The public health stakes are considerable. Stroke ranked as the third leading cause of mortality and the fourth leading cause of disability-adjusted life-years worldwide in 2021, with roughly 11.9 million incident cases annually, the global burden driven predominantly by ischemic stroke. Despite continuing advances in acute clinical interventions, the absolute disease burden keeps increasing with population aging, and global stroke-related DALYs are projected to reach 189 million by 2050. Up to 29% of stroke survivors sustain permanent disabilities, which underscores the urgency of primary prevention. Optimizing dietary structures, such as adopting a Mediterranean-style diet, is an established preventive strategy for cardiovascular and cerebrovascular events, and longitudinal studies indicate that higher intake of nutrient-rich foods is associated with a lower risk of such events.

Yet the clinical trial record tells a different story, a discrepancy often called the antioxidant paradox. Randomized controlled trials evaluating single antioxidant supplements, such as vitamin E or vitamin C, have largely failed to demonstrate significant protective effects against stroke, and some trials even indicate an elevated risk of specific adverse cardiovascular and cerebrovascular events. Vitamin E supplementation under particular intervention strategies has been linked to an increased risk of hemorrhagic stroke. The authors attribute this mismatch partly to methodological limitations in traditional epidemiological approaches: dietary nutrients do not function in isolation in vivo but constitute a whole food matrix characterized by complex collinearity and interactions, and stroke populations in observational cohorts typically exhibit pronounced clustering of baseline risk factors such as advanced age and hypertension, which conventional linear covariate adjustments cannot rigorously eliminate.

To address these challenges, the research team constructed a nationally representative cohort of 101,316 NHANES participants and applied sequential exclusions for missing stroke status, absent dietary records, and incomplete covariates, leaving 37,746 eligible adults from nine cycles (2001–2002 through 2017–2018), of whom 1,376 reported a physician-diagnosed stroke and 36,370 had no such history. Dietary exposure was assessed using the standardized 24-hour dietary recall interviews, extracting only first-day data to ensure a uniform standard across survey cycles, and the nutrient variables reflected intakes exclusively from foods and beverages, excluding dietary supplements. The six target nutrients were vitamin A, vitamin C, vitamin E, β-carotene, selenium, and zinc, treated as continuous variables and also categorized into quartiles.

Baseline imbalances between the stroke and non-stroke groups were severe. Age showed the largest discrepancy, with a standardized mean difference of 1.100, as the stroke group had a mean age of 65.99 years versus 48.67 years in the non-stroke group. Hypertension prevalence (81.7% versus 42.0%, SMD = 0.895) and diabetes (35.5% versus 13.9%, SMD = 0.519) were also markedly higher among those with stroke history. To control for this systematic confounding, the researchers applied 1:1 nearest-neighbor propensity score matching, pairing 1,375 participants with stroke history against non-stroke controls to yield a matched sub-cohort of 2,750 individuals. After matching, the SMD for age dropped from 1.100 to 0.016, hypertension from 0.895 to 0.017, and diabetes from 0.519 to 0.068, all below the predefined 0.10 threshold for optimal covariate balance. A small residual imbalance in racial composition (SMD = 0.130) remained below the 0.25 acceptability limit and was retained as an adjustment covariate in all downstream multivariable models.

Within the matched sub-cohort, multivariable logistic regression revealed that participants in the highest intake quartile for vitamin A, vitamin C, vitamin E, β-carotene, and selenium had significantly lower odds of stroke compared with those in the lowest quartile, and these inverse associations remained robust after adjusting for multiple comparisons using the Benjamini–Hochberg false discovery rate procedure. Zinc intake did not demonstrate a significant independent inverse association. Restricted cubic spline analyses indicated non-linear associations for selenium (P for non-linearity = 0.013) and β-carotene (P = 0.001), and two-piecewise logistic regression identified an approximate slope-change point of 79.8 μg/day for selenium, above which the inverse association attenuated, though the post-breakpoint slope remained weakly negative. For β-carotene, the estimated breakpoint was 0.87 mg/day, after which the slope was no longer statistically significant, with the fitted curve reaching approximately 90% of its total decline by 3.0 mg/day. In an additional sensitivity analysis adjusting for total energy intake, the inverse associations for vitamin A, selenium, and β-carotene remained statistically significant, whereas those for vitamins C and E attenuated to marginal significance.

