Characterizing Important Dietary Exposure Sources of Perfluoroalkyl Acids in Inuit Youth and Adults in Nunavik Using a Feature Selection Tool.

BACKGROUND: Previous studies have identified the consumption of country foods (hunted/harvested foods from the land) as the primary exposure source of perfluoroalkyl acids (PFAA) in Arctic communities. However, identifying the specific foods associated with PFAA exposures is complicated due to correlation between country foods that are commonly consumed together.

METHODS: We used venous blood sample data and food frequency questionnaire data from the Qanuilirpitaa? (“How are we now?”) 2017 (Q2017) survey of Inuit individuals ≥16 y of age residing in Nunavik (𝑛= 1,193). Adaptive elastic net, a machine learning technique, identified the most important food items for predicting PFAA biomarker levels while accounting for the correlation among the food items. We used generalized linear regression models to quantify the association between the most predictive food items and six plasma PFAA biomarker levels. The estimates were converted to percent changes in a specific PFAA biomarker level per standard deviation increase in the consumption of a food item. Models were also stratified by food type (market or country foods).

RESULTS: Perfluorooctanesulfonic acid (PFOS), perfluorodecanoic acid (PFDA), and perfluoroundecanoic acid (PFUnDA) were associated with frequent consumption of beluga misirak (rendered fat) [14.6%; 95% confidence interval (CI): 10.3%, 18.9%; 14.6% (95% CI: 10.1%, 19.0%)], seal liver [9.3% (95% CI: 5.0%, 13.7%); 8.1% (95% CI: 3.5%, 12.6%)], and suuvalik (fish roe mixed with berries and fat) [6.0% (95% CI: 1.3%, 10.7%); 7.5% (95% CI: 2.7%, 12.3%)]. Beluga misirak was also associated with higher concentrations of perfluorohexanesulphonic acid (PFHxS) and perfluorononanoic acid (PFNA), albeit with lower percentage changes. PFHxS, perfluorooctanoic acid (PFOA), and PFNA followed some similar patterns, with higher levels associated with frequent consumption of ptarmigan [6.1% (95% CI: 3.2%, 9.0%); 5.1% (95% CI: 1.1%, 9.1%); 5.4% (95% CI: 1.8%, 9.0%)]. Among market foods, frequent consumption of processed meat and popcorn was consistently associated with lower PFAA exposure.

CONCLUSIONS: Our study identifies specific food items contributing to environmental contaminant exposure in Indigenous or small communities relying on local subsistence foods using adaptive elastic net to prioritize responses from a complex food frequency questionnaire. In Nunavik, higher PFAA biomarker levels were primarily related to increased consumption of country foods, particularly beluga misirak, seal liver, suuvalik, and ptarmigan. Our results support policies regulating PFAA production and use to limit the contamination of Arctic species through long-range transport.

Perfluoroalkyl acids (PFAAs), a subset of per- and polyfluoroalkyl substances (PFAS), are environmentally persistent chemicals used in various consumer products and industrial processes for their oil- and water-repelling properties. Despite the lack of any known major contamination sites in northern communities, PFAAs have been measured at highly elevated concentrations in Arctic communities,[1-3] deeming PFAAs as chemicals of emerging Arctic concern (CEACs).[4] PFAAs and their precursors reach the Arctic via long-range atmospheric and oceanic transport from southern latitudes.[2] In Arctic conditions, these chemicals accumulate and biomagnify in marine and, to a lesser extent, terrestrial food webs,[5],[6] including species regularly consumed by Inuit.

Due to international regulations and voluntary phaseouts, concentrations of perfluorooctane sulfonic acid (PFOS) and perfluorooctanoic acid (PFOA) have decreased in Inuit communities in Nunavik, northern Quebec since 2004.[3],[7] However, the concentrations of longer-chain PFAAs-perfluorononanoic acid (PFNA), perfluorodecanoic acid (PFDA), and perfluoroundecanoic acid (PFUnDA)-are increasing.[3] The Nunavik population is exposed to higher PFAA concentrations than the general Canadian population.[1] For example, PFNA concentrations were 7-fold higher in a population-representative sample of Inuit uQanuilirpitaa? ("How are we now?") 2017 Nunavik Inuit Health Survey (Q2017)] in comparison with concentrations in participants from the Canadian Health Measures Survey Cycle 5 2016-2017.[1] Understanding specific exposure sources is crucial given these elevated PFAA concentrations in Inuit populations.

Consumption of country foods-foods hunted and harvested from the land, rivers, and sea-are a key source of PFAA exposures because of their high concentrations in various Arctic species.[4],[6],[8],[9] Country foods remain an integral part of Inuit culture, despite an increased dependence on market foods (imported foods purchased from grocery stores) since the late 20th century.[10] Approximately 13% of Inuit in Nunavik consume predominantly country foods, and an additional half consume some country foods regularly.[11] The most commonly consumed country foods include caribou meat, wild berries, Arctic char, and beluga mattaaq (beluga skin and fat).[11] In a prior study, Inuit with dietary profiles defined by frequent country food consumption also had higher concentrations of plasma PFAAs.[12]

However, few studies have examined the specific food items linked to PFAA biomarker concentrations in Arctic populations. Previous environmental studies largely analyzed PFAA concentrations in liver samples only, and epidemiology studies on country foods and PFAAs faced challenges in handling correlations between commonly consumed country food variables.[3],[13-16] For example, Inuit in Nunavik who regularly consume seal are also likely to consume beluga. Similarly, different parts of the animals may be consumed, including the meat, skin, and blubber, and this distinction has not been accounted for in previous studies. Marine mammal meats are also regularly dipped in rendered beluga fat (misirak). Not accounting for these correlations properly can result in inflated associations and false positives. Although machine learning methods on feature selection can address these correlation issues, they have not been applied in this context. Moreover, even if a lack of multicollinearity exists, including a wide range of food items into one model can be problematic, particularly when the sample size is limited. This is particularly the case in Indigenous communities and rural or coastal communities relying on local food for subsistence, where limited sample sizes are common due to small community sizes.

