Early Pregnancy Essential and Non-Essential Metal Mixtures and Gestational Glucose Concentrations in the 2nd Trimester: Results from Project Viva

Abstract

Metals are involved in glucose metabolism, and some may alter glycemic regulation. However, joint effects of essential and non-essential metals on glucose concentrations during pregnancy are unclear. This study explored the joint associations of pregnancy exposures to essential (copper, magnesium, manganese, selenium, zinc) and non-essential (arsenic, barium, cadmium, cesium, lead, mercury) metals with gestational glucose concentrations using1,311 women enrolled 1999–2002 in Project Viva, a Boston, MA-area pregnancy cohort. The study measured erythrocyte metal concentrations from 1st trimester blood samples and used glucose concentrations measured 1 hour after non-fasting 50-gram glucose challenge tests (GCT) from clinical gestational diabetes screening at 26–28 weeks gestation. Bayesian Kernel Machine Regression (BKMR) and quantile-based g-computation were applied to model the associations of metal mixtures—including their interactions—with glucose concentrations post-GCT. We tested for reproducibility of BKMR results using generalized additive models. The BKMR model showed an inverse U-shaped association for barium and a linear inverse association for mercury. Specifically, estimated mean glucose concentrations were highest around 75th percentile of barium concentrations [2.1 (95% confidence interval: −0.2, 4.4) mg/dL higher comparing to the 25th percentile], and each interquartile range increase of erythrocyte mercury was associated with 1.9 mg/dL lower mean glucose concentrations (95% credible interval: −4.2, 0.4). Quantile g-computation showed joint effects of all metals, essential-metals, and non-essential metals on gestational glucose concentrations were all null, however, we observed evidences of interaction for barium and lead. Overall, we found early pregnancy barium and mercury erythrocytic concentrations were associated with altered post-load glucose concentrations in later pregnancy, with potential interactions between barium and lead.

1. INTRODUCTION

Gestational diabetes mellitus (GDM) affects approximately 5–9% of pregnancies in the U.S. (DeSisto et al. 2014; Lavery et al. 2017) and can lead to poor pregnancy outcomes as well as long-term morbidity for both the mother and the child (Bellamy et al. 2009; Dabelea 2007; Kim 2014). Even milder degrees of hyperglycemia may increase risks of these adverse health outcomes (Hapo Study Cooperative Research Group et al. 2008; Lowe et al. 2019; Scholtens et al. 2019). Therefore, it is critical to identify modifiable risk factors associated with glucose concentrations during pregnancy.

Environmental factors may contribute to the development of glucose intolerance in pregnancy (Ehrlich et al. 2016; Metzger et al. 2007). In particular, metals and metalloids are important to consider because many are known to be involved in glucose metabolism and exposures are modifiable through dietary management [arsenic and selenium are metalloids, but for simplicity hereafter we refer to the mixture of metals and metalloid as “metal” mixture]. Essential metals, such as magnesium (Mg), manganese (Mn), Se, and zinc (Zn), help to sustain glucose homeostasis through stimulating insulin secretion, increasing insulin sensitivity, regulating hormone concentrations, and/or acting as antioxidants (Iavicoli et al. 2009; Kaur and Henry 2014; Kong et al. 2016; Mwiti Kibiti and Jide Afolayan 2015). In fact, Zn and Se are studied as potential preventive and treatment agents for type 2 diabetes (El Dib et al. 2015; Karamali et al. 2015; Stranges et al. 2007). Non-essential metals, such as As, cadmium (Cd), mercury (Hg), and lead (Pb), on the other hand, are known toxicants and can induce hyperglycemia. Mechanistically, the pancreatic β-cells have high expression of metal transporter and low expression of antioxidants, thus they are especially prone to oxidative stress induced by metals (Chen et al. 2009; Hectors et al. 2011; Kuo et al. 2013); furthermore, many evidence indicated that As induces oxidative stress and might impair glucose metabolism by altering signaling transcription factors and affecting the insulin-stimulated glucose uptake (Chen et al. 2009); exposure to Cd may increase gluconeogenesis, alter glucose transport and disrupt the function of pancreatic islets (Edwards and Ackerman 2016); and both organic and inorganic Hg are well known to induce cellular damages in various cell types including pancreatic islet β-cells (Chen et al. 2009).

Evidence on the association between metals exposure and glycemic status during pregnancy is sparse. A number of studies have evaluated Mg, Se and Zn as they are related to GDM/hyperglycemia risks; results showed relatively consistent inverse associations for Mg and Se but inconsistent findings for Zn (Goker Tasdemir et al. 2015; Kong et al. 2016; Mishu et al. 2019; Nabouli et al. 2016; Wilson et al. 2016; Zheng et al. 2008; Zheng et al. 2020). Studies on non-essential metals, such as As, Barium (Ba), Cd, Pb, and Hg, also have had inconsistent results (Ettinger et al. 2009; Farzan et al. 2016; Peng et al. 2015; Rahman et al. 2016; Shapiro et al. 2015; Soomro et al. 2019; Valvi et al. 2017; Varshavsky et al. 2020) and little is known on the effects of metals and metal mixtures on glycemic status during pregnancy; except for blood As which had consistently shown to be associated with risk of GDM. Furthermore, most of the available literature comes from cross-sectional analyses and, few studies accounted for concurrent essential and non-essential metal exposure, as well as their potential interactions.

