Dietary characteristics associated with plasma concentrations of per- and polyfluoroalkyl substances among adults with pre-diabetes: cross-sectional results from the Diabetes Prevention Program Trial

Abstract

Diet is assumed to be the main source of exposure to per- and polyfluoroalkyl substances (PFAS) in non-occupationally exposed populations, but studies on the diet-PFAS relationship in the United States are scarce. We extracted multiple dietary variables, including daily intakes of food group, diet scores, and dietary patterns, from self-reported dietary data collected at baseline (1996–1999) from adults with pre-diabetes enrolled in the Diabetes Prevention Program, and used linear regression models to evaluate relationships of each dietary variable with plasma concentrations of six PFAS (perfluorooctane sulfonic acid (PFOS), perfluorooctanoic acid (PFOA), perfluorohexane sulfonic acid (PFHxS), 2-(N-ethyl-perfluorooctane sulfonamido) acetic acid (EtFOSAA), 2-(N-methyl-perfluorooctane sulfonamido) acetic acid (MeFOSAA), perfluorononanoic acid (PFNA) adjusting for covariates. Participants (N=941, 65% female, 58% Caucasian, 68% married, 75% with higher education, 95% nonsmoker) had similar PFAS concentrations compared to the general U.S. population during 1999–2000. Using a single food group approach, fried fish, other fish/shellfish, meat and poultry had positive associations with most PFAS plasma concentrations. The strongest effect estimate detected was between fried fish and PFNA [13.6% (95% CI: 7.7, 19.9) increase in median concentration per SD increase]. Low-carbohydrate and high protein diet score had positive association with plasma PFHxS. Some food groups, mostly vegetables and fruits, and the Dietary Approaches to Stop Hypertension diet score had inverse associations with PFOS and MeFOSAA. A vegetable diet pattern was associated with lower plasma concentrations of MeFOSAA, while high-fat meat and low-fiber and high-fat grains diet patterns were associated with higher plasma concentrations of PFOS, PFHxS, MeFOSAA and PFNA. We summarized four major dietary characteristics associated with variations in PFAS plasma concentrations in this population. Specifically, consuming more meat/fish/shellfish (especially fried fish, and excluding Omega3-rich fish), low-fiber and high-fat bread/cereal/rice/pasta, and coffee/tea was associated with higher plasma concentrations while dietary patterns of vegetables, fruits and Omega-3 rich fish were associated with lower plasma concentrations of some PFAS.

1. Introduction

Per- and polyfluoroalkyl substances (PFAS) are a group of synthetic organofluorine compounds with potential adverse health effects on obesity and diabetes risks (Cardenas et al. 2017; Cardenas et al. 2018; Liu et al. 2018). PFAS have desirable properties for industrial uses and can be used in many household products, such as non-stick cookware, food packaging, and stain repellants (Vestergren et al. 2008; Wang et al. 2014). Many PFAS have long biological half-lives, e.g., 2.7 years for perfluorooctanoic acid (PFOA); 5.3 years for perfluorohexane sulfonic acid (PFHxS) (Li et al. 2018). PFAS are ubiquitously detected in human blood samples; >95% of the U.S. population (12 years and older) sampled in National Health and Nutrition Examination Survey (NHANES) 1999–2016 had detectable serum concentrations of perfluorooctane sulfonic acid (PFOS), PFOA, PFHxS, and perfluorononanoic acid (PFNA) (Calafat et al. 2007a; Centers for Disease Control and Prevention 2019).

Diet has been considered the main source of PFAS exposure for the general population (Vestergren and Cousins 2009; Vestergren et al. 2008), among other routes of exposure, including air, house dust, and water (Domingo 2012; Fraser et al. 2013; Haug et al. 2011; Schwanz et al. 2016; Vestergren and Cousins 2009). The Canadian Total Diet Study estimated dietary exposure to PFOS at 250 ng/day, accounting for more than 60% of the total exposure (Tittlemier et al. 2007). Studies on the dietary exposure to PFAS are limited in U.S. populations (Domingo and Nadal 2017). Analyses using the 2003–2008 NHANES data found 10.4%, 15.75, 17.3%, 21.3% of the variations in serum concentrations of PFHxS, PFOS, PFNA, and PFOA, respectively, could be explained by dietary variables composed of 17 food groups (Jain 2014). Two other published studies on the association between diet and blood PFAS concentrations in the USA have only focused on fish consumption (Christensen et al. 2017; Hu et al. 2018). PFAS contamination in food could occur via contact with food packaging or cookware containing PFAS, or directly from bioaccumulation in aquatic and terrestrial food chains. Given the complex nature of the human diet, examining dietary patterns in addition to individual food items can represent a broader picture of food consumptions, and thus provide a better evaluation of dietary characteristics associated with variations in PFAS concentrations. To our knowledge, only one published paper described the association of dietary patterns with PFAS concentrations; Sjogren et al. examined serum concentrations of PFAS in Swedish elders (N=844) and found positive correlations with high adherence to a Mediterranean dietary pattern and negative associations with Word Health Organization (WHO) recommended diet and with a Low-Carbohydrate High-Protein dietary patterns (Sjogren et al. 2016). Evaluating different types of dietary variables (food intake and dietary pattern) in the same study can provide a multifaceted evaluation of dietary characteristics (Jacobs et al. 2009), and thus provide a more comprehensive description of the diet-PFAS relationship, which is important for PFAS research with diet-related health outcomes.

The goal of this study was to use multiple dietary variables to describe dietary characteristics and examine their associations with PFAS plasma concentrations. We hypothesized that fish and meat intakes would have a positive association with PFAS plasma concentrations, while a diet with higher fruit and vegetable consumption would be inversely associated with PFAS concentrations.

