Association of Prenatal Exposure to Endocrine-Disrupting Chemicals With Liver Injury in Children

Key Points Question Is prenatal exposure to endocrine-disrupting chemicals (EDCs) associated with liver injury and hepatocellular apoptosis in school-aged children? Findings In the Human Early-Life Exposome population-based cohort study of 1108 mother-child pairs from 6 European countries, prenatal exposures to EDC mixtures, including organochlorine pesticides, polybrominated diphenyl ethers, perfluoroalkyl substances, and metals, significantly increased the risk for liver injury and/or hepatocellular apoptosis in school-aged children. Meaning These findings suggest that exposure to mixtures of EDCs during the sensitive pregnancy period may increase the risk for liver injury and hepatocellular apoptosis in childhood, potentially contributing to the current epidemic of pediatric nonalcoholic fatty liver disease.


eMethods. Description of HELIX Subcohort and Inclusion Criteria, Liver Outcome Assessment, Details on Statistical Methods, and Postnatal EDC Exposure Assessment
The description of the HELIX sub-cohort and the eligibility criteria for inclusion has been obtained from Maitre et al., 2018 BMJ Open 1 and presented here verbatim.

HELIX sub-cohort:
"From the entire cohort, a sub-cohort of mother-child pairs was selected to be fully characterized for a broad suite of environmental exposures and 'omics' data, to be clinically examined and to have biological samples collected. A new follow-up visit was organized for these mother-child pairs between December 2013 and February 2016. Subcohort subjects were recruited from within the entire cohorts such that there were approximately 200 mother-child pairs from each of the six cohorts. Sub-cohort recruitment in the EDEN cohort was restricted to the Poitiers area and in the INMA cohort to the city of Sabadell." "Eligibility criteria for inclusion in the sub-cohort was (a) age 6-11 years at the time of the visit, with a preference for ages 7-9 years if possible; (b) sufficient stored pregnancy blood and urine samples available for analysis of prenatal exposure biomarkers; (c) complete address history available from first to last follow-up point; (d) no serious health problems that may affect the performance of the clinical testing or impact the volunteer's safety (e.g., acute respiratory infection). In addition, the selection considered whether data on important covariates (diet, socioeconomic factors) were available. Each cohort selected participants at random from the eligible pool in the entire cohort and invited them to participate in this sub-cohort until the required number of participants was reached. In total, 1301 mother-child pairs with complete questionnaire and clinical examination data, and urine and blood samples, were included in the HELIX sub-cohort."

Liver Outcome Assessment
ALT, AST, and GGT in serum were measured at the Biochemistry Laboratory of the Clínica Universidad de Navarra, using homogenous enzymatic colorimetric methods on a Colorimetry Cobas 8000 analyzer (as described in the manufacturer's instructions (Roche Diagnostics GmbH, Mannheim, Germany)). CK-18 (a plasma caspase-generated fragment and novel marker of hepatocyte apoptosis, in children 2 ) in serum was measured using ELISA (M30 Apoptosense® ELISA, PEVIVA).

