Association of Low-Moderate Arsenic Exposure and Arsenic Metabolism with Incident Diabetes and Insulin Resistance in the Strong Heart Family Study

The SHFS is a prospective family based cohort study designed to identify genetic and environmental factors for cardiovascular disease and their risk factors in American Indians from 13 communities residing in Arizona, Oklahoma, North Dakota, and South Dakota. Details about design and methodology for SHFS have been published (North et al. 2003). In the SHFS, 3,838 men and women from 96 families have baseline data that were collected in 1998–1999 and 2001–2004, and follow-up data in 2001–2004 (for some participants recruited between 1998–1999) and 2006–2009 (Figure S1). The protocol was approved by the institutional review boards of the Indian Health Service and the participating Indian tribes. All participants gave informed consent.

Participants free of diabetes at baseline and with available urine arsenic measurements were selected for this study ($n=\mathrm{2,453}$). Due to tribal request, data from one of the original tribes was not used ($n=504$). We further excluded participants missing diabetes status at follow-up ($n=38$), urine creatinine measurements ($n=1$), baseline values of homeostasis model assessment for insulin resistance (HOMA2-IR) ($n=25$), and other relevant covariates such as baseline body mass index (BMI), waist circumference, estimated glomerular filtration rate (GFR), smoking status, and AS3MT genotype ($n=47$). As a result, 1,838 participants were included in the present study. Included participants were similar to those who were excluded because of missing data (not shown).

We determined two study outcomes at the follow-up visit: a) incident type 2 diabetes status (yes/no); and b) HOMA2-IR (continuous). Incident type 2 diabetes was defined as fasting plasma glucose $\ge 126\mathrm{mg}/\text{dL}$, self-reported physician diagnosis or self-reported use of insulin or oral diabetes treatment. Similar to other studies (Chow et al. 2013; Juraschek et al. 2013), we estimated the date of diagnosis under the assumption that glucose levels increased at a linear rate between study visits for participants diagnosed based on glucose levels. Impaired fasting glucose (IFG) and normal fasting glucose (NFG) were defined as fasting glucose concentrations between $100\text{and}126\mathrm{mg}/\text{dL}$ and $<100\mathrm{mg}/\text{dL}$, respectively. Baseline and follow-up HOMA2-IR values were calculated with the computed solved model for HOMA2-IR (Levy et al. 1998) using fasting glucose and insulin values. HOMA2-IR at follow-up was estimated only among people free of incident diabetes because HOMA-IR correlates well with insulin sensitivity in the SHS nondiabetic population (Resnick et al. 2002).

Spot urine samples collected the morning of the baseline visit were stored at $-70º\mathrm{C}$. Total urine arsenic was measured by inductively coupled plasma mass spectrometry (ICPMS) and arsenic species (iAs, MMA, DMA, and arsenobetaine) were measured by high-performance liquid chromatography-ICPMS (HPLC-ICPMS) at the Trace Element Laboratory of Graz University, Austria. The limit of detection (LOD) for all arsenic species was $0.1\mathrm{\mu }\mathrm{g}\mathrm{As}/\mathrm{L}$. Among the 1,838 participants, 197 (10.7%), 57 (3.1%), and 111 (6.0%) participants had urine iAs, MMA, and arsenobetaine concentrations below the LOD, respectively. No participants had DMA concentrations below the LOD. We imputed the concentrations of iAs, MMA, and arsenobetaine in 221 participants with only one of the species undetected using the equation $totalarsenic=iAs+MMA+DMA+arsenobetaine$. For 64 individuals with two arsenic species undetected, we estimated the arsenic species concentrations as the LOD divided by the square root of 2. Those 64 participants were excluded for arsenic metabolism analyses because it is difficult to estimate arsenic metabolism if arsenic exposure itself is very low and imputation as the LOD divided by the square root of more than one of the species would lead to wrong estimates of the arsenic methylation patterns for those individuals. Therefore, only 1,774 participants were included in arsenic metabolism analyses.

We used the sum of iAs, MMA, and DMA ($\mathrm{\Sigma }\mathrm{As}$) as a measure of inorganic arsenic exposure and the relative proportions of iAs, MMA, and DMA over the sum of the three (expressed as iAs%, MMA%, and DMA%) as biomarkers of arsenic metabolism.

Information on age, sex, study region (Arizona, Oklahoma, North, Dakota, and South Dakota), educational level, and smoking status was provided in a personal interview. Height, weight, and waist circumference were collected by physical examination using a standardized protocol. BMI was calculated dividing the weight in kilograms by the square of height in meters. Estimated GFR was obtained using the chronic kidney disease epidemiology equation. Estimates of macro- and micronutrients, including data on folate and other B vitamins (B1, B2, B6, and B12), were measured through a Block 119-item food frequency questionnaire (FFQ). Detailed information about the FFQ has been previously published (Fretts et al. 2012; Spratlen et al. 2017). Information on vitamins B1 and B12, however, was not used for this study because they were not available for most of participants. Urine creatinine levels were measured by an automated alkaline picrate reagent method. We studied effect modification of the associations by rs12768205 in AS3MT, the single nucleotide polymorphism (SNP) with the strongest association in a MetaboChip association analysis with iAs%, MMA%, and DMA% in the SHFS (Balakrishnan et al. 2016). SNP genotyping details have been previously published (Balakrishnan et al. 2016).

