Characterizing Chemical Exposure Trends from NHANES Urinary Biomonitoring Data

Abstract

Background:

Xenobiotic metabolites are widely present in human urine and can indicate recent exposure to environmental chemicals. Proper inference of which chemicals contribute to these metabolites can inform human exposure and risk. Furthermore, longitudinal biomonitoring studies provide insight into how chemical exposures change over time.

Objectives:

We constructed an exposure landscape for as many human-exposure relevant chemicals over as large a time span as possible to characterize exposure trends across demographic groups and chemical types.

Methods:

We analyzed urine data of nine 2-y cohorts (1999–2016) from the National Health and Nutrition Examination Survey (NHANES). Chemical daily intake rates (in milligrams per kilogram bodyweight per day) were inferred, using the R package bayesmarker, from metabolite concentrations in each cohort individually to identify exposure trends. Trends for metabolites and parents were clustered to find chemicals with similar exposure patterns. Exposure variation by age, gender, and body mass index were also assessed.

Results:

Intake rates for 179 parent chemicals were inferred from 151 metabolites (96 measured in five or more cohorts). Seventeen metabolites and 44 parent chemicals exhibited fold-changes $\ge 10$ between any two cohorts (deltamethrin, di-$n$-octyl phthalate, and di-isononyl phthalate had the greatest exposure increases). Di-2-ethylhexyl phthalate intake began decreasing in 2007, whereas both di-isobutyl and di-isononyl phthalate began increasing shortly before. Intake for four parabens was markedly higher in females, especially reproductive-age females, compared with males and children. Cadmium and arsenobetaine exhibited higher exposure for individuals $>65$ years of age and lower for individuals $<20$ years of age.

Discussion:

With appropriate analysis, NHANES indicates trends in chemical exposures over the past two decades. Decreases in exposure are observable as the result of regulatory action, with some being accompanied by increases in replacement chemicals. Age- and gender-specific variations in exposure were observed for multiple chemicals. Continued estimation of demographic-specific exposures is needed to both monitor and identify potential vulnerable populations. https://doi.org/10.1289/EHP12188

Introduction

Human biomonitoring studies make up an important area of exposure science that aids in public health assessments. Chemical metabolites or biomarkers are measured by collecting biological samples, typically blood/serum or urine, from individuals at a given point in time. Biomonitoring studies allow for individual- and population-level assessment of chemical exposure and risk. The importance of biomonitoring data has been emphasized by the U.S Government Accountability Office, which recommended that the US Environmental Protection Agency (EPA) develop a strategy to categorize existing biomonitoring data (e.g., determine which chemicals present in humans are lacking toxicological data¹) identify limitations in analytic approaches, and prioritize gaps.² The National Research Council of the US National Academies of Science, Engineering, and Medicine also called for increased use of biomonitoring data to support chemical risk assessment activities.^3–5

Traditional exposure modeling uses data that characterize different pathways of exposure, such as from use of consumer products or bodily uptake of chemicals in the environment, to build models that estimate a daily chemical intake rate. However, this is a difficult problem when it comes to modeling the complete exposome for an individual because of many unknowns. These include activity patterns of the individual, as well as various chemical-specific data (fate and transport, use in manufacturing/consumer goods, and release into the environment). It is even more challenging to collect or estimate these data consistently and accurately from year to year.⁶ However, biomonitoring studies capture the aggregate exposure from all pathways to each individual in the study and can be collected at multiple time points.^6–9

One particularly rich biomonitoring resource is the US Centers for Disease Control and Prevention (CDC) National Health and Nutrition Examination Survey (NHANES), which is designed to assess the health and nutritional well-being of children and adults in the United States. Since 1999, NHANES has become a continuous survey, performed in 2-y cohorts. Each cohort is designed to be statistically representative of the entire US population, with about 5,000 participants each year. The NHANES biomonitoring component can be treated as temporal concentration data given the nature of the continuous survey design and relatively large overlaps in the biomarker panels selected each in each cohort. Therefore, NHANES can be used to obtain a big-picture view of the chemical exposure landscape in the US population over time, as well as to identify trends for individual chemicals and various population groups.

A number of previous works have used the NHANES data (blood, serum, or urine concentrations) to look at chemical trends, spanning anywhere from 3 to 14 y. Some examined a single chemical in a single population group.^10,11 One study looked at the difference in blood concentrations of a single chemical between two distinct population groups.¹² Two studies assessed multiple chemicals of the same type or class for one population group,^13,14 whereas four studies compared trends across multiple populations (different combinations of age, sex, race/ethnicity, and smoking status).^15–18 Although these works effectively use the longitudinal nature of the NHANES survey, they are in most cases limited to a small number of chemicals, usually of the same class, and a subset of the NHANES participants. One study, by Nguyen et al., used all NHANES participants to characterize age-based trends from urine biomarker concentrations across 141 of the NHANES chemicals.¹⁹ However, this study, and all the aforementioned studies, employ the blood/serum/urine concentrations without making any estimates of the external exposure (e.g., daily intake rates) to the environmental chemicals that produce the biomonitoring concentrations.

Over the past 25 y, approaches have been developed for estimating daily intake exposure rates of specific parent chemicals based on the concentrations of those chemicals or their metabolites in biomonitoring studies.^20–25 Recently these approaches have been generalized (i.e., made chemical independent).^21,26,27 Using these generalized approaches, we can obtain time series of population intake rates of chemicals across all NHANES urine biomarkers.

Reyes and Price used all NHANES participants and employed a reverse-dosimetry approach⁹ to estimate exposures to a number of demographic groups, but their analysis was limited to six phthalate metabolites.²⁸ Prior to the work outlined here, there has not been a systematic effort to analyze the entire NHANES urine chemical biomarker data (all metabolites, cohorts, and individuals) for temporal trends in parent compounds across cohorts or how those trends might vary across different demographics.

In this work, we set out to construct a large-scale human exposure landscape from biomonitoring data. To best use the urine biomarker concentration data of the continuous NHANES survey, five objectives were formulated: a) devise an approach to systematically obtain exposure estimates for a panel of chemicals measured in urine, b) characterize overall exposure trends, c) group chemicals by similar patterns of exposure over time, d) assess exposure patterns within and across various demographic groups, and e) identify more robust, longer-term exposure patterns at the decades timescale. To address these objectives, we built an analysis pipeline to estimate chemical exposures for each 2-y cohort of the NHANES continuous survey. The approach used in this work is a systematic way to estimate and characterize aggregate human exposure from all chemicals measured in urine, allowing for improved understanding of how human exposure differs by chemical, by population, and over time.

Methods

All data, methods, and analyses performed in this work are depicted in the workflow diagram in Figure 1. We first employed the R package “bayesmarker” to generate exposure estimates for all chemicals in each cohort, as well as estimates where cohorts have been combined by decade. Chemicals were then filtered and processed based on data quality and the availability to identify chemicals with similar exposure trends, as well as chemical- and demographic-specific patterns. This workflow allowed for multiple assessments of the data, resulting in a better characterization of chemical exposures.

