Interpretable machine learning-based insights into early-life endocrine disruptor exposure and small vulnerable newborns

Abstract

Early-life exposure to endocrine-disrupting chemicals (EDCs) may contribute to small vulnerable newborns, including conditions such as being small for gestational age (SGA) and preterm birth (PTB), yet evidence remains limited. This study, which is based on 739 mother–infant pairs in the Chinese Jiashan Birth Cohort (2016–2018), including 39 SGA and 38 PTB cases, employed interpretable machine learning to elucidate the isolated effects of 34 EDCs on SGA and PTB risk and sex interactions in a multi-substance exposure context. Extra Trees and CatBoost classifiers performed best for SGA and PTB, respectively, achieving sensitivities of 0.60 and 0.73 and specificities of 0.82 and 0.97. For SGA, key predictors included bisphenol A (2,3-dihydroxypropyl) glycidyl ether (BADGE-H₂O), benzophenone (bZp), bisphenol A bis(2,3-dihydroxypropyl) ether (BADGE-2H₂O), propyl paraben (PrP), and 2-methylthio-benzothiazole (2-Me-S-BTH). Lower exposures to BADGE-H₂O, bZp, and BADGE-2H₂O (concentrations below 0.21, 4.22, and 0.93 μg·g⁻¹ creatinine, respectively) and higher exposure to 2-Me-S-BTH (above 0.15 μg·g⁻¹ creatinine) were both associated with increased SGA risk. Notably, BADGE-H₂O, BADGE-2H₂O, and PrP showed significant interactions with fetal sex. For PTB, key predictors included ethyl paraben (EtP), methyl paraben (MeP), bZp, BADGE-H₂O, and 1H-benzotriazole (1-H-BTR). Lower BADGE-H₂O and higher EtP and bZp exposures increased PTB risk (< 0.10 and > 0.01 and 0.60 μg·g⁻¹ creatinine, respectively). Male fetuses appeared more susceptible to EtP and MeP, and female fetuses were more susceptible to 1-H-BTR. Bayesian kernel machine regression was performed to compare the results. This study demonstrated the potential of interpretable machine learning in environmental epidemiology.