Although country foods have been the primary exposure source examined in Nunavik, market foods are considered a major contributor to PFAA exposures (e.g., PFOA and PFOS) in southern latitudes,[17],[18] with seafood considered the primary source of dietary PFAA exposure in adults.[18],[19] PFAS compounds are also used in food packaging to increase packaging resistance to humidity and grease and may contaminate market or fast foods.[20] Approximately 42% of Inuit in Nunavik consume predominantly market foods.[12]

Identifying the dietary exposure sources to chemical contaminants is critical for targeted interventions and policy initiatives. Thus, this study expands on our previous work by identifying the key country and market food items associated with PFAA exposure using machine learning feature selections to better pinpoint the most important sources of PFAA exposures in Nunavik.

We used data from the Q2017 survey conducted in Nunavik, northern Quebec, Canada. Q2017 is a health survey that assessed the health status of Nunavik permanent residents (Nunavimmiut). The survey targeted residents 16 y of age and over and implemented a proportional stratified sampling method based on the three ecological regions (Ungava Bay, the Hudson Strait, and the Hudson Bay) of residence and age category (16-19, 20-30, 31+ y). Further details on the survey have been described in a publicly available report.[21] Q2017 recruited 1,326 individuals, ∼10% of the Nunavik population,[22] and is the largest survey of its kind in Arctic communities. The total participation rate was 42% for people over 30 y of age and 31% for those under 30 y of age. Lower participation rates were driven by noncontact; however, among those contacted, 80% agreed to participate. Participants from all 14 villages in Nunavik completed the survey aboard the Amundsen, a Canadian Coast Guard icebreaker, from 19 August 2017 to 5 October 2017. Data collection included questionnaires, clinical measurements, and human biological sampling of urine and blood.

Ethical approval was received from the Comité d'éthique de la recherche du Center Hospitalier Universitaire de Québec- Université Laval (No. 2016-2499). All participants provided informed consent. Several organizations from Nunavik collaborated closely in the preparation, implementation, and assessment of the survey (including but not limited to the Nunavik Regional Board of Health and Social Services, Kativik Regional Government, and Makivik Corporation). Q2017 was also governed by the OCAP principles (Ownership, Control, Access, and Possession). The Data Management Committee (DMC), heavily represented by Inuit colleagues and partners, was integral to the discussion and interpretation of the Q2017 survey results.

Study participants completed a food frequency questionnaire (FFQ) regarding the frequency of their consumption of a wide range of market and country foods in the previous 3 months. These were described previously.[11] In brief, consumption frequency of each food item was rated on a scale of 1 (consumed never or less than once a month), 2 (1-3 times a month), 3 (Once a week), 4 (2-6 times a week), 5 (once a day), 6 (2-3 times a day), and 7 (4 times or more a day). Few food items were rated 6 or 7, so we collapsed the last three ratings (5, 6, and 7) into one category (5). This approach led to a new scale of 1 (item consumed never or less than once a month) to 5 (item consumed daily or more). For ease of interpretation, we converted these consumption categories to estimated number of meals consumed in a month: 0.5 (consumed never or less than once a month), 2 (1-3 times a month), 4 (once a week), 16 (2- 6 times a week), and 28 (item consumed daily or more). Statistical analyses included food items that are consumed at least once a month by more than 20% of the population (Excel Table S1). Spearman correlation coefficients between individual food items are presented in Figure S1 and Excel Table S2.

Venous blood samples were drawn from participants at enrollment, collected in K2-EDTA vacutainers, and processed within 90 min on board the Amundsen. Blood samples were centrifuged at 2,000 × g for 10 min, and the plasma was transferred into a 2-mL polypropylene tube for storage at -20°C until time of analysis.

Nine PFAA congeners were quantified in plasma samples at the Center de toxicologie du Québec (CTQ) of the Institut national de santé publique du Québec (INSPQ) (accredited by the Canadian Association for Environmental Analytical Laboratories and ISO 17025). PFAA congeners included perfluoro-n-butanoic acid (PFBA), perfluoro-hexanoic acid (PFHxA), PFOA, PFNA, PFDA, PFUnDA, perfluorobutane sulfonic acid (PFBS), perfluorohexanesulphonic acid (PFHxS), and PFOS. Compounds were extracted using a weak anion exchange solid-phase extraction on a 96-well plate after acidification of the samples. The extracts were evaporated, dissolved into the mobile phase, and then analyzed using ultra-performance liquid chromatography (Waters Acquity) coupled with tandem mass spectrometry (Waters Xevo TQ-S) with electrospray negative ionization in multiple reaction monitoring mode.

Internal reference materials were used to control the quality of the analyses using the certified reference material SRM-1958 from the National Institute of Standards and Technology (NIST) and inhouse quality controls. Limits of detection (LOD) were 0.06 μg/L for PFHxS, 0.07 μg/L for PFOA and PFBS, 0.08 μg/L for PFBA and PFHxA, 0.09 μg/L for PFDA, 0.10 μg/L for PFNA and PFUnDA, and 0.40 μg/L for PFOS. For concentrations < LOD, we imputed a value equal to LOD/2. Although ≥97.5% of PFOA, PFNA, PFDA, PFUnDA, and PFOS values were above the LOD, only 0.1%-11% of PFHxA, PFBA, and PFBS concentrations were above the LOD (Excel Table S3).

Descriptive analysis of population covariates was conducted, followed by bivariate PFAA distributions by age, sex, education, marital status, smoking status, and alcohol consumption in all Q2017 participants and in a complete dataset without missing data. PFAA concentrations were log10-transformed for further analysis to approach normality. PFHxA, PFBS, and PFBA were excluded due to a high number of nondetects (>80%) in the sample population.