Figure 1.

Analysis workflow used to produce all figures. Data from NHANES and the literature (metabolite map) were curated to create the bayesmarker input file, consisting of three tables (Excel Tables S1–S3). The bayesmarker pipeline was run in the combine data mode to compare exposures between the 2000s and the 2010s and in the individual cohort mode to generate a landscape of exposure spanning from 1999 to 2016. Chemical filtering and data imputation were used to carry out time-series analysis for chemicals with sufficient data. Population-specific patterns were identified as were clusters of chemicals with similar exposure patters. Clusters were assessed for enrichment related to chemical class and use. Note: ACToR, Aggregated Computational Toxicology Resource; BWT, bodyweight; CrI, Credible Interval; DEMO, demographic; FUse, Functional Use Database; GM, geometric mean; LOD, Limit of Detection; MW, molecular weight; NHANES, National Health and Nutrition Examination Survey.

Bayesmarker Package

The R package bayesmarker was developed and described in full detail by Stanfield et al.²⁶ and is publicly available on GitHub (https://github.com/USEPA/CompTox-HumanExposure-bayesmarker). Bayesmarker uses Bayesian methodology to infer the geometric mean (GM) daily intake rates for parent chemicals that are consistent with metabolite concentrations measured in urine samples. Bayesmarker uses data from urine because it is the human media in which the most chemicals are measured, procedures for adjusting urine concentration data have been extensively studied, and the metabolites measured in urine tend to have shorter half-lives and, therefore, better represent recent exposures. A mass-balanced approach was used to characterize the impact of two sources of uncertainty in the estimated parent chemical intake rates based on biomonitoring data. The first is that urine biomarkers often are the metabolic products of the parent chemicals to which we are attempting to estimate exposure. Given that some parent chemicals can be metabolized into multiple different metabolites, it can be difficult to know the proportion of the parent chemical that is transformed into each metabolite. Further, one biomarker might result from an unknown combination of multiple parent chemicals. The second distinct contributor to uncertainty is the detection limits of the analytical techniques used in NHANES. For many metabolites, a large fraction of biomonitoring samples is below the limit of detection (LOD). Data censoring techniques²⁹ (i.e., in this context, making use of the knowledge of the distribution of detectable levels to predict the levels that are below the LOD) are used to address this issue and inform the certainty with which we can estimate GM intake rate.

The bayesmarker pipeline can be described in three stages. Stage one calculates metabolite GM concentrations from estimates of population quantiles generated using the individual measurements of the study participants. This stage assumes the biomonitoring data are distributed log-normally. Stage two converts the GMs and distributions from units of concentration to units of exposure. This is achieved via a relatively simple toxicokinetic model that assumes steady-state exposure; that is, the amount of chemical leaving the body through urine is immediately matched by intake of additional chemical.^30,31 To handle variability in the time since the last urine void (because the concentration of chemical depends on the total volume of urine and therefore the time since last void), the metabolite concentrations use creatinine excretion, which is also assumed to occur at a constant rate. Creatinine excretion rate is estimated for study individuals using a predictive model trained on NHANES data that incorporates gender, ethnicity, age, and bodyweight (BW).^26,31 Stage three then propagates the metabolite exposures to parent chemicals based on literature-reported parent–product relationships. This mapping contains links from metabolites to parents based on known metabolic reactions. When there is a one-to-many relationship between a given metabolite and its parent chemicals (or vice versa), different possible combinations of metabolite proportions are considered by the Bayesian statistical model. All combinations are consistent with preserving mass balance of parents and metabolites.

Inputs for the Bayesmarker Package

To run the bayesmarker package, a single input file is needed. This file should be in the form of an Excel workbook that contains three sheets/tables: a) all the metabolites of interest across all cohorts in which they were measured (Excel Table S1), b) the relevant NHANES data file names (urine concentrations are part of the Laboratory Data), such as demographic (DEMO.xpt files part of the Demographics Data), bodyweight (BMX.xpt files part of the Examination Data), and urine (ALB_CR.xpt and UCFLOW.xpt files part of the Laboratory Data) data files corresponding to the metabolites included in sheet 1 (Excel Table S2), and c) a mapping of known parent chemicals to each metabolite in sheet 1 (for a description of how this mapping was obtained, see Stanfield et al.²⁶; Excel Table S3).

NHANES Data

NHANES is an elaborate study run by the CDC at multiple locations throughout the United States (all data and information can be found at https://www.cdc.gov/nchs/nhanes/index.htm). Among many other end points, NHANES includes measurement of chemical indicators of exposure to environmental chemicals—that is, exposure biomarkers.^32,33 In NHANES, concentrations of biomarkers are obtained by analyzing “spot” biological samples (e.g., blood and urine at single time points). Chemical exposure biomarker data are stratified by age group, gender, and race/ethnicity and published in the National Reports on Human Exposure to Environmental Chemicals.³³ Since 1999, NHANES has expanded from about two dozen biomarkers of exposure to currently $>150$ with the survey being conducted in 2-y cohorts. The cohort design is used to help ensure sufficient statistical power to characterize key biometrics for various population groups of interest. The total number of individuals that participate in a given NHANES cohort varies but is roughly 10,000 (5,000/y). To provide sufficient material for chemical analysis, testing for specific chemicals is conducted in only one of three subsets of the total cohort, with between 2,000 and 3,500 individuals—each subset is representative of the total US population.³⁴ Thus, we chose to use the NHANES data for obtaining time-series exposure estimates. Public release of the biomarker data from a cohort can take up to several years to be completed owing to various timelines of experimental measurement procedures; at the time of writing, the most recent data available was from the 2015–2016 cohort. NHANES metabolite concentrations are reported as raw data, at select percentiles (e.g., 50th, 95th), and as a GM. For this analysis, quantiles of the total population and nine demographic groups [male, female, 6–11 y old, 12–19 y old, 20–65 y old, $>65$ y old, reproductive-age female (18–45 y old), body mass index (BMI) $\le 30{\mathrm{kg}/\mathrm{m}}^2$, and BMI $>30{\mathrm{kg}/\mathrm{m}}^2$] were calculated from the raw urinary metabolite concentrations in the CDC NHANES data files across all individuals. These specific demographic groups were chosen because they have been shown to exhibit different exposure profiles in previous works^26,35 and cover potentially vulnerable populations. Data on race/ethnicity and pregnancy status was self-reported. Relevant NHANES data file names used in this work are provided in Excel Table S2.

Obtaining Exposure Trends

All chemicals measured in one or more of the NHANES surveys (151 metabolites) were identified and the preferred name, Chemical Abstract Service (CAS) Registry Number, and DSSTox Substance Identifier (DTXSID) were extracted from the CompTox Chemicals Dashboard (https://comptox.epa.gov/dashboard/).³⁶ All relevant file information (metabolite concentrations, demographic data, and body measurements) from NHANES that was used to populate the bayesmarker input file (see Excel Tables S1–S3) was identified for each cohort in which the chemicals were measured. The function bayesmarker::get_NHANES_data was used with this input file to fetch and locally download all the relevant data from NHANES. The remaining steps in the pipeline (Figure 1) were then run for each individual NHANES cohort to obtain exposure estimates for all parent chemicals linked to the cohort-specific metabolite list via the metabolite mapping. Thus, time-series exposure estimates, in units of milligrams per kilogram BW per day, were calculated by running the bayesmarker pipeline for the NHANES cohorts 1999–2000 through 2015–2016. A simplified version of this analysis is outlined in the bayesmarker vignette called calculateExpInferences.Rmd.

