practice, dietary recommendations should remain individualized and consider patients’ cardiovascular status, nutritional requirements, medication use, and lifestyle context. Future prospective studies and intervention trials are needed to determine whether targeted dietary modifications can influence RLS symptoms and clinical outcomes among patients with CHD.

Several limitations should be acknowledged. As a cross-sectional study, the results cannot establish causality, and reverse causation remains possible. For example, RLS symptoms or CHD severity may influence dietary choices, reducing consumption of beneficial foods such as vegetables due to functional limitations or appetite changes. In addition, the relatively small number of RLS events compared with the number of candidate predictors represents an important limitation for machine-learning modelling. Although we applied multiple strategies to reduce overfitting, including Boruta feature selection, cross-validation, and temporal validation, flexible algorithms may still be sensitive to sample size constraints. Consequently, the identified predictors and SHAP rankings should be interpreted cautiously and require validation in larger prospective datasets. In addition, several dietary factors identified by SHAP analysis, including caffeine, calcium, vitamin B1, milk, and cheese, did not show significant differences between RLS and non-RLS groups in baseline comparisons. Therefore, their directional interpretation relies primarily on model-derived feature contributions rather than conventional epidemiological associations. Given the limited number of RLS events, these findings should be considered exploratory and should not be directly translated into clinical dietary recommendations. Dietary micronutrient estimates derived from 24-hour recall may also be influenced by memory bias, particularly among individuals who consume nutrient-dense foods infrequently, resulting in under- or overestimation. Identification of RLS based on medication use may also introduce misclassification. Although participants were classified according to medications specifically indicated for RLS (ICD-10-CM G25.81), some RLS-related medications, including dopamine agonists and gabapentinoids, may have alternative indications such as neuropathic pain or movement disorders. ⁴⁷ Moreover, because NHANES does not include a standardized RLS diagnostic questionnaire or clinical confirmation, the sensitivity and specificity of this medication-based algorithm cannot be directly determined. The relatively low prevalence observed in our study compared with symptom-based epidemiological estimates likely reflects the identification of treated and clinically recognized cases rather than the entire RLS population. Therefore, our findings should be interpreted cautiously, and future studies incorporating validated diagnostic criteria are required to confirm these associations. The absence of an independent external cohort with comparable RLS assessment and dietary measurements represents an important limitation. Although the MyPyramid Equivalents Database provided a complementary food-group-based representation of NHANES dietary data and temporal validation demonstrated model robustness across survey periods, these approaches do not constitute external validation. Therefore, the generalizability of the models requires confirmation in independent populations. These limitations underscore the need for prospective studies to establish temporal relationships and validate the findings across diverse populations.

Conclusion

In conclusion, SHAP-based interpretation identified several dietary components, including lycopene, alpha-carotene, vitamin C, folate, magnesium, red and orange vegetables, total vegetables, and white potatoes, as features inversely associated with predicted RLS probability, whereas milk, meat, added sugars, calcium, and vitamin B1 were positively associated with predicted RLS probability. However, these findings represent exploratory model-derived associations rather than causal dietary effects, particularly because many individual dietary variables showed no significant differences in baseline comparisons and the number of RLS events was limited. These findings highlight the potential relevance of vegetable-centered, whole-food dietary patterns in this comorbid population and demonstrate the utility of machine-learning approaches for exploring complex nutrient–disease associations. Future prospective cohort studies and randomized controlled trials are needed to confirm these associations, determine temporal causality, and evaluate whether targeted dietary interventions can reduce RLS symptom burden and improve quality of life in CHD patients.

ORCID iD

Fei Yang https://orcid.org/0009-0001-1628-2423

Ethical considerations

This study used publicly available, de-identified data from the National Health and Nutrition Examination Survey (NHANES). The NHANES protocol was approved by the National Center for Health Statistics (NCHS) Research Ethics Review Board, and written informed consent was obtained from all participants during the original survey. Because the present study involved only secondary analysis of anonymized data without access to identifiable participant information, additional ethical approval was waived.

Data Availability Statement

Availability of data and materials Publicly available datasets were analyzed in this study. The raw data can be found in the National Health and Nutrition Examination Survey (NHANES) database (https://www.cdc.gov/nchs/nhanes/) and the MyPyramid Equivalents Database. The specific datasets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.*