2. Methods

2.1. Study population

This analysis used baseline data collected from participants enrolled in a multi-center randomized controlled trial, the Diabetes Prevention Program (DPP, ClinicalTrials.gov number, NCT00004992). The time of recruitment was between July 1996 and May 1999. Inclusion criteria for the DPP study were adults with pre-diabetes who were at least 25 years old, had a body mass index (BMI) of 24 kg/m² or greater (22 or higher for Asians), and a serum glucose concentration of 95 to 125 mg/dL fasting and 140 to 199 mg/dL 2 h after a 75-gram oral glucose load. Participants were randomized into three treatment arms upon recruitment: lifestyle intervention, metformin, or medication placebo. A detailed description of the DPP study has been published previously (Diabetes Prevention Program Research Group 1999). For this cross-sectional analysis, we included participants from the lifestyle and placebo arms who had available PFAS plasma measurements and self-reported dietary intake at baseline (N=941, see Supplementary Material, Figure S1, for study flowchart). This study used de-identified data that did not contain any information on participants’ geographic location from the February 2008 Full Scale Data Release available from the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) Central Repository (Diabetes Prevention Program 2008) and was approved by the Institutional Review Board at Harvard Pilgrim Health Care. The involvement of the Centers for Disease Control and Prevention (CDC) laboratory did not constitute engagement in human subject’s research.

2.2. Dietary assessment

The DPP used a semi-quantitative food frequency questionnaire (FFQ) with 117 items to measure participants’ dietary intake during the previous year (Mayer-Davis et al. 2004). All FFQs were administered by trained and certified interviewers (Spanish bilingual interviewers were available). The DPP FFQ was developed based on a validated FFQ used in the Insulin Resistance Atherosclerosis Study (Mayer-Davis et al. 1999) and the National Cancer Institute’s Health, Habits, and History Questionnaire (HHHQ) (Block et al. 1986), had comparable validity and reproducibility across subgroups of ethnicity (non-Hispanic white, Hispanic, and African American), weight and diabetes status, education level, and urban versus rural community dwellers in the USA (Mayer-Davis et al. 1999). Participants reported consumption frequency (9 response categories) and portion size (3 response categories) for each item. The DPP Nutrition Coding Center (NCC) summarized daily intakes into six food pyramid groups (US Dept of Agriculture 1992) and further subcategorized them into 27 food groups based on an expanded list of food groups developed by the NCC (University of Minnesota 2019); the list took into account of the low-fat diet recommendation of the DPP lifestyle intervention (Diabetes Prevention Program 2008). Daily intakes of food groups were derived in units of servings per day; they included bread/cereal/rice/pasta (FG1–3), vegetable (FG4–7), fruit & fruit juice (FG8–9), dairy (FG10–11), fish (FG12–14), dried beans (FG15), meat (FG17–18), poultry (FG19–20), sweet & desserts (FG21), fats & oils (FG22), soy products (FG23), nuts and seeds (FG24), coffee and tea (FG25), meal replacements (instant breakfast/slimfast) (FG26), and alcohol (FG27), see the Supplementary Material (Table S1) for detailed description of food items in each of the food groups. Estimates of daily nutrient (macronutrients and micronutrients) and food intakes were done on HHHQ/DietSys software with nutrient values from the Nutrition Data System (NDS-R, Nutrition Coordinating Center, University of Minnesota, Minneapolis MN, Database version 2.6/8A/23). The estimated daily nutrient intake did not account for self-reported use of vitamins/supplements. Raw data of the 117 individual food items was not available for analysis via the NIDDK data repository.

We calculated a-priori diet scores based on food and nutrient intake levels, including Healthy Diet Indicator (HDI), Mediterranean-like diet score (MDS), Low-Carbohydrate and High Protein (LCHP), and Dietary Approaches to Stop Hypertension (DASH). A detailed description of diet score is presented in the Supplementary Materials (Appendix 1).

We used principal component analysis (PCA) with orthogonal (varimax) rotation to derive a-posteriori dietary patterns based on standardized daily intakes of the 27 food groups. Factor loadings (correlations between the derived PCA components and each input variable) greater than 0.2 were considered when describing dietary patterns. We used Scree plots to determine the appropriate number of components to select. For comparison, we also used a supervised method, reduced rank regression (RRR), to derive dietary patterns while considering variations of all PFAS plasma concentrations (Hoffmann et al. 2004).

2.3. Per- and Polyfluoroalkyl Substances Plasma Concentrations

We retrieved plasma samples stored at the NIDDK Central Repository for PFAS quantification using solid-phase extraction followed by high-performance liquid chromatography couple to isotope dilution-tandem mass spectrometry at CDC (Kato et al. 2011; Kato et al. 2018). The limit of detection (LOD) was 0.1 ng/mL for all analytes which included linear PFOS (n-PFOS); sum of perfluoromethylheptane sulfonic acid isomers (Sm-PFOS); linear PFOA (n-PFOA), sum of perfluoromethylheptanoic and perfluorodimethylhexanoic acids (Sb-PFOA); PFHxS, N-ethyl-perfluorooctane sulfonamido acetic acid (Et-PFOSA-AcOH; also known as EtFOSAA), N-methyl-perfluorooctane sulfonamido acetic acid (MePFOSA-AcOH; also known as MeFOSAA), and perfluorononanoic acid (PFNA). n-PFOS, Sm-PFOS, n-PFOA were detected in all plasma samples; for other isomers with nondetectable (i.e., <LOD) concentrations (16.6% for Sb-PFOA, 0.1% for PFHxS, 3.3% for EtFOSAA, 2.6% for MeFOSAA, and 6.8% for PFNA), we imputed concentrations <LOD with LOD/√(2)(Hornung and Reed 1990) for statistical analysis. We calculated total PFOS (PFOS=n-PFOS+Sm-PFOS) and total PFOA (PFOA=n-PFOA+Sb-PFOA) summing concentrations of branch and linear isomers, and imputed concentrations <LOD before summation; this method was consistent with the Fourth National Report on Human Exposure to Environmental Chemicals (Centers for Disease Control and Prevention 2019).