Chemical Filtering

The bayesmarker pipeline produces a probability distribution for the GM of the intake rate (in milligrams per kilogram BW per day) that are consistent with the available data. The pipeline was applied to both metabolites and parent chemicals. We represent the uncertainty in the GM estimate using a 95% credible interval (CrI). This uncertainty is the result of the limited number of samples, the potentially high fraction of samples that are below the LOD (Figure S2), and the complexity of parent–metabolite relationships.

Estimates of chemicals with high uncertainty were excluded from further analysis using two criteria. Data on measurements of metabolites in a cohort were excluded if 98% of the concentration measurements from the survey population’s urine samples in the cohort were below the LOD. Doses of parent chemicals that were estimated based on measured metabolites were excluded if the estimate of the GM of a cohort had a 95% CrI that spanned 5 orders of magnitude. We chose 5 orders of magnitude because the uncertainty distribution was somewhat bimodal in that most chemicals had CrIs with a range spanning $<2$ or 3 orders of magnitude, whereas a minority of chemicals’ inferred exposure was very uncertain owing to complex metabolite stoichiometry. This threshold removed these obviously uncertain chemicals and resulted in a number of parent chemicals that was similar to the number of metabolites after the filtering step.

Clustering Chemical Exposure Trends and Enrichment

To obtain an unsupervised (i.e., algorithmic) grouping of chemical exposure trends by similarity, the R package dtwclust (version 5.5.10) was used.³⁷ Unsupervised clustering helped identify similarities between chemical exposure trends without assuming anything about those chemicals (e.g., class annotation, structure) and was therefore used to reveal unique patterns. After additional chemical filtering, which included removing chemicals missing data in six or more of the nine total cohorts, hierarchical clustering was performed for parents and metabolites separately using the tsclust function [dissimilarity metric = Global Alignment Kernel (GAK),³⁸ centroid = “shape_extraction”]. The data were preprocessed using a $z$-score scaling. The optimal number of clusters for both parents and metabolites was determined by setting $k\left(\text{number\hspace{0.17em}of\hspace{0.17em}clusters}\right)=5:20$ and evaluating them on the internal cluster validity indices (CVI) metrics. This resulted in 15 metabolite clusters and 8 parent chemical clusters.

To assess the extent to which the identified clusters contained chemicals with similar characteristics, three different forms of chemical data were used: a) chemical class annotation, b) the Aggregated Computational Toxicology Resource (ACToR) database,³⁹ and c) the Functional Use Database (FUse).⁴⁰ The annotated chemical classes of the NHANES metabolites were obtained from the full list of chemicals in the CDC’s National Report on Exposure to Environmental Chemicals (https://www.cdc.gov/exposurereport/pdf/Report_Chemical_List-508.pdf). Metabolite class was propagated to parent chemicals using the metabolite map. In cases where a parent had metabolites from more than one chemical class, the parent class was labeled as not applicable (NA). This class information is used here for enrichment but also as an annotation throughout the manuscript for additional context. Both ACToR and FUse contain use-related data (in other words, how chemicals are used in products and manufacturing processes), with 16 and 103 unique uses, respectively. Chemical identifiers in ACToR and FUse were matched to DTXSIDs via the US EPA’s CompTox Chemicals Dashboard.³⁶ Not every chemical had a match in ACToR and FUse, and those without a match were dropped from the enrichment analysis. Chemicals in FUse are assigned functional uses based on manufacturer/trade sources that report a chemical and its use. Phillips et al.⁴⁰ built quantitative structure–use relationship (QSUR) machine learning models trained on FUse to predict functional uses for chemicals without a reported use based on their structural features. Functional use predictions (a numeric probability between 0 and 1) for chemicals with probability $\ge 80\%$ were retained for the enrichment calculation. Data from FUse and the QSUR predictions were used together (note that this combined data is referred to as just “FUse” in tables and figures) to map functional uses to as many metabolites and parent chemicals as possible.

Overrepresentation, or cluster enrichment, was calculated using a one-tailed hypergeometric test. Enrichment analysis lets one assign a $p$-value to a given annotation among a group of items with a certain characteristic. For example, assume 100 chemicals are being analyzed, 5 of which are known to be phthalates. Say these chemicals are then clustered on some numerical data (e.g., exposure trend, occurrence in certain media) which results in 10 clusters of about the same size (10 chemicals in each cluster). If all 5 of the phthalates appear in the same cluster, this would be associated with a very small $p$-value, meaning there is a small probability that the overrepresentation of phthalates in that cluster occurred by chance when the overall fraction of phthalates in the 100 total chemical set is considered. That set is then enriched for phthalates. In this way we can assign a meaningful measure ($p$-value) of shared characteristics among each exposure trend cluster we identify. Clusters of size 1 were not assessed for enrichment.

Data Imputation

NHANES selects a specific set of metabolites to measure during each cohort of the survey based on a number of factors (e.g., those chemicals prioritized in recent toxicology and exposure literature or widely used consumer product chemicals), meaning that not all metabolites are screened in every cohort (Figure S1 and Excel Table S14). Thus, many chemicals have a limited number of exposure time points, reducing the ability of the data to support trend analyses. To overcome this limitation, where appropriate, a straightforward imputation approach designed to be trend-neutral was used. For metabolites and parent chemicals missing data before a certain cohort back to the first cohort, the first (oldest) measurement value for that chemical was used. For example, a metabolite that was first measured in 2003–2004 (with no data for 1999–2000 and 2001–2002), would be assigned the observed value in the 2003–2004 cohort for the 1999–2000 and 2001–2002 cohorts. Similarly, for data missing after a certain cohort up to the last cohort, the imputed value for all later cohorts would be equal to that of the most recently measured cohort. When data was missing between two or more cohorts, the median of the two surrounding cohorts with measurement data was used. This approach ensures that the existing data is the main driver of each trend, with the imputed data providing minimal influence.

Decades Comparison

The NHANES survey is designed such that measurement data from multiple cohorts can be combined to increase the sample size for each metabolite, leading to more confident concentration distributions and less uncertainty. The survey data are combined by scaling the 2-y Mobile Examination Center (MEC) weights by the number of cohorts being grouped (see https://wwwn.cdc.gov/nchs/nhanes/tutorials/Weighting.aspx). To obtain a more robust, long-term change in exposure for each chemical, data from NHANES cohorts were combined into two time periods, the 2000s decade (including cohorts 2001–2002, 2003–2004, 2005–2006, 2007–2008, and 2009–2010) and the 2010s decade (including cohorts 2011–2012, 2013–2014, and 2015–2016). Analyzed metabolites were limited to those having data in at least one cohort for both decades, which resulted in 92 metabolites that corresponded to 106 parent chemicals. The bayesmarker R package automatically performs data combination when the cohort argument for the readNHANES function contains more than one cohort and creates separate directories for each set of combined data. Exposure estimates and their 95% CrIs (obtained using the quantiles function) were obtained for each decade and all survey individuals (total population) were used for the comparison.