This study aimed to use a prospective study design to evaluate the associations between erythrocyte metal concentrations and glycemic status during pregnancy, accounting for individual and joint effects of essential and non-essential metals, as a mixture, employing novel statistical approaches, including the quantile based g-computation (Keil et al. 2020) and Bayesian Kernel Machine Regression (BKMR) (Bobb et al. 2018; Bobb et al. 2015), which allow the inclusion of a relatively large number of metals and examines their potential interactions. We hypothesized that concentrations of metal mixtures among pregnant women during early pregnancy would be positively associated with glucose concentrations later in mid-pregnancy and there would be potential interactions between metals.

2. METHODS

2.1. Study population

Women in this study were participants in Project Viva, a prospective observational cohort study designed to examine prenatal diet and other health factors in relation to pregnancy and child health outcomes (Oken et al. 2015). The cohort was recruited during participants’ first prenatal visit at eight obstetrical offices of Atrius Harvard Vanguard Medical Associates in eastern Massachusetts between 1999 and 2002. Inclusion criteria included singleton gestation, ability to answer questions in English, and gestational age < 22 weeks at recruitment. Project Viva included a total of 2,670 pregnancies; 2,128 had live births and continued follow-ups. There were 28 participants with multiple enrollments from different pregnancies and this analysis only included their first enrollment record, leaving 2,100 unique women in the cohort. For this analysis, we selected participants who (1) had blood sample collected at the first trimester for metals analysis, (2) had gestational glucose tolerance test results at the second trimester, (3) had no preexisting diabetes, and (4) had no missing information on the important covariates. A total of 1,325 women met these criteria. After removing pregnant women whose metal measurements were 5 times below or above the standard deviation of the mean on the log scale (N=14), 1,311 women were included for the main analyses. We presented the flow-chart of the study in Figure 1. All women provided written informed consent at recruitment. The Institutional Review board of Harvard Pilgrim Health Care reviewed and approved all study protocols.

Figure 1.

Study flow chart

Abbreviations: GDM, gestational diabetes mellitus; SD, standard deviation; As, arsenic; Cd, cadmium; Cu, copper; Mg, magnesium; Se, selenium; Zn, zinc.

2.2. Metals concentrations

Previous study on blood As and GDM had suggested that early pregnancy is a possible window of susceptibility (Xia et al. 2018), therefore, we used first trimester blood as the biometric for our analysis. Project Viva collected a blood sample at enrollment (median: 10 weeks gestation). The blood sample was centrifuged at 2000 rpm for 10 minutes at 4°C to separate erythrocytes from plasma. Limited by sample availability, we had only erythrocyte samples for metal analysis but did not retain any whole blood. Erythrocyte samples were stored at −70°C prior to metals analyses, which were done at Mount Sinai CHEAR Network Laboratory. The trace element panel included As, Ba, Cd, Hg, Mg, Mn, Se, Pb, Zn, Aluminum (Al), Cobalt (Co), Chromium (Cr), Cesium (Cs), Copper (Cu), Molybdenum (Mo), Nickel (Ni), Antimony (Sb), Tin (Sn), Timonium (Tl), and Vanadium (V). They measured concentrations of Hg using Direct Mercury Analyzer 80 (Milestone Inc.,) and the other metals using triple quadrupole Inductively Coupled Plasma Mass Spectrometry (Agilent 8800 ICP-QQQ) on a single run using appropriate cell gases and internal standards. The 8800 ICP-QQQ, has higher sensitivity and lower background interference than traditional ICP-MS, providing overall better performance than single quadrupole ICP-MS instruments. All sample handlings for metals analysis were done in an ISO class 6 clean room with an ISO class 5 laminar flow clean hood. All machines underwent initial and continuous calibration verification, each run included procedural blanks and we repeated analysis for 2% of the samples. Matrix-appropriate certified reference materials were analyzed once per study and Seronorm-Blood L3 was analyzed once daily as quality control (QC) sample to monitor accuracy. One sample each of the Blind QC pools at high and low levels were run per batch, with a total of 52 duplicate samples included.

Recoveries were within QC standards (90% −110% for all selected metals). Intraday coefficient of variation (CV) was calculated using analysis of inhouse QC pools at three different concentration levels before and after every 10 samples (N=7). Intraday CV was <5% for most of the analytes except Se (<10%). Interday CV for all elements was <15% except for concentrations near the limit of detection (LOD).

We report erythrocyte metals concentrations in ng/g. We selected 11 of the metals that were detected in more than 80% of the samples and whose QC duplicated samples had intra-class correlations above 0.60. Metals concentrations were right-skewed; thus, we performed natural-log transformation and then standardization before statistical modeling. We used the instrument estimated concentrations for values below the LOD; however, since some instrument values were negative (N=20 for As, N=4 for Ba), we right-shifted all metals concentrations by 1 ng/g before log-transformation.

2.3. Gestational glucose tolerance

Around the late 2^nd trimester (median: 28 weeks gestation, range: 12.7 to 36.0), all women underwent a clinical, non-fasting 50g oral glucose challenge test (GCT) as the first step of GDM diagnosis. This study used the continuous glucose concentrations 1-hour post-GCT as the primary outcome.

As a secondary outcome, we also looked a binary outcome of glucose intolerance defined as having either impaired glucose tolerance (IGT) or GDM. As part of usual clinical prenatal care, women with a glucose concentrations ≥140 mg/dL 1-hour post-GCT underwent additional screening for GDM using a fasting 100g oral glucose tolerance test (OGTT). Normal glycemia thresholds were defined by Carpenter-Coustan criteria (American Diabetes Association 2018): fasting < 95 mg/dL, 1 hour < 180 mg/dL, 2 hour < 155 mg/dL, 3 hour < 140 mg/dL. GDM was defined as having 2 or more abnormal OGTT glucose values and IGT was defined as having only 1 abnormal OGTT glucose value. The composite of IGT and GDM was used because of the relatively small sample size for GDM in this study population (N=66).