2.4. Covariates

We selected a priori potential confounders which may influence both self-reported dietary intake and PFAS plasma concentrations. They included age, sex, race/ethnicity, education level, household income, marital status/living arrangement, current smoking status, and waist circumference. All variables were categorical, except for sex which was binary (male/female) and waist circumference (continuous). Age was categorized into 5-year groups. Race/ethnicity categorization was based on the 1990 US census questionnaire including Caucasian, African American, Hispanic (of any race), and all other. Education level were based on year of school completed and categorized into <high school (elementary/junior high), high school/GED, college, and graduate school. Annual household income from all sources was reported in US dollars, responses included <$20,000, $20,000 to <$35,000, $35,000 to <$50,000, $50,000 to <$75,000, ≥$75,000, and refused to report. Marital status/living arrangement, categorized as married/cohabitating, single, divorced/separated, widowed, could potentially account for some variations in dietary behaviors of eating take-out and processed food. Current smoking status included noncurrent smokers (former smoker or nonsmoker) and current smoker. Waist circumference, measured in centimeters by research staff, was a proxy for body size. We used it instead of BMI because BMI was only available as categorical variable at baseline in the dataset. Other potential confounders considered but not available in the dataset included parity, breast feeding history, water consumption level, geographic location, and occupation. None of the covariates contained missing value for this analysis.

2.5. Statistical analysis

We performed analyses using SAS (version 9.4; SAS Institute Inc., Cary, NC, USA) and R (version 3.5). We summarized dietary intake using descriptive statistics in units of serving/day or standardized in per standard deviation (SD) increase per day; reported median and interquartile range (IQR) for PFAS plasma concentrations and compared them using Kruskal-Wallis test across participants’ characteristics. We calculated Spearman rank coefficients to examine the strength of correlations between 27 food groups and between PFAS. We checked for normality among continuous variables using a Shapiro-Wilk test. PFAS concentrations were skewed, and thus log (base 10) transformed for linear regression analyses.

We examined the bivariate relationship between each dietary variable (six pyramid food groups, 27 food groups, diet scores, and dietary patterns) and PFAS plasma concentrations (PFOS, PFOA, PFHxS, EtFOSAA, MeFOSAA, and PFNA) using separate univariate and multivariable linear regression models, accounting for potential confounders. We considered one dietary variable at a time and adjusted for multiple testing using the false discovery rate (FDR<0.2) approach to account for type I errors (Catelan and Biggeri 2010). We summarized coefficients from multiple models in a feature expression heat map for better visual depiction and comparison (Haarman et al. 2015; Lin et al. 2017a). Briefly, both exposure (food group) and outcomes (PFAS plasma concentrations) were standardized (as z-scores), and effect size and significance levels were graphically summarized using correlation plots.

We transformed the parameter estimates of linear regression model into relative percent change in median PFAS plasma concentrations for each SD increase in intake per day to provide a more useful interpretation of effect size using a method previously described (Barrera-Gomez and Basagana 2015). In these models, food group intakes were adjusted for daily calorie intake using the residual method (Willett et al. 1997). In all models, we tested for effect modification by age and sex using the method of adding a multiplicative interaction term of PFAS and covariate in the models and considered evidence for effect modification at P_interaction < 0.15.

All models were complete case analyses. We treated the income category “refused to report” (n=76, 8.1%) as a separate category, and also performed multiple imputation as a sensitivity analysis. We assessed the robustness of results by removing participants whose daily calorie intake were above the 95^th percentile or below the 5^th percentile.

3. Results

3.1. Study population

This sample of individuals with pre-diabetes (N=941) was composed of a majority of women (65%), Caucasian (58%), married or cohabitating (68%), college-educated or higher (75%), not current smokers (95%) (Table 1). The mean ± SD waist circumference was 104.7 ± 14.3 cm. The FFQ-estimated daily energy intake was 2091 ± 975 kcal. PFAS plasma concentrations (collected 1996–1999) were comparable to the concentrations among US adults in NHANES 1999–2000 (Calafat et al. 2007b; Centers for Disease Control and Prevention 2019). Median (IQR) concentrations in our study population were 32.6 (17.5–40.3) ng/mL (PFOS), 5.7 (3.5–6.7) ng/mL (PFOA), 3.6 (1.4–3.8) ng/mL (PFHxS), 1.9 (0.6–2.1) ng/mL (EtFOSAA), 1.3 (0.6–1.7) ng/mL (MeFOSAA), and 0.7 (0.4–0.8) ng/mL (PFNA) (Table 2). We detected differences in plasma concentrations of some PFAS across participants’ characteristics. For example, males had higher PFOS, PFOA, PFHxS, and PFNA compared to females; African Americans had higher PFOS and PFNA compared to other races and ethnicities; and persons with high school/GED education level had higher concentrations of PFOS and PFOA compared to the other groups, but the trend across education-level is non-monotonic (see Supplementary Material Table S2 for strata-specific concentrations). Strong correlations (r>0.6) were observed between PFOS and PFOA as well as PFOS and EtFOSAA (see Supplementary Material, Figure S2).

Table 1.

