Probabilistic Modeling of Dietary Arsenic Exposure and Dose and Evaluation with 2003–2004 NHANES Data

Background Dietary exposure from food to toxic inorganic arsenic (iAs) in the general U.S. population has not been well studied. Objectives The goal of this research was to quantify dietary As exposure and analyze the major contributors to total As (tAs) and iAs. Another objective was to compare model predictions with observed data. Methods Probabilistic exposure modeling for dietary As was conducted with the Stochastic Human Exposure and Dose Simulation–Dietary (SHEDS-Dietary) model, based on data from the National Health and Nutrition Examination Survey. The dose modeling was conducted by combining the SHEDS-Dietary model with the MENTOR-3P (Modeling ENvironment for TOtal Risk with Physiologically Based Pharmacokinetic Modeling for Populations) system. Model evaluation was conducted via comparing exposure and dose-modeling predictions against duplicate diet data and biomarker measurements, respectively, for the same individuals. Results The mean modeled tAs exposure from food is 0.38 μg/kg/day, which is approximately 14 times higher than the mean As exposures from the drinking water. The mean iAs exposure from food is 0.05 μg/kg/day (1.96 μg/day), which is approximately two times higher than the mean iAs exposures from the drinking water. The modeled exposure and dose estimates matched well with the duplicate diet data and measured As biomarkers. The major food contributors to iAs exposure were the following: vegetables (24%); fruit juices and fruits (18%); rice (17%); beer and wine (12%); and flour, corn, and wheat (11%). Approximately 10% of tAs exposure from foods is the toxic iAs form. Conclusions The general U.S. population may be exposed to tAs and iAs more from eating some foods than from drinking water. In addition, this model evaluation effort provides more confidence in the exposure assessment tools used.

Human exposure to arsenic (As) can occur via different routes. A wellknown early medical report about As exposure and adverse health effects discussed cancer associated with der mal exposure to Ascontaining medication used for treating some forms of skin diseases (Hutchinson 1887). Later studies on occupa tional populations exposed to As compounds in industrial environments demonstrated that respiratory inhalation is a primary route of occupational As exposure, but ingestion and dermal exposure can be significant in specific situations (Occupational Safety and Health Administration 2005;World Health Organization 2004).
Compared with the simpler As chemistry and easily identified As exposure in medical and occupational fields, As chemistry and exposure routes for the general population are much more complex. General population As exposure varies according to local geochem istry, environmental pollution, living condi tions, lifestyles, and activity patterns of the exposed populations. Better characterization of environmental As levels and human activ ity patterns is critical for accurately assess ing the human exposure to As in the general population and the related health risks.
Many efforts in studying As exposure of and regulating As intake by the general popu lation have been focused on the ingestion of Ascontaminated water (Abernathy et al. 1999(Abernathy et al. , 2003Anetor et al. 2007;Chen et al. 1988aChen et al. , 1988bChiou et al. 2001;National Research Council 2001;Tchounwou et al. 2003). This drinking water-focused As regu lation also reflects a common understanding that inorganic As (iAs) is more harmful than organic As (oAs) (Tchounwou et al. 2003). A recent publication concluded that typical and highend background exposures to iAs in the U.S. population do not present elevated risks of carcinogenicity (Boyce et al. 2008). However, other reports show significant dietary intake of iAs via food and even show food as a greater source of iAs intake than is drinking water (Meacher et al. 2002;Schoof et al. 1999aSchoof et al. , 1999b. Yost et al. (2004) esti mated dietary intake of iAs in U.S. children as 3.2 µg/day on average. A recent study shows that, in three U.S. counties, the food intake pathway is the dominant contributor to total As (tAs) exposure and dose (Georgopoulos et al. 2008).
In this study we extend findings from the previous studies by a) assessing the dietary tAs and iAs exposure using the peerreviewed U.S. Environmental Protection Agency (EPA) Stochastic Human Exposure and Dose Simulation (SHEDS) model (SHEDS 2007), b) using more recent and larger databases rep resentative of the U.S. population for food consumption and As concentrations in food and drinking water, and c) conducting model evaluation using duplicate diet and biomarker data. We used a populationbased dietary exposure model, one module of the SHEDS model (SHEDS 2007;Xue et al. 2006;Zartarian et al. 2006), to estimate the expo sure of As (tAs and iAs) from both food and drinking water. We linked the total predicted exposure with the Modeling ENvironment for TOtal Risk with Physiologically Based Pharmacokinetic Modeling for Populations (MENTOR3P) system (Georgopoulos and Lioy 2006) to estimate the speciated As in urine. We compared the model results with biomarkers of tAs and As species measured in the National Health and Nutrition Examination Survey (NHANES 2003. Using large data sets of food consumption from NHANES, As concen trations in drinking water and various foods from the U.S. Food and Drug Administration (FDA) and Natural Resources Defense Council databases, and urinary biomark ers from NHANES (same individuals as for food consumption data), we demonstrate that dietary exposure can be a significant route for human exposure to both tAs and iAs. volume 118 | number 3 | March 2010 • Environmental Health Perspectives