Results

Overall Exposure Trends

Figure 2 shows the GM daily intake rate estimates (in milligrams per kilogram BW per day) for all chemicals across all cohorts from 1999 to 2016. There were 151 metabolites (Figure 2A) linked to 179 (Figure 2B) parent chemicals, with data in at least one NHANES cohort (Excel Table S4). The chemical with the highest overall daily intake rate was nitrate, in the 2015–2016 cohort at $0.798\mathrm{mg}/\mathrm{kg}/\text{day}$. Most metabolites showed relatively minor differences in exposure over time [median fold change across all cohorts (maximum exposure/minimum exposure) was 1.88]. However, 17 metabolites exhibited an exposure change of $\ge 10$-fold (Excel Table S5). The chemical 2-isopropoxyphenol showed the largest difference, a 191-fold (or roughly 2.25 orders of magnitude) decrease from the 1999–2000 cohort to the 2003–2004 cohort. Exposure changes for the parent chemicals showed a markedly higher average fold change of 3.22, with 44 parents having an exposure change of $\ge 10$-fold (Excel Table S6). The greatest fold change in median exposure for parents was achieved by deltamethrin, which increased from the 1999–2000 to the 2013–2014 cohort by 369-fold (about 2.57 orders of magnitude).

Figure 2.

Estimated log–mean exposures in milligrams per kilogram bodyweight per day (mg/kg/day) of (A) urinary metabolites and (B) parent chemicals by 2-y cohort from 1999 to 2016. The bayesmarker package was run for each NHANES cohort (columns in heatmap) starting in 1999 to obtain geometric mean estimates in units of exposure for 151 metabolites and 179 parent chemicals. A gray cell represents missing data for that chemical in that cohort. Chemical class was obtained for metabolites from the NHANES report and propagated to the parent chemicals using the parent–metabolite map. Rows (chemicals) are first grouped by chemical class and then, within each group, clustered by similarity of exposure trend. The row grouping order matches the class order within the figure legend, and white space is allotted to allow easier delineation of the different chemical classes. The NA class is given to any metabolite without an assigned class and any parent chemical with one metabolite lacking a class label or with two metabolites of different class. The corresponding data for this figure is covered by Excel Table S4. Note: NA, not applicable; NHANES, National Health and Nutrition Examination Survey.

Figure 2 also shows the available data for each metabolite and indicates exposure patterns across cohorts. There is a clear variability in the amount of data for each chemical. Chemical–cohort combinations not examined by NHANES are indicated by the gray cells. Twelve chemicals (7 metals, which are not metabolized but are directly measured in urine by NHANES, and 5 phthalate metabolites) were measured in all NHANES cohorts, whereas 21 chemicals (mostly heterocyclic amines and herbicides) had data for only one cohort. In addition, the range of intake rates spans just over 9 orders of magnitude for both metabolites and parent chemicals.

Chemicals in Figure 2 are grouped and sorted by chemical class, as well as clustered by similarity of exposure trend within each class grouping. Therefore, an assessment of chemical variability within various chemical classes could be performed. For example, the sulfonyl urea herbicides (second chemical group from the bottom) exhibited very little variability in exposure, as did fungicides (third group from the top), carbamate pesticides (top group), and organochlorine pesticides (ninth group from the top). Alternatively, chemicals in the classes of metals (seventh group from the top), personal care and consumer products (seventh group from the bottom), and volatile organic compounds (VOCs; bottom group) showed greater variability in daily intake rates. Supplemental Excel Tables S7 and S8 cover the full exposure landscape of the NHANES continuous survey (GM intake rates and corresponding 95% CrIs for the metabolites and parent compounds, respectively, of all cohorts and all population groups).

Identifying Exposure Trend Categories

We performed trend analysis of groups of chemicals starting with the metabolites. 24 metabolites were filtered (removed from analysis) for having $\le 2\%$ measurements above the LOD. An additional 55 metabolites were filtered out owing to having $<4$ data points across time (in other words, being measured in fewer than four of the nine total cohorts). The 96 remaining metabolites were then grouped into 15 clusters based on trend similarities over time (Figure 3A). Table 1 displays significant cluster enrichments across key clusters. The largest cluster, cluster 4, had 31 chemicals exhibiting a steady decrease in exposure over time. This cluster was enriched for (i.e., chemicals inside this cluster were more likely to have been identified as) personal care and consumer product chemicals ($p=0.0173$) according to NHANES classification. ACToR was more likely to identify chemicals in this cluster as having CONSUMER.USE ($p=0.0145$). Cluster 8 displayed decreasing exposure occurring only in the later cohorts. Cluster 8 was enriched for the chemical class perchlorates (2 of the 3 perchlorate chemicals). Clusters 1 and 7 exhibited two different increasing trends, with the former increasing quickly to level off, and the latter, steady for the first cohorts and increasing quickly in the later cohorts. Cluster 1 was enriched for phytoestrogens and organophosphorus insecticides, whereas cluster 7 was enriched for VOCs. Cluster 6, consisting of 5 metabolites, showed a pattern of decreasing sharply for a middle cohort followed by an exposure increase toward the later cohorts. Cluster 6 was enriched for chemicals used as antimicrobials, in chemical industrial processes, and as pesticides. Finally, metabolite cluster 2 was composed of two VOCs, $N$-acetyl-$S$-($N$-methylcarbamoyl)cysteine and $S$-benzyl-$N$-acetyl-l-cysteine, with an extremely similar exposure pattern that stepped up to peak exposure followed by a sharp decrease in the last NHANES cohort.

Figure 3.

Exposure trend clusters of (A) urinary metabolites and (B) parent chemicals. Solid colored lines represent individual chemical trend lines and gray dashed lines represent the median exposure trend for all chemicals appearing in each individual cluster (represented by the numbered boxes/panels). Each exposure trend line for a chemical in each cluster is plotted in a different color. The dtwclust package was used to perform hierarchical clustering (tsclust function) after cluster number optimization (cvi function). The y-axis shows median exposure values after $z$-score normalization and the x-axis represents time. There are nine cohorts, going from the oldest (1999–2000) to the most recent (2015–2016), meaning the 5.0 mark on the x-axis represents the midway point of the NHANES continuous survey, which is the 2007–2008 cohort. The corresponding data for this figure and Table 1 is covered by Excel Table S9. Note: NHANES, National Health and Nutrition Examination Survey.

Table 1.

Enrichment of chemical uses (ACToR and FUse) and chemical classes (obtained from NHANES) within the identified metabolite and parent chemical clusters (Figure 3).