We implemented a machine learning tool known as adaptive elastic net as a feature selection tool to identify the food items most predictive of PFAA biomarker levels, while controlling for a wide range of food items (𝑛= 79). We chose adaptive elastic net because it handles correlated variables better in comparison with other linear methods such as LASSO and elastic net.[23] We conducted a series of adaptive elastic net models with the log10-transformed PFAA measurements as the outcome variable and the main predictors as the individual food items. To be able to compare the associations between the food item and PFAA measurements across all the different food items, we normalized all food items by converting the number of meals consumed per month into a 𝓏-score, where the mean is 0 and the standard deviation (SD) is 1. We included a table to interpret the meaning of a SD in 𝓏-score in terms of the number of meals per month (Excel Table S4). For example, a 1 SD in consumption of Arctic char is 7.08 meals per month. For ease of interpretation, the coefficients from the models were converted to percent changes [10coefficient - 1 × 100. For the food item, a percent change of d% is interpreted as the percent change in the biomarker levels of a PFAA for every 1-SD increase in consumption of the food item.

We adjusted for age (continuous), age2 (continuous), sex [categorical; females (reference group), male], education [categorical; Grade 1-8 (reference group), Grade 9-11 higher education], marital status [categorical; single, separated, divorced, or widowed (reference group), married or in relationship], alcohol consumption [categorical; Drink ≤2 times/wk (reference), drink 3-7 times/wk] and smoking status [categorical; never smoked or smoked 1-99 cigarettes in a lifetime (reference), ex-smoker (at least 100 cigarettes in a lifetime), current frequent smoker]. We used age and age2 to represent the nonlinear associations between age and PFAA levels.[1],[24] We used education and marital status[25] as surrogates for socioeconomic status instead of income, because more participants have data available for education and marital status in comparison with income, which is difficult to assess in northern communities due to economic differences in the north and cultural sensitivities. Waist circumference was not included as a covariate because previous analyses did not find evidence of confounding by body size.[1] Occupation was not included as a covariate due to how occupation explained only 2% of biomarker levels of PFAAs in a prior study.[26] Smoking status and alcohol were further included based on associations identified in the literature.[27],[28]

We were concerned that the association between country foods and PFAA levels would mask the associations between market foods and PFAA levels, so we conducted a stratified analysis by the type of food items. Therefore, we ran two series of adaptive elastic net models with the main predictors as only the market foods or only the country foods as a sensitivity analysis. To assess the difference in the results from the full model vs. the stratified models, we compared the PFAA percent changes for all food items between the analysis with all food items as the main predictors vs. the stratified analysis by calculating correlation coefficients.

The adaptive elastic net model does not provide the statistical significance of the associations between the food items and PFAA biomarker levels, so we ran generalized linear regression models to obtain the 𝑝-values. Generalized linear models perform in a manner that is similar to post model-selection inference methods, such as Post-Selection Inference (PoSI) and Exact Post-Selection Inference (EPoSI); moreover, when there are many features, the computational cost of PoSI and EPoSI can be high.[30] Thus, we ran a generalized regression model with log10-transformed PFAA measurements as the outcome variable, the food items that were selected by adaptive elastic net (i.e., food items with nonzero coefficients from the adaptive elastic net analysis) as the main predictors, and the same covariates as in the adaptive elastic net model. To identify statistically significant food items predicting PFAA biomarker levels while maintaining a low false positive rate, we applied the false discovery rate (FDR) method to the 𝑝-values of the associations between the food items and a given PFAA, from the generalized linear regression model.[31] Associations with FDR-adjusted 𝑝-values <0.05 are considered significant.

We identified groups of food items that have similar associations across all PFAA congeners by performing hierarchical clustering analysis on the dataset of percent changes pertaining to the food items. We used Euclidean distance as our distance measure to help measure the dissimilarity between each pair of food items. We used the complete linkage method to help determine how the food items should cluster together based on similarity in PFAA levels.[32] We displayed the results of the clustering analysis in a heat map.

For comparison and to test the difference in results from multiple linear regression models with and without a variable selection tool, we also analyzed the associations between each individual PFAA and food item controlling only for the sociodemographic variables listed above (age, age2, sex, education, marital status, alcohol consumption, and smoking status). We excluded beluga misirak and wild berries from models as a sensitivity analysis because neither food is anticipated to be associated with PFAAs. A heat map was created for these results for easier interpretation.

The statistical analyses are summarized in Figure 1. We performed all analyses using R (version 4.0.3; R Development Core Team).

The study population included in the study (𝑛= 1,193) was relatively evenly distributed by age, with the exception of a smaller group of participants 60 y of age and over (10.6%) (Table 1). More females (65.6%) vs. males (34.4%) participated in the survey. The prevalence of currently smoking and drinking alcohol three or more times per week was 79% and 19.3%, respectively. Distributions by demographics and lifestyle habits were similar after dropping individuals with missing data (𝑛= 133).

Only six PFAA congeners (PFOA, PFNA, PFDA, PFUnDA, PFHxS, and PFOS) were detected in >75% of the participants (Excel Table S3). PFBA, PFHxA, and PFBS were dropped from the remaining analyses.