Materials and Methods
Food consumption data. We used NHANES (2003 data for model inputs regarding the amount of food and water consumed by individuals. This database contains 16,934 persondays of realtime dietary consumption data-that is, amounts of food and drinking water recorded instantly by individuals for each separate eating occasion. The average number of eating occasions is approximately 4.8 times per person per day. The U.S. EPA's Food Consumption Intake Database (FCID) containing recipe files with 553 food com modities was applied where needed to break down NHANES food reported into raw agri cultural commodities (RACs).
tAs and iAs concentrations in food and drinking water. We used tAs residue data from the FDA's ongoing Total Dietary Survey (TDS), also known as the market basket study (FDA 1991(FDA -2004. TDS collects and ana lyzes approximately 280 foods for pesticide residues, industrial chemicals, and toxic and nutrient elements. Foods collected in the TDS are prepared as "table ready," that is, as would be consumed, for realistic estimates of dietary intake of those targeted components. As water concentrations recorded in the Natural Resources Defense Council database (Natural Resources Defense Council 2000) were used and assumed to be tAs. This database reported average and maximum As concentrations (a total of 8,970 records) in water from 25 U.S. states. The As drinking water concentration data were weighted by population and fitted for the best distribution, to yield a lognormal distribution with 1.03 ppb as the geometric mean and 4.06 ppb as the geometric standard deviation. We derived iAs concentration in each food commodity by using iAs percent age in the same food category as reported by Schoof et al. (1999aSchoof et al. ( , 1999b. Biomarker data for As exposure. We compared urinary biomarker data from the same individuals for consumption data (2,573 records) from the NHANES with model pre dictions during the same time period as the consumption data were collected. Detection rates for tAs, dimethylarsinic acid (DMA), arsenobetaine, and monomethylarsonic acid (MMA) were 98.9%, 87.4%, 66.7%, and 36.2%, respectively. Because the detection rates for iAs and other species were very low (1-7%), our model evaluation study using biomarker data focused primarily on tAs.  Figure 1). The SHEDSDietary model can use residues for food items as consumed, as well as residues of RACs. The reported NHANES food items were matched with food items in the TDS where pos sible ( Figure 1, step 1). If TDS residues for As were available for a particular food (e.g., rice, chicken), then SHEDSDietary ran domly drew a TDS tAs or iAs residue from that corresponding residue distribution of the same food. Otherwise, the model applied the FCID recipe files to the NHANES food items and randomly selected a residue for each of the RAC ingredients according to the recipe ( Figure 1, step 2).
Through the recipe files, the unmatched foods consumed were matched by RAC so that residues for those foods could be calcu lated. The SHEDSDietary model drew the same residue value if that RAC was found in the same foods. Assignment of residues for nondetect values depended on the commod ity: if there was at least one detection, half the limit of detection was assigned; if no As values were detected, zero values were assigned. For each NHANES food diary, SHEDSDietary was applied using Monte Carlo simulation by selecting a residue value from an empiri cal distribution for each TDS food or RAC. Although a particular commodity may be used in multiple foods, the cooking method may differ, so it will have a different food form. Process factors can then be applied ( Figure 1, step 3). These factors account for food changes and related concentration changes due to dilu tion, drying, and so on, but were not used here because of the lack of sufficient such informa tion for our study. Each simulated individual's exposure for each commodity was calculated by multiplying total daily consumption with corresponding residues. Aggregate daily expo sure was calculated by summing exposures across all commodities: Exposure (mass chemical/eating occasion) = Σamount of food item consumed (mass) × As concentration in the food item (mass chemical/mass food) × process factors. [1] Summation of As exposures from every eating occasion for 1 day yielded the individual's daily tAs exposure (Figure 1, step 4). In principle, both food residues and drinking water con centrations may vary by eating occasion and/ or across foods consumed within an eating occasion.
For modeling drinking water As expo sures, we used the NHANES data to assess   Food consumption data from NHANES Total As residue data by food item or commodity from TDS Fittings of residue data into suitable statistical distribution Food products people in the survey consumed such as pizza, raw apple Raw agriculture commodity (RAC) Pesticide usage percentages by RAC from USDA Concentration or dilution factors due to processes of food from RAC into food products Data base for percents of various RACs for the food products Dietary intake estimation Modified matched data