Chemical set	Data type	Cluster ($n$)	Cluster size	Category within cluster ($n$)	Category (total chemicals mapped to data type) ($N$)	$p$-Value	Category
Metabolites
	ACToR	4	31	11	41 (86)	0.0145	CONSUMER.USE
	ACToR	6	5	2	8 (86)	0.008	ANTIMICROBIAL
	ACToR	6	5	4	49 (86)	0.032	CHEMICAL.INDUSTRIAL.PROCESS
	ACToR	6	5	3	27 (86)	0.0426	PESTICIDE
	FUse	15	3	2	7 (55)	0.0398	Flame_retardant
	Class	1	17	4	8 (91)	0.0374	Organophosphorus insecticide
	Class	1	17	4	5 (91)	0.0039	Phytoestrogen
	Class	2	2	2	17 (91)	0.0332	VOCs
	Class	4	31	6	8 (91)	0.0173	Personal care and consumer product chemical
	Class	7	12	6	17 (91)	0.0084	VOC
	Class	8	9	2	3 (91)	0.025	Perchlorates
Parents
	ACToR	1	26	14	23 (63)	0.0166	PERSONAL.CARE.PRODUCT
	ACToR	3	9	2	2 (63)	0.0184	TOXIN
	FUse	4	5	2	3 (54)	0.0202	Monomer
	FUse	4	5	2	2 (54)	0.007	Solvent
	Class	1	26	6	7 (63)	0.0166	Personal care and consumer product chemical
	Class	4	7	4	12 (63)	0.0205	VOC
	Class	8	2	1	1 (63)	0.0317	Herbicide

Note: The table consists of only significant cases of enrichment (upper tailed hypergeometric test; $p<0.05$) for metabolite and parent clusters. The number in the parentheses in column 6 changes by data type; this is because not all chemicals have a known use or chemical class. ACToR, Aggregated Computational Toxicology Resource; FUse, Functional Use Database; NHANES, National Health and Nutrition Examination Survey; VOC, volatile organic compound.

For the parent chemicals, 71 parents were filtered out because they exhibited at least one data point with very high uncertainty in median exposure value (95% CrI around the median spanning $\ge 5$ orders of magnitude). Of the 108 remaining parents, 44 were filtered out owing to having $<4$ data points. The 64 remaining chemicals grouped into 8 clusters (Figure 3B). As was the case for the metabolites, the largest cluster, cluster 1 for the parents (26 chemicals), exhibited a steady decrease in exposure over time and contained 14 chemicals with the ACToR use PERSONAL.CARE.PRODUCT and 6 with the class annotation “personal care and consumer product chemical” (Table 1). Alternatively, cluster 3 exhibited a pattern of marginally increasing exposure over time and was enriched for the ACToR use TOXIN, containing both of the two toxicants (arsenocholine and trimethylarsine oxide) in the full set of parent chemicals with ACToR data. Cluster 4, which showed a pattern similar to that of metabolite cluster 2 (slow or stepwise exposure increase until one of the last cohorts followed by a sharp decrease in exposure in the last one or two cohorts), contained four VOCs and enrichment for the FUse labels of monomer and solvent. Excel Table S9 contains the data for Figure 3 and data relevant to Table 1.

Chemical and Demographic-Specific Trends

Phthalates.

There were 10 phthalates in the parent chemical set, and the exposure trends for 4 of these were relatively dynamic (large exposure changes over time) and variable (multiple different trend patterns). Two of these 4 phthalates showed a decrease in exposure over time. Diethyl phthalate (Figure 4A) exposure decreased after the 2003–2004 cohort, whereas di-2-ethylhexyl phthalate (DEHP) exhibited a sharp increase between the 1999–2000 and 2007–2008 cohorts, at which point its exposure steadily declined (Figure 4B). Unlike diethyl phthalate and DEHP, both di-isononyl phthalate and di-isobutyl phthalate exhibited exposure increases between multiple cohorts (Figure 4C,D; Excel Table S10 contains all the data for Figure 4). For the remaining 6 phthalates, dibutyl phthalate was filtered out owing to the CrI span threshold; dicyclohexyl and benzylbutyl phthalate exhibited slight decreases in exposure over time, but the 95% CrI spanned the magnitude of these decreases; di-$n$-octyl and di-isodecyl phthalate showed steady exposure until the 2015–2016 cohort (decrease); and dimethyl phthalate also exhibited relatively steady exposure over all cohorts except for a dip in exposure during the 2005–2006 and 2007–2008 cohorts.

Figure 4.

Exposure trends by population group for select phthalates. These four phthalates were selected from the 10 phthalate parent chemicals because they exhibited the largest changes in exposure over time and complementary trends. Geometric mean exposure (point) estimates in units of milligrams per kilogram bodyweight per day with 95% credible intervals shown as error bars for 10 population groups. Each demographic trend line is marked by a relevant letter or point symbol. Missing data for cohorts were imputed (points designated by an opaqueness of 0.5) based on surrounding nonmissing data. The corresponding data for this figure is covered by Excel Table S10. Note: <, BMI_le_30 (BMI $\le 30$); >, BMI_gt_30 (BMI $>30$); A, adults (20–65 years of age); BMI, body mass index; C, children (6–11 years of age); F, female; J, juveniles (12–19 years of age); M, male; R, ReproAgeFemale (reproductive-age females); S, seniors ($\ge 66$ years of age); T, total.

Parabens.

Four parabens studied in this work exhibited markedly higher exposure in females, even more so for reproductive-age females, and lower exposure to males and children (Figure 5; Excel Table S11). $N$-propyl paraben, butyl paraben, and methyl paraben all showed a decrease in exposure over time and seem to be converging with the total population trend line to various degrees. Ethyl paraben alternatively showed an initial exposure increase from the 2005–2006 cohort to the 2009–2010 cohort, followed by a decrease in exposure in the later cohorts (Figure 5A). Parabens are primarily used as preservatives and antibacterial agents in cosmetics and other personal care products,^41,42 which supports the observed higher exposure for females.

Figure 5.

Exposure trends by population group for parabens. Geometric mean exposure (point) estimates in units of milligrams per kilogram bodyweight per day with 95% credible intervals shown as error bars for 10 population groups. Each demographic trend line is marked by a relevant letter or point symbol. Missing data for cohorts were imputed (points designated by an opaqueness of 0.5) based on surrounding nonmissing data. The corresponding data for this figure is covered by Excel Table S11. Note: <, BMI_le_30 (BMI $\le 30$); >, BMI_gt_30 (BMI $>30$); A, adults (20–65 years of age); BMI, body mass index; C, children (6–11 years of age); F, female; J, juveniles (12–19 years of age); M, male; R, ReproAgeFemale (reproductive-age females); S, seniors ($\ge 66$ years of age); T, total.

Diverging population groups.