Concentrations of detected PFAAs generally increased with age; however, PFOA and PFNA concentrations in those 16-19 y of age (𝑛= 202) were comparable to those 30-39 y of age (𝑛= 195) and 40-49 y of age (𝑛= 190) (Table 2). PFOA, PFHxS, and PFOS concentrations were higher among males (1.31, 0.86, and 5.43 μg/L, respectively) vs. females (0.83, 0.48, and 4.63 μg/L, respectively). Conversely, PFDA and PFUnDA concentrations were higher in females (0.71 and 0.76 μg/L, respectively) vs. males (0.64 and 0.63 μg/L, respectively). Participants with low education (up to grade 8) had the highest PFAA concentrations, whereas participants in the middle education category (Grade 9- 11) had the lowest concentrations of PFAAs. Finally, in comparison with single participants, married participants had higher concentrations of PFNA (3.92 vs. 3.51 μg/L), PFDA (0.78 vs. 0.60 μg/L ), PFUnDA (0.69 vs. 0.63 μg/L), PFHxS (0.65 vs. 0.52 μg/L), and PFOS (5.67 vs. 4.14 μg/L). Current frequent smokers tend to have lower levels of PFAA in comparison with those who never smoked, smoked 1-99 cigarettes in their lifetime, and those who are ex-smokers. Concentrations of PFAAs were similar between participants who drink less than twice per week vs. those who drink 3-7 times per week except for PFNA (3.74 vs. 3.65 μg/L) and PFOS (4.93 vs. 4.71 μg/L).

PFOA concentrations increased by 6.1% [95% confidence interval (CI): 3.2%, 9.0%] with every SD increase in consumption frequency of ptarmigan, which translates to ∼4 meals or times per month (Figure S2; Excel Table S5). Per SD increase in suuvalik (fish roe mixed with berries and seal/beluga fat or vegetable oil) consumption which is ∼7 meals per month, participants had a 3.2% (95% CI: 0.2%, 6.2%) higher PFOA concentration, but this was not statistically significant after adjusting for multiple comparisons. PFOA concentrations were also 4.7% higher (95% CI: 1.5%, 7.8%), with an SD increase in fruits (fresh/frozen) (10.3 times/ month). The consumption of canned fish was significantly associated with higher PFOA concentrations (4.5%; 95% CI: 1.6%, 7.4%). PFOA was also 2.9%-3.7% higher with SD increases in ice cream, other beverages, cheese, ketchup, and seaweed, but these associations were not significant after adjusting for multiple comparisons. The consumption of carbonated beverages (-4.5%; 95% CI: -7.4%, -1:6%) (12.0 times/month) and processed meat (-4.4%; 95% CI: -7.4%, -1.4%) (7.3 times/month) was significantly associated with lower PFOA concentrations. The consumption of milk, yogurt, rice, other fish, fast-food chicken, and fries was also associated with 3.7%-3.1% lower PFOA concentrations but was not significant after adjusting for multiple comparisons.

PFNA concentrations were 7.2% (95% CI: 3.3%, 11.1%) and 6.3% higher (95% CI: 2.6%, 10.5%) with a SD increase in beluga misirak (7.6 times/month), which is rendered blubber, and wild berries (9.1 times/month), respectively (Figure 2; Excel Table S5). Higher levels of PFNA were associated with increased consumption of ptarmigan (5.1%; 95% CI: 1.1%, 9.1%) (4.2 times/month), canned fish [4.4% (95% CI: 0.4%, 8.4%)] (3.9 times/month), and cereal [4.1% (95% CI: 0.1%, 8.2%)] (10.4 times/month) but were not significantly associated after adjusting for multiple comparisons. Similar to those of PFOA, PFNA concentrations were lower with higher consumption of processed meats (6.0%; 95% CI: -10.1%, -1.9%) (7.3 times/month). It is interesting to note that the foods that were associated with higher PFNA concentrations were predominantly country food items, whereas those associated with lower PFNA concentrations were all market food items.

PFDA and PFUnDA concentrations followed similar trends (Figure 3; Figure S3; Excel Table S5). PFDA and PFUnDA were higher by 14.6% (95% CI. 10.3%, 18.9%) and 14.6% (95% CI. 10.1%, 19.0%), with a SD increase in beluga misirak (7.6 times/ month); 9.3% (95% CI. 5.0%, 13.7%) and 8.1% (95% CI. 3.5%, 12.6%), with a SD increase in seal liver (3.1 times/month); and 6.0% (95% CI. 1.3%, 10.7%) and 7.5% (95% CI. 2.7%, 12.3%), with a SD increase in suuvalik (7.9 times/month). Conversely, consumption of popcorn [(-6.6%; 95% CI. -11.1%, -2.1%); -5.4%; 95% CI. -10.1%, -0.7%], processed meats [(-5.4%; 95% CI. -9.9%, -0.9%); -5.5%; 95% CI. -10.2%, -0.7%], and sea trout [(-4.9%; 95% CI. -9.4%, -0.4%); -5.4%; 95% CI. -10.2%, -0.6%] were associated with lower levels of PFDA and PFUnDA. The PFUnDA associations with processed meats, popcorn, and sea trout and the PFDA associations with sea trout were not statistically significant after adjusting for multiple comparisons.

PFHxS concentrations were significantly higher by 10.2% (95% CI. 6.7%, 13.7%) with a SD increase in beluga misirak and 5.4% (95% CI. 1.8%, 9.0%) with a SD increase in ptarmigan (Figure S4; Excel Table S5). Increased consumption frequency of fruit, other beverages, and seaweed was also significantly associated with 4.7-5.3% higher levels of PFHxS. PFHxS concentrations were significantly lower, with an SD increase in bottled water (-5.1%; 95% CI. -8.7%, -1.6%) and processed meat (-4.9%; 95% CI. -8.7%, -1.1%). In addition, increased consumption of popcorn, rice, and carbonated beverages was also significantly associated with 4.9%-4.3% lower levels of PFHxS.

PFOS (similar to PFDA and PFUnDA) was significantly associated with higher consumption of beluga misirak (14.4%; 95% CI. 10.0%, 18.8%) and seal liver (8.6%; 95% CI. 4.2%, 13.0%) (Figure 4; Excel Table S5). PFOS levels were higher by 5.7%, 5.5%, and 5.0% with consumption of wild berries, canned fruit, and suuvalik, respectively, but this was not significant after adjustment for multiple comparisons. In contrast, consumption of popcorn (-6.9%; 95% CI: -11.5%, -2.4%) and bottled water (-6.7%; 95% CI: -11.0%, -2.3%) were significantly associated with lower levels of PFOS. Increased consumption frequency of processed meat, salad dressing, and tea was also associated with 5.2-5.1% lower levels of PFOS but was not significant after adjusting for multiple comparisons.

