Association between metal mixture exposure and abnormal glucose metabolism in multiple mixture exposure models: Evidence from NHANES 2015–2016 Authors: Jiamin Zhu, Shiman Hu, Shanshan Wang, Yuting Zhang, Qingyi Zhu, Mingzhi Zhang, Zhonghua Shi Published: 2023 ABSTRACT Graphical abstractKeywords: Metals, Co-exposure, Abnormal glucose metabolism, Diabetes, Prediabetes, Machine learning Introduction Type 2 diabetes is defined as a severe endocrine disorder characterized by hyperglycemia and insulin resistance. In addition, prediabetes is used increasingly to describe people with raised glucose levels above normal but below the diagnostic criteria for type 2 diabetes, including impaired fasting glucose, a slightly higher HbA1c level, and impaired glucose tolerance (ADA, 2018). According to the IDF Diabetes Atlas 10th Edition, the prevalence of diabetes contributes to about 10.5 % around the world, of which the United States ranks 4th at 32.2 million adults with type 2 diabetes and the number is increasing significantly every year (Magliano et al., 2021). Moreover, it is estimated that there are 96 million adults (more than a third) in the US who have prediabetes but 80 % of them are in ignorance of it (CDC, 2022b). Diabetes is generally considered to cause a series of severe chronic complications such as cardiovascular diseases, diabetic retinopathy, diabetic nephropathy, diabetic foot, and diabetic ketoacidosis (Hendryx et al., 2021). Approximately 6.7 million deaths in 2021 can be attributed to diabetes, which is responsible for 11.5 % of health expenditure globally (Magliano et al., 2021). Prediabetes is in a high-risk state of developing type 2 diabetes and diabetes-related complications. Early identification and control play a critical role in diabetes healthcare and the reasons need to be urgently addressed. Therefore, we consider combining prediabetes and type 2 diabetes as AGM to discuss the effect of metal mixture exposure on it based on the guideline of the Chinese Diabetes Society (2017 edition) (Weng, 2016). Well-recognized risk factors for type 2 diabetes include family history, obesity, aging, ethnicity, and physical inactivity (Fletcher et al., 2002). However, the current prevalence of type 2 diabetes is dramatically increasing and it is important to identify other potential contributing factors to curb the diabetes epidemic. In recent years, the role of environmental pollutants, particularly metals, in the type 2 diabetes has received increasing attention (Hectors et al., 2011, Javaid et al., 2021). Metals exposure is ubiquitous in the surroundings and easy access to human bodies through diverse routes including dermal absorption, gastrointestinal intake, and respiratory inhalation (Al Osman and Yang, 2019). Additionally, mixed metals exposure has a greater impact on health than single metal exposure due to the cumulative effect of metals toxicity (Xia et al., 2022). Long-term exposure to metals can lead to their accumulation in the liver, kidney, and pancreas, resulting in harmful effects on glucose metabolism and its related pathways, such as glycolysis, glycogenesis, and gluconeogenesis, ultimately leading to the development of type 2 diabetes (Javaid et al., 2021). More specifically, some toxic metals may consume antioxidants and enzymes and then produce excessive reactive oxygen species (ROS) in cells, which can disrupt normal antioxidant defenses and induce oxidative stress (OS) (Ercal et al., 2001). The impairment of pancreatic islets beta cells caused by OS plays a critical role in the development of diabetes, characterized by insulin resistance and impaired secretion (Darenskaya et al., 2021). In addition, there is a study indicating that metals can interfere with protein folding and aggregation (Tamás et al., 2014). Considering the pathogenesis of T2DM, the misfolding and aggregation of islet amyloid polypeptid damage the islet cells and contribute to the development of diabetes (Ahmed et al., 2019). Previous studies evaluated the association between single metal exposure and type 2 diabetes, but these results remain conflicting. Some epidemiologic studies have suggested that chronic exposure to cadmium (Cd) could elevate the level of fasting blood glucose among adults and progress to type 2 diabetes, which may be attributed to the mechanism of oxidative stress (Xiao et al., 2021). However, another study reported that no association between Cd exposure and the incidence of type 2 diabetes (Ji et al., 2021). Similarly, although