Four other chemicals showed potential age-dependent exposure patterns. Cadmium and arsenobetaine, exhibited sustained differences in age-dependent exposures (Figure 6A,B). Cadmium has had a number of uses over the last few decades and arsenobetaine is an organoarsenical compound commonly found in fish.⁴³ Another observed pattern of interest occurred for the chemicals $N$, $N$-dimethylformamide (DMF) and acrylonitrile, which showed greater exposure for males and adults and lower exposure to children and to individuals $>65$ years of age (Figure 6C,D; Excel Table S12 contains all the data for Figure 6). Both chemicals are used in production processes, acrylonitrile for chemical and polymer production,⁴⁴ and DMF, a solvent, for a wide variety of products and practices, including preparation of polyacrylonitrile.⁴⁵ Again, these chemicals suggest potential age-dependent trends, which may or may not directly translate to population-specific vulnerability.

Figure 6.

Exposure trends by population group for four chemicals with notable differences across populations. Geometric mean exposure (point) estimates in milligrams per kilogram bodyweight per day with 95% credible intervals shown as error bars for 10 population groups. Each demographic trend line is marked by a relevant letter or point symbol. Missing data for cohorts were imputed (points designated by an opaqueness of 0.5) based on surrounding nonmissing data. The corresponding data for this figure is covered by Excel Table S12. Note: <, BMI_le_30 (BMI $\le 30$); >, BMI_gt_30 (BMI $>30$); A, adults (20–65 years of age); BMI, body mass index; C, children (6–11 years of age); F, female; J, juveniles (12–19 years of age); M, male; R, ReproAgeFemale (reproductive-age females); S, seniors ($\ge 66$ years of age); T, total.

Exposure Changes from the 2000s to 2010s

In total, 106 parent chemicals met the criterion of having at least one measured metabolite in at least one cohort from both decades (Figure 7; Excel Table S13). The labeled chemicals in Figure 7 are those that exhibited a change in exposure across the two decades that was above a preset threshold. These chemicals met two criteria: a) the 95% CrIs on the exposure mean from the two decades did not overlap, and b) the ${\log }_2$-fold change between the two individual decade exposure means was $\ge 0.5$ (i.e., an exposure increase from the 2000s to the 2010s) or $\le -0.5$ (exposure decrease from the 2000s to the 2010s). These criteria help identify quantifiable differences in exposure between decades. Forty-two chemicals met criteria 1 (12 increases and 30 decreases from the 2000s to the 2010s), whereas 33 chemicals had a ${\log }_2$-fold change on the median $\ge 0.5$, with 22 having a ${\log }_2$-fold change of $\le -0.5$. This translated to 20 parent chemicals meeting both criteria. Thirteen of these (3 metals, 1 organophosphorus insecticide, 5 personal care/consumer product chemicals, 3 phthalates, and 1 VOC) exhibited a decrease in exposure from the 2000s to the 2010s. These include chemicals that have some level of regulation or are generally recognized as potentially harmful, including lead, mercury, triclosan, and bisphenol A. The chemicals trichloroethene (an industrial solvent), deltamethrin (a pyrethroid insecticide), benzene (a major chemical intermediate that is also present in gasoline and other petroleum products), two phthalates (di-isononyl phthalate and dimethyl phthalate; plasticizers commonly used in a large variety of plastic items), and two polycyclic aromatic hydrocarbons (naproanilide and pyrene; chemicals that occur naturally in coal, crude oil, and gasoline) all exhibited an increase in exposure that met the two threshold criteria. The other 86 chemicals failed to meet the threshold criteria, with the majority of these (55 chemicals) meeting criteria 2 but not criteria 1 (in other words, the median exposure change was not small, but the uncertainty on those medians was too large to say a clear change in exposure occurred). This was most common for organophosphorus insecticide chemicals (true for 25 out of 36 chemicals).

Figure 7.

Decade comparison, 2000s vs. 2010s, for estimated exposure means of parent chemicals (in units of milligrams per kilogram bodyweight per day) inferred from urinary metabolite concentrations. NHANES cohorts were combined into two time periods, the 2000s decade (including cohorts 2001–2002, 2003–2004, 2005–2006, 2007–2008, and 2009–2010) and the 2010s decade (including cohorts 2011–2012, 2013–2014, and 2015–2016) using the 2-y survey weights. Error bars indicate 95% credible intervals (CrIs). Chemicals on the dotted identity line indicate no change in exposure across decades, above the dotted line indicates higher exposure in the 2010s (increase), and below indicates higher exposure in the 2000s (decrease). Chemical class was obtained for metabolites from the NHANES report and propagated to the parent chemicals using the parent–metabolite map. Chemicals were labeled if they met two criteria: a) The 95% CrIs around the exposure means did not overlap between decades, and b) the ${\log }_2$-fold change of the exposure means was either $\ge 0.5$ or $\le -0.5$. The corresponding data for this figure is covered by Excel Table S13. Note: NHANES, National Health and Nutrition Examination Survey.

Discussion

In this work, a Bayesian inference approach was used to estimate chemical intake rates from the NHANES continuous survey. Urinary metabolite concentrations from NHANES between the years of 1999 and 2016 were used to infer exposure to their parent chemicals using a toxicokinetic model and a mapping between parent chemicals and their known metabolic products. Our analysis resulted in what is essentially an exposure landscape for the US population. We first examined in a high-resolution manner by clustering chemicals with similar exposure trends. We identified chemical classes and uses enriched in these various clusters. We combined data from multiple NHANES cohorts to observe a more robust (in other words, less influenced by various issues/characteristics of the data) exposure change between two decades. We also observed specific trends for certain chemicals and population groups, many of which could seemingly be explained. We believe this provides a large-scale summary of exposure to the US population at the metabolite (151 chemicals) and parent chemical (179) level, spanning an 18-y time frame. These findings will be helpful for evaluating and benchmarking various exposure models, as well as understanding how exposures change over time.

Corroborating Observed Exposure Patterns

When making exposure predictions, it is important to validate those predictions by showing consistency with results from other estimation approaches or alignment with relevant literature sources. Owing to the large number of chemicals analyzed in this work, we cannot explain every exposure pattern observed. However, there are a number of chemicals that have sufficient information for comparison. We focused on the chemicals in Figures 4–6 owing to their distinct trends.

Figure 4 shows the exposure trends for four phthalates, a group of chemicals that have been studied extensively in the context of toxicity and exposure changes over time.^28,46,47 Phthalates are used as plasticizers in a wide range of consumer goods (e.g., food packaging, cosmetics, children’s toys), and are not strongly bound to the polymers they plasticize, which leads to leaching of the compounds into certain types of products.⁴⁸ Because phthalates are incorporated in such products, it is important to closely monitor the exposures of these chemicals over time because they may present an increased risk to certain populations, particularly children. Diethyl phthalate and DEHP both underwent exposure decreases in later cohorts (Figure 4A,B). The US Consumer Product and Safety Commission (CPSC) permanently prohibited the use of any children’s toy or child-care article containing concentrations of $>0.1\%$ of DEHP in 2008,⁴⁹ which is supported by the observed exposure decrease after the 2007–2008 cohort for these two phthalates, as well as for three known metabolites of DEHP: mono-(2-ethyl-5-oxyohexyl) phthalate, mono-(2-ethyl-5-hydroxyhexyl) phthalate, and mono-(2-ethyl-5-carboxypentyl) phthalate. Alternatively, di-isononyl phthalate and di-isobutyl phthalate (Figure 4C,D) exhibited exposure increases, with higher exposure in the later cohorts. The rise in dominance for these two phthalates has been found in other works using NHANES.^{28,46,47,50,51} These chemicals may have been phased into the production of plastics,⁵² including children’s toys or child-care articles (particularly di-isobutyl phthalate given that it showed markedly higher exposure in 6- to 11-y-olds), which would support the observed increases in exposure. However, the CPSC issued an update in 2017 to the list of prohibited phthalates⁵³ to include di-isononyl phthalate and di-isobutyl phthalate, suggesting that their exposure in more future NHANES cohorts will decrease.