No
Step 1 Step 3 Step 2 Step 4

Yes Yes
Matched data

Usage factors
Process factors the timing and amounts of direct and indirect drinking water intake within a simulated per sonday. Total drinking water consumed (both direct and indirect water, from tap, bottled, and other sources) was assumed to contain the same concentration level; that is, only one concentration value was selected in the Monte Carlo simulation for each eating occasion. Water used in cooking is one example of indi rect water. The modeled drinking water expo sure algorithm in SHEDSDietary is similar to that used for food exposure (Equation 1). One residue value is randomly selected and multi plied by total water intake to obtain drinking water exposures. Although SHEDSDietary can be used to model longitudinal dietary exposure as well as crosssectional exposure, we addressed only the crosssectional exposure based on singleday data. PBPK modeling. We used MENTOR3P to represent absorption, distribution, metabo lism, and excretion processes of As inside the human body by lumping together similar tis sues as a set of physiologic compartments. A ''flowlimited'' PBPK formulation, represent ing a simplification of a generalized PBPK model of MENTOR3P (Figure 2), was adopted here. This simplified PBPK model for As employed the model parameters in the work of Yu (1999aYu ( , 1999b, including frac tional blood flow rates, metabolism param eters, and tissue/blood partition coefficients. The modification of calculating tissue volumes and blood flow rate based on body weight was added to this simplified population oriented PBPK model (see Georgopoulos et al. 2008 and references therein), such that the interindividual variability of these physiologic parameters can be captured. The dynamics of four As circulating species in body compart ments (arsenates, arsenites, and the As metab olites MMA and DMA) were captured using this PBPK model. Also characterized were the corresponding biomarker levels in urine.
Model results evaluation. We conducted two types of model evaluation: a) SHEDS Dietary predictions were compared with National Human Exposure Assessment Survey (NHEXAS) duplicate diet data; and b) linked SHEDS-MENTOR predictions were compared with NHANES biomonitor ing data. Duplicate food study subjects in NHEXAS (n = 156) were matched by age, sex, and location with modeled results from SHEDSDietary (based on NHANES con sumption diaries). To account for variability, we ran the model 200 times for 156 matched subjects, and selected three cumulative distri bution functions according to the 5th, 50th, and 95th percentiles of the 200 simulations. Modeled estimates of tAs dose from the linked SHEDS-MENTOR predictions were com pared with the NHANES urinary biomarker data for tAs. For the matched NHANES dietary consumption with NHANES bio marker data, 2,355 records were available.

Results
Using the SHEDSDietary model, we cal culated that the tAs exposure from food is 0.36, 1.28, and 1.40 µg/kg/day for the mean, SD, and 95th percentile, respectively, for the entire simulated population (Table 1). The tAs exposure from food for young children (≤ 5 years of age) is higher (means ranged between 0.54 and 0.62 µg/kg/day) than that shown for other age groups (means ranged between 0.25 and 0.37 µg/kg/day) ( Table 1). Based on mean values in Tables 1 and 2, the tAs exposure from food predicted by SHEDS Dietary is, on average, approximately 14 times higher than the tAs exposure from drinking water. iAs exposures from drinking water are 0.025, 0.104, and 0.107 µg/kg/day for the mean, SD, and 95th percentile, respectively (Table 2). There is no clear age group differ ence in the drinking water As exposure.
The iAs exposure from food for young children (≤ 5 years of age) is higher (means ranged between 0.08 and 0.23) than that shown for other age groups (means ranged between 0.03 and 0.04) ( Table 1). The iAs exposure from food predicted by SHEDS Dietary model (Table 1) is on average two times higher than the tAs exposure from drinking water (Table 2). Thus, even if we assume all As in the drinking water exists in the iAs forms, the dietary food iAs exposure by the modeled general U.S. population is still greater than the drinking water exposure. Summarizing the iAs contribution by food commodities, we estimate that about 10% of tAs exposure from foods is the toxic iAs form.
Among biomarkers analyzed for As expo sure in the NHANES subjects, arsenobetaine and DMA had high concentrations, with means of 8.4 and 5.4 µg/L, respectively, whereas the mean concentration for tAs in the urine was 18.4 µg/L [see Supplemental Material, Table 4s (available online (doi:10.1289/ehp.0901205. S1 via http://dx.doi.org)].
Compared with the NHEXAS duplicate diet data, our SHEDSDietary modeling of tAs exposure from foods performed reasonably well ( Figure 3). Among 156 paired compari sons, the mean ± SD of SHEDSDietary esti mates for tAs exposure from food was 0.192 ± 0.561 µg/kg/day, compared with 0.185 ± 0.3  (Table 3).
The linked SHEDS-MENTOR model also predicted well the tAs in urine (Figure 4). The SAS (version 9.2; SAS Insitute Inc., Cary, NC) regression analysis showed a good fit with a slope of 1.4 and R 2 of 0.91 for the logarithmictransformed predicted and meas ured values. The means of model predictions and NHANES urine measurements of tAs are 18.32 and 18.06 µg/L, respectively (Table 3).