2.4. Covariates

Participants provided sociodemographic information, personal and family medical history, and lifestyle factors at their initial study visit during prenatal care using structured questionnaire. Based on previous literature (Bouthoorn et al. 2015; Galtier 2010; Hedderson et al. 2010) and using directed acyclic graphs (DAG), we selected the following variables a priori as potential confounders and/or strong predictors of glucose intolerance: maternal age (continuous), self-identified race/ethnicity (white, Black, Hispanic, Other), pre-pregnancy body mass index (BMI) (continuous), GDM in a prior pregnancy (yes, no, nulliparous), maternal education (college graduate, yes or no), and history of diabetes in participants’ biological mother (yes, no, missing/don’t know). Since concentrations of metals may vary throughout pregnancy (Gulson et al. 1997; Vukelic et al. 2012), we also included gestational week at blood collection for metals measurements (continuous) in all models. For sensitivity analysis, we additionally considered first trimester hemoglobin levels (continuous), which were clinical lab values abstracted from medical record; gestational age at the time of glucose tolerance test (continuous), environmental tobacco smoke exposure per week (0 to less than 1 hour, 1–4 hours, 5–8 hours, 9–12 hours, >12 hours); and diet, which included fish consumption (continuous) and n-3 fatty acid intake (continuous). Diet information was collected using a semiquantative Food Frequency Questionnaire (FFQ) after the initial pregnancy visit in the first trimester (median: 12 weeks gestation) (Rifas-Shiman et al. 2009). Fish consumption questions included canned tuna, shellfish, dark meat fish, and other fish.

2.5. Statistical Analysis

There were three parts of our statistical analysis. First, we applied BKMR to evaluate the nonlinearity of effects and interactions among the metal mixture and estimated the individual association of metals on glucose concentrations. Second, we used supplemental methods including GAMs and multivariable linear regressions to confirm findings from BKMR and formally test the effect estimates for individual metals. Third, we estimated the joint associations using BKMR and quantile g-computation models. For the quantile g-computation model, we included interaction and nonlinearity identified by BKMR and grouped metals into all metals, essential, and non-essential. Detailed descriptions of each of the methods follow.

We used the “bkmr” R package (Bobb et al. 2018) and chose component-wise selection where each metal was selected individually as they were not highly correlated. We used the Gaussian kernel with the default parameters, as it has been shown to work well in both simulation studies and environmental exposure-response scenarios (Bobb et al. 2015; Valeri et al. 2017). We ran each Markov Chain Monte Carlo (MCMC) sampler for 100,000 iterations with the default noninformative priors specified in the R package, and used the Geweke’s diagnostic (R package “coda”) (Geweke 1992; Plummer 2006) to evaluate convergence; specifically, we compared the means of the first 10% and the last 50% of the Markov chain and found no evidence of lack of convergence. From the model fit, we derived summary statistics to quantify the relevant features of the exposure-response function, including differences in glucose concentrations associated with changes in the metals concentrations as well as pointwise 95% credible intervals. Details of the BKMR method and implementation can be found elsewhere (Bobb et al. 2018; Bobb et al. 2015).

To evaluate the reproducibility of the results from BKMR and to test the statistical significance of potential interactions, we ran additional multivariable GAMs and multivariable linear models. For GAMs (R packages “gam”), we fitted the main effects of all metals and suggestive pairwise interactions among the metals from the results of BKMR as smooth functions. We used thin plate regression splines (Wood 2017) to represent the smooth terms, setting the maximum possible degrees of freedom allowed for each metal at four. The smoothing parameters were estimated by maximum likelihood (ML). An extra penalty was added to each term through automatic term selection. For the multivariable linear model, we selected important metals and pairwise interactions based on previous results (the main association or the interaction terms with p-value ≤ 0.1 from the GAM) in order to reduce the number of parameters needed to be estimated, and accommodated potential nonlinear associations by categorizing the metals concentrations into tertiles.

We performed quantile-based g-computation (Keil et al. 2020) to estimate the joint associations of all metals, essential metals, and non-essential metals with gestational glucose concentrations. This method relaxes the directional homogeneity assumption and allows for individual components in the mixture to contribute either a positive or a negative weight to a mixture index. The method then estimates the effect of the mixture index on an outcome which can be interpreted as mean difference in outcome across quantile range of the exposure variables as a mixture. Model building was informed by results from BKMR, GAM and linear regression models to include nonlinear and interaction effects. We included the quantile-based g-computation in order to provide effect estimates for comparison with future analyses. We ran the model using the “gqcomp” R package (Keil et al. 2020) and presented the joint association with overall model confidence bounds (estimated by 10,000 bootstraps).