As for parabens, the main trend characteristic was that women exhibited the highest exposures, whereas men exhibited the lowest. A similar result for pregnant women in Puerto Rico was observed by Ashrap et al.,⁵⁴ with higher paraben concentrations being associated with use of cosmetics and lotions. This specific exposure variation by demographic is interesting because, typically, most chemicals exhibit the highest exposure for children and lowest for individuals $\ge 66$ years of age or those having a BMI of $>30$.²⁶ Another characteristic was that for all four parabens except ethyl paraben, the greater exposure to female populations seemed to diminish in the later cohorts. This was also observed by Ashrap et al.⁵⁴ Such a phenomenon may be due to increased social awareness of the presence of parabens in products and their potential to be harmful.

In Figure 6 we showed chemicals with potential age-specific exposures. Cadmium and arsenobetaine showed higher exposure in older populations and lowest exposure in children. Cadmium had been heavily used for plating and in pigments, but its use for these purposes declined in the 1980s and 1990s.⁵⁵ It was then primarily used in batteries, with restrictions limiting the cadmium content.⁵⁶ Furthermore, reports on high levels of cadmium use in children’s jewelry in 2010 led to a CPSC investigation and subsequent recall notices for jewelry in select stores.^57,58 Therefore, the exposure patterns seen in Figure 6A (slow, marginal decrease in exposure) may be due to the decreasing presence of cadmium in various products, resulting in age-separated exposure trends, specifically the more pronounced decrease in exposure for 6- to 11-y-olds after the 2009–2010 cohort. As for arsenobetaine, we previously mentioned that it is commonly found in fish. A potential explanation for its observed exposure across population groups may be that children in the United States generally do not consume much fish in their diet. Alternatively, older individuals are often advised to eat more fish or take fish oil supplements for better heart health.⁵⁹ Another potential explanation for these two chemicals could be attributed to smoking, which has been found to be a major source of exposure for heavy metals.⁶⁰ Last, $N$, $N$-dimethylformamide and acrylonitrile are known to be involved in manufacturing processes, which might indicate that the population exposure differences for these two chemicals were a result of workplace exposures, given that exposure was highest for adults and males and generally lowest for children.

The most similar study, in terms of number of NHANES metabolites (141) and cohorts (8), to the work presented here was Nguyen et al.¹⁹ However, we calculate daily intake rates for parent chemicals, whereas Nguyen et al. focused on the metabolite concentration level and employed regression with an age-centered concentration metric. Nguyen et al. stratified the NHANES survey participants into 11 age groups compared with the 4 used in this work. These differences do not allow for a direct comparison of results. Nguyen et al. also focused on incorporating chemical half-life and evaluating the impact of restriction information in their analysis. The objectives of the work here were to characterize the coverage of chemicals by NHANES and categorize exposure trends in regard to chemical classes and population demographics.

Tracking Exposure in the Future

Exposure assessments help us understand how the changes in chemical use are affecting the general population’s exposures to various chemicals. Where changes in chemical use are the result of regulation or an increased perception of the risks posed by specific chemicals, such data provide insight on the effectiveness of risk assessment and risk management actions. Exposure estimates must be revised as NHANES and other biomonitoring efforts continue or begin. Obtaining exposure estimates for the chemicals presented in this work will require only small, partially manual updates to the input file for the bayesmarker package. Namely, the codes and weights tables to list the chemical, cohort, units, and file names. However, by referencing the chemical exposure inferences in Excel Table S8, obtaining and comparing the latest exposures will be relatively fast and simple. As biomonitoring data for new chemicals are introduced, updates to the parent–metabolite linkage map may be needed, but otherwise the process will be similarly straightforward.

Collection of Complimentary Data

Large-scale time-series data are quite limited in the field of exposure science.⁶¹ Data are often lacking in at least one of three dimensions: a) the time examined, b) the number of chemicals studied, and c) the diversity of the individuals making up the sample size. NHANES is a rare and valuable study because it examines dozens of the same metabolites in thousands of diverse individuals over multiple years, with 107 metabolites having data for at least four cohorts (Figure S1). Therefore, moving forward, it is important to collect data that will help us draw more distinctive conclusions from the inferred chemical exposures. For example, chemicals that are used in consumer products can change over time based on many factors including production cost, ease of use in manufacturing, or even government regulation on use limits. These factors are not as well known to the scientific community as they are to industry corporations, which can make validating observed changes in exposure difficult. Data that include more complete product ingredient information to help track how different chemicals are being phased in and out of common consumer products may explain increases or decreases in chemical exposure, potentially even for specific population groups (like decreased phthalate exposure in young children after their use in children’s toys was limited). Emerging technologies like nontargeted analysis (NTA)⁶² can help augment manufacturer reporting. Product ingredient data (like the US EPA’s Chemical and Products Database or CPDat⁶³) can work particularly well with consumer use or purchasing data (e.g., the Nielsen company survey described by Stanfield et al.⁶⁴) to record how much of specific products are being used over time. Last, a record of reporting and regulatory actions for various chemicals of concern by authorities in news, government, or regulatory bodies would allow for the development of models that estimate how the regulation of chemicals effects changes in exposure. Such a model could help explain future exposure patterns for chemicals of interest.

Managing Uncertainty

Inference of intake rate from urinary biomarker data is complicated by many factors, two of which were addressed here: a) issues related to the LOD in biomonitoring data, and b) complexity in chemical parent–metabolite relationships. The former contributor of uncertainty is based on the number of individuals in the biomonitoring study who have a concentration measurement below the LOD, see Figure S2. The more samples below the LOD, the less certain we can be in the estimated population distribution (specifically, the median and variance of concentration) for a chemical. The other major contributor to uncertainty, chemical metabolism, results from not always knowing what fraction of parent chemicals will undergo each potential metabolism pathway. For example, if parent chemical A is known to have two products, metabolite A.1 and A.2, there are two expected pathways of metabolism. However, given certain biological and environmental circumstances, the concentration of A.1 compared with that of A.2 may have a large range. Here, we assume that the fraction of parent A that metabolizes into A.1 plus the fraction of parent A that metabolizes into A.2 equals 1, but we do not know these fractions beforehand. We can however infer the fractions from the biomonitoring data. For example, if we see the concentration of A.1 is twice as much as A.2 and the only parent is A, then we assume 66% metabolizes into A.1 and 33% to A.2. With a Bayesian framework it would be possible to incorporate additional data apportioning fractions metabolized as prior knowledge, but these fractions can even vary between individuals. When there are multiple parent chemicals for a metabolite, the uncertainty to which we can estimate median exposure to each parent chemical increases significantly because any given parent could be contributing some fraction to the metabolite.