numerous studies have shown that lead (Pb) exposure has a deleterious impact on the blood glucose level based on epidemiologic studies and animal experiments (Tyrrell et al., 2017, Leff et al., 2018), other studies have found that no statistical relationship between Pb exposure and diabetes or its related diagnostic indicators (Moon, 2013, Cai et al., 2022). These inconsistencies may be attributed to the fact that the effect of mixed metals exposure was not taken into account. In reality, individuals are exposed to various metals simultaneously through different pathways, and the types and doses of metal exposure vary greatly. Thus, assessing the impact of single metal exposure on diabetes may induce serious bias. In addition, attention also should be paid to the interactions among metals. However, the combined effect of multiple metals is not simply the sum of the effects of each metal exposure, and the multicollinearity of metals needs to be considered. Conventional regression models cannot provide accurate interpretation and prediction for the mixed metals exposure. Consequently, it is crucial to employ appropriate approaches to investigate the co-exposure effect of mixed metals. In this cross-sectional study, we utilized population data from the NHANES 2015–2016 to comprehensively and precisely describe the mixed effect of metal exposure on AGM using multiple mixture exposure models. Materials and methods Study population NHANES has been conducted across the United States since the 1960s as a cross-sectional research using complex stratified multi-stage sampling. More detailed methodologies have been provided in the official description previously (CDC, 2022a). We used the public data from 2015 to 2016 cycle year, of which 9971 participants were interviewed. All participants have given their informed consent. The demographics data, diabetes questionnaire, urinary metals, and related laboratory test indicators were included and matched by the unique survey participant identifier, and then 1871 cases remained. The participants missing the results of fasting plasma glucose (FPG), oral glucose tolerance test (OGTT), and glycosylated hemoglobin type A1C (HbA1c) at the same time (N = 80) were deleted then there were 1791 cases left. We excluded those with missing data on smoking (N = 1), BMI (N = 11), urinary creatinine (N = 2), and diagnosed with type 1 diabetes (N = 2), resulting in 1775 participants for our further analysis (Fig. 1). Flow chart of participants included (N = 1775) in NHANES 2015–2016. Exposure assessment Thirteen urinary metals were detected in NHANES 2015–2016 cycle year, including barium (Ba), Cd, cesium (Cs), cobalt (Co), manganese (Mn), molybdenum (Mo), Pb, antimony (Sb), strontium (Sr), thallium (Tl), tin (Sn), tungsten (W), and uranium (U). The concentrations of urinary metals were assayed by inductively coupled plasma dynamic reaction cell-mass spectrometry (ICP-MS). The lower limit of detection (LLOD) for the urinary metals was set based on the results of detection concentrations and the value below the LLOD was replaced by (LLOD/2). Mn was excluded because of its low detection frequency (29.070 %). Ba, Cd, Cs, Co, Mo, Pb, Sb, Sr, Tl, Sn, W, and U were ultimately included in the subsequent analysis. Outcome assessment The initial information on impaired glucose metabolism was obtained from the diabetes questionnaire and laboratory data. AGM was considered as a combination of diabetes and prediabetes. Among them, type 2 diabetes was defined as previously diagnosed diabetes (participants answering “Yes” to the question, “Doctor told you have diabetes”) OR current hypoglycemic medications in use (participants answering “Yes” to the question, “Take diabetic pills to lower blood sugar”) OR current insulin therapy in use (participants answering “Yes” to the question, “Taking insulin now”) OR FPG ≥ 7.0 mmol/L OR OGTT ≥ 11.1 mmol/L OR HbA1c ≥ 6.5 %. Prediabetes was defined as below the diagnostic threshold of diabetes but with 6.1 mmol/L ≤ FPG ≤ 6.9 mmol/L OR 7.8 mmol/L ≤ OGTT < 11.1 mmol/L OR 5.7 % ≤ HbA1c ≤ 6.4 %. The healthy was defined as FPG < 6.1 mmol/L AND OGTT < 7.8 mmol/L AND HbA1c < 5.7 %. Patients under 30 years old who required insulin therapy were diagnosed with type 1 diabetes and excluded from our study population. Covariates Age, urinary creatinine, the ratio of family income to poverty, and BMI were considered as continuous variables, and gender, race, educational level, physical activity, hypertension, smoking, and drinking