For sensitivity analyses, we first assessed the influence of extreme observations; we fitted the BKMR model and the GAM with the inclusion of the 14 women who were removed in the main analyses due to having extreme metals concentrations. Second, we additionally adjusted for hemoglobin concentration, as some of the metals tend to bind to hemoglobin (deSilva 1984; Goyer 1997). Third, we further adjusted for early pregnancy fish and n-3 fatty acid consumption in the BKMR model and the GAM, as they may be related to certain metals (e.g. Hg, Pb) as well as be protective of glycemia (Santangelo et al. 2016; Schoenaker et al. 2016); our previous study among the same cohort of pregnant women also showed statistically significant positive associations between fish consumption and erythrocyte As and Hg concentrations (Lin 2021). We also tested fish consumption as an effect modifier by using an interaction term between fish consumption and metal concentrations in the model. Fifth, we fitted the GAM with GDM and IGT combined as a clinically relevant binary outcome, adjusting for the same covariates as the main analysis. Lastly, given BKMR is sensitive to the choice of model prior, we included BKMR models with alternative prior to test the robustness of the finding; in particular, we tested different degrees of smoothness (lower b=50 and higher b=200) that control the exposure response function using methods previously reported (Bauer et al. 2020; Howe et al. 2020; Valeri et al. 2017).

We performed the statistical modeling using R version 3.6.0 (R Core Team 2013). Reporting of the manuscript follows the guidelines and checklist for Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) (Von Elm et al. 2007).

3. RESULTS

3.1. Study population characteristics

Characteristics of the pregnant women included in the analysis are shown in Table 1. Among the 1,311 Project Viva mothers, the mean (standard deviation, SD) age at study enrollment was 32.3 (4.6) years, 73% self-identified as white, and pre-pregnancy BMI was 24.9 (5.4) kg/m². The median and interquartile range (IQR) of erythrocyte metals concentrations were 0.84 (0.44, 1.53) ng/g for As, 3.18 (1.99, 5.90) ng/g for Ba, 0.40 (0.28, 0.56) ng/g for Cd, 0.53 (2.03, 3.18) ng/g for Cs, 563 (515, 617) ng/g for Cu, 3.25 (1.64, 6.39) ng/g for Hg, 41.1 (37.0, 46.1) μg/g for Mg, 16.1 (13.1, 20.3) ng/g for Mn, 17.6 (13.5, 23.6) ng/g for Pb, 248 (220, 282) ng/g for Se, and 10.4 (9.27, 11.6) ng/g for Zn, which were all within the recommended reference range (see Table S2 for distribution and normal reference ranges).”All metals were weakly correlated (Figure S1, Spearman correlation coefficients < 0.40), except for four moderately correlated pairs: Cu-Zn, As-Hg, Mg-Se, Cu-Mg (Spearman correlation coefficients = 0.58, 0.56, 0.43, 0.40, respectively). The mean (SD) glucose concentrations post-GCT was 114 (27) mg/dL; 221 (17%) participants had glucose concentrations ≥ 140 mg/dL and received additional screening for GDM using a fasting 100g oral OGTT and 66 (5%) women were subsequently diagnosed with GDM, and another 38 (3%) met criteria for IGT. The Spearman correlations between glucose concentrations (continuous) and GDM/IGT were 0.43 (p<0.001).

Table 1.

Characteristics of pregnant women in Project Viva included in this study (N = 1311)

	N (%) or mean (SD)
Age (year), mean (SD)	32.3 (4.6)
Race/ethnicity, n (%)
White	962 (73)
Black	165 (13)
Hispanic	87 (7)
Other	97 (7)
Pre-pregnancy BMI (kg/m2), mean (SD)	24.9 (5.4)
GDM in a prior pregnancy, n (%)
No	640 (49)
Yes	23 (2)
Nulliparous	648 (49)
DM in participant’s mother, n (%)
No	1070 (92)
Yes	89 (8)
Missing/don’t knowa	152
Smoking habit, n (%)
Never	899 (68)
Before pregnancy	258 (20)
During pregnancy	154 (12)
Household income > $70,000/year, n (%)	797 (61)
College graduate and above, n (%)	912 (70)
Fish consumptionb (servings/week), mean (SD)	1.6 (1.4)

Note:

^aThe category “Missing/don’t know” was not included in the total population when calculating the percentages of “Yes” and “No”

^bDuring past 3 months of women’s initial pregnancy visit (median: 12 weeks gestation) estimated from self-administered Food Frequency Questionnaire

Abbreviations: BMI, body mass index; GDM, gestational diabetes mellitus; DM, diabetes mellitus; SD, standard deviation.

The 789 Project Viva participants excluded from this analysis had higher proportion of non-white women, lower household income and education level (Table S1). However, the distributions of the outcomes were not statistically different between the included and excluded population, see Supplementary Material Appendix A for detailed DAG of the relationship between the exposure, outcome and the selection criteria.

3.2. Bayesian Kernel Machine Regression Analyses

The BKMR model showed Ba had the highest posterior inclusion probability (PIP: 0.6), followed by Hg and Se (0.5), suggesting that these were the most relevant metals associated with glucose concentrations. We examined the univariate exposure-response association of each metal, fixing all other metals at their medians (Figure 2, Table 2). The association between Ba and Se and gestational glucose concentrations appeared to be an inverse U-shaped pattern. For example, the estimated glucose concentrations were on average 1.3 mg/dL (95% credible interval: −1.2, 1.3) higher comparing women with Ba concentrations at the 75^th percentile of the study population to those at the 25^th percentile, holding the covariates constant and other metals at their median. However, the association became inverse when the women’s Ba concentrations were above the 75^th percentile in the population (5.90 ng/g), with an estimated decrease of 1.1 mg/dL (95% credible interval: −0.9, 1.0) for women with Ba concentrations at the 90^th percentile compared to those at the 75^th percentile. We observed a linear inverse association between Hg and gestational glucose concentrations; the estimated glucose concentrations were 0.6 mg/dL lower comparing women with Hg concentrations at the 75^th percentile to those at the 25^th percentile (95% credible interval: −2.0, 0.7). The associations for other metals were weaker.

Figure 2.

