289 - Regression-based Machine Learning Reveals Dynamic Windows of Fetal Growth Susceptibility to Prenatal Endocrine-Disrupting Chemicals

Publication Number: 2279.289

Kee Hyun Cho, Kangwon National University, Chuncheon, Kangwon-do, Republic of Korea; Payam Hosseinzadeh Kasani, Kangwon National University, Chuncheon, Kangwon-do, Republic of Korea; Woo Jin Kim, Kangwon National Univeristy, Chuncheon, Kangwon-do, Republic of Korea

Background: Endocrine-disrupting chemicals (EDCs) are pervasive environmental contaminants that can disrupt hormonal regulation during pregnancy, yet their cumulative and time-varying effects on fetal growth remain poorly understood due to exposure intercorrelation and non-linear dose–response patterns

Objective: To develop and apply advanced interpretable machine learning models to predict gestational age–standardized birth weight z-scores from complex mixtures of EDC exposures across pregnancy.

Design/Methods: We analyzed data from 4,274 mothers enrolled in the Korean Children’s Environmental Health Study (Ko-CHENS), a multicenter nationwide cohort, in which biomarkers of phthalates, phenols,and parabens were quantified during early ( < 20 weeks) and late (>28 weeks) pregnancy. Eight regression algorithms were implemented, spanning penalized linear models, ensemble trees, and boosting-based approaches. Performance was evaluated using R², root mean square error (RMSE), and mean absolute error (MAE). Interpretability was enhanced through Shapley Additive Explanations (SHAP), partial dependence plots, and correlation-based network analysis.

Results: CatBoost consistently achieved the strongest predictive performance, explaining up to 12–15% of the variance in birth weight z-scores across trimesters and yielding lower error compared to other approaches (early pregnancy: R² = 0.12, RMSE = 0.74, MAE = 0.58; late pregnancy: R² = 0.15, RMSE = 0.71, MAE = 0.56). In early pregnancy, predictions were driven primarily by MEHHP, bisphenol S, ethyl paraben, and cotinine, implicating endocrine disruption and lifestyle-related exposures as early determinants of fetal growth. In late pregnancy, bisphenol A, mono-butyl phthalate, MEOHP, and triclosan dominated, pointing toward mechanisms related to placental transfer, rapid somatic expansion, and immune–endocrine interactions. Network analysis further revealed that phthalates and maternal covariates such as BMI and parity occupied structurally central positions within the exposure matrix, even when not top-ranked by SHAP, suggesting that they exert influence both directly and through network-mediated pathways.

Conclusion(s): Regression-based machine learning identified early pregnancy as sensitive to low-level lifestyle exposures and late pregnancy to cumulative high-intensity exposures, offering interpretable trimester-specific insights for mechanistic research and maternal–child health policy.