Demographic characteristics of study participants (N=941)

Characteristics (N=941)	N (%) or mean ± SD
Sex
Male	329 (35.0)
Female	612 (65.0)
Age
< 40	110 (11.7)
40–44	107 (11.4)
45–49	209 (22.2)
50–54	164 (17.4)
55–59	136 (14.5)
60–64	104 (11.1)
≥ 65	111 (11.8)
Race
Caucasian	545 (57.9)
African American	177 (18.8)
Hispanic of any race	177 (18.8)
All other	42 (4.5)
Marital status
Married/cohabitating	638 (67.8)
Single	110 (11.7)
Divorced/separated	150 (15.9)
Widowed	43 (4.6)
Education status
< High school	44 (4.7)
High school/GED	196 (20.8)
College	462 (49.1)
Graduate school	239 (25.4)
Income
<$20,000	117 (12.4)
$20,000 – <$35,000	165 (17.5)
$35,000 – <$50,000	191 (20.3)
$50,000 – <$75,000	181 (19.2)
≥ 75,000	211 (22.4)
Refused to answer	76 (8.1)
Smoking Status
Noncurrent smoker	889 (94.5)
Current smoker	52 (5.5)
Waist circumference (cm)a	104.7 ± 14.3
Daily calorie intake (kcal)	2091± 975

Note: SD=standard deviation.

^aSex specific mean ± SD waist circumference: 108.2 ±12.7 cm for men and 102.9 ± 14.8 cm for women.

Table 2.

PFAS plasma concentrations (in ng/mL) of study participants (N=941)

	Adults from Diabetes Prevention Program (N=941, 1996–1999)			U.S. population aged 40–59 years (N=295, NHANES 1999–2000) (Calafat et al. 2007b)
PFAS	Median (IQR)	Note	% below LOD	Geometric means (95% CI)
PFOS	32.6 (17.5, 40.3)	a,d,g	0%	30.4 (27.1–33.9)
PFOA	5.7 (3.5, 6.7)	a,b,c,d,e,g	0%	5.2 (4.7–5.7)
PFHxS	3.6 (1.4, 3.8)	a,b,c,d,g	0.1%	2.1 (1.9–2.4)
EtFOSAA	1.9 (0.6, 2.1)	a,b,d	3.3%	0.6 (0.6–0.7)
MeFOSAA	1.3 (0.6, 1.7)	b,d,e,g	2.6%	1.0 (0.8–1.1)
PFNA	0.7 (0.4, 0.8)	a,b,d,e,f,g	6.8%	0.5 (0.5–0.7)

Note: PFOS: Perfluorooctane sulfonic acid [sum of linear and branched isomers]; PFOA: perfluorooctanoic acid [sum of linear and branched isomers]; PFHxS: perfluorohexane sulfonic acid; EtFOSAA: N-ethyl-perfluorooctane sulfonamido acetic acid; MeFOSAA: N-methyl-perfluorooctane sulfonamido acetic acid; PFNA: perfluorononanoic acid.

Group means of PFAS plasma concentrations significantly differed by sex^a, age^b, marital status^c, education^d, income^e, smoking status^f, and race/ethnicity^g (Kruskal–Wallis one-way analysis of variance, p<0.05). See the Supplementary Material Table S2 for strata-specific concentrations.

3.2. Food groups

We visually summarized associations between food groups and the six PFAS plasma concentrations in Figure 1 and presented specific effect estimates in the Supplementary Material (Table S3). High intakes of vegetables, fruit and fruit juice, dairy (low fat), high Omega-3 fish, dried beans, soy products, and nuts/seeds were associated with lower concentrations of PFOS and MeFOSAA while high intakes of bread/cereal/rice/pasta (low-fiber and high-fat), fried fish, other fish/shellfish, meat (high fat), meat (low fat), poultry (high fat), poultry (low fat), sweet/dessert, coffee/tea were associated with higher concentrations of most PFAS (Figure 1). The strongest positive associations came from fried fish [median concentrations of PFNA 13.6% (95% CI: 7.7, 19.9) higher per SD (0.1 serving/day) increase] and other fish/shellfish [median concentrations of PFNA 19.1% (95% CI: 13.0, 25.4) per SD (0.2 serving/day) increase] while cruciferous vegetables (such as broccoli, cauliflower, Brussels sprouts) [median concentration of plasma MeFOSAA 11.5% lower (95% CI: −16.1, −6.6) per SD (0.3 serving /day) increase] and dairy (low fat, such as yogurt and low-fat milk) [median concentration of plasma EtFOSAA 11.8% lower (95% CI: −17.6, −5.5) per SD (1.2 serving /day) increase] had the strongest negative associations. We did note some differences in associations between separate PFAS with certain food groups, for example, low fat dairy had a positive association with PFHxS but inverse associations with other PFAS (Figure 1). We did not observe significant effect modification by sex or age. All significant findings remained consistent after removing intake levels above 95% and below 5%.

Figure 1.

Feature expression heatmap on associations of 27 food groups and 6 pyramid food groups (G1-G6) with PFAS plasma concentrations (N=941).

Note: PFOS: Perfluorooctane sulfonic acid [sum of linear and branched isomers]; PFOA: perfluorooctanoic acid [sum of linear and branched isomers]; PFHxS: perfluorohexane sulfonic acid; EtFOSAA: N-ethyl-perfluorooctane sulfonamido acetic acid; MeFOSAA: N-methyl-perfluorooctane sulfonamido acetic acid; PFNA: perfluorononanoic acid.

M1 represents univariate linear regression model with food group as the independent variable and PFAS plasma concentration as the dependent variable. M2 represents multivariate linear regression model adjusting for sex, age, marital status, education, income, waist circumference, and smoking status. M3 represents M2 with additional adjustment for daily caloric intake. Effect size represents change in PFAS plasma concentration (in standardized z-score) per standard deviation increase in daily intake and are depicted by color intensity (blue for negative effect and red for positive effect). Size of the circles indicates statistical significance (p-value) and circles with a dot in the middle represent significance below the 0.2 false discovery rate (FDR).