Associations between first trimester erythrocyte concentrations of individual metals and gestational glucose concentrations at second trimester from 1-hr glucose load test for 1311 pregnant women in Project Viva estimated using Bayesian Kernel Machine Regression (BKMR) fixing other metals in the mixture at their medians.

Note: Models were adjusted for maternal age, self-identified race/ethnicity, pre-pregnancy BMI, GDM in prior pregnancy, smoking, maternal education, diabetes status of biological mother, and gestational week at blood collection for metals measurements. Black lines indicate the predictor-response function and the grey areas represent 95% credible intervals.

Table 2.

Differences in mid-gestational glucose concentrations (mg/dL) associated with interquartile (25^th to 75^th) changes of each first-trimester metal concentration, with all other metals fixed at their medians, estimated by Bayesian Kernel Machine Regression (BKMR). Results from 1311 women in Project Viva.

	Difference in glucose (mg/dL) (95% credible interval)
Arsenic	−0.2 (−1.9, 0.8)
Barium	1.3 (−1.2, 1.3)
Cadmium	−0.2 (−1.1, 0.5)
Cesium	−0.1 (−1.3, 0.6)
Copper	0.0 (−1.0, 0.5)
Magnesium	0.1 (−0.9, 0.5)
Manganese	−0.2 (−1.4, 0.6)
Lead	−0.5 (−1.6, 0.6)
Selenium	0.5 (−1.5, 1.0)
Zinc	0.2 (−1.1, 0.6)
Mercury	−0.6 (−2.0, 0.7)

Note:

Models were adjusted for maternal age, self-identified race/ethnicity, pre-pregnancy BMI, GDM in prior pregnancy, smoking, maternal education, diabetes status of biological mother, and gestational week at blood collection for metals measurements.

We explored potential pairwise interactions between metals by plotting bivariate exposure-response curves; specifically, the exposure-response associations of each metal with glucose concentrations was estimated holding a second metal at different percentiles (10^th, 50^th, 90^th) and the rest fixed at their medians (Supplementary Material Figure S2). The exposure-response curves of several metals, including As, Hg, Mg, Pb, Se were not parallel at low, median, and high Ba concentrations, suggesting potential interactions for these metals with Ba. For example, we observed a stronger positive association between As and glucose at higher concentration of Ba.

Validating results from BKMR

We tested for linearity of associations using multivariable GAM incorporating all metals and these interaction terms between Ba-As, Ba-Hg, Ba-Mg, Ba-Pb and Ba-Se. The result confirmed a similar inverse U-shaped association between Ba and glucose concentrations (p = 0.004), and an inverse linear association for Hg (Figure S4); the association between Se and glucose did not significantly deviate from linearity (the spline term with 1.26 degree of freedom had p=0.103). To estimate the effect of interactions, we used a multivariable linear regression model with selected metals (As, Ba, Mg, Pb, Se, Hg, where Ba was categorized into tertiles) and their pairwise interactions with Ba suggested by BKMR. The results were consistent in directions with the interaction suggested in BKMR models (Figure S2). For example, we observed stronger associations between As and gestational glucose concentrations with increased concentrations of Ba. Specifically, the effect per 1 SD increase of As (on the log-transformed scale) on gestational glucose concentrations was negative at the 1st tertile of Ba [−0.82 (95% CI: −3.98, 2.35)] but the effect became more positive with increasing concentration of Ba: 1.53 (95% CI: −2.76, 5.81) mg/dL higher comparing the 2^nd to the 1^st tertile concentrations of Ba, and 3.01 (95% CI: −1.35, 7.37) mg/dL higher comparing 3^rd to the 1^st tertile of Ba. Among all the interactions included in the model, we observed statistically significant findings only between Pb and Ba; in particular, the effect of per SD increase of Pb on gestational glucose concentrations was 3.95 (95% CI: 0.28, 7.61, p=0.04) mg/dL higher comparing the 3^rd to the 1^st tertile of Ba (Table 3).

Table 3.

Differences in mid-gestational glucose concentrations (mg/dL) associated with 1 standard deviation increase of first trimester metals concentrations on log-transformed scale, estimated by multivariable linear regression models with interaction between Ba and other metals (As, Hg, Mg, Pb, and Se). Results from 1311 women in Project Viva.

	Estimate (95% confidence interval)
Ba tertile 2 vs 1	3.32 (−0.16, 6.80)
Ba tertile 3 vs 1	1.86 (−1.63, 5.36)
As in Ba tertile 1	−0.82 (−3.98, 2.35)
As in Ba tertile 2 vs 1	1.53 (−2.76, 5.81)
As in Ba tertile 3 vs 1	3.01 (−1.35, 7.37)
Hg in Ba tertile 1	−1.88 (−4.68, 0.92)
Hg in Ba tertile 2 vs 1	0.67 (−3.37, 4.71)
Hg in Ba tertile 3 vs 1	0.01 (−4.11, 4.14)
Mg in Ba tertile 1	1.19 (−1.54, 3.91)
Mg in Ba tertile 2 vs 1	−0.33 (−4.39, 3.72)
Mg in Ba tertile 3 vs 1	−1.82 (−5.89, 2.26)
Pb in Ba tertile 1	−2.64 (−5.30, 0.00)
Pb in Ba tertile 2 vs 1	0.24 (−3.39, 3.87)
Pb in Ba tertile 3 vs 1	3.95 (0.28, 7.61)
Se in Ba tertile 1	2.50 (−0.74, 5.74)
Se in Ba tertile 2 vs 1	−2.84 (−7.52, 1.85)
Se in Ba tertile 3 vs 1	−2.40 (−6.80, 2.01)