These two forms of uncertainty can also impact each other given that uncertainty with a metabolite’s concentration due to LOD issues will be propagated to its known parent chemical. Likewise, when a parent chemical has multiple metabolites, the likelihood that one of its metabolites has a fair amount of uncertainty increases, which will lead to higher uncertainty in the parent exposure distribution even if uncertainty in the other metabolites is relatively low. This is one potential explanation as to why fewer clusters and less enrichment were seen for the parent chemicals (multiple ways for high uncertainty to occur). Characterizing and keeping track of uncertainty is important for estimating ranges of exposure given that some individuals may be at greater risk owing to a number of factors (e.g., age, gender, race), but uncertainty introduces another caveat when looking at exposure trends specifically. That is higher uncertainty results in larger CrIs, which makes assessing the significance of changes in exposure over time more difficult because the CrIs across time points are more likely to span the observed differences in exposure. This was evident with the organophosphorus insecticides in the comparison between decades. Uncertainty was high for these chemicals because there were only a handful of metabolites in this class whose urine concentrations were being propagated to dozens of parent chemicals (see Stanfield et al.²⁶ for a detailed analysis on contributors to uncertainty in these exposure inferences). However, combining data across cohorts allowed for more confident assessments of exposure changes over coarser timescales for several other chemicals. It will be important to perform similar analyses as more NHANES cohorts are completed. We chose to approach uncertainty in two ways in this work. The first was by using a statistical framework that is good at incorporating and keeping track of uncertainty, Bayesian inference, and the second was by filtering out chemicals with particularly large uncertainties. Carefully quantifying uncertainty allows confidence that in some cases there are clear changes in chemical exposure occurring over time.

Limitations

Owing to the complicated problem of building a reverse-dosimetry approach that can be applied to many diverse chemicals, there are multiple potential limitations to the method implemented in this work. First, it relies on the biomonitoring study itself. The set of parent chemicals for which we can estimate exposure trends depend on which metabolites are measured by NHANES, consistent monitoring of metabolites across cohorts, and the quality of the data collected by NHANES (e.g., small sample size or fraction of measurements below the LOD). For example, it is important to note that $\sim 54\%$ of metabolites and 47% of parent chemicals were missing data from five or more NHANES cohorts. Due to numerous factors that contribute to exposure and individual metabolite concentrations, it is not possible to confidently state that the missing concentration from a cohort would fall near or between the concentrations reported in the surrounding cohorts. However, it is possible that chemicals with consistently low or steadily decreasing exposures were dropped by NHANES after this observation was made. This is an acceptable case of missing data given that those metabolites are very likely to be of low concern. Examples of this can be seen in Figure 2, such as for sulfonyl urea herbicides (second group from the bottom) and heterocyclic amines (fifth group from the top). These chemicals exhibited lower exposures in earlier cohorts and do not have data for the latest cohorts.

Another limitation stems from most biomonitoring studies employing a spot urine sampling procedure (a single urinary aliquot at a point in time) owing to the extensive time and resources needed to perform 24-h urine sampling on such a large scale. Given the many factors that influence exposure (e.g., different magnitudes, routes, timings of exposure scenarios), the same spot urine concentration could occur under multiple circumstances.⁹ This complexity, along with the intra-individual and interindividual variability in repeated spot urine samples and chemical elimination half-lives, means that biomonitoring data from spot urine sampling can only be used to accurately estimate the central tendency and not the tails of the exposure distribution, such as in highly exposed individuals at the 95th percentile.^26,65,66 Assuming steady-state conditions is another potential limitation. This assumption is not true for all chemicals, particularly those that are quickly metabolized. With enough high-throughput pharmacokinetic data available for chemicals used in biomonitoring studies, it will become possible to assign confidence in the inferred exposures in the context of factors such as frequency and duration of exposure, metabolism pathways, and excretion half-life. Together with sufficient chemical use information, it may allow the use of more chemical-specific toxicokinetic methods. Last, we combined 2-y cohorts into decades to observe longer-term exposure patterns. Choosing such a time grouping is somewhat arbitrary because human exposures change on much shorter timescales and can vary within the decades examined here, suggesting some information loss can occur when combining data in this way. However, these shorter fluctuations in exposure were extensively examined by the other analyses in this work. In addition, being able to identify metabolites with a marked change between decades indicates that the long-term exposure trend is greater than the within-decade variation, which adds insights and context to the other results in this work.

Application of Results

This work can be of use in several ways moving forward. Implications for it could include a) phasing out existing biomonitoring of metabolites exhibiting long-term consistency in exposure at values well below doses that are known to result in risk, b) phasing out of chemicals showing extended decreases in exposure over recent cohorts, c) continued monitoring of metabolites with increasing exposure trajectories, d) screening of new metabolites of current parent chemicals, or e) screening of metabolites of additional chemicals with exposures potentially relevant to public health. The inferred exposure estimates can also be used in conjunction with other exposure models to either calibrate high-throughput approaches [see the Systematic Empirical Evaluation of Models (SEEM) method⁷] or make comparisons in intake rates to characterize variability across approaches applied to the same chemicals. Last, the inferred exposure values from bayesmarker are in the same units (in intake rates of milligrams per kilogram per day) as most toxicity metrics, such as the no observed adverse effect level or can be converted to hazard-based metrics, such as bioactivity exposure ratios, using toxicokinetic methods.

Conclusions

A transparent, straightforward analysis pipeline was developed to generate time-series exposure data for environmental chemicals from a panel of urine biomarkers with concentration data. Using this pipeline, we were able to characterize overall exposure trends by building a time-series exposure landscape. Multiple groups of chemicals exhibiting similar exposures over time were identified in an unsupervised manner and attributed to different chemical classes and uses. For example, increasing exposure trends were observed for organophosphorus insecticides, phytoestrogens, and VOCs, whereas decreasing exposure trends were observed for personal care and consumer product chemicals, perchlorates, and flame retardants. These patterns in exposure were also corroborated for several chemicals based on regulations of those chemicals in the market (e.g., restrictions in products for phthalates and cadmium; reduced use of hazardous chemicals, such as lead, mercury, triclosan, and bisphenol A). Exposure differences by population were observed by plotting exposure trend lines for each demographic group. Higher exposures were seen for females, specifically reproductive-age females, to parabens; children, to certain phthalates; individuals $>65$ years of age, to cadmium; and adults and males, to chemicals used in manufacturing processes. Finally, we identified robust, longer-term exposure trajectories by examining the differences in intake rate for chemicals across two decades. By developing a chemical-independent reverse-dosimetry approach, we can make use of the vast data collected by NHANES to produce exposure estimates for hundreds of chemicals in the environment over time periods of several years, which provides multiple insights into the human exposure landscape.