The NHANES, conducted by the National Center for Health Statistics (NCHS), employs stratified multistage sampling to collect nationally representative data on health and nutritional status among US civilians (12). The original protocols received NCHS Ethics Review Board approval with documented informed consent from all participants. As this study involved secondary analysis of the NHANES public data (available at https://wwwn.cdc.gov/nchs/nhanes/default.aspx), no additional ethical review was required.

The individuals in this study were included from NHANES 2007–2016 (n = 50,588). Participants with missing DR assessment data were excluded (n = 46,799). The Diabetes Questionnaire (DIQ) item DIQ080, administered only to individuals with diabetes to assess if diabetes has affected the eyes or caused retinopathy, categorized recorded answers as valid (‘Yes’/'No’) or missing (‘Refused’/'Do not know’/'Missing’). Furthermore, participants lacking dietary metal intake data were excluded from the analysis (n = 399). Additionally, participants with no or invalidated data on covariates required for subsequent analyses were also excluded (n = 568), including sex, age, race, educational level, family poverty-income ratio (FPIR), body mass index (BMI), marital status, smoking status, drinking status, hypertension, physical activity, and dietary information. Finally, a total of 2,822 individuals with diabetes were included in this study analysis with 556 individuals with diabetes diagnosed with DR (Figure 1).

In the questionnaire data of NHANES, the diabetes section (prefix DIQ) encompasses data obtained through personal interviews covering diabetes, prediabetes, insulin/oral hypoglycemic drug use, and diabetic retinopathy. It also incorporates self-reported accounts of diabetes risk factor awareness along with associated medical treatments and personal care routines. In this study, participants were determined to have DR if their physicians informed them that diabetes had affected their eyes or that they had retinopathy (13).

In each NHANES cycle, participants submitted comprehensive dietary records across two non-consecutive 24-h periods, enabling estimation of energy expenditure, nutrient absorption, and food constituent intake. Initial dietary interviews were conducted face-to-face during examinations, with subsequent telephone-based recalls performed 3–10 days thereafter. For these analyses, the average intake of food-derived metals (Zn, Cu, Mg, Fe, Ca, and Se) was calculated over the two recall periods; if only the first day’s data was available, that figure was used instead of an average. Participants were also asked about their supplement use during the same two 24-h periods, and the intake of metals from supplements was averaged over the 2 days if possible. The dietary metals intake was determined by summing both supplement and food-derived sources.

Various sociodemographic information incorporating sex, age, race, educational level, FPIR, and marital status were comprehensively documented. The FPIR, derived from the ratio of family income to poverty thresholds adjusted for family size, year, and state of residence, utilized the Department of Health and Human Services’ poverty criteria and was categorized into three groups: low (≤1), middle (>1 to ≤4), and high (>4) (14). The smoking and drinking status were categorized as never, former, and current based on previous study (15). BMI (kg/m2) was calculated as weight divided by height squared. Hypertension was identified as having a systolic blood pressure of >130 mm Hg or diastolic blood pressure of >80 mm Hg from an average of 3 measurements or a history of high blood pressure or a history of oral antihypertensive medications (16). Physical activity, gained by the Global Physical Activity Questionnaire (17), was categorized into two groups based on whether participants met the 2018 physical activity guidelines (18). Detailed records of protein, total fat, carbohydrate, dietary fiber, and energy intake derived from foods and supplement consumption were obtained using two 24-h dietary recall administrations. Bilirubin (BIL), albumin (ALB), and γ-glutamyl transferase (GGT) were designated as oxidative stress indicators, while C-reactive protein (CRP), white blood cell (WBC), lymphocyte count, absolute neutrophil count, and alkaline phosphatase (ALP) served as inflammatory markers. These parameters have been extensively applied in NHANES research for quantifying oxidative stress and inflammatory responses (19–22).