Note:

The multivariable linear regression model had As, Ba (tertile), Hg, Mg, Pb, Se, and interactions As*Ba, Mg*Ba, Se*Ba, Hg*Ba, and Pb*Ba as independent variables and controls for maternal age, self-identified race/ethnicity, pre-pregnancy BMI, GDM in prior pregnancy, smoking, maternal education, diabetes status of biological mother, and gestational week at blood collection for metals measurements. Ba concentrations were categorized as tertiles using log-transformed and standardized machine read value; Ba concentration (ng/g) for tertile 1 [<LOD, 2.30], tertile 2 (2.31, 4.77), tertile 3 (4.78, 59.20). “Ba tertile 1” showed the absolute effect estimates for each metal in the first tertile of Ba, while the estimates “Ba tertile 2 vs1” and “Ba tertile 3 vs 1” were the relative difference in effect estimates compared to tertile 1.

Abbreviations: Se, selenium; Hg, mercury; As, arsenic; Ba, barium; Mg, magnesium; Pb, lead.

Joint associations between metal mixtures and gestational glucose concentrations

The overall effect of the metal mixture estimated by BKMR was shown in the Supplementary Material Figure S3; gestational glucose concentrations increased 0.29 (95% CI: −0.86, 141) mg/dL as concentrations of the metal mixture increased from the 50^th to the 75^th percent quantile, and decreased 0.70 (95% CI: −2.87, 1.46) mg/dL as metal mixture decrease from 50^th to the 25^th percent quantile. We also estimated the overall joint association using a quantile g-computation model. As informed by the result from BKMR, we used a non-linear term for Ba and included an interaction between Ba and Pb (see Supplementary Material); Figure S5 showed the joint associations of the first trimester erythrocyte concentrations of all metals (Figure S5a), essential metal (Figure S5b), and non-essential metals (Figure S5c) on second trimester gestational glucose concentrations. The joint effects of all metals, essential-metals, and non-essential metals on gestational glucose concentrations were 0.54 (95% CI: −6.91, 7.98), 1.05 (95% CI: −9.83, 3.07), −0.35 (95% CI: −7.81, 7.12) mg/dL per quartile increase in metal concentrations, respectively. All confidence intervals were wide and included the null. The difference in glucose concentrations comparing the 4^th to the 1st quartile of the mixture with all metals after adjusting for covariates was 2.17 (SE: 5.51) mg/dL, and 3.14 (SE: 3.12) mg/dL and −0.11 (SE: 4.92) mg/dL for essential and non-essential metal mixtures, respectively (Table S3). We did detect evidence of interaction between Pb and Ba using this method as well, specifically, the effect for each SD increase of Pb on glucose concentrations was 4.65 (95% CI: 1.13, 8.18) mg/dL higher comparing the 4^th to the first quartile of Ba (Table S3).

3.3. Sensitivity analyses

All results remained robust when we included 14 pregnant women with extreme metals concentrations, and the effect estimates did not change substantially when we further adjusted for hemoglobin concentrations, fish consumption or n-3 fatty acid consumption, gestational age at the glucose tolerance test and environmental tobacco smoke exposure. We did not observe significant effect modification by fish consumption. We observed similar findings when using the combined IGT and GDM criteria as the outcome (binary) for this, the effect estimates appeared to be in the same direction as results from analyses using glucose as a continuous outcome variable; however, no associations reached significance (results not shown). Sensitivity analyses with varying BKMR model priors showed findings were not sensitive to difference smoothing parameters.

4. DISCUSSION

We assessed the individual and joint effects of both essential and non-essential metals exposures using data from the Project Viva pregnancy cohort. Among individual metals, we observed the strongest associations from Ba and Hg. First trimester erythrocyte Ba concentration exhibited a modest inverted U-shaped association with gestational glucose concentrations at late 2^nd trimester, where Ba was associated with higher gestational glucose concentrations in concentrations below the 75^th percentile (5.90 ng/g) and was associated with lower glucose concentrations in concentrations above 75^th percentile. The association between Hg and gestational glucose was inverse and the effect size remained similar even with additional adjustment for the beneficial effect of fish consumption, however, the 95% confidence interval crossed the null. Interestingly, we saw suggestive evidence of pairwise interactions between Pb and Ba which had not been reported previously. However, the joint effects of mixtures of all metals, essential-metals and non-essential metals on gestational glucose concentrations were null.