The distribution of $\mathrm{\Sigma }\mathrm{As}$ was divided by urine creatinine to account for urine dilution and log-transformed for the analyses. iAs%, MMA%, and DMA% were analyzed in the original scale. Because iAs%, MMA%, and DMA% sum to 100%, we used a diagram of three axes (triplot) to describe the compositional means of baseline iAs%, MMA%, and DMA% in participants with normal fasting glucose, impaired fasting glucose, and type 2 diabetes at follow-up.

We estimated the hazard ratios of diabetes incidence and the geometric mean ratios (GMR) of HOMA2-IR by urine arsenic exposure levels and arsenic metabolism patterns. Hazard ratios (HR) were evaluated using Cox proportional hazard models with age as time scale and age at baseline treated as staggered entries. For HOMA2-IR analyses, we conducted multilevel models (MLM) in which both HOMA2-IR values at baseline and at follow-up were treated as the outcome and the linear predictor included the interaction of arsenic and time since baseline (in years). Specifically, the time variable included two values for each participant: $\text{time}=0$ and $\text{time}=\text{follow}-\text{up}$ duration (in years). This analytical strategy allows estimating the GMR of HOMA2-IR by arsenic levels at baseline ($\text{time}=0$), the corresponding GMR at follow-up (for instance considering 5 y of follow-up, $\text{time}=5$), and the mean change on the GMR per each year of follow-up (increasing time one unit). In order to account for the lack of independence among family members, we used mixed effects Cox proportional hazard models for HR and linear regression models with generalized estimating equations for GMR. $\mathrm{\Sigma }\mathrm{As}$ was introduced in the models as continuous (comparing participants in the 75th vs. the 25th percentile) and as tertiles (comparing participants in the two highest tertiles vs. the lowest one). The role of arsenic metabolism was evaluated in two ways. First, we entered one arsenic species percentage in the models (conventional approach) and the associations were estimated for a 5% increase in each species in separate models. To address the difficult interpretation of the traditional approach given that a percentage increase of one arsenic species yields to a percent decrease in one or two of the other arsenic species, we entered two arsenic species percentages in the same model (leave-one-out approach), and reported the associations per a 5% increase of one of the included percentages, meaning that the levels of the second arsenic species in the model are fixed and the not-included arsenic species decreases a 5%.

Study region was introduced in all models as a staggered variable. Model 1 was adjusted for sociodemographic variables: sex, age at baseline visit (continuous), and education ($<12\mathrm{y}$, $\ge 12\mathrm{y}$). Model 2 was further adjusted for traditional cardiovascular risk factors: BMI, waist circumference, smoking status (never, former, current), estimated GFR, and fasting glucose levels at baseline (normal, impaired). Model 3 was further adjusted for relevant determinants of arsenic metabolism that we wanted to evaluate in exploratory analyses: estimated dietary vitamin B2, B6, and folate (continuous) and allelic dosage of rs12768205. All arsenic metabolism models were also adjusted for $\mathrm{\Sigma }\mathrm{As}$ levels. The use of 12 y as a benchmark for educational level categories has been used in other studies conducted in American Indian populations (Dickerson et al. 2012; Moon et al. 2013). To allow for flexible associations, we modeled $\mathrm{\Sigma }\mathrm{As}$ and the arsenic species percentages using restricted cubic splines. We also explored whether the associations of arsenic exposure and metabolism with incident diabetes and HOMA2-IR are modified by participant subgroups by including in the models the interaction term between the arsenic variable and the corresponding subgroup variable.

Several sensitivity analyses were conducted. We reanalyzed models for the association between diabetes incidence and arsenic exposure stratifying by fasting glucose status (normal vs. impaired) at baseline and in models further adjusted for baseline HOMA2-IR levels (shown in main results). Models further adjusted for urine arsenobetaine (log-transformed), intake of certain food groups (meat, rice and cereals intake), or cigarette packs-per-year showed consistent results (not shown). We also checked the robustness of the findings using other ways to account for urine dilution. In particular, analyses with treating urine arsenic in $\mathrm{\mu }\mathrm{g}/\mathrm{L}$ and adjusting for specific gravity or urine creatinine in statistical models showed almost identical results (not shown). Finally, analyses excluding participants with undetectable iAs or MMA concentrations resulted in nondifferent results (not shown). All analyses were performed with R software (version 3.3.1; R Development Core Team). The statistical significance level was set at $\mathrm{\alpha }=0.05$.