3.3. Fish intake

Fried fish and other fish/shellfish (trout, sole, halibut, poke, grouper, shrimp, lobster, crab, oysters, mussels, etc.) had positive associations with plasma concentrations of PFOS, PFHxS, and PFNA (Supplementary Material, Table S3). For example, each SD (0.1 serving/day) greater intake of fried fish was associated with 13.6% (95% CI: 7.7, 19.9, p<0.01) higher median plasma concentration of PFNA and 5.5% (95% CI: 1.1, 10.2, p<0.01) higher PFOS. Additional adjustments of proxies for healthy (HDI score and fiber intake) and less-healthy (saturated fat intake) eating habits in multivariable linear regression models did not change parameter estimates significantly.

We found a different direction of association for high Omega-3 fish (salmon, tuna, sardine) compared to the other two fish food groups (fried fish and other fish/shellfish), even though intakes of all three fish groups were positively correlated (r=0.15 to 0.30, p<0.01) (Figure S3). Greater intake of high Omega-3 fish was associated with lower rather than higher PFAS plasma concentrations [e.g. 6.2% (95% CI: −9.6, −2.8, p<0.01) lower median PFOA concentration per SD (0.1 serving/day) increase in intake], after accounting for participant characteristics (Supplementary Material, Table S3). The effect estimates did not change significantly with additions of proxies for healthy and less-healthy dietary habits in the models.

3.4. Diet scores

We summarized associations between diet scores and PFAS plasma concentrations in Table 3. Higher HDI was associated with lower PFAS concentrations in univariate but not multivariate analysis. MDS did not have clear pattern of associations with plasma concentrations of PFAS, and all were null after accounting for baseline characteristics. LCHP (score range 2–20) had positive association with PFHxS; 0.68% (95% CI: 0.17, 1.20) increase in median concentration per LCHP score increase (Table 3). DASH diet score (range 0–9) had inverse associations with plasma concentration of PFOS (−1.57%), PFOA (−1.04%), EtFOSAA (−3.33%), and MeFOSAA (−3.36%) (Table 3). We did not observe significant effect modification by sex or age.

Table 3.

Percent relative difference (95% confidence interval) in the median of plasma PFAS concentrations per one unit increase in diet scores

	Percent relative difference (95% CI) in median plasma concentrations of PFAS
Diet scores	PFOS	PFOA	PFHxS	EtFOSAA	MeFOSAA	PFNA
Crude modela
HDI	−0.77 (−2.14, 0.61)	−0.45 (−1.67, 0.79)	−0.04 (−1.82, 1.77)	−2.12 (−4.26, 0.05)	−2.37 (−4.17, −0.54)*	0.27 (−1.51, 2.07)
MDS	−0.43 (−1.40, 0.56)	0.18 (−0.70, 1.06)	1.61 (0.33, 2.91)*	0.19 (−1.37, 1.77)	−0.45 (−1.76, 0.87)	0.85 (−0.42, 2.13)
LCHP	−0.18 (−0.58, 0.23)	0.03 (−0.33, 0.39)	0.56 (0.03, 1.08)*	−0.30 (−0.94, 0.34)	−0.02 (−0.56, 0.53)	0.07 (−0.44, 0.60)
DASH	−2.08 (−3.15, −0.99)*	−0.72 (−1.70, 0.27)	1.23 (−0.22, 2.69)	−3.96 (−5.63, −2.27)*	−4.70 (−6.08, −3.29)*	−0.43 (−1.85, 1.00)
Adjusted modelb
HDI	−0.72 (−2.11, 0.69)	−0.51 (−1.74, 0.73)	−0.87 (−2.59, 0.89)	−1.45 (−3.65, 0.79)	−0.83 (−2.64, 1.01)	−0.39 (−2.08, 1.33)
MDS	−0.37 (−1.35, 0.62)	−0.05 (−0.92, 0.83)	0.78 (−0.46, 2.03)	0.54 (−1.04, 2.15)	−0.30 (−1.59, 0.99)	0.81 (−0.40, 2.03)
LCHP	−0.07 (−0.48, 0.34)	−0.02 (−0.38, 0.34)	0.68 (0.17, 1.20)*	−0.40 (−1.05, 0.25)	−0.27 (−0.81, 0.26)	0.42 (−0.08, 0.92)
DASH	−1.57 (−2.71, −0.41)*	−1.04 (−2.05, −0.01)*	0.83 (−0.63, 2.32)	−3.33 (−5.11, −1.51)*	−3.36 (−4.81, −1.88)*	0.08 (−1.34, 1.52)

Note: HDI: healthy diet index; MDS: Mediterranean Diet Score; LCHP: Low Carbohydrate and High Protein; DASH: Dietary Approaches to Stop Hypertension (see Supplementary Material, Appendix 1 for detail description of diet scores); PFOS: Perfluorooctane sulfonic acid [sum of linear and branched isomers]; PFOA: perfluorooctanoic acid [sum of linear and branched isomers]; PFHxS: perfluorohexane sulfonic acid; EtFOSAA: N-ethyl-perfluorooctane sulfonamido acetic acid; MeFOSAA: N-methyl-perfluorooctane sulfonamido acetic acid; PFNA: perfluorononanoic acid.

^aUnivariate linear regression models with log-transformed plasma PFAS concentrations as the dependent variable and diet scores as the independent variable

^bMultivariable linear regression models with additional adjustment for sex, age, marital status, education, income, smoking status, and waist circumference

^*p<0.05

3.5. Dietary patterns

PCA derived 10 dietary patterns, explaining 60.8% of the variance in intake (see Figure 2 for all 10 patterns). The first component (vegetable diet, PCA1) had high positive loadings on vegetables, fruits and fruit juice, and explained 11.3% of the variance. Other notable dietary patterns included high-fat meat diet (PCA2, 7.8% variance), lean poultry and fish diet (PCA3, 7.2% variance), and low-fiber and high-fat grains diet (PCA6, 5.4% variance). Vegetable diet (PCA1) was inversely associated with plasma concentrations of all PFAS, but the effect estimate was only statistically significant with MeFOSAA [−4.13% (95% CI: −6.39, −1.81), Figure 2]. High-fat meat diet (PCA2) had positive association with PFOS, PFHxS, MeFOSAA, and PFNA and low-fiber and high-fat grains diet (PCA6) had positive associations with PFOS PFOA, and EtFOSAA. Lean poultry and fish diet (PCA3) was associated only with higher PFNA but no other PFAS. We did not observe significant effect modification by sex or age.

