The NHANES is a national, biennial cross-sectional survey to assess the health and nutritional status of the general populations in the U.S., using a multistage, probability-sampling design to generate population-level estimates. It examines a nationally representative sample of 5,000 persons selected from 15 different locations. All study protocols and results are available on the website of the American Centers for Disease Control and Prevention (CDC).¹ The survey was approved by the NCHS Research Ethics Review Board and all participants provided informed consent.

Data from three cycles (2011–2012, 2013–2014, and 2015–2016) of NHANES were combined in the study. Participants who were not tested for heavy metals were excluded. Dependent variable was the self-reported history of stroke. Participants with missing information on self-reported stroke outcomes were also excluded. Finally, 11,803 participants aged ≥20 years were involved in the final analysis.

Pb, Cd, Hg, and Mn content of whole blood specimens were measured using mass spectrometry after a simple dilution sample preparation step. Inductively coupled plasma dynamic reaction cell mass spectrometry (ICP-DRC-MS) was used to measure the entire panel of the serum concentrations of Cu, Se, and Zn. According to the NHANES standard, the limit of detection (LOD) divided by the square root of two was used to replace the values below the LOD.

Covariates included demographics such as age, gender (male or female), race (Mexican American, Non-Hispanic White, Non-Hispanic Black, Non-Hispanic Asian, Other Hispanic), family income to poverty ratio(a measure of family income relative to poverty guidelines specific to the survey year), body mass index (kg/m²), smoking status (≥100 or < 100 cigarettes/entire life), alcohol consumption (<12 or ≥ 12 drinks/year), and disease histories (yes or no) of coronary heart disease and diabetes. Data of covariates were collected through home interview, laboratory measurements, and questionnaires.

According to the NHANES standard, we used the 2011–2016 sample weights of NHANES to calculate the new weights of the study samples. The PROC SURVEY procedure was employed to account for the intricate sampling design of NHANES, which included weight, cluster, and strata statements, to incorporate sampling weights. Data on seven metals were natural log 10 transformed to normalize their distribution for analysis. Statistical analyses were performed with SAS software (V.9.4).

To explore the non-linear relationship, restricted cubic spline (RCS) with three knots coupled with a logistic regression model was to assess the dose–response relationship between single and multiple metals (continuous variables) and stroke. In addition, the RLM function was used to test the nonlinear effect of the model. The R package (“rms”) was used to achieve this.

WQS regression was adopted to estimate the association between multiple-metal exposure and stroke (17). We used the R package (“gWQS”) to calculate the WQS index comprised of weighted sums of individual metal concentrations. Heavy metals are often correlated with one another, and traditional regression models can struggle with multicollinearity, leading to unreliable estimates. It is important to note that all heavy metals in the model are assumed to have the same direction of association with the outcome. Qgcomp, an adaptive modeling method with weighted quantile sum (WQS) regression, unlike the WQS model, does not assume a fixed direction and can evaluate the differing directions of mixed effects for individual exposures (18). To investigate the difference of heavy metal mixing effect between gender, age and BMI, stratified analysis was used (BMI was divided into two groups by 24 kg/m², and age was divided into two groups by 60 years). The R package (“qgcomp”) was used to achieve this.

All models were adjusted for gender, age, race/ethnicity, family income-to-poverty ratio, drinking alcohol status, smoking status, body mass index, diabetes, and coronary heart disease. Missing covariate values were processed using multiple imputation. p value <0.05 was considered significant.