Median [interquartile range (IQR)] age of study participants was 36 (24–47) y and 60% (1,122) were women (see Table S1). Over 10,327 person-years of follow-up, 255 (13.7%) participants developed diabetes (incidence of 24.7 per 1,000 person-years), with no difference by sex. Compared with nondiabetes participants, individuals who developed diabetes were older and more likely to be obese and to have impaired fasting glucose and higher HOMA2-IR at baseline. Diabetes participants also showed a higher dietary intake estimate of folate than participants without diabetes. In particular, overall median levels of estimated intake of vitamin B2, B6, and folate were 1.6, 1.6, and $336\mathrm{mg}/\mathrm{d}$, respectively. Median (IQR) urine $\mathrm{\Sigma }\mathrm{As}$, iAs%, MMA%, and DMA% was $4.4\mathrm{\mu }\mathrm{g}/\mathrm{g}$ creatinine ($2.9–7.2$), 9.5% ($6.3–13.8$), 14.4% ($11.0–18.1$), and 75.6% ($68.5–81.7$), respectively. The median (IQR) of $\mathrm{\Sigma }\mathrm{As}$ before urine creatinine correction was $5.9\left(3.6–9.9\right)\mathrm{\mu }\mathrm{g}/\mathrm{L}$. Participants with incident diabetes had higher baseline urine $\mathrm{\Sigma }\mathrm{As}$ levels and a metabolic profile characterized by lower MMA% and higher DMA% compared with those without diabetes over the follow-up. The compositional means of iAs%, MMA%, and DMA% showed that individuals with incident diabetes and participants with impaired fasting glucose at follow-up had lower MMA% and higher DMA% levels compared with participants with normal fasting glucose at follow-up (Figure 1). Data on the median (IQR) of $\mathrm{\Sigma }\mathrm{As}$, iAs%, MMA%, and DMA% on participants subgroups are described in Figures S2 and S3.

The fully adjusted HR [95% confidence interval [CI)] of incident diabetes comparing participants in the 75th versus the 25th percentile of urine $\mathrm{\Sigma }\mathrm{As}$ was 1.16 (95% CI: 0.94, 1.42) in the complete sample, 1.57 (95% CI: 1.18, 2.08) among participants with normal fasting glucose at baseline, and 0.92 (95% CI: 0.67, 1.21) among participants with impaired fasting glucose at baseline (Table 1, Model 3). Modeling $\mathrm{\Sigma }\mathrm{As}$ as tertiles and restricted cubic splines showed positive and linear associations with incident diabetes that were suggestive but nonsignificant in the complete sample, and significant among normal fasting glucose participants at baseline (Table 1 and Figure 2). No associations were found between arsenic metabolism and diabetes incidence in fully adjusted models (see Table S2 and Figure S4). In interaction analysis, the association between arsenic exposure and incident diabetes was modified by fasting glucose levels at baseline ($p-\text{interaction}=0.003$) but not by other participant characteristics (see Figure S5). We found no effect modification of the association between arsenic metabolism and incident diabetes by any participant characteristics (see Figure S6).

We found that baseline $\mathrm{\Sigma }\mathrm{As}$ was positively associated with baseline HOMA2-IR, but negatively associated with HOMA2-IR at follow-up (Table 2 and Figure 3). In particular, in fully adjusted models comparing an IQR increase in $\mathrm{\Sigma }\mathrm{As}$, the GMR (95% CI) of HOMA2-IR was 1.04 (95% CI: 1.01, 1.08) at baseline and 0.95 (95% CI: 0.92, 0.98) after 5 y of follow-up. For arsenic metabolism, higher MMA% was associated with lower HOMA2-IR both baseline and follow-up. In particular, the fully adjusted GMR (95% CI) of HOMA2-IR after 5 y of follow-up per 5% increase in arsenic metabolism biomarkers when entered individually in the model (conventional approach) was 0.97 (95% CI: 0.95, 0.99) for iAs%, 0.93 (95% CI: 0.91, 0.95) for MMA%, and 1.04 (95% CI: 1.02, 1.05) for DMA% (Table 2). Using the leave-one-out approach, we confirmed that higher MMA% was associated with decreased HOMA2-IR levels both at baseline and follow-up. The GMR (95% CI) of HOMA2-IR after 5 y of follow-up per 5% increase in MMA% was 0.93 (95% CI: 0.90, 0.96) when iAs% decreased a 5%, and 0.93 (95% CI: 0.91, 0.95) when DMA% decreased a 5%. Models with restricted cubic splines showed the dose–response of these associations and confirmed these findings (Figure 3).

The inverse association between MMA% and HOMA2-IR at follow-up was stronger in men ($p-\text{interaction}=0.01$; Figure 4) and in participants with higher intake of vitamin B2 ($p-\text{interaction}=0.01$), vitamin B6 ($p-\text{interaction}=0.002$), and folate ($p-\text{interaction}=0.03$). Moreover, the association between iAs% and DMA% with HOMA2-IR was modified by rs12768205 ($p-\text{interaction}=0.03$ in both cases).