Figure 2.

Associations between dietary patterns and PFAS plasma concentrations among prediabetic adults in the Diabetes Prevention Program (N=941). (A) Loading values for factor score of dietary patterns derived using Principal Component Analysis; (B) Relative percent change in median PFAS plasma concentrations per standard deviation increase in factor score.

Note: PFOS: Perfluorooctane sulfonic acid [sum of linear and branched isomers]; PFOA: perfluorooctanoic acid [sum of linear and branched isomers]; PFHxS: perfluorohexane sulfonic acid; EtFOSAA: N-ethyl-perfluorooctane sulfonamido acetic acid; MeFOSAA: N-methyl-perfluorooctane sulfonamido acetic acid; PFNA: perfluorononanoic acid.

The top 10 PCs explained 60.8% in the variances of baseline self-reported intakes. Loading value were multiplied by 100 and rounded to the nearest integer. Values greater than 40 are colored in dark blue. Effect estimates statically differed from 0 were marked with * (p < 0.05).

3.6. Alternative approach to describe dietary characteristics and sensitivity analyses

Reduced rank regression further supported our findings. Specifically, the first dietary pattern (high-fat meat diet, RRR1), which explained the most variances of both dietary intakes and PFAS plasma concentrations, had main loadings on food items which had strong associations found in the individual food group analysis. Additionally, there was high resemblance between several dietary patterns derived from PCA and RRR, i.e., high-fat meat diet (PCA2 and RRR1) and low-fiber and high-fat grain diet (PCA6 and RRR2), see Supplementary Material Figure S5 for details. The findings were robust to outliers; statistically significant associations remained consistent after removing participants whose daily calorie intake were above the 95th percentile or below the 5th percentile. We observed consistent findings between the complete case analysis and the multiple imputation approaches.

4. Discussion

In this population of 941 adults with pre-diabetes enrolled in the Diabetes Prevention Program (1996–1999) who had PFAS plasma concentrations comparable with those in the U.S. general population (NHANES 1999–2000), we found several dietary variables significantly associated with PFAS plasma concentrations, even after accounting for baseline characteristics. Putting results from different dietary variables (nutrients, food group, diet scores, and dietary patterns) together, we found four distinct dietary characteristics that were associated with PFAS concentrations.

First, participants with high consumption of meat, fried fish, and other fish/shellfish (but not Omega-3 rich fish) had higher plasma concentrations of PFOS, PFHxS, MeFOSAA and PFNA (3–5% per SD change in intake). Exposure associated with these foods likely comes from PFAS contamination in the food themselves. Food groups of animal origin, especially seafood and shellfish, have been identified as the main dietary source of PFAS exposure in other populations (Carlsson et al. 2016; Clarke et al. 2010; Haug et al. 2010; Squadrone et al. 2015; Tittlemier et al. 2007). Sampling of 91 food samples in 2017 as part of the Total Diet Study in the USA showed detectable levels of PFOS in pork, beef, chicken and fish samples (US Food and Drug Administration 2019). Several studies had shown that self-reported meat and fish intakes were associated with plasma concentrations of PFOS, PFOA, PFHxS, and PFNA, as well as other PFAS not examined in our study, including perfluorodecanoic acid, perfluoroundecanoic acid, and perfluorododecanoic acid (Christensen et al. 2017; Halldorsson et al. 2008; Jain 2014; Liu et al. 2017; Tian et al. 2018). We also noted that LCHP had positive associations with plasma PFHxS concentration.

Second, we noticed participants with high intake of vegetables, fruit and high Omega-3 fish (salmon/tuna) had slightly lower plasma concentration of some PFAS (1–5% lower per SD increase in intake). Inverse associations with DASH diet score, fiber, vitamin C and carotenes intakes further supported this finding (Supplementary Material, Figure S4). Even though dietary exposure assessments in Europe detected some PFAS in vegetables samples (European Food Safety Authority 2018), our finding showed that participants who ate more vegetables did not have higher PFAS body burden, which is consistent with some previous literature (Herzke et al. 2013). Studies among pregnant women in Norway (N=1076) and China (N=981) also found inverse associations between vegetable intake and PFAS concentrations (Halldorsson et al. 2008; Tian et al. 2018). Though not statistically significant after adjustment for baseline demographics, study participants with higher HDI diet score, i.e. stronger adherence to the WHO Dietary Recommendation Guideline (WHO 1991), had lower PFAS plasma concentrations (see Appendix 1 in Supplementary Material for the detail dietary guideline). Diets characterized by greater consumption of plant-based foods, low saturated fatty acid intake, high dietary fiber intake, and low intakes of refined carbohydrate have many health benefits, including protection against cardiovascular diseases, cancer and overall mortality (McEvoy et al. 2012). Additional studies are needed to evaluate if there is a beneficial effect on body burden of PFAS.