We showed which groups of food items had similar associations with PFAA from the hierarchical clustering analysis by using a heat map of percent changes in biomarker levels of PFAAs (Figure 5). Similar to the results described above, Zone I showed significantly higher consumption of beluga misirak, seal liver, and suuvalik was associated with higher biomarker levels of PFUnDA, PFDA, and PFOS. The associations for the same aforementioned food items but for PFHxS, PFNA, and PFOA were lower in magnitude and/or null (Zone II). Higher consumption of ptarmigan and fresh fruits was significantly associated with higher levels of PFHxS and PFOA (Zone III). Zone IV was all white, indicating that the food items in this zone were deemed by the adaptive elastic net model to not be important in predicting levels for any of the studied PFAAs. Processed meat, popcorn, and bottled water showed some of the highest negative and significant associations with PFDA, PFUnDA, and PFOS biomarker levels (Zone V). The associations for the same food items and PFHxS, PFNA, and PFOA were lower in magnitude and/or null (Zone VI). Rice and carbonated beverages showed significant associations with lower PFHxS, but only carbonated beverages showed a significant association with lower PFOA (Zone VII).

Stratifying the models by country foods mimicked the results observed in the model with all food items (Figure S5; Excel Table S6). Frequent consumption of beluga misirak was associated with all PFAAs, except PFOA. Frequent consumption of seal liver and suuvalik was associated only with higher concentrations of PFDA, PFUnDA and PFOS. On the other hand, frequent consumption of ptarmigan was associated with higher concentrations of PFOA, PFNA and PFHxS, albeit some associations were not statistically significant. Finally, frequent consumption of sea trout was negatively associated with PFDA, PFUnDA, and PFOS.

Stratifying models by market foods highlighted the association between PFAA consumption and fruit, ice cream, and nondairy coffee creamers (Figure S6; Excel Table S7). In addition, there was indication of a negative association between frequent consumption of popcorn and processed meat with long-chain PFAAs.

There was high correlation between the percent changes from the analysis with the food items all together as the main predictors and the percent changes from the stratified analysis separating the market foods from the country foods (Figure S7; Excel Table S6 and S7). This finding suggested that the PFAA associations for the food items are similar between the global analysis and the stratified analysis. Moreover, the PFAA associations with the market foods were not entirely masked by the associations with the country foods, although the associations were weaker when combined with country foods.

Associations between individual PFAA and food items differed from the adaptive elastic net model results (Figure S8; Excel Table S8). Without controlling for other food items, all PFAAs were associated with a much wider range of country food items, including (but not limited to) mollusks, game eggs, seaweed, goose, lake trout, caribou meat, and wild berries. Strong associations with market foods, on the other hand, were not observed.

Removing beluga misirak and wild berries from the models yielded similar results (Figure S9). PFNA is not significantly associated with other foods, although the association with suuvalik (a food made from fish roe, fat, and berries) and canned fruit are stronger (6% and 5% vs. 3%). The association between PFOA and PFHxS and ptarmigan and canned fruit was slightly stronger after dropping misirak and fruit.

Our study examined the associations between six PFAA congeners and various food items/groups (country and market food) in an Inuit population in Nunavik, northern Quebec. To our knowledge, this is the first application of feature selection to identify the key dietary sources of chemical contaminants. Of note, frequent consumption of beluga misirak (rendered beluga fat), seal liver, and suuvalik (a mixture of fish roe, berries, seal/beluga fat or vegetable oil) were most strongly associated with increased concentrations of plasma PFDA, PFUnDA and PFOS. Frequent consumption of ptarmigan was associated with increased concentration of PFOA and PFHxS, and PFNA with less confidence. Decreased PFAA concentrations were also associated with frequent consumption of processed meat, popcorn, and fruits. Overall, country foods exhibited stronger associations with PFAAs in comparison with market foods.

The strongest associations in our study were between PFDA and PFUnDA and some selected country foods. These congeners were detected at exceptionally higher concentrations in an Inuit population in Nunavik in comparison with the general Canadian population[1] and were previously associated with a profile defined by consumption of seal and beluga meat, seal liver, goose, wild eggs, and diverse seafood.[13] This finding reaffirms that consuming local foods harvested in Nunavik, particularly food derived from marine mammals, is an important exposure source of these two longer-chain PFAAs. A similar epidemiological study in Greenland found that PFNA, PFDA, PFUnDA, PFHxS, and PFOS were all associated with consuming marine mammals, seabirds, and dried fish, but no associations were observed with PFOA.[16] We observed positive associations between these congeners and specific marine mammal parts, namely seal liver and beluga misirak, although not all were statistically significant after adjusting for multiple comparisons. Although PFNA showed positive associations with various country foods in our study, few were statistically significant, and the primary PFNA exposure source in Nunavik remains unclear. Future studies should explore other exposure sources such as indoor dust or other foods not captured in the food frequency questionnaire. marine mammal parts, namely seal liver and beluga misirak, although not all were statistically significant after adjusting for multiple comparisons. Although PFNA showed positive associations with various country foods in our study, few were statistically significant, and the primary PFNA exposure source in Nunavik remains unclear. Future studies should explore other exposure sources such as indoor dust or other foods not captured in the food frequency questionnaire.

