Diabetes mellitus (DM) has emerged as one of the most prevalent and serious chronic diseases (1, 2), with patients facing a significantly elevated risk of cardiovascular disease (CVD), which remains the leading cause of mortality in this population (3–5). The mechanisms underlying diabetes-associated cardiovascular disease involve oxidative stress, inflammatory responses, metabolic disturbances, mitochondrial dysfunction, accumulation of advanced glycation end products (AGEs), insulin signaling abnormalities, endoplasmic reticulum stress, and cardiomyocyte death (6, 7). Epidemiological studies indicate that individuals with diabetes have a higher risk of CVD compared to non-diabetic individuals (8). The Framingham Heart Study demonstrated that the attributable risk of CVD due to diabetes increased from 5.4% during 1952–1974 to 8.7% in 1975–1998 (9). Haffner et al. further conducted a 7-year follow-up study on cardiovascular mortality, reporting a mortality rate of 15.4% among diabetic patients without a history of myocardial infarction (MI) and 42.0% among those with MI, compared to 2.1 and 15.9%, respectively, in non-diabetic individuals (10). A 10-year follow-up study by van Hateren et al. also showed that the risk of CVD-related mortality in diabetic patients increased annually (11). Given the substantial clinical burden of CVD complications in diabetic patients, integrated management of diabetes and CVD has become a major focus. Central to this management is effective blood glucose control. While intensive glucose control has proven beneficial in preventing microvascular complications and CVD in type 1 diabetes (12, 13), its role in reducing cardiovascular risk in type 2 diabetes mellitus remains contentious (14–16). Consequently, preventing macrovascular complications requires a comprehensive approach, addressing multiple risk factors such as blood glucose management, a healthy diet, smoking cessation, regular physical activity, blood pressure control, and treatment of dyslipidemia (17, 18).

In recent years, dietary factors, particularly the potential cardiovascular protective effects of dietary antioxidants, have garnered increasing attention (19, 20). Antioxidants in the diet, such as flavonoids, vitamins, and polyphenols, may lower the risk of CVD in diabetic patients by reducing oxidative stress, exerting anti-inflammatory effects, improving vascular function, and regulating metabolic processes. Due to differences in mechanisms of action, metabolism, and bioavailability among antioxidants (21, 22), identifying the most protective compounds may inform more targeted dietary interventions for individuals with diabetes.

This study utilizes data from the National Health and Nutrition Examination Survey (NHANES) and applies machine learning (ML) methods to investigate the potential relationship between dietary antioxidant intake and cardiovascular disease in diabetic patients. Compared to traditional statistical approaches, machine learning techniques are better equipped to manage large, complex datasets and identify intricate relationships among health features, thus enabling more accurate predictions of disease risk (23). To ensure model reliability, we conducted benchmark testing to compare different models and employed SHapley Additive exPlanations (SHAP) values to enhance model interpretability, highlighting the specific contributions of various dietary antioxidants in disease prediction. While previous studies have investigated the cardiovascular effects of individual antioxidants, total antioxidant intake, or antioxidant scores (24–26), they predominantly relied on traditional statistical methods, limiting the ability to assess the relative importance of each antioxidant in disease risk. This study innovates by integrating machine learning with SHAP analysis, enhancing prediction accuracy and precisely quantifying the independent contribution of each antioxidant to cardiovascular disease risk, offering valuable insights for personalized nutrition interventions and risk stratification.

This study, utilizing data from the 2007–2010 and 2017–2018 US NHANES cycles, systematically investigates the relationship between dietary antioxidant intake and CVD in diabetic patients through explainable machine learning techniques. Among the eight machine learning models developed, XGBoost demonstrated superior performance, achieving an average AUC of 0.955, highlighting its exceptional and consistent classification ability. The SHAP method, based on game theory, was employed to elucidate the contribution of each feature to the model’s predictions. The findings revealed that Daidzein, Mg, Isorhamnetin, Pelargonidin, Epigallocatechin 3-gallate, and Se were key influencing factors. To the best of our knowledge, this study is the first to integrate various antioxidants with baseline characteristics to build and validate a model for predicting CVD risk in diabetic patients. While the focus is on dietary antioxidants, the model also incorporates traditional features such as demographic factors, lifestyle, and health status, significantly enhancing prediction accuracy. Furthermore, systematic benchmarking was conducted to ensure a fair comparison and robust results across different models.