The third dietary characteristic, mainly described by PCA6 and high intakes of bread/cereal/rice/pasta (low-fiber and high-fat), which included fried rice, baked goods, enchiladas, tacos, nachos, corn bread, crackers, potato chips, corn chips, popcorns, and sweets/dessert, was associated with higher PFOS, PFOA and EtFOSAA plasma concentrations (2–5 % per SD increase in intake). PFAS contaminations in the ingredients of these food items were likely low—based on data from Europe (2010–2016), only 0.4% of grains and grain-based products and 2% of starchy roots and tubers had detectable PFOA and PFOS (European Food Safety Authority 2018), and the relative contribution from starchy roots and tubers to overall PFAS exposure was only 0.2–2.1% (European Food Safety Authority 2012). Additionally, because the bread/cereal/rice/pasta (low-fiber and high-fat) food group did not have strong correlations with other food groups, the observed positive associations with variations in PFAS plasma concentrations may likely be a result of contact with food packaging and of the preparation process. Birth cohort studies in Korea, Norway, Denmark, and Canada reported the frequency of eating snack, junk foods (microwave popcorn, potato chip), dining out, ordering take-outs, and eating heated packed foods had positive associations with PFAS concentrations measured in serum, plasma, or breastmilk (Averina et al. 2018; Halldorsson et al. 2008; Hu et al. 2018; Lee et al. 2018). These food items are often prepared, packaged and served in containers that may be coated with PFAS (Begley et al. 2008), though the specific PFAS used might have changed over the years. Chemical migration test of food contact materials found detectable levels of dialkyl polyfluorinated phosphate ester surfactants, precursor chemicals that can be transformed to PFOA, in microwave popcorn bags (0.2–1.4 mg/kg food) and burger box (Begley et al. 2008; Trier et al. 2011). Risk assessments on PFOA showed fast food consumption and food packaging containing PFAS were influential parameters for overall exposure (Vestergren et al. 2008). In the DPP study, information on fast food consumption and use of food packaging was not available, as PFAS exposure was not part of the original DPP study aims. Nevertheless, our data suggested that intake of these food items could be used to develop a proxy that describe this consumption behavior.

Last, we found that the food group, coffee/tea, which did not have significant correlation with other food groups, had positive association with PFOA plasma concentration (5% higher per SD increase in intake). Coffee/tea intake contributed strong loadings to several dietary patterns, especially PCA7 and RRR4, which also had positive association with PFOA plasma concentrations. PFOA exposure from drinking coffee/tea could come from (1) ingredients themselves (water, coffee bean, tea leaves, milk/creamer), (2) the brewing and preparation process, or (3) contact with the container/cups. It is unlikely that this association is driven by intake of milk and creamer as they were considered in the dairy food group of our analysis and showed inverse rather than positive associations. PFOA and other PFAS can be used in the polystyrene coating of paper cups, but the amount transferrable to food was likely low (Trier et al. 2011). A controlled experiment conducted in the Netherlands showed that coffee (N=12) brewed with coffee machine and tap water (3.7 ± 0.45 ng/L PFOA) contained 4.4 ± 3.3 ng/L of PFOA (Eschauzier et al. 2013). The study found no significant leaching of PFOA from the coffee machine and concluded tap water and coffee beans to be the main source of PFAS. Water consumption can contribute significantly to PFAS exposure, especially in areas with contaminated water sources (Sunderland et al. 2018). Our study did not capture water consumption information, thus we were unable to evaluate the extent of contribution from water intake. NHANES (2005–2010) data suggested coffee/tea accounted for more than a quarter of the total water consumption among adults >50 years old in the USA (Drewnowski et al. 2013).

Our study comprehensively describes the relationship between diet and PFAS plasma concentrations among a prediabetic population in the USA. Results were consistent with previous findings from Singapore, Korea, China, Denmark, and Spain (Brantsaeter et al. 2013; Halldorsson et al. 2008; Lee et al. 2018; Liu et al. 2017; Tian et al. 2018). However, we were not able to replicate the previously observed positive associations of HDI and MDS with PFAS plasma concentrations (Sjogren et al. 2016), possibly because of differences in study population and the composition of diet scores. Specifically, PFAS exposure patterns for U.S. middle-aged adults (1996–1999) could be different than those of Swedish elders above 70 years of age (2001). In general, fish consumption per capita is higher in Northern Europe compared to the USA (Guillen et al. 2019). In addition, Sjogren’s paper additionally accounted for fish intake in HDI which our paper did not include as it was not part of the WHO’s dietary recommended guideline. Our finding on fish consumption was consistent with other studies in the USA (Christensen et al. 2017; Hu et al. 2018), and supports the hypothesis that higher intake of fish/shellfish is associated with higher plasma concentration of some PFAS, with the exception of fish high in Omega-3 fatty acids. The inverse association, also noted by the salmon intake from the 2007–2014 NHANES data (Christensen et al. 2017), could be related to a healthier dietary pattern with higher vegetable and fruit consumption, but more study is warranted. While HDI did not show significant association with plasma concentration of the six PFAS examined, we did note significant inverse relationships of DASH diet score with PFOS, PFOA, EtFOSAA and MeFOSAA. Both DASH and HDI diet emphasize vegetables, fruits and low-fat dairy foods, but the calculation of the DASH diet score uses stricter cut-offs for saturated fat and cholesterol intakes, a higher percent of energy from protein, and includes intakes of sodium, calcium, potassium, and magnesium (see Appendix 1 in the Supplementary Material for detail). We speculated that the inclusion of sodium, calcium, potassium and magnesium intakes in the DASH diet score could potentially provide a better capture of the variation in the dining behavior of eating processed food versus fresh whole food compared to HDI, thus resulting in different associations observed. Analyses on nutrient intake did show sodium intake had a positive association, while calcium, potassium and magnesium intakes had negative associations with these four PFAS (Supplementary Material, Figure S4).

We also noted certain PFAS were associated with specific food sources, i.e., PFHxS and PFNA with fish/shellfish and meat, and PFOA with coffee/tea, reaffirming PFAS composition profiles in blood may provide valuable information on identifying sources of exposure as suggested by previous literature (Hu et al. 2018). However, comparisons should be done across similar time periods and demographic groups, as mixtures of PFAS could be highly dynamic across time and regions.