Ba is a non-essential metal that naturally occurs in most foods, surface waters and in public drinking water supplies. While the general population is usually exposed to Ba at low levels, workers in barium mining industries and individuals who reside near such industries may be exposed to relatively high levels through the inhalation of dust containing Ba compounds (ATSDR 2008). An animal study reported decreased blood glucose concentrations in barium-exposed rats (Tarasenko et al. 1977). Human studies that evaluated the association between Ba and glycemic status are limited, most were done among non-U.S. populations and/or workers occupationally exposed to higher levels of toxic metals (Cancarini et al. 2017; Feng et al. 2015; Li et al. 2017; Liu et al. 2016), and yielded different findings. Cancarini et al. found higher Ba concentrations in the tear fluid of 47 adults with diabetes compared to 50 without diabetes in Italy, but observed no significant difference in their serum Ba concentrations (Cancarini et al. 2017). Li et al. performed a case-control study in China using 122 newly diagnosed type 2 diabetes patients and 429 matched controls and found positive associations between plasma Ba with diabetes risk after controlling for age, gender, BMI, family history, smoking and drinking status. The adjusted odd ratio (OR) of diabetes was 6.2 (95% CI: 3.4–11.1) comparing the highest to the lowest tertile of plasma Ba concentrations (median: 5.1 μg/L and quartile range: 4.6 μg/L) (Li et al. 2017). Using urinary Ba as the biomarker, Feng et al. found null cross-sectional associations with altered glucose concentrations and diabetes risk among 2242 Chinese adults in a community-based cohort study (Feng et al. 2015). On the other hand, Liu et al observed a positive cross-sectional association between urinary Ba and hyperglycemia among 1,493 coke oven workers, when comparing the third (>7.14 μg/L) and first (≤4.16 μg/L) tertiles of urinary Ba [adjusted OR: 1.54 (1.22–2.23)], adjusting for gender, age, BMI, smoking status, drinking status, physical activity, education levels, urinary creatinine, hypertension, hyperlipidemia, and urinary polycyclic aromatic hydrocarbons levels (Liu et al. 2016). One study based on the 1999–2010 National Health and Nutrition Examination Survey, a representative sample of the U.S. civilian population, found that higher levels of urinary Ba were associated with greater HOMA-insulin resistance, but not with risk of diabetes (Menke et al. 2016). Most of these previous studies did not account for confounding by diet which could lead to the inconsistent findings. For our study, a previous diet-wide association study of erythrocyte metals among the same cohort of pregnant women did not identify any specific food item with strong association with Ba (Lin 2021), and additional adjustment for dietary variables in the sensitivity analysis also did not change the findings, therefore, the associations we observed between Ba and glucose concentrations were unlikely due to confounding by diet. To our knowledge, we are the first to report the potential interaction between blood Ba and Pb concentrations on gestational glucose concentrations. Literature on blood Pb and gestational glucose concentrations was limited; the French EDEN study (N=623 pregnant women) found borderline significance between second-trimester blood Pb and GDM [OR 1.65 (95% CI 0.82–3.34)] (Soomro et al. 2019), however the Canadian birth cohort, Maternal–Infant Research on Environmental Chemicals (MIREC), did not find significant associations between first-trimester blood Pb with GDM and IGT (Shapiro et al. 2015). Possible mechanisms of Ba in relation to glucose concentrations during pregnancy and its interaction with Pb remain unclear and more investigation is warranted.

On the other hand, several studies—both basic and epidemiological studies—have evaluated the association between Hg and glucose concentrations. Animal and in vitro studies show that exposure to Hg can cause pancreatic beta cell apoptosis and interfere with lipid peroxidation, possibly accelerating progression to diabetes (Chen et al. 2006; Moreira et al. 2012). However, a review of epidemiologic studies examining the association between Hg exposure and glycemic status outside of pregnancy suggested that findings were inconsistent (Roy et al. 2017). As far as we are aware, only three studies have evaluated Hg exposure in association with glycemic status during pregnancy; specifically, a retrospective nested case-control study in China found significantly higher Hg concentrations in newborns’ meconium sample among GDM mothers compared to non-GDM mothers (Peng et al. 2015). A study of pregnant women in Faroe Islands showed an inverse association between maternal hair Hg concentrations (collected at parturition) and GDM [OR: 0.79 (95% CI: 0.62, 0.99) per doubling of exposure], but null associations between cord blood Hg and GDM [OR: 0.87 (95% CI: 0.66, 1.15) per doubling of exposure] (Valvi et al. 2017). The MIREC study found null associations between whole blood Hg concentrations (measured in first trimester) and GDM (Shapiro et al. 2015). The inconsistencies in results could arise from several reasons. First, the findings can be impacted by the biosamples used to estimate Hg exposure applied in different studies. In our study, we used total Hg measured in erythrocytes collected during early pregnancy. Erythrocyte Hg is a good indicator of total Hg exposure due to the uniform distribution of Hg in tissues and that a high proportion (over 90%) of Hg stably resides in erythrocytes (National Research Council 2000; Nordberg 1976). Blood Hg reflects both inorganic and organic forms of Hg. However, it is unclear whether the distribution of Hg in the blood is altered by early pregnancy blood glucose concentrations, as women with higher blood glucose concentrations have enhanced glycation of hemoglobin (Sacks 2013). Hair and nail Hg concentrations reflect organic Hg exposure, such as that found in fish and rice. The toxicity profile may therefore differ from blood which also contains inorganic Hg. Second, the findings can be affected by unadjusted or not fully adjusted confounding factors, as well as the measurement for confounding factors. Healthy dietary patterns that includes fish consumption – one of the major sources of methyl Hg exposure – may lower glucose concentrations and reduce the risk of developing GDM (Santangelo et al. 2016; Schoenaker et al. 2016). However, Project Viva women with higher self-reported intake of n-3 fatty acid during pregnancy had higher odds of GDM (OR 1.11 [95% CI 1.02, 1.22] per each 300 mg/day) after adjusting for maternal age, pre-pregnancy body mass index, race/ethnicity, previous gestational diabetes, history of diabetes in participant’s mother, and smoking during pregnancy (Radesky et al. 2008). In our study, the association between erythrocyte Hg and glucose concentrations remained similar after further adjusting for fish consumption or n-3 fatty acid consumption estimated from FFQ. It was likely that the fish consumption estimates could be affected by reporting bias, and/or it may not fully capture the beneficial factors relevant to glucose metabolism (e.g. specific types of fish). Alternatively, it was also possible that inorganic Hg in erythrocyte was the major player driving the inverse association, thus, adjusting Hg for fish intake had no impact on this. Third, the findings may reflect associations between Hg exposure and glycemic status only in certain exposure ranges. Hg exposure is expected to be low in this study population, given the generally low habitual fish intake and low environmental contamination (Oken et al. 2016). Therefore, it is possible that the inverse Hg association seen in the present study is only applicable in lower Hg exposure settings.