status were considered as categorical variables. Urinary creatinine was considered as a covariate to interpret the differences in urinary dilution. The prevalence of hypertension in diabetic patients is twice as high as that in the general population (Petrie et al., 2018). People with cigarette smoking and alcohol intake suffered more risk of developing diabetes (Crandall et al., 2009, U.S. Department of Health and Human Services, 2014). Thus, these predisposing factors were considered as covariates. The values of covariates with an 8 %-10 % missing rate were filled with the sequential random forest method with missForest package in R (version 4.1.1) (Alsaber et al., 2021). Statistical analysis The differences between two groups were calculated by student’s t test for continuous variables and chi-square test for categorical variables. The basic characteristics of metals data, including detection frequency, arithmetic mean, geometric mean, percentile, and LLOD were described in Table S2. The differences in urinary metals distribution between included and excluded population as well as AGM and the healthy were calculated in Table S3-S4. Before the statistical analysis, ln-transformed was introduced for metal concentrations to correct the right skewness. Pearson correlation analysis was conducted among urinary metals in order to test the correlations (Table S1). To enhance the early recognition and control of AGM, we further investigated the association between mixed metals exposure and prediabetes employed four statistical models. GLM The logistic regression model was fitted to evaluate the effect of single metal exposure on AGM. Model 1 was not adjusted; Model 2 was adjusted for age, educational level, and urinary creatinine and model 3 was adjusted completely. The OR with a 95 %CI was calculated to determine the risk of AGM associated with the metal mixture exposure. WQS regression model The WQS model was used to estimate the metal mixture exposure and single metal exposure load. Model 1 was not adjusted; Model 2 was adjusted for age, educational level, and urinary creatinine and model 3 was adjusted completely. The WQS index was derived by bootstrapping 1000 times which represented the mixture exposure effect on AGM. Also, individual metal weights were calculated to represent the contribution of single metal exposure on AGM. LASSO regression model The LASSO model was conducted to select metals that impacted significantly on AGM. 100-fold cross-validation was used to choose the optimal LASSO lambda of a standard error with the minimum mean squared error to fit the model. The selected variables were analyzed by the logistic regression with fully adjusted covariates to evaluate their effects. BKMR model The BKMR model was used to fit the overall effect on AGM as well as the nonlinear relationship and interaction among metals. According to Pearson correlation analysis, urinary metals were divided into two groups. Markov chain Monte Carlo method was used to calculate the group posterior inclusion probability (groupPIP) and conditional posterior inclusion probability (condPIP) with 5000 iterations. All statistical analysis was performed in R (version 4.1.1), and P < 0.05 was considered significant. Results Demographic characteristics The demographic characteristics of participants were shown in Table 1. A total of 1775 participants were eligible for this study, and the frequency of AGM was 48.113 %. Significant differences were displayed in age, gender, race, educational level, smoking status, hypertension, the ratio of family income to poverty, BMI, and physical activity between AGM group and the healthy group. AGM group was higher in mean age (56.90 years vs 40.33 years) and BMI (30.78 kg/m2 vs 28.04 kg/m2) than the healthy group. In addition, AGM group accounted for the most in the smoking status (46.49 % vs 37.57 %) and hypertension (53.28 % vs 16.94 %) compared to the healthy group. There was no significant difference in the drinking status. Table 1Demographics characteristics of study population (N = 1775).CharacteristicsAGMThe healthyP valueN = 854N = 921Age (mean ± SD) (year)56.90 ± 15.6640.33 ± 16.45<0.001Gender0.022 Male449 (52.58 %)433(47.01 %) Female405 (47.42 %)488 (53.99 %)Race0.006 Mexican American169 (19.79 %)146 (15.85 %) Other Hispanic128 (14.99 %)120 (13.03 %) Non-Hispanic White252 (29.51 %)328 (35.61 %) Non-Hispanic Black187 (21.90 %)171 (18.57 %) Non-Hispanic Asian95 (11.12 %)118 (12.81 %) Other Race - Including Multi-Racial23 (2.69 %)38 (4.13 %)Education<0.001 Less than 9th grade140 (16.39 %)69 (7.49 %) 9-11th grade111 (13.00 %)103 (11.18 %) High school graduate198 (23.19 %)205 (22.26 %) Some college or AA degree220 (25.76 %)281 (30.51 %) College graduate or above185 (21.66 %)263 (28.56 %)Ratio of Family Income to Poverty (mean ± SD)2.30 ± 1.452.61 ± 1.55<0.001Smoking<0.001 Yes397 (46.49 %)346 (37.57 %) No457 (53.51 %)575 (62.43 %)Drinking0.041 Yes570 (66.74 %)657 (71.34 %) No284 (33.26 %)264 (28.66 %)Hypertension<0.001 Yes455 (53.28 %)156 (16.94 %) No399 (46.72 %)765 (83.06 %)BMI (mean ± SD) (kg/m2)30.78 ± 7.1628.04 ± 6.40<0.001Physical activity<0.001 Vigorous259 (30.33 %)467 (50.71 %) Moderate248 (29.04 %)223 (24.21 %) No or lower347 (40.63 %)231 (25.08 %) Demographics characteristics of study population (N = 1775). Distribution of urinary metals Mo and Sr were presented with the highest detection frequency (100.000 %). Mn was excluded in the subsequent analysis because of the low detection frequency (29.070 %). The detection frequency of Ba, Cd, Co, Cs, Pb, Sn, and Tl were higher than 90.000 % (93.521 % − 99.944 %) and the rest contained about 80.000 % (79.493 % − 85.577 %). Additionally, the concentration of Sr was the highest and the U was the lowest. Pearson correlation analysis showed there was a significantly positive correlation between any two metals (Fig. 2). Sr had the highest correlation with Ba, while Cs had a higher correlation with Pb, Tl, Co, Mo, and Sr. The correlation coefficient of U, Sb, and Sn with other metals was low. Pearson correlation analysis of urinary metals exposure. All the correlations were significant (P value < 0.001). Independent associations between urinary metals exposure and AGM Logistic regression was used to explain associations between single metal exposure and AGM (Table 2). In model 3, Cd exposure exhibited a positive association with AGM (OR: 1.060, 95 %CI: 1.032–1.089, P value < 0.001). Tl exposure was also positively associated with AGM (OR: 1.039, 95 %CI: 1.004–1.075, P value = 0.031). No significant associations were found in the other metals. Table 2The association between single metal exposure and AGM using logistic regression.MetalsModel 1Model 2Model 3OR (95 %CI)P valueOR (95 %CI)P valueOR (95 %CI)P valueBa0.979 (0.958,1.001)0.0630.998 (0.976,1.019)0.8221.011 (0.990,1.032)0.296Cd1.115 (1.091,1.139)<0.0011.037 (1.011,1.064)0.0061.060 (1.032,1.089)<0.001Co0.984 (0.956,1.012)0.2600.985 (0.955,1.016)0.3311.004 (0.974,1.035)0.808Cs1.017 (0.984,1.051)0.3090.997 (0.959,1.037)0.8931.019 (0.982,1.058)0.318Mo1.023 (0.997,1.049)0.0871.024 (0.994,1.055)0.1111.028 (0.999,1.058)0.058Pb1.061 (1.035,1.088)<0.0010.980 (0.952,1.009)0.1681.003 (0.975,1.032)0.828Sb1.004 (0.977,1.032)0.7631.036 (1.005,1.068)0.0241.026 (0.996,1.056)0.091Sn1.059 (1.038,1.080)<0.0011.018 (0.997,1.040)0.0991.008 (0.988,1.030)0.430Sr0.987 (0.961,1.013)0.3160.996 (0.971,1.022)0.7751.024 (0.998,1.050)0.069TI0.968 (0.938,1.000)0.0491.028 (0.992,1.065)0.1321.039 (1.004,1.075)0.031W0.996 (0.974,1.019)0.7581.031 (1.006,1.056)0.0131.022 (0.998,1.046)0.069U1.023 (0.999,1.048)0.0561.018 (0.994,1.043)0.1441.022 (0.999,1.047)0.064OR: odds ratio; CI: confidence interval. Model 1 was not adjusted; Model 2 was adjusted for age, educational level, and urinary creatinine; Model 3 was adjusted for gender, age, race, educational level, ratio of family income to poverty, smoking status, drinking status, hypertension, BMI, physical activity, and urinary creatinine. The association between single metal exposure and AGM using logistic regression. OR: odds ratio; CI: confidence interval. Model 1 was not adjusted; Model 2 was adjusted for age, educational level, and urinary creatinine; Model 3 was adjusted for gender, age, race, educational level, ratio of family income to poverty, smoking status, drinking status, hypertension, BMI, physical activity, and urinary creatinine. WQS regression analysis of associations between metal mixture exposure and AGM The results of WQS regression were presented in Table 3. In model 3, the positive WQS index for metal mixture exposure was significantly associated with AGM (OR: 1.358, 95 %CI: 1.007–1.832, P value = 0.045). Cd exposure had the largest weight in not and completely adjusted models (Model 1: 0.712; Model 3: 0.330), contributing the most to the positive correlation between metal mixture exposure and AGM. The estimated WQS weight of each metal for three models was displayed in Fig. 3. Table 3The association between metal mixture exposure and AGM using