Another strength of our study is the utilization of multiple approaches to evaluate dietary characteristics. Diet is an important source of exposure to many environmental chemicals and toxicants, such as PFAS, mercury, polychlorinated biphenyls and arsenic. However, environmental epidemiologists typically pay relatively limited attention on studying chemical-diet relationship, mainly because of the complexity of capturing dietary information in large population studies. Diet scores are useful for comparing diet to known dietary patterns or dietary recommendation guidelines, but they have limited ability to identify exposure sources as compositions of diet scores must be pre-specified. Our study demonstrated the concurrent use of single food group and dietary pattern analysis provided corroborating information to evaluate the diet-chemical relationship. Food group analysis allows for exploration of potential sources of exposure. Dietary patterns analysis summarizes and reduces food groups into meaningful patterns which could reflect dietary behaviors that may influence chemical exposure pathways. Newer exposome methods, such as high-resolution metabolomics, could provide more information on diet compositions and individual variations in metabolism. As demonstrated in our study, diet can be associated with PFAS plasma concentrations in either positive or negative directions, future studies of PFAS with diet relevant health outcomes should account for confounding by diet.

Our study has some limitations. First, we were not able to accurately account for PFAS exposure from water sources. Many sites in the USA are known to be affected by PFAS contamination in drinking water (Hu et al. 2016) and a previous study showed that PFAS level in water is a significant predictor for plasma concentrations of PFAS among individuals who consume ≥8 cups of water per day (Hu et al. 2019). However, because of privacy protection concerns, the NIDDK data repository could not provide participants’ geographic location and we did not have either an accurate measure of participants’ water consumption habits. We recognize that water may be a greater contributor to plasma PFAS concentrations relative to food in areas with PFAS contamination in the water source, however, our study was not able to evaluate the specific contribution of drinking water. Second, we also did not have information on occupation or reproductive history (parity and lactation) which could result in some residual confounding. Third, FFQs are not accurate in estimating absolute intake and are at risk of recall bias, although many studies showed FFQs can still provide adequate ability to rank relative intakes (Lin et al. 2017b; Molag et al. 2007; Willett and Lenart 2012). Intake levels estimated by the DPP FFQ were consistent with other published data in the USA and were correlated with levels measured by other dietary instruments that do not require memory recall (r=0.5) (Mayer-Davis et al. 1999). We also attempted to control for measurement errors in reporting by statistical adjustment for total energy using the residual method. Even though the pre-specified food list in the FFQ might not comprehensively reflect the specific eating patterns of some participants in the study, the DPP FFQ had been tested to be valid and reliable across a culturally diverse group in the USA (Mayer-Davis et al. 1999). Last, industrial production and population-level body burden of specific PFAS compounds have changed over time. For example, US manufactures initiated phase-out of legacy PFAS (e.g., PFOS, PFOA) since the 2000s but shifted production to other PFAS (i.e., GenX, short-chain PFAS) (Gomis et al. 2017; Sunderland et al. 2018; Wang et al. 2017). Therefore, generalizability of our study findings is limited by the study population, study period, and type of PFAS examined; more recent biomonitoring data would be useful to evaluate current diet-PFAS relationship.

5. Conclusion

We performed a comprehensive assessment on the relationship between diet and PFAS plasma concentrations among adults with pre-diabetes in the USA. Dietary sources of PFAS exposure might originate from bioaccumulation in the food chain and transfer from contact materials used in food processing and packaging. We demonstrated using multiple approaches to characterize diet was useful in identifying potential exposure pathways in a large prospective study. Diet could influence body burden by increasing exposure dose and/or potentially by affecting absorption and elimination. Specifically, we found consuming more meat/fish/shellfish (especially fried fish, and excluding Omage3-rich fish), low-fiber and high-fat bread/cereal/rice/pasta, and coffee/tea was associated with higher plasma concentration of certain PFAS while a diet of vegetables, fruits and Omega-3 rich fish was associated with lower plasma concentrations of other PFAS.

Highlights.

• We evaluated dietary characteristics among 941 prediabetic adults in the USA

• Vegetable diet pattern had inverse association with plasma MeFOSAA concentration

• Meat/fish/shellfish pattern had positive associations with most plasma PFAS

• Low-fiber and high-fat grains intake had positive associations with some PFAS

• Coffee/tea consumption had positive association with plasma PFOA concentration

Abbreviations

CDC

Centers for Disease Control and Prevention

CI

confidence interval

DASH

Dietary Approaches to Stop Hypertension

DHA

docosahexaenoic acids

DPP

Diabetes Prevention Program

DPPOS

Diabetes Prevention Program Outcome Study

EPA

eicosapentoenoic acid

EtFOSAA

2-(N-Ethyl-perfluorooctane sulfonamido) acetic acid

FDR

false discovery rate

FFQ

food frequency questionnaire

HDI

Healthy Diet Indicator

HHHQ

Health, Habits, and History Questionnaire

IQR

interquartile range

LCHP

Low-Carbohydrate and High Protein

LOD

limit of detection

MDS

Mediterranean-like diet score

MeFOSAA

2-(N-Methyl-perfluorooctane sulfonamido) acetic acid

NHANES

National Health and Nutrition Examination Survey

NIDDK

National Institute of Diabetes and Digestive and Kidney Diseases

n-PFOA

linear PFOA

n-PFOS

linear PFOS

PCA

principal component analysis

PFAS

per- and polyfluoroalkyl substances

PFHxS

perfluorohexane sulfonic acid

PFNA

perfluorononanoic acid

PFOA

perfluorooctanoic acid (sum of linear and branched isomers)

PFOS

perfluorooctane sulfonic acid (sum of linear and branched isomers)

RRR

reduced rank regression

Sb-PFOA

sum of perfluoromethylheptanoic and perfluorodimethylhexanoic acids

SD

standard deviation

Sm-PFOS

sum of perfluoromethylheptane sulfonic acid isomers

WHO

Word Health Organization