Statistical Analyses
We used triangulation of available state-of-the-art statistical methods for exposure mixture analysis in environmental health studies to confirm the robustness of associations independently of the statistical approach used and identify consistent findings across all methods, as each method presents unique strengths and limitations. Firstly, to model group mixture associations under the assumption of linearity and additivity, Bayesian Weighted Quantile Sum (BWQS) regressions 3 were fitted where all chemicals belonging to a particular group were modelled at once. This method is useful when there is no prior information about the direction of the mixture-outcome association. BWQS does not select a priori the directionality of the association, rather utilizes a Dirichlet prior which is naturally incorporated in this Bayesian analysis. Further, the estimated group mixture coefficient is then mapped to the corresponding contributions of each of the individual exposures. These contributions or weights identify the relative importance of each individual exposure within the group mixture association. Note that, in each chemical group, the estimated weights sum up to one, and are constrained to be positive.
Secondly, after relaxing the assumptions of linearity and additivity in group mixture models, Bayesian Kernel Machine Regression (BKMR) 4 was used to identify more flexible non-linear group associations. This tool utilizes the Kernel machine regression framework (also called Gaussian process regression) to flexibly model the exposureoutcome association without assuming any stringent structure. BKMR is a blend of usual machine learning framework and classical statistics, which allows to flexibly model the exposure-response function after controlling for the effect of the confounders. To control for multi-collinearity among the exposures and variable selection, it uses the so-called "Spike and Slab" prior. Since, it doesn't assume any structure for the exposure-response function, BKMR doesn't provide an overall estimate for group association. But similar to BWQS, the individual contribution to the exposure-response curve can be estimated using the Posterior Inclusion Probabilities (PIPs). Note that, in each chemical group, the estimated PIPs were scaled to sum up to one, to denote the relative inclusion of each chemical exposure to the overall group mixture association. This facilitates direct comparison with weights derived from BWQS.
Thirdly, we used Generalized Linear Mixed Effect Regression (GLMER) models to evaluate the associations of each individual chemical with the liver outcomes while avoiding multi-collinearity. All models were adjusted for confounders. For ease of interpretation and comparison of estimates between the methods, all the prenatal chemical exposures (which were log transformed before) were converted to quartiles. See eTable 3 for a summary of advantages and limitations of using each one of the three statistical approaches. BWQS was used as the main model for its simplicity while interpreting the results and computational speed.
A single imputation of imputed biomarker concentrations below the limit of detection was conducted using a quantile regression approach for the imputation of left-censored missing data. The urinary biomarker concentrations were divided by creatinine concentration and the haemal lipophilic biomarker concentrations were standardized and expressed in ng/g of total lipids in serum or plasma. Further, the missing data for exposure biomarker concentrations were imputed for 100 imputed datasets via a chained equations algorithm. See Appendix Table 5 of 5 for more details.

Postnatal EDC exposure assessment for Sensitivity Analysis
Postnatal EDCs exposures were measured in blood and urine samples of children collected during the HELIX sub-cohort follow-up visit at ages 6-11 years. Blood draws were performed after a median (5th, 95th percentile) fasting time of 3.3 (2.2, 5.9) hours. All blood samples were collected and processed after following identical predefined standardized protocols for all the six sub-cohorts. To provide slightly long-term exposure assessment than could be achieved with one spot urine sample, a combination of urine samples from the night before the visit and from the first morning void on the day of the visit was used 6 . See 1, 6, 7 for more details on postnatal EDC exposure measurements.

GLMER results
Analysis of individual chemicals using GLMER indicated DDT was associated with increased odds of liver injury (

EFFECT MODIFICATION BY SEX
We tested for effect modification and stratified analyses by sex (eFigure 3 and eTable 11). We observed differences in the effect sizes for liver injury by sex for OC pesticides and metals with stronger associations suggested in males compared to females (

COLLECTION FOR NON-PERSISTENT CHEMICALS AND CHILD BMI Z-SCORE
The strength and directionality of all EDC associations with liver injury remained unchanged after adjusting the statistical models for EDC exposures measured in child blood or urine at age ~8 years (eTable 12) and maternal diet (consumption of fish, fruits and vegetables; in times per week) (eTable 13). For the associations with CK-18, the PCBs effect estimate was slightly attenuated and did not remain statistically significant after adjustment for post-natal PCB exposures (eTable 12). Adjusting models for child BMI z-scores did not meaningfully change effect estimates for any of the associations of interest (<10% change in estimates) (eTable 14). Further, sensitivity analysis was also conducted for shorter half-life chemicals stratifying analyses between those cohorts who measured non-persistent EDCs in urine in the second pregnancy trimester versus INMA (that collected urine around 34 (SD:1.3) weeks of gestation) and we did not find any significant differences in the effect estimates (eFigure 4). Overall, we did not observe meaningful changes in effect estimates after controlling for postnatal EDC exposures, maternal diet or child BMI in pregnancy in statistical models.

Bayesian Weighted Quantile Sum (BWQS) Regression
• Estimates the mixtureoutcome association and the contribution of each chemical to the mixture • Direction of the association is estimated based on data • Does not consider nonadditive and non-linear relationships • Possibility of overfitting due to lack of penalization

Bayesian Kernel Machine Regression (BKMR)
• Estimate the association between chemicals and the outcome allowing nonadditive, and non-linear relationships     The estimates are presented as (OR (95% Credible Interval) or Beta (95% Credible Interval)). All models were adjusted for sub-cohort, maternal age, maternal pre-pregnancy BMI, maternal education level parity, child sex and child age.