Baseline characteristic data were grouped by diabetic retinopathy (DR) status. Continuous variables were presented means with standard deviation (SD) and categorical variables as frequencies with percentages. Group comparisons were conducted using t-tests for continuous variables and chi-squared tests for categorical variables.

The intake of dietary metals was log10-transformed due to skewed distributions then divided into quartiles. Multivariable logistic regression was utilized to evaluate associations between dietary metal intakes and DR prevalence. Three hierarchical models were constructed: Model 1 was adjusted for sex, age, race, educational level, and marital status; Model 2 further adjusted for FPIR, BMI, smoking status, drinking status, and hypertension; Model 3 further adjusted for physical activity, protein, total fat, carbohydrate, fiber, and energy intake. The first quartile (Q1) was set as the control group.

To assess the combined effects of the six dietary metals on DR prevalence, a WQS regression was applied by calculating a weighted linear index and assigning corresponding weights. A 1000-time bootstrapping procedure was used to construct WQS indices in both positive and negative directions. When the WQS index was significantly nonzero, corresponding weights were calculated to identify the relative contribution of each dietary metal within the index to the incidence of DR. The dataset was randomly split into training set (40%) and validation set (60%) according to previous research (22).

Additionally, the Bayesian variable selection framework was also utilized to explore the overall effects of six dietary metals mixtures on the odds of DR. To be specific, the BKMR model assessed the overall impact of mixtures levels at specific quartiles compared to the medians. All metal intakes were log10-transformed and standardized before analysis. The posterior inclusion probability (PIP) was calculated to quantify the relative importance of each metal (PIP > 0.5 indicating strong contribution). Moreover, univariate and bivariate exposure-response functions were applied to assess both the individual effects and interactions of dietary metals, while simultaneously considering the other dietary metals at the 25th, 50th, and 75th percentiles. This model conducted 10,000 iterations via the Markov Chain Monte Carlo algorithm after accounting for all covariates.

Restricted Cubic Spline (RCS) regression was conducted to examine potential dose–response relationships between log10-transformed dietary Zn intake and DR incidence (23). Subgroup analyses were performed stratified by age, sex, race, educational level, FPIR, marital status, smoking status, drinking status, BMI, hypertension, and physical activity.

Linear regression models were used to assess the associations between log10-transformed dietary Zn intake and the biomarkers of inflammation and oxidative stress. Mediation analysis was performed to estimate direct and indirect effects of dietary Zn on DR risk via inflammatory and oxidative stress pathways, with 1,000 bootstraps after adjusting all covariates.

In this study, data were not weighted as adjustments for demographic factors were already made (24). The statistical significance level was set at a p value of < 0.05. R 4.4.2 software was utilized for all statistical analyses.

After excluding ineligible participants, a total of 2,822 individuals remained for further analysis. All included participants had diabetes, and 556 of them were diagnosed with DR. Among the qualified participants, As shown in Table 1, there were significant differences in sex, race, educational level, family poverty income ratio (FPIR), hypertension, and physical activity among the qualified participants (all p < 0.05).

To assess the relationship between six dietary metals intake and the odds of DR among individuals with diabetes, we employed univariate and multivariate logistic regression. As shown in Table 2, an inverse relationship between log10-transformed dietary Zn and Cu intake and the risk of DR was consistently observed across the crude model, model 1, model 2, and model 3. In model 3, the increase in log10-transformed dietary Zn (OR = 0.53, 95% CI 0.35–0.80, p = 0.003) and Cu (OR = 0.52, 95% CI 0.32–0.82, p = 0.005) intake was significantly correlated with a decreased prevalence of DR. Furthermore, when participants were categorized by the quartiles of log10-transformed dietary metal intake, individuals in the highest quartile group of dietary Zn intake (OR_{Q4 vs. Q1} = 0.62, 95% CI 0.45–0.86, p = 0.004, p for trend = 0.005) and dietary Cu intake (OR_{Q4 vs. Q1} = 0.63, 95% CI 0.45–0.88, p = 0.007, p for trend = 0.01) had a lower risk of DR compared to those in the lowest quartile group through logistic regression model 3. The rest of the four dietary metals intake had no significance to the occurrence of DR in individuals with diabetes.