WQS regression.WQS indexOR95 % CIP valueModel 11.490(1.309,1.696)<0.001Model 21.321(0.994,1.755)0.055Model 31.358(1.007,1.832)0.045OR: odds ratio; CI: confidence interval. Model 1 was not adjusted; Model 2 was adjusted for age, educational level, and urinary creatinine; Model 3 was adjusted for gender, age, race, educational level, ratio of family income to poverty, smoking status, drinking status, hypertension, BMI, physical activity, and urinary creatinine. The association between metal mixture exposure and AGM using WQS regression. OR: odds ratio; CI: confidence interval. Model 1 was not adjusted; Model 2 was adjusted for age, educational level, and urinary creatinine; Model 3 was adjusted for gender, age, race, educational level, ratio of family income to poverty, smoking status, drinking status, hypertension, BMI, physical activity, and urinary creatinine. Estimated metal weights for WQS index of three models. Model 1 (A) was not adjusted; Model 2 (B) was adjusted for age, educational level, and urinary creatinine; Model 3 (C) was adjusted for gender, age, race, educational level, ratio of family income to poverty, smoking status, drinking status, hypertension, BMI, physical activity, and urinary creatinine. LASSO regression analysis of associations between metal mixture exposure and AGM The Cross validation plot for the penalty term of LASSO regression and the LASSO regression coefficients based on the different values of the penalty parameter of 12 metals were shown in Fig. 4. Cd, Tl, Co, Sn were selected with different LASSO regression coefficients (Cd, 0.392; Tl, −0.177; Co, −0.104; Sn, 0.069). According to the results of completely adjusted logistic regression, Cd and Tl exposure were significantly positively associated with AGM. Results of LASSO regression. A. LASSO regression coefficients plot; B. Cross validation plot for the penalty term. BKMR analysis of associations between metal mixture exposure and AGM Based on the Pearson correlation analysis, Ba, Cs, Pb, Tl, Co, Mo, and Sr with high correlation were classified as group 1, and the others were classified as group 2 for the subsequent BKMR analysis. The groupPIP and condPIP of associations between urinary metals exposure and AGM were calculated. Group 2 had a larger groupPIP (1.000), of which the condPIP of Cd exposure accounted for the largest at 1.000. When the other metals were fixed at the median, the dosimetric response curve for Cd exposure was significantly elevated (Fig. 5A). There was no interaction effect between Cd and Tl exposure on AGM based on the bivariate exposure–response functions (Fig. 5B). The overall effect on AGM exhibited a significantly increasing trend. (Fig. 5C). When the other metals were set at the 25th, 50th, and 75th percentile respectively, Cd exposure was associated positively with AGM (Fig. 5D). BKMR results of the healthy and AGM. A. univariate exposure–response functions for single metal exposure. B. interaction effect of Cd and Tl exposure. C. overall effect of metal mixture exposure. D. effect of single metal exposure. The model was adjusted for gender, age, race, educational level, ratio of family income to poverty, smoking status, drinking status, hypertension, BMI, physical activity, and urinary creatinine. The associations between metal mixture exposure and prediabetes Cd (OR: 1.053, 95 %CI: 1.021–1.086, P value = 0.001), Tl (OR: 1.041, 95 %CI: 1.001–1.082, P value = 0.045), and Sr (OR: 1.032, 95 %CI: 1.003–1.063, P value = 0.033) exhibited positive associations with prediabetes based on the completely adjusted logistic regression (Table S5). In model 3 of WQS (OR: 1.509, 95 %CI: 1.068–2.132, P value = 0.019), the positive index for metal mixture exposure was significantly associated with prediabetes (Table S6). In addition, Cd played the most contributing role (0.254), followed by Tl (Fig.S1). In the LASSO regression model, Cd, Tl, and Co (coefficients: Cd, 0.389; Tl, −0.047; Co, −0.140) were selected (Fig.S2), while Cd and Tl exposure were significantly positively associated with prediabetes according to the results of completely adjusted logistic regression. According to the BKMR model (Fig. S3), the overall effect displayed an upward trend, of which Cd exposure contributed the most. No interaction effect was observed between Cd and Tl exposure. Discussion and conclusions To our knowledge, our research was the first study to estimate the association between metal mixture exposure and AGM through multiple mixture exposure models. Four regression