Literature on blood Pb and gestational glucose concentrations was limited; the French EDEN study (N=623 pregnant women) found borderline significance between second-trimester blood Pb and GDM [OR 1.65 (95% CI 0.82–3.34)] (Soomro et al. 2019), however the MIREC study in Canada did not find significant associations between first-trimester blood Pb with GDM and IGT.

Our study has some limitations. First, a single measurement of metals may not represent the average concentrations of the metals during early pregnancy. For example, since Ba in the blood is readily eliminated within a few days (ATSDR 2008; Choudhury 2001), the erythrocyte Ba concentrations may vary largely based on a woman’s day-to-day intake. However, this is less of a concern for metals with longer half-lives in the blood such as Pb, Hg and Cd (Jarup et al. 1983; Kioumourtzoglou et al. 2016; Rabinowitz et al. 1976). Second, the glucose concentrations used as the main outcome of this study were from the non-fasting 50-gram GCT used as a part of GDM screening, which can be affected by the timing and compositions of the last meal. As a secondary analysis, we also evaluated combined IGT and GDM as the outcome; however, we were limited by the sample size to have power to detect any statistically significant associations. Third, although we adjusted for major demographic, lifestyle and reproductive health information, our findings may be subject to residual confounding, such as exposure to other environmental toxicants or endocrine disruptors. It is also possible that some non-environmental factors, such as genetic and epigenetic variations, may modify the associations between metals and gestational glucose concentrations, however, we did not have data for such assessment. Fourth, our study population included women living in eastern Massachusetts, who all had access to prenatal care and were relatively well educated. Therefore, results from this study may not be generalizable to other populations, especially those with lower socioeconomic status, at higher risk of developing pregnancy complications, and with differing dietary and metals exposure source patterns. We measured metal concentrations in the erythrocytes rather than in whole blood (which we did not retain), which could make our results harder to compare and translate. However, as discussed in our other publication which included a more in-depth review for individual metals (Lin 2021), erythrocytes could be a better biomarker for some metals such as Cu, Cd, Mg, Mn (Vormann 2003). Last but not least, we measured only total As but not organic As. Total As reflects both the more toxic inorganic arsenic and its metabolites as well as nontoxic arsenicals; this might explain why we did not observe the association between As exposure and increased risk for GDM/glucose intolerance reported by previous studies. These limitations warrant careful interpretation and further investigation of the findings, particularly on any biological mechanism that may lead to observed associations and interactions between metals.

Despite the limitations, our study has several strengths. We employed BKMR, a nonparametric novel method to evaluate the associations between metal mixtures and gestational glucose concentrations. This statistical approach allowed us to overcome many limitations that traditional approaches have, such as model misspecification, violation of linearity, and multiple comparison. In addition to providing estimations of the associations between metals and gestational glucose concentrations, BKMR also served as an exploratory tool for constructing parametric models in GAMs and linear regression models. In this study, where a relatively large number of metals (n = 11) were assessed simultaneously, using parametric approaches directly can be challenging as they require estimating a much larger number of parameters in order to allow for non-linear exposure-response relationships and interactions among the exposures. Other strengths of the study include its relatively large sample size, the fact that we examined the associations between metals and glucose concentrations prospectively, our ability to investigate these associations adjusting for major confounders, and the use of g-computation to assess overall metal mixture as well as essential and non-essential metal mixtures.

5. CONCLUSIONS

Among a prospective cohort of Massachusetts pregnant women, we found evidence for an inverse U-shaped nonlinear association between early pregnancy erythrocyte Ba and gestational glucose concentrations, and an inverse association for erythrocyte Hg. We also saw suggestive evidence of interactions between Ba and Pb. While these exploratory findings suggest that essential and non-essential metals exposures may individually and jointly play a role in glucose metabolism during pregnancy, caution should be taken when interpreting the results as this is one of the first studies to examine these complex associations using erythrocyte metals. If replicated and well quantified in different biomarkers and among different pregnancy populations, these findings may provide implications for the prevention of impaired glucose tolerance during pregnancy and gestational diabetes.

Abbreviations

As

arsenic

Ba

barium

BKMR

Bayesian Kernel Machine Regression

BMI

body mass index

Cd

cadmium

CHEAR

Children’s Health Exposure Analysis Resource

Cs

cesium

Cu

cupper

CV

coefficient of variation

DAG

directed acyclic graphs

FFQ

Food Frequency Questionnaire

GAM

generalized additive models

GCT

glucose challenge test

GCT

glucose challenge tests

GDM

gestational diabetes mellitus

Hg

mercury

ICP-MS

Inductively Coupled Plasma Mass Spectrometry

ICP-QQQ

triple quadrupole Inductively Coupled Plasma Mass Spectrometry

IGT

impaired glucose tolerance

IQR

interquartile range

LOD

limit of detection

MCMC

Markov Chain Monte Carlo

Mg

magnesium

MIREC

Maternal–Infant Research on Environmental Chemicals

ML

maximum likelihood

Mn

manganese

OGTT

oral glucose tolerance test

OR

odd ratio

Pb

lead

PIP

posterior inclusion probability

QC

quality control

SD

standard deviation

Se

selenium

Zn

zinc