ML models have been increasingly utilized to explore dietary factors associated with cardiovascular disease. For instance, Ravi V. Shah and colleagues, using data from 2,259 young white and black adults in the CARDIA cohort, employed multivariate analysis and penalized machine learning techniques to identify metabolite features linked to dietary intake, encompassing 17 food groups, 2 nutrient types, and the Healthy Eating Index (HEI-2015). Their study highlighted that metabolic features associated with unhealthy dietary patterns, such as high intake of red/processed meat and refined grains, were stronger predictors of long-term diabetes and cardiovascular disease risks than traditional dietary scores (39). Similarly, Orly Ben-Yacov and colleagues applied machine learning algorithms to evaluate the effects of personalized postprandial target (PPT) diets compared to the Mediterranean diet in adults with prediabetes, showing that the PPT diet significantly improved cardiometabolic markers by modulating the gut microbiota, emphasizing the value of personalized nutrition strategies (40). Yue Wang and colleagues analyzed data from 90,167 participants in the UK Biobank using four machine learning algorithms, with the XGBoost model revealing that potassium, vitamin E, and vitamin C were significant predictors of CVD risk (41). Subgroup analyses identified calcium intake as a key factor for CVD risk in older adults and those with high BMI, while vitamin B6 was closely linked to CVD risk in women (41). Agustin Martin-Morales and colleagues compared multiple ML models, including logistic regression, support vector machines, RF, XGBoost, and LightGBM, selecting RF as the most effective model. SHAP analysis was used to identify critical factors such as age, systolic blood pressure, fiber, calcium, and vitamin E in predicting cardiovascular mortality (42). These studies illustrate the application of ML in examining the link between dietary factors and disease, offering valuable insights for the fields of cardiovascular disease in diabetes. However, many of these studies have limitations, such as the lack of systematic comparisons of model performance and benchmark evaluations, despite utilizing multiple machine learning models.

We selected several models, including RPART, RF, K-KNN, NB, LightGBM, XGBoost, MLP, and SVM, to develop a prediction system for cardiovascular disease in diabetes, and assessed each model’s performance through benchmarking to identify the most suitable approach. Compared to traditional statistical methods, such as logistic regression, machine learning offers notable advantages. First, machine learning is capable of capturing complex nonlinear relationships, whereas traditional methods typically assume linearity, limiting their effectiveness in addressing complex issues. Second, machine learning can automatically identify and select key predictive features, eliminating the subjective and labor-intensive feature selection required in traditional methods. Additionally, machine learning makes fewer assumptions about data distribution and can handle numerical and categorical data, as well as missing values, unlike traditional methods, which often rely on strict distribution assumptions. Machine learning, particularly with ensemble techniques like random forests and XGBoost, demonstrates robust predictive power by effectively preventing overfitting and improving model generalization. Thus, machine learning is more adaptable and efficient for handling complex datasets, offering more accurate decision-making support for disease prediction and personalized healthcare.

Our results demonstrate that the XGBoost model exhibited superior performance among the machine learning algorithms evaluated. As an advanced gradient boosting technique, XGBoost provides distinct advantages in processing complex, high-dimensional datasets. By aggregating multiple weak learners through decision trees and incorporating key strategies such as regularization, column sampling, and gradient-based optimization, XGBoost achieves both high predictive accuracy and strong generalization ability. Its inherent resistance to overfitting further ensures reliable performance across diverse datasets. Owing to these strengths, XGBoost has been extensively applied in the analysis of electronic health records for the development of robust disease prediction models (43–45).

In this study, dietary antioxidants are categorized into three primary types: vitamins, minerals, and polyphenols, each exerting antioxidant effects through distinct mechanisms. Vitamin C directly scavenges free radicals, regenerates oxidized vitamin E, and inhibits metal ion-induced oxidation reactions, thereby protecting cells from oxidative damage (46). Vitamin E, embedded in cell membranes, prevents lipid peroxidation by halting free radical chain reactions, thus preserving membrane integrity (47). Vitamin A and β-carotene effectively neutralize singlet oxygen, safeguarding epithelial cells and preventing lipid oxidation (48). Selenium, as a crucial component of glutathione peroxidase (GPx), reduces the harmful effects of hydrogen peroxide and lipid peroxides, while also synergizing with vitamin E to enhance antioxidant defense (49). Zinc scavenges free radicals by promoting metallothionein expression and serving as a cofactor for superoxide dismutase (Cu/Zn-SOD), thereby maintaining cellular redox balance and stabilizing cell membranes to protect against oxidative damage (50). Magnesium mitigates oxidative stress-induced cellular damage by regulating redox balance, supporting ATP synthesis, and maintaining mitochondrial function (51). Dietary polyphenols, known for their potent antioxidant properties, neutralize reactive oxygen species (ROS) and reactive nitrogen species (RNS) such as superoxide anions (O₂•⁻), hydroxyl radicals (•OH), and hydrogen peroxide (H₂O₂) through their phenolic hydroxyl groups (52). Additionally, polyphenols activate the Nrf2 pathway to increase the expression of endogenous antioxidant enzymes like superoxide dismutase (SOD), glutathione peroxidase (GPx), and glutathione (GSH) (53–55), while reducing chronic inflammation by modulating inflammatory pathways such as NF-κB (56), thereby offering combined antioxidant and anti-inflammatory effects.