The core of the analysis, however, lay in the mixture models designed to handle the moderate correlations among nutrients, such as selenium with zinc (r = 0.66) and β-carotene with vitamin A (r = 0.62), which may destabilize coefficients in traditional mutually adjusted single-nutrient models. Weighted quantile sum (WQS) regression, constrained to the negative direction based on an a priori hypothesis, showed that each quantile increment in the mixed exposure index was associated with 23.2% lower odds of stroke (OR = 0.768, 95% CI: 0.679–0.870, p < 0.001). Quantile g-computation (qgcomp), which imposes no directional constraints, validated the robustness of the joint effect, yielding 26.2% lower odds per quantile increment (OR = 0.738, 95% CI: 0.665–0.819, p < 0.001). Both models consistently assigned the largest weights to selenium and β-carotene, and these weight distributions remained essentially unchanged in energy-adjusted sensitivity analyses.

A Bayesian kernel machine regression (BKMR) framework, implemented to account for the correlated exposure structure and to flexibly evaluate non-linear relationships and interactions, corroborated these findings. The overall mixture effect analysis demonstrated a significant inverse association with odds of stroke when joint exposure levels exceeded the median. Posterior inclusion probabilities identified selenium (PIP = 0.988) and β-carotene (PIP = 0.975) as the components with the highest probabilities for the inverse association, whereas the contributions of the remaining components were marginal, with PIPs below 0.37. The interaction assessment showed that when background nutrients were fixed at different quantiles, the effect estimates of individual components exhibited a parallel trend, with no statistical evidence of departure from additivity within the investigated intake ranges. The univariate exposure-response functions revealed that selenium and β-carotene demonstrated non-linear inverse associations characterized by slope attenuation at higher exposure ranges, while the dose-response curves for the remaining nutrients were relatively flat.

Machine learning analyses provided a further cross-validation. Four algorithms—LightGBM, Random Forest, support vector machine, and XGBoost—were coupled with the SHapley Additive exPlanations (SHAP) approach to quantify feature contributions within the matched sub-cohort, using a 70% training set and a 30% independent test set with hyperparameters tuned by five-fold cross-validation. Global feature importance analysis revealed that, across all nutrient-only models, selenium and β-carotene consistently ranked as the top two features based on mean absolute SHAP values, and SHAP beeswarm plots showed that high feature values for these nutrients were predominantly clustered along the negative SHAP axis, indicating that higher intakes are associated with lower odds of stroke. In the comprehensive XGBoost model integrating the six nutrients with demographic and clinical covariates, selenium and β-carotene retained their positions as the two most important features. Notably, test-set AUCs were close to 0.5 and Brier scores ranged from 0.248 to 0.252, close to the value corresponding to uninformative prediction in this balanced test set. The authors explain that discrimination close to 0.5 is an expected property of the matched design, in which baseline clinical covariates had been balanced between groups and thus contributed little between-group information for classification; the machine learning analyses were intended to quantify feature importance rather than develop a clinical prediction tool.

The biological rationale for why selenium and β-carotene might dominate the mixture rests on an in vivo cross-phase complementary defense network. Lipophilic β-carotene accumulates within low-density lipoprotein and cellular lipid bilayers, whereas the selenium-dependent glutathione peroxidase system localizes in aqueous environments such as the cytosol. This compartmentalization necessitates functional complementarity: after scavenging lipid-phase radicals, β-carotene requires aqueous-phase reducing systems as electron donors for its regeneration, without which accumulated pro-oxidant intermediates could exacerbate vascular injury. Theoretical models suggest β-carotene exerts its antioxidant capacity primarily through quenching of singlet oxygen and scavenging of peroxyl radicals, potentially delaying the formation of oxidized LDL, shielding endothelium-derived nitric oxide from quenching by reactive oxygen species, and downregulating pro-inflammatory signaling pathways such as nuclear factor-κB. On the selenium side, experimental models indicate that adequate selenium upregulates glutathione peroxidase 4, and targeted delivery can arrest the ferroptotic cascade in neurovascular cells to facilitate neurological recovery. Selenium deficiency also impairs the activity of key antioxidant enzymes such as thioredoxin reductase, inducing oxidative stress that disrupts the maturation of critical interneurons and perineuronal nets.