Discussion
It is challenging to study As exposure in the general human population because many vari ables affect the processes, and obtaining rel evant information has numerous limitations.
Unlike the study of occupational As expo sure, where populations are relatively homo geneous, As compounds are easy to identify, and exposure routes are limited. As exposure in the general population is complicated with subject heterogeneity, different As species, and multiple exposure routes. Some informa tion easily obtainable from industrial settings may be difficult or too expensive to obtain in general environmental settings. Another chal lenge is that As from the diet exists in many forms, most as oAs, which is much less toxic than iAs. Thus, it is important to consider the different As species in As exposure and risk analysis. Using some modeling approaches to estimate general human exposure to As and to identify some data gaps or assumption deficiencies is helpful for understanding As exposure in the general population.
Previous studies have shown that, for most people in the general population, diet may be the largest source of exposure to As (MacIntosh et al. 1996). For example, MacIntosh et al. (1997) reported that mean dietary intakes of tAs is 50.6 µg/day for females and 58.5 µg/ day for males. Some recent studies suggested that dietary exposure to As may exceed the maximum As intake from drinking water in areas where elevated As levels were found in rice (Williams et al. 2007). Other studies have shown a greater intake of toxic iAs from food compared with that from drinking water (e.g., Meacher et al. 2002). Schoof et al. (1999aSchoof et al. ( , 1999b estimated that intake of iAs in the U.S. diet ranges from 1 to 20 µg/day, with a mean of 3.2 µg/day. An estimation of dietary iAs intake by U.S. children was 3.2 µg/day on average, with a range of 1.6-6.2 µg/day (Yost et al. 2004). These estimations are close to values reported in another study that showed average iAs intake ranges from 1.34 µg/day in infants to 12.54 µg/day in 60 to 65yearolds (Tao and Bolger 1998). However, these stud ies of dietary As exposure are usually based on the same assumed food intake values per per son, so they lack characterization of interindi vidual variability of exposures. Lack of data about the actual amount of food consumed accounted for at least 80% of the total uncer tainty for As exposure estimation (MacIntosh et al. 1996). MacIntosh et al. (1997 also pointed out that the food consumption-food composition approach adopted in their earlier study (MacIntosh et al. 1996) did not capture all the As exposure as reflected in the empiri cally weighted toenail As concentration data used for validation.
In the present study we used data from NHANES, thus far the most comprehensive survey including food intakes, which has the unique advantage of containing biomarker information for the same subjects in the sur vey (NHANES 2003(NHANES -2004. Biomarkers of exposure are independent measurements that   can be used to evaluate the validity of dietary assessment methods and food composition data. Using the biomarker data from the same survey for model evaluation is more reliable, because it does not suffer from other complica tions such as differences between study groups related to location, lifestyle, living conditions, and other potential confounding factors. The NHANES data are also more recent than data such as the Continuing Survey of Food Intakes by Individuals 1994(Agricultural Research Service 2009 used in previous studies. Compared with previous As exposure modeling, the SHEDSDietary model we used in this study performed food item matching and incorporated usage factors in the modeling. We also based the dietary intake estimation on actual eating occasions (Figure 1).
Our modeling approach yielded estimates that are very compatible with the duplicate diet data (Figure 3). The mean and 95th percentile of modeled tAs exposure (0.192 and 0.723 µg/kg/day, respectively) were very comparable to As intakes from the NHEXAS duplicate food study (0.185 and 0.612 µg/kg/ day, respectively) for the same age, sex, and location. The combination of the SHEDS Dietary model with MENTOR3P also pre dicted urine tAs concentrations that compared well with biomarker monitoring data in the NHANES (slope = 1.4 and R 2 = 0.91 with logarithmictransformed data) ( Figure 4). Thus, it seems that our modeling approach has overcome some previous deficiencies and yielded more reliable estimates.