We applied the WQS model to examine the association between the combined effects of the six dietary metals intake and the prevalence of DR in individuals with diabetes. As shown in Table 3, the WQS index indicated that the combined effects of dietary metals intake were negatively associated with the prevalence of DR (Crude model: OR = 0.83, 95% CI 0.70–0.95, p = 0.005; Model 1: OR = 0.83, 95% CI 0.70–0.96, p = 0.008; Model 2: OR = 0.82, 95% CI 0.69–0.95, p = 0.008; Model 3: OR = 0.79, 95% CI 0.61–0.97, p = 0.024). In the fully adjusted covariates WQS regression (Figure 2), dietary Zn intake received the highest weight of 0.364 for DR risk in the negative direction, compared to weights of 0.209, 0.203, 0.109, 0.062, and 0.052 for Cu, Mg, Fe, Ca, and Se, respectively. In the positive direction, the combined effects of dietary metals intake showed no significant association with the odds of DR after fully adjusting for covariates (Crude model: OR = 0.95, 95% CI 0.83–1.06, p = 0.379; Model 1: OR = 0.93, 95% CI 0.79–1.06, p = 0.263; Model 2: OR = 0.92, 95% CI 0.78–1.05, p = 0.224; Model 3: OR = 0.93, 95% CI 0.75–1.12, p = 0.499), with dietary Zn intake receiving the lowest weight of 0.009 for DR risk, as shown in Table 3 and Supplementary Figure 1.

In the BKMR model, the risk of DR among individuals with diabetes was decreased for the combined intake of six dietary metals mixtures above the 50th percentile compared to the medians (Figure 3A). Figure 3B illustrated an inverse association between dietary Zn intake and the risk of DR, with the intake levels of all other dietary metals held constant at the median. Supplementary Table 1 presented a summary of the PIPs from the BKMR model. Among the dietary metals, Zn intake displayed the highest PIP (0.8574) in relation to the prevalence of DR in individuals with diabetes. Additionally, Supplementary Figure 2 possibly indicated a negative association between dietary Zn intake and DR prevalence compared to other dietary metals, when controlling for the 25th, 50th, and 75th percentiles of other dietary metals, although this association did not reach statistical significance. Furthermore, Supplementary Figure 3 indicated that there were no potential interactions among dietary metals intake.

In the RCS regression model, the log10-transformed dietary Zn intake was found to have an inverse association with the risk of DR in a linear dose–response manner (p for non-linear = 0.2063, p for overall = 0.0041), as shown in Figure 4. The subgroup analysis results are presented in Figure 5, indicating that the inverse correlation between dietary Zn intake and DR was not modified by various factors such as age, sex, race, educational level, marital status, FPIR, smoking status, drinking status, BMI, hypertension and physical activity (all p for interaction > 0.05). As shown in Supplementary Table 2, the results of the linear analysis suggested that dietary Zn intake was significantly correlated with the concentration of serum ALB with all covariates controlled (Model 3: β = 2.25, 95% CI 1.33–3.82, p = 0.003). Additionally, the increase of serum ALB could reduce the incidence of DR in individuals with diabetes (Model 3: OR = 0.95, 95% CI 0.92–0.97, p < 0.001) (Supplementary Table 3). According to Table 4, the results of the mediation analysis showed that serum ALB level mediated the negative association between dietary Zn intake and DR prevalence after fully adjusting for covariates (Total effect = −0.128321, 95% CI -0.232569--0.047959, p = 0.002; Indirect effect = −0.008067, 95% CI -0.015724--0.002561, p = 0.002; Mediation proportions = 0.062866, 95% CI 0.018701–0.163004, p = 0.004).