models were fitted and combined to verify the robustness of our findings. With the WQS and LASSO models, we found that metal mixture exposure was associated with AGM in a positive manner, of which Cd and Tl were recognized as the leading metals, aligned with the results from GLM. The BKMR model also demonstrated an overall positive effect on AGM, with Cd exposure being particularly prominent. No interaction effect between Cd and Tl exposure was observed. Furthermore, the results of prediabetes were also dominated by Cd and Tl exposure, consistent with the findings for AGM. In our analysis, Cd and Tl exposure were significantly related to AGM in logistic regression, respectively. However, it was unable to capture the non-linear relationship and interaction effect of mixture exposure. In the WQS model, the bootstrap approach was used to empirically calculate the weight index. However, it only allowed for the analysis of either positive or negative effects at a time, potentially leading to neglect of information. Nevertheless, the overall effect demonstrated a positive association with AGM, with Cd showing the highest weight, followed by Tl. The LASSO model selected the optimal variables by minimizing the other coefficients to 0, which was available for multicollinearity data. Cd and Tl were selected based on the LASSO coefficients and logistic regression. To sum up, GLM, WQS, and LASSO models all indicated that Cd and Tl exposure were positively associated with AGM. The BKMR model was fitted to identify the interaction effect and non-linear relationship between any two metals. The overall effect showed a significantly elevated trend, with Cd exposure being significantly positively associated with AGM. Importantly, no interaction effect between Cd and Tl exposure was observed, suggesting that Cd and Tl exposure independently contributed to the metal mixture exposure. One study found that co-exposure to Cd, arsenic, nickel, selenium, and zinc was associated with the elevated fasting blood glucose in community-based Chinese adults (Li et al., 2019). In addition, epidemiological research displayed generally positive associations of metal mixture exposure with diabetes, and with HbA1c levels, including Mn, selenium, Pb, zinc, Cd, copper, and mercury (Wu et al., 2022). The toxicity effect of metal mixture exposure is additive and even be magnified due to competitive absorption and bioavailability (Wu et al., 2016). Co-exposure to low doses of metals in the surroundings is usually grossly neglected thus we should attach importance to the effect of metal mixture exposure. According to our results, Cd and Tl exposure among metal mixture exposure exhibited a significantly positive association with AGM. A growing number of researchers have investigated that Cd exposure was related to the hyperglycemia states (Treviño et al., 2015, Xiao et al., 2021, Wang et al., 2022). Notably, a meta-analysis found that Cd exposure was positively associated with the risk of both type 2 diabetes and prediabetes with moderate-certainty evidence (Filippini et al., 2022). However, a cross-section study from NHANES suggested Cd exposure upper the normal level was not associated with diabetes in the general people from 1999 to 2020 (Gong et al., 2022). The reason for the inconsistency with our results may be the neglect of synergistic toxic effects from other metals and the prediabetes not included in the analysis. Therefore, results supported by multiple mixture exposure models may be more reliable. Cd tends to accumulate in the kidney, liver, and pancreas through metal transporters (Filippini et al., 2022). Within the pancreas, Cd accumulated preferentially in islets, which may damage beta cells directly and decrease the secretion of insulin (Himeno et al., 2019). Additionally, Cd also caused the rise in enzymes activity for gluconeogenesis and the decrease of glucose transporters expressions (Chapatwala et al., 1982, Han et al., 2003). Overall, Cd may exert several diabetogenic effects through the dysfunction of islet, an increase of gluconeogenesis, and down-regulation of glucose transporters (Joshua and Christopher, 2016), leading to hyperglycemia states. Research on the association between Tl and type 2 diabetes was scarce, and more recent studies focused on gestational diabetes mellitus. A birth cohort study suggested that Tl exposure was associated with gestational diabetes mellitus in both a dose-dependent manner (Zhang et al., 2021) and an age-dependent manner (Zhu et al., 