We observed surprisingly strong associations between PFDA, PFUnDA, PFOS, PFNA, and PFHxS and frequent consumption of beluga misirak, a traditional Inuit food made from aging rendered beluga blubber, even though PFAAs are not lipophilic. Beluga misirak is a staple in Nunavik[10] and consumed with other dishes, such as dipping cooked, frozen, or dried meat (hunted or storebought) in misirak. Although the adaptive elastic net model accounts for the correlation between food variables, the model may still favor beluga misirak over meat consumption due to its widespread use (almost 60% of study participants reported consuming misirak). Inuit who consumed more beluga misirak also consumed beluga and seal meat, seal liver, caribou meat, wild birds, and fish.[12] In addition, beluga misirak may be a component of suuvalik, [33] another country food associated with various PFAA congeners in the present study. Seal misirak, prepared in a manner similar to that of beluga misirak-but less commonly consumed in Nunavik-was not associated with any PFAA congeners. A 2009 study in Nunavut detected PFOS, PFNA, PFDA, and PFUnDA (but not PFOA) in beluga blubber but not ringed seal blubber.[34] The PFAA concentrations detected in the Nunavut beluga blubber samples were half those in beluga meat, but these results still provide evidence for a difference in PFAA concentrations in beluga blubber vs. seal blubber. Further studies directly measuring PFAA concentrations in beluga blubber and associated food products in Nunavik are urgently needed.

Unlike other persistent organic pollutants, PFAAs bind to proteins instead of fats and accumulate in the blood and wellperfused organs, such as the liver, kidney, lung, and heart.[6],[35-39] This binding is in line with our findings of strong associations between frequent seal liver consumption and PFAA concentrations. Notably, beluga liver is not commonly consumed in Nunavik and was not included in the questionnaire. Other environmental studies have similarly detected high concentrations of PFDA, PFUnDA, and PFOS in the livers of ringed seals across the Arctic and of killer whales in Greenland.[6],[8],[40] In previous studies, researchers found that PFNA concentrations in marine mammal livers varied across the Arctic regions, whereas PFOA and PFHxS concentrations were low.[6],[8],[40] Such findings are similar to our findings of weaker associations between PFOA, PFNA, and PFHxS and seal liver consumption. Given the protein binding capacity of PFAAs, we anticipated observing associations between PFAAs and wild meats. However, we did not observe associations between frequent consumption of beluga meat and PFAAs. Previous studies from 2009 and 2014 reported varying PFAA concentrations measured in marine mammal meats including beluga, narwhal, whale, and seal; however, of the PFAA detected, PFUnDA and PFOS concentrations were most commonly detected.[34],[41] Our results add to the complexity of understanding multiple contaminant exposure sources in the Arctic context.

Suuvalik was associated with higher concentrations of PFDA, PFUnDA, PFOS, and PFHxS. It is a traditional dish prepared by mixing fish roe (eggs) and beluga/seal fat or vegetable oil/shortening with wild berries, (i.e., blueberries or crowberries/blackberries), depending on food availability and preferences.[33] PFAAs have been globally detected at high concentrations in fish roe.[42] PFOS (but not PFOA and PFHxS) were detected in lake whitefish and brown trout eggs from Lake Huron and Lake Superior in Michigan and aquatic organisms in Taihu Lake, China, at concentrations higher than those in the liver, indicating preferential binding of PFOS to fish egg albumin.[43] Moreover, PFOS, PFDA, and PFUnDA concentrations were comparable to or higher than concentrations detected in various types of fish liver.[44] High concentrations of PFOS were also detected in the eggs of farmed and wild aquatic species.[45],[46] Although the FFQ in the Q2017 survey did not specifically inquire about fish egg consumption, our findings, coupled with previous literature, suggest that the robust associations between suuvalik and PFAAs may be related to the high PFAA concentrations in fish eggs. The small (albeit largely insignificant) associations with wild berries may also be related to suuvalik consumption because both are consumed at similar rates across dietary profiles.[12]

Fish consumption has been identified as one of the main dietary sources of PFAAs in other populations.[47-51] Consequently, populations relying on subsistence fishing may be at higher risk of PFAA exposure (particularly PFNA, PFDA, PFUnDA, and PFOS) in comparison with populations that consume frequent fish/seafood purchased from markets. This difference is likely because subsistence fishing involves consuming the whole fish, including its organs and roe, whereas populations obtaining fish from markets often consume only the fillets.[52-54]

Increased consumption of ptarmigan was associated with an approximate 5%-6% increase in PFOA, PFHxS, and PFNA. Ptarmigan is frequently consumed because it is easy to hunt and requires minimal equipment, and it is sometimes referred to as "emergency food" among Inuit.[55] It was one of the most frequently consumed country foods among Inuit with a dietary profile associated with food insecurity.[11] Ptarmigan consume largely vegetation, including twigs, willow, berries, and insects (depending on the season), and are less migratory than other Arctic birds.[56] PFOA and PFNA, as shorter chain PFAAs, have greater atmospheric deposition in comparison with PFDA and PFUnDA,[57] and PFOA is less likely to bioaccumulate in highertrophic animals than PFOS.[5] As such, exposure to PFOA and PFNA may be of larger concern to terrestrial animals vs. marine animals. PFNA was the dominant PFAS congener detected in terrestrial herbivores in Canada, including caribou and reindeer.[58] Similarly, only PFNA was detected in ptarmigan liver samples from Greenland and the Faroe Islands.[59] The unique patterns of PFAAs detected in ptarmigan vs. seabirds (like murre or gulls) suggest a unique exposure pathway,[6] potentially emphasizing the importance of atmospheric deposition or sea spray deposition in terrestrial environments.[60] PFOA and PFNA also dominated distribution patterns of PFAAs in plants (particularly willow and grass).[6] Given this, the finding of higher concentrations of PFOA and PFNA is in line with increased consumption of ptarmigan. PFHxS is highly correlated with PFOA,[1] and this may be indicating similar environmental distribution patterns (or perhaps lifestyle patterns) not captured in this study. Future studies should consider measuring ptarmigan samples to test these results.