Population-level evidence aligns with these mechanisms. A meta-analysis demonstrated that omitting selenium from an antioxidant mixture paradoxically resulted in a significant increase in all-cause mortality associated with the mixture’s intake, suggesting that maintaining adequate selenium levels may be integral to the vascular protective role of the antioxidant network. Increased circulating selenium levels have been associated with a lower incidence of ischemic stroke and lower stroke mortality, though the association between selenium and vascular outcomes appears bimodal, with excessive intake linked to glucose metabolism disorders and hypertension. For β-carotene, the 31-year prospective Alpha-Tocopherol, β-Carotene Cancer Prevention Study cohort demonstrated that higher serum concentrations were associated with significantly lower stroke-specific and all-cause mortality, although null findings from Mendelian randomization studies imply that elevated circulating levels driven solely by genetic background may not directly translate into clinical protection, pointing instead to a complex endogenous metabolic network influenced by interactions between genetic susceptibility and overall lifestyle factors.

The authors are careful to frame the numerical thresholds as hypothesis-generating reference points rather than prescriptive recommendations. The current Recommended Dietary Allowance for selenium is 55 μg/day in the United States, with a tolerable upper intake level of 400 μg/day, while the 2023 European Food Safety Authority assessment derived a lower upper limit of 255 μg/day from selenosis-related endpoints. The inflection point of dietary selenium intake identified in the study, 79.8 μg/day, exceeded the RDA but remained far below these upper limits, suggesting that the inverse association operates largely within the range of physiological adequacy, beyond which additional benefit may be limited. No official RDA or upper limit has been established for β-carotene, whose breakpoint of 0.87 mg/day lies within the range readily achievable through habitual fruit and vegetable consumption; background dietary intake has not been linked to adverse health effects, whereas limiting supplemental use is recommended.

Several methodological limitations temper the conclusions. The inherent cross-sectional design precludes causal inference, and a single 24-hour dietary recall primarily reflects short-term intake and is susceptible to recall bias, with differential reporting according to health status unable to be excluded. The NHANES database cannot distinguish between ischemic and hemorrhagic stroke, preventing evaluation of subtype-specific heterogeneity. Furthermore, while 1:1 propensity score matching strengthened internal validity, it compromised external validity: the matched sub-cohort’s stroke prevalence no longer represents that of the original NHANES sample, so the findings should be interpreted strictly as covariate-adjusted associations within a controlled epidemiological framework rather than nationally representative estimates, and the strict matching restricted natural variation such that potential residual effect modification by matched clinical variables such as age and hypertension cannot be completely ruled out. The authors conclude that higher combined dietary intake of the six antioxidant-related nutrients was associated with lower odds of self-reported physician-diagnosed stroke, with selenium and β-carotene contributing most consistently across WQS regression, quantile g-computation, BKMR, and machine learning-based feature interpretation, and they call for further prospective studies integrating serological biomarkers and multi-omics to confirm these observations.

Subject of Research: Dietary antioxidants and trace elements in relation to stroke risk

Article Title: Dietary Selenium and β‐Carotene Contribute to the Inverse Association Between Antioxidant Mixtures and Stroke: Evidence From NHANES Using Mixture Models and Machine Learning

Article References: Zhong, A., Deng, X., Tan, Y., Lai, R., Lv, X., Zhou, X., & Zhu, J. (2026). Dietary Selenium and β‐Carotene Contribute to the Inverse Association Between Antioxidant Mixtures and Stroke: Evidence From NHANES Using Mixture Models and Machine Learning. Food Science & Nutrition, 14(10), Article e72442. https://doi.org/10.1002/fsn3.72442

Image Credits: AI Generated

DOI: 10.1002/fsn3.72442

Keywords: stroke, selenium, beta-carotene, dietary antioxidants, NHANES, mixture models, weighted quantile sum regression, Bayesian kernel machine regression, machine learning, SHAP, propensity score matching, oxidative stress