Daidzein and Mg are two of the most critical antioxidants examined in this study. Daidzein, an isoflavone primarily found in soy and its derivatives, including tofu, soy milk, soybeans, and bean sprouts, has been clinically confirmed for its potential role in the prevention and treatment of cardiovascular diseases. For example, a cross-sectional study by D. Goodman-Gruen et al. demonstrated that postmenopausal women who consumed high amounts of soy isoflavones, such as Genistein, had significantly lower body mass index (BMI), waist circumference, and fasting insulin levels compared to those who did not consume isoflavones (57). Additionally, isoflavone intake was positively correlated with high-density lipoprotein cholesterol (HDL-C) levels and negatively correlated with postprandial insulin levels, suggesting that dietary soy may have protective effects on cardiovascular health in postmenopausal women (57). In a 16-week randomized controlled trial, Lea Tischmann et al. observed that soy nuts reduced low-density lipoprotein cholesterol (LDL-C) and mean arterial pressure (MAP), while significantly improving endothelial function in healthy elderly individuals (58). In a randomized crossover trial, K.E. Wangen et al. found that a high soy isoflavone diet significantly lowered LDL cholesterol and the LDL/HDL cholesterol ratio in postmenopausal women, indicating potential benefits in improving lipid profiles and reducing the risk of coronary heart disease (59). Moreover, a meta-analysis revealed that isoflavone intake significantly reduced triglyceride (TG) levels and moderately increased HDL-C levels in postmenopausal women, with more pronounced effects observed in women under the age of 65 (60). Another meta-analysis involving 2,305 postmenopausal women showed that soy protein containing isoflavones and soy isoflavone extracts significantly reduced total cholesterol and triglyceride levels while moderately increasing HDL-C, further supporting the potential benefits of soy-based products in improving lipid metabolism and reducing cardiovascular risk (61). Similarly, Daidzein has demonstrated potential cardiovascular benefits in several preclinical studies. Its mechanisms of action include antioxidant properties that reduce free radical generation and alleviate oxidative stress, thereby mitigating endothelial cell damage and lowering the risk of atherosclerosis (62, 63). Moreover, Daidzein plays a role in regulating lipid metabolism by decreasing total cholesterol, LDL-C, and triglyceride levels, while simultaneously increasing HDL-C levels, leading to improved lipid profiles (64). Additionally, Daidzein has anti-inflammatory effects, inhibiting the expression of pro-inflammatory factors, which contributes to enhanced vascular health (65). It also promotes the synthesis of nitric oxide (NO), which enhances endothelial function, improves vasodilation, and supports vascular elasticity, ultimately aiding in blood pressure regulation and improving blood flow (66). Furthermore, Daidzein exhibits antithrombotic properties by reducing platelet aggregation and enhancing fibrinolytic activity, thus lowering the risk of thrombosis (62). Mg ranks second in importance according to SHAP values. As the most abundant divalent cation in cells, Mg is essential for maintaining cellular physiological functions and metabolism. It acts as a cofactor for numerous enzymes, regulates ion channels, and supports energy production (67). In the cardiovascular system, Mg plays a critical role in neuronal excitability, cardiac conduction, and myocardial contraction by modulating ion transport proteins, such as potassium and calcium channels (67, 68). Research has shown that low serum magnesium levels or inadequate dietary intake are closely linked to an increased risk of hypertension (69), atherosclerosis (70), coronary artery disease (71), arrhythmias (72), and heart failure (73).