Some market foods were also associated with PFAAs in Nunavik. A study using a text mining and database fusion reported that although fish and seafood represent the largest source contribution of PFOS, vegetables, meats, and grains may be more important dietary-exposure sources of PFOA.[61] We also observed more consistent positive associations between market foods and PFOA in comparison with PFOS. In contrast, a study with preconception Chinese women found positive associations between the consumption of leafy vegetables and PFDA, PFUnDA, and PFOS. Frequent consumption of fruits was only associated with increased PFOA, and no associations were observed with rhizome vegetables (such as potatoes).[62] Similarly, we also observed an association between PFOA and PFHxS with frequent fruit consumption. More studies are needed to document PFAAs in imported fruits (not produced in Nunavik). Our results also speak to the importance of also considering PFAA exposure from imported market foods when describing the Arctic context.

In a US study on pregnant women, meat consumption, particularly poultry and fish, was associated with increased PFOS, PFDA, and PFUnDA concentrations,[63] because of PFAA' ability to bind to proteins.[35],[39] Other studies also reported elevated PFAAs with meat consumption.[64-66] A surprising result was that our study observed that consuming processed meats is associated with lower PFAA concentrations, likely reflecting the replacement of country food with imported market options.

Unlike other studies, we observed negative associations between PFAAs and most dairy products (except for ice cream). In the United States, maternal PFNA levels were associated with increased consumption of dairy products such as milk and cheese during pregnancy.[63] An association between PFNA and milk consumption was also observed in Anishinabe and Innu children in southern Quebec.[67] Specifically, frequent consumption of cheese was associated with increased concentrations of PFOA, PFNA, PFHxS, and PFOS (but not PFDA and PFUnDA); frequent consumption of yogurt was associated with increased PFOA, PFHxS, and PFUnDA. Greater intake of ice cream was associated with an increase in PFOA, PFNA, PFDA, MeFOSAA, PFHxS, and PFOS in children living in Boston, Massachusetts,[68] and this may be because of the packaging used in ice cream[69],[70] or the industrial phase separation of cream from milk potentially contributing to higher PFAS concentrations.[71]

PFAAs are regularly used in food packaging materials for their grease-resistant properties and could leach into foods.[72],[73] Market foods in Nunavik are highly packaged to protect them during the long delivery process and to help increase their shelf life. In addition, the extremely elevated costs of market food in Nunavik lead to increased consumption of processed and packaged foods (energy-dense vs. nutrient-dense).[74] In a study of Brazilian mothers in their third trimester, umbilical cord blood concentrations of PFOA but not PFOS were elevated with regular consumption of ultraprocessed foods.[75] However, as in our results, no elevated concentrations of maternal serum PFOA, PFNA, PFDA, PFUnDA, PFHxS, or PFOS were observed among 509 US pregnant women with frequent consumption of fast food, French fries, pizza, or store-bought foods.[63]This finding contrasts with studies reporting elevated PFOA levels after consuming fast food or pizza,[73] likely due to PFAA being in the packaging.[72] Another US study identified elevated concentrations of PFOA, PFNA, PFDA, PFHxS, and PFOS in 548 children with a diet defined by frequent consumption of packaged foods and fish.[68] We observed negative associations between popcorn consumption and PFAA concentrations, differing from studies that linked microwaved popcorn bags with PFAAs.[72],[73] The Q2017 survey did not specify whether the popcorn consumed was microwavable, possibly explaining the discrepancy.

Municipal drinking water is the main source of drinking water in Nunavik (83%), followed by bottled water.[29] Though drinking water is not considered an important exposure source of PFAA in Nunavik, drinking water samples will be taken in a follow-up study to measure PFAA concentrations in water consumed by community members.

To our knowledge, the Nunavik Q2017 is the largest study conducted on PFAAs and diet exposure in the Arctic. Our use of a population-based sample and a wide range of market and country food items pinpoint potential key dietary sources. Our application of feature selection reduces confounding while accounting for multicollinearity. Particularly, our application of adaptive elastic net narrowed down the dietary exposures associated with PFAAs and can be used in future studies for exposure source identification. Our results can help identify key foods to be analyzed for PFAS measurements and guide potential public health interventions for reducing PFAS exposure in vulnerable groups.

However, our study has limitations. First, because Q2017 is a cross-sectional survey, we cannot make claims that the food items are causal factors of PFAA biomarker levels. Second, the one-time recording of food consumption may not fully represent participants' eating habits, but the persistent nature of PFAAs allows plasma measurements to reflect several months of exposure. Third, some foods may not have been captured in the FFQ, and the questionnaire did not capture the quantity consumed in each meal (only frequency). Fourth, we assumed a linear association between food consumption and PFAA levels, suggesting the need for future studies using nonlinear feature selection methods[76] to account for any nonmonotonic relationships.

Our study identified strong associations between PFAAs and specific country foods, including seal liver, suuvalik, ptarmigan, and potentially beluga misirak. In addition, frequent consumption of processed meat, popcorn, fruits, and bottled water was also associated with several PFAA congeners. Our implementation of a feature selection tool allowed us to control for the consumption of other foods and pinpoint the key foods most strongly correlated with PFAAs. Associations with the largest percentage changes point to a more important contribution of select country food consumption and exposure to PFDA and PFUnDA, two long-chain PFAA congeners. Our findings provide further evidence of PFAA contamination in country foods due to the longrange transfer of these compounds and their precursors. The implementation of feature selection methods, such as adaptive elastic net, could be used in other studies to determine key correlated dietary sources of chemical contaminants.

The authors are grateful to all Nunavimmiut who participated to the 2017 Qanuilirpitaa? Nunavik Health Survey and to all of those who planned and carried out this survey. The authors are also grateful to the Qanuilirpitaa Steering Committee who revised this manuscript and the INSPQ for conducting the chemical analyses.