Because of the low detection rates of iAs (1-7%) in the NHANES urine data, the evaluation of SHEDS-MENTOR model ing results for iAs could not be conducted. However, the Yu et al. PBPK model adapted for MENTOR3P has been validated with experimental observations from the literature for urinary biomarker levels of speciated arse nic such as in Buchet et al. (1981), Pomroy et al. (1980), and Johnson and Farmer (1991) as described by Yu (1999aYu ( , 1999b. Because the TDS study provided only tAs concentra tions in foods, we used the iAs percentage in the same food category as reported by Schoof et al. (1999aSchoof et al. ( , 1999b to derive iAs food con centrations. This assumption could result in uncertainties of estimated iAs exposure from foods, which could be carried into the subse quent PBPK modeling analysis for estimating target tissue doses and biomarker levels of iAs. Our results in general are consistent with those reported in previous studies. For example, a duplicate diet study of children in Germany showed weekly As intake as 2.31 µg/ kg body weight/week, which is equivalent to 0.33 µg/kg/day and is close to our estimate of 0.39 µg/kg/day (Wilhelm et al. 2003). These are compatible with our estimates of 7.2 and 3.5 µg/day for 1 to 2yearolds and 10.8 and 4.1 µg/day for 3 to 5yearolds. Another study showed that average intake of tAs for the gen eral U.S. population estimated by the Dietary Exposure Potential Model is 0.653 µg/kg/day (Moschandreas et al. 2002), which is similar to our result of 0.39 µg/kg/day for the same popu lation. Even when iAs is specifically considered, our results are also within the wider range of iAs exposures reported in previous such stud ies. For example, Schoof et al. (1999aSchoof et al. ( , 1999b estimated the iAs intake from U.S. diet to be 1-20 µg/day with a mean of 3.2 µg/day, and Tao and Bolger (1998) reported it as 1.34 µg/ day in infants and 12.54 µg/day in adults 60-65 years of age. Our results of the major food contributors to As exposure are consistent with the As levels measured in various foods in U.S. markets (Tao and Bolger 1998).
Our modeling assessment advances the sci ence by using the large and recent databases from NHANES, TDS, NHEXAS, and the Natural Resources Defense Council to esti mate As intake for the U.S. general population from food and drinking water. Other unique aspects of research presented in this article are    evaluation of tAs intake estimates using dupli cate food survey data from NHEXAS, and using urine biomarker data from NHANES to evaluate the SHEDS-MENTOR model predictions. The integrated exposure and dose modeling application presented in this arti cle for As has not been attempted before for a large general population (e.g., the U.S. general population), to our knowledge, in the expo surerelated literature. The SHEDSDietary model and the linked SHEDSDietary-MENTOR3P model predictions compared well with the measured duplicate diet data and urine biomarker data, respectively; thus, this was an important model evaluation effort to provide more confidence in these predictive exposure assessment tools.

Conclusions
The relationship between As intake from drink ing water and related health effects has been well studied previously. Using rich data sets and stateofthescience models, we found that the general U.S. population may be exposed to tAs and toxic iAs through the dietary route more from eating some Ascontaining foods than from drinking Ascontaining water. The major food contributors to tAs exposure were fish, shellfish, rice, fruit juices and fruits, and meats; the major food contributors to iAs exposure were vegetables, fruit juices and fruits, rice, beer and wine, and flour, corn, and wheat. Approximately 10% of tAs exposure from foods is the toxic iAs form.
Our study reinforces and expands on previ ous observations that dietary As exposure via food is an important route for As intake by the general population and that in some cases it can be even a greater source of As exposure than drinking water. Thus, for complete expo sure analysis and risk assessment in the gen eral population, iAs intake from food should be considered in addition to iAs intake from drinking water.