2019). However, a study from NHANES demonstrated no association between urinary Tl and diabetes using the logistic regression model (Menke et al., 2016). This was inconsistent with our results, which may be due to their diagnosis of diabetes using only HbA1c and inadequate covariates (Menke et al., 2016). Tl was considered cumulative in human body by diverse routes and the pancreas was one of the organs containing the highest level (André et al., 2010, Cvjetko et al., 2010). Tl exposure could induce the formation of reaction oxygen species and damage the mitochondria (Hanzel and Verstraeten, 2006). Oxidative stress was considered a potential mechanism for beta cells dysfunction to cause the insulin resistance and even diabetes (Ceriello and Motz, 2004). Thus, we preliminarily speculated that Tl exposure may lead to prediabetes state and progress to diabetes under the insulin disorders induced by chronic inflammation. Our research is the first to report the positive association between Tl exposure among mixed metals and AGM. Further studies are required to figure out more mechanisms of Tl exposure on impaired glucose metabolism. Taken together, our analysis had the following advantages. First of all, we took the combined effect of metal mixture into account and fitted four statistical models to assess the independent and combined aspects of their effects. The use of multiple exposure models helped assess the overall effect and identified the leading metals, which reduced the bias of single metal exposure. Secondly, the diagnostic criteria for diabetes and prediabetes were comprehensive, encompassing not only the results of diabetes questionnaires but also the WHO diagnostic criteria including FPG, OGTT, and HbA1c. The establishment of sound diagnostic criteria helped to improve the sensitivity of disease identification and avoid the neglect of AGM. Finally, we also considered prediabetes as the outcome to fit four statistical models, which helped to explore more sensitive and incipient effects. More early intervention can be exerted to prevent the development of impaired glucose metabolism. Apart from these strengths, our study had several limitations. Firstly, we used the unweighted data. instead of the recommended weighted data from NHANES. However, the unweighted data are preferred when covariates were included in regression models to avoid over-adjustment bias. Secondly, the measurement of urinary metals was based on a single time point instead of 24-hour continuous collection in NHANES, which may not accurately describe the daily exposure (Menke et al., 2016). Moreover, our research was based on a cross-sectional survey. Further prospective studies as well as multi-center studies including multi-regional populations could be performed to validate the underlying causal association. In summary, we fitted four models to assess the effect of metal mixture exposure on AGM. The overall effect was significantly associated with AGM, of which Cd and Tl exposure accounted for the most, consistent with results of prediabetes. Therefore, Cd and Tl exposure played vital roles in AGM, especially in the early onset. Our study provides a novel and preventable perspective on the association of metal mixture exposure and AGM, providing the foundation for the development of applicable environmental policies. CRediT authorship contribution statement Jiamin Zhu: Data curation, Software, Writing – original draft, Writing – review & editing. Shiman Hu: Visualization, Validation, Writing – original draft. Shanshan Wang: Methodology, Investigation. Yuting Zhang: Data curation, Investigation. Qingyi Zhu: Data curation, Methodology. Mingzhi Zhang: Conceptualization, Project administration, Supervision, Writing – review & editing. Zhonghua Shi: Project administration, Supervision, Writing – review & editing, Funding acquisition. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This work was supported by National Key R&D Program of China [grant number 2022YFC2703503]; National Natural Science Foundation of China [grant number 82371697; 81971410]; Open Project of Jiangsu Biobank of Clinical Resources [grant number TC2021B024]. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.crtox.2023.100141. Contributor Information Mingzhi Zhang, Email: mingzhizhang@njmu.edu.cn. Zhonghua Shi, Email: jesse_1982@163.com. Supplementary data The following are the Supplementary data to this article: Supplementary data 1 References Associated Data Supplementary Materials Supplementary data 1