A total of 11,803 participants (438 were diagnosed with stroke) with complete data on heavy metals and stroke were included in this study. The survey-weighted descriptive statistics are presented in Table 1. The weighted mean age of participants was 49.14 years, and 48.45% (n = 5,781) were male. Overall, age, ethnicity, drinking alcohol status, smoking status, family PIR, and comorbidities (diabetes and coronary heart disease) were statistically significant differences between stroke and non-stroke participants (p < 0.05). Stroke participants had higher proportion of diabetes, coronary heart disease prevalence, older age, and smoking status with ≥100 cigarettes/entire life. The non-stroke participants had higher proportion of drinking alcohol status with ≥12 drinks/year, and higher family PIR. The participants’ weighted mean concentrations of Cd, Pb, Hg, Mn, Cu, Se, and Zn was 0.34 ± 0.49 μg/L, 1.19 ± 1.499 μg/L, 1.17 ± 2.169 μg/L, 10.38 ± 3.919 μg/L, 118.85 ± 29.189 μg/L, 187.32 ± 26.419 μg/L, and 81.80 ± 15.229 μg/L, respectively. The detection rate of all metals was greater than 75%.

The individual effects of seven metals on stroke using the survey-weighted logistic regression are provided in Table 2. After adjusting for the potential confounders, the log-transformed Cd and Zn level was associated with stroke (Cd: OR = 1.13, 95% CI: 1.01, 1.26; Zn: OR = 0.17, 95% CI: 0.05, 0.58). The highest exposure quantile of blood Cd increased the risk of stroke compared to quantile 1 (OR = 1.54, 95% CI: 1.02, 2.30; P trend = 0.004). While highest exposure quantile of blood Zn decreased the risk of stroke compared to quantile 1 (OR = 0.47, 95% CI: 0.24, 0.92; P trend = 0.024). Besides, there was an elevated stroke risk in Q3 of Cd (OR = 1.88, 95% CI: 1.29, 2.73), Q2 of Hg (OR = 1.98, 95% CI: 1.17, 3.34), compared to Q1. And there was a lower stroke risk in Q2 of Mn (OR = 0.68, 95% CI: 0.47, 0.96), compared to Q1.

Figures 1a–f showed the results of the restricted cubic spline analyses. The results showed that except Cd, the other six heavy metals had nonlinear association with stroke risk (P Zn for nonlinearity = 0.004, P Se for nonlinearity = 0.007, P Cu, Mn, Hg and Pb for nonlinearity<0.001). Stroke risk decreased with increasing Se, while increasing Cu, Hg, and Mn can increase stroke risk. We also found a suggestion of U-shaped associations between Zn concentrations and stroke risk.

WQS regression analysis showed associations between all seven metals (Pb, Cd, Hg, Mn, Cu, Se, and Zn) co-exposure and stroke (p < 0.001). According to the logistic results, the preset beta values was positive. The estimated weights for the seven metals are provided in Figure 2. Blood Cd was the highest weighted metal in general population, followed by blood Cu.

Qgcomp analyzed the estimated exposure weights in the positive and negative directions, respectively. The joint weighted value of exposure was significantly associated with stroke risk (OR = 1.10, 95%CI: 1.01, 1.21; p < 0.05). The metal of largest positive weight in stroke risk was Cd (42.4%), followed by Cu (40.6%), while Zn (84.8%) was the largest negative weight metal in stroke risk (Figure 3). And there was a significant correlation between metals and stroke risk in smoke group (OR = 1.38, 95%CI: 1.22, 1.55; p < 0.05), young group (OR = 1.46, 95%CI: 1.23, 1.73; p < 0.05), and female group (OR = 1.17, 95%CI: 1.03, 1.33; p < 0.05).

In smoke group, Cu (weighted 49.1%) was the most positive weighted metals contributing to the stroke risk, followed by Cd (weighted 32.0%). Zn (weighted 70.5%) was the most negative weighted metals (Figure 4). In young group, Mn and Cd (weighted 42.6 and 21.9%) were the most positive weighted metals contributing to the stroke risk, while Zn (weighted 58.1%) was the most negative weighted metals, followed by Se (weighted 41.9%) (Figure 5). In female group, Zn (weighted 55.5%) was also the most negative weighted metals contributing to the stroke risk, the metal of largest positive weight in stroke risk was Cu (weighted 36.9%) (Figure 6). The results of other subgroups were not statistically significant (Supplementary Figures S1–S5).