The Q2017 survey was funded by the Nunavik Regional Board of Health and Social Services, the Institut national de santé publique du Québec, the Kativik Regional Government, the Makivik Corporation, Kativik Ilisarniliriniq, the Ministère de la santé et des services sociaux du Québec, ArcticNet, the Amundsen Science Ship Fund, and the Northern Contaminants Program (NCP) of the Crown-Indigenous Relations and Northern Affairs Canada (CIRNAC). E.C.-B. and A.A. received a salary from the Littoral Research Chair (2019-2020 and 2020-2024, respectively), which is mainly funded by Sentinel North and NCP. A.A. also received a Sentinel North scholarship (2021-2022). M.L. is a member of Quebec Océan and also received a salary grant from the Fonds de recherche du Québec-Santé (FRQS): Junior 1 (2015-2019) and Junior 2 (2019-2023).

Note: -, no data.

Note: PFAA, perfluoroalkyl acid; PFDA, perfluorodecanoic acid; PFHxS, perfluorohexanesulphonic acid; PFNA, perfluorononanoic acid; PFOA, perfluorooctanoic acid; PFOS, perfluorooctanesulfonic acid; PFUnDA, perfluoroundecanoic acid; Ref, reference.

GRAPH: Figure 2. Forest plot showing percent change in PFNA concentration and all food items in the Qanuilirpitaa? Nunavik Inuit Health Survey (Q2017), Nunavik, northern Quebec, Canada (𝑛= 1,093). Results are derived from generalized linear models with the outcome variable as biomarker concentrations of PFNA and the main predictors as the consumption of the food items normalized as 𝓏-scores. The models are adjusted for age, age2, sex, marital status, education, alcohol consumption, smoking, and food items selected by adaptive elastic net. Number of asterisks indicate statistical significance of the percent change: * 𝑝-value ϵ (0.01, 0.05), **𝑝-value ϵ (0.001, 0.01), and * * *𝑝 ≤ 0.001. The 𝑝-values are corrected for multiple comparisons with the Benjamini-Hochberg false FDR procedure of 5%. Percent changes are represented by points, and 95% CIs are represented by the corresponding vertical lines. Corresponding results in Excel Table S5. Note: CI, confidence interval; FDR, false discovery rate; PFNA, perfluorononanoic acid.

GRAPH: Figure 3. Forest plot showing percent change in PFDA concentration and all food items in the Qanuilirpitaa? Nunavik Inuit Health Survey (Q2017), Nunavik, northern Quebec, Canada (𝑛= 1,086). Results are derived from generalized linear models with the outcome variable as biomarker concentrations of PFDA and the main predictors as the consumption of the food items normalized as 𝓏-scores. The models are adjusted for age, age2, sex, marital status, education, alcohol consumption, smoking, and food items selected by adaptive elastic net. Number of asterisks indicate statistical significance of the percent change: * 𝑝-value ϵ (0.01, 0.05), **𝑝-value ϵ (0.001, 0.01), and * * *𝑝 ≤ 0.001. The 𝑝-values are corrected for multiple comparison with the Benjamini-Hochberg FDR procedure of 5%. Percent changes are represented by points, and 95% CIs are represented by the corresponding vertical lines. Corresponding results in Excel Table S5. Note: CI, confidence interval; FDR, false discovery rate; PFDA, perfluorodecanoic acid.

GRAPH: Figure 4. Forest plot showing percent change in PFOS concentration and all food items in the Qanuilirpitaa? Nunavik Inuit Health Survey (Q2017), Nunavik, northern Quebec, Canada (𝑛 = 1,093). Results are derived from generalized linear models with the outcome variable as biomarker concentrations of PFOS and the main predictors as the consumption of the food items normalized as 𝓏-scores. The models are adjusted for age, age2, sex, marital status, education, alcohol consumption, smoking, and food items selected by adaptive elastic net. Number of asterisks indicates statistical significance of the percent change: * 𝑝-value ϵ (0.01, 0.05), **𝑝-value ϵ (0.001, 0.01), and * * *𝑝 ≤ 0.001. The 𝑝-values are corrected for multiple comparison with the Benjamini-Hochberg FDR procedure of 5%. Percent changes are represented by points, and 95% CIs are represented by the corresponding vertical lines. Corresponding results in Excel Table S5. Note: CI, confidence interval; FDR, false discovery rate; PFOS, perfluorooctanesulfonic acid.

GRAPH: Figure 5. Heatmap of percent changes of the six detectable PFAAs by all food items in the Qanuilirpitaa? Nunavik Inuit Health Survey (Q2017), Nunavik, northern Quebec, Canada. Results are derived from generalized linear models with the outcome variable as biomarker levels of PFAAs and the main predictors as the consumption of the food items normalized as 𝓏-scores. The models are adjusted for age, age2, sex, marital status, education, alcohol consumption, smoking, and food items selected by adaptive elastic net. The dendrogram of the food items and PFAAs are defined based on hierarchical clustering with using the complete linkage function with Euclidean distance. Number of asterisks indicates the statistical significance of the percent change: * 𝑝-value ϵ (0.01, 0.05), **𝑝-value ϵ (0.001, 0.01), and * * *𝑝 ≤ 0.001. The 𝑝-values are corrected for multiple comparisons with the Benjamini-Hochberg FDR procedure of 5%. Corresponding results in Excel Table S5. Note: CI, confidence interval; FDR, false discovery rate; PFAA, perfluoroalkyl acids.

DIAGRAM: Figure 1. A schematic overview describing the analytical pipeline to identify the most important dietary sources predicting the PFAA biomarker levels among the Qanuilirpitaa? Nunavik Inuit Health Survey (Q2017) participants (𝑛= 1,093), Nunavik, northern Quebec, Canada. Our pipeline includes the curation process of PFAA congeners and food items, the inclusion criteria of participants, and the statistical methods used to characterize the associations between food items and PFAA biomarker levels. Note: PFAA, perfluoroalkyl acid.