Our findings also suggest that dietary antioxidants, including EGCG, pelargonidin, vitamin A, and theaflavin 3′-gallate, play crucial roles in CVD prevention through distinct mechanisms. EGCG demonstrates potent ROS scavenging, metal ion chelation, inhibition of lipid peroxidation and oxidative enzymes, and activation of the Nrf2-ARE pathway, thereby enhancing cellular antioxidant defenses (74). It also reduces atherosclerosis risk by improving endothelial function, lowering inflammatory cytokines, and regulating blood pressure and lipid levels (74, 75). Pelargonidin, through its phenolic hydroxyl groups, scavenges free radicals, alleviates oxidative stress, reduces lipid accumulation, and enhances lipid profiles and endothelial function, thus decelerating atherosclerosis progression (76, 77). Vitamin A, a vital fat-soluble antioxidant, stabilizes cell membranes, modulates gene expression linked to endothelial repair and inflammation, and inhibits arterial remodeling and vascular aging via nuclear receptor mechanisms (78). Theaflavin 3′-gallate effectively prevents LDL oxidation, reduces vascular inflammation, and exhibits anti-platelet, lipid-lowering, and antihypertensive effects, thus disrupting multiple cardiovascular risk pathways (79, 80). These natural bioactive compounds offer multi-targeted antioxidant, anti-inflammatory, lipid-regulating, and vascular-protective effects, providing a comprehensive approach to CVD prevention.

Our study holds certain clinical application value. Firstly, the developed predictive model demonstrated strong performance in assessing CVD risk, indicating the potential for future non-invasive risk stratification in diabetic patients through dietary intake assessments. Secondly, this study identifies a significant association between several antioxidant nutrients and CVD risk in diabetic patients, providing evidence to inform clinical dietary recommendations. Consistent with the guidelines from the American Diabetes Association (ADA) and other relevant nutritional frameworks, the findings offer specific guidance for dietary interventions targeting CVD risk in diabetic individuals. The study highlights that antioxidants such as soy isoflavones (e.g., Daidzein), Mg, isorhamnetin, pelargonidin, epigallocatechin gallate (EGCG), and Se are strongly correlated with CVD risk. Consequently, it is recommended that diabetic patients incorporate moderate amounts of soy products (such as soy milk and tofu) to achieve a daily intake of 25–50 mg of isoflavones; consume 310–420 mg of magnesium daily from sources like leafy vegetables, nuts, and whole grains; increase the consumption of fruits and vegetables rich in isorhamnetin and pelargonidin (e.g., apples, onions, and berries), aiming for at least 400 grams per day; drink 1–2 cups of green tea per day to supplement EGCG; and ensure an intake of approximately 55 μg of selenium, primarily from natural sources like Brazil nuts and seafood. By optimizing the intake of these antioxidant nutrients, diabetic patients may further reduce their CVD risk in addition to blood glucose control. Lastly, although various small-molecule antioxidants have shown promise in preclinical research, clinical trials have yielded inconsistent or unsatisfactory outcomes. Our findings may offer valuable insights for future mechanistic studies and the refinement of evidence-based nutritional intervention strategies.

This study has several limitations. First, the diagnosis of diabetes and cardiovascular disease was partially based on self-reported data from the NHANES interview questionnaire, which may introduce information bias due to recall bias or cognitive limitations. Second, variations in dietary habits across different populations and regions could influence the model’s predictions, but further analysis was not possible due to the lack of relevant data. As cross-sectional data were used, this study is unable to establish causal relationships, and future longitudinal studies will be necessary to validate the model’s effectiveness. Although the use of the nationally representative NHANES dataset, along with the inclusion of factors such as gender, race, income, lifestyle, and health status, enhances the generalizability of the results, differences in dietary habits and health conditions across countries and regions may limit the external validity of the findings. Future research should aim to validate the model in diverse countries and dietary contexts. In addition, the complexity and limited interpretability of the model may impact its reproducibility and practical utility. Although SHAP values facilitate the assessment of feature contributions, they rely on the assumption of feature independence and may be affected by residual inter-feature correlations. While highly collinear variables were excluded in this study, the interpretation of feature importance should be approached with caution. Future efforts toward more rigorous feature selection may improve model robustness, though this must be balanced against potential information loss. Finally, the study observed a significant age difference between the DM without CVD group and the DM + CVD group, as well as a higher proportion of females in the CVD group. These factors may influence the results. While machine learning methods can partially adjust for these differences, future research should conduct more detailed analyses of age, gender, and other potential confounders, and use more representative samples to minimize bias.

In conclusion, we developed and validated a cardiovascular disease prediction model for diabetic patients using eight different algorithms: RPART, RF, K-KNN, NB, LightGBM, XGBoost, MLP, and SVM. Of these, XGBoost exhibited the highest discrimination and accuracy in predicting cardiovascular disease in diabetes. SHAP value analysis further elucidated the roles and contributions of various antioxidants, with Daidzein and Mg emerging as the key antioxidants in the model.