Interpretable machine learning identifies ozone as a key predictor of expanded small-for-gestational-age phenotype across specific gestational window

Prenatal ambient air pollution has been associated with impaired fetal growth; however, the gestational window most informative for predicting expanded small for gestational age (SGA) phenotype remains unclear. Previous studies have often relied on whole-pregnancy exposure, focused on outcomes other than SGA, or emphasized prediction without integrating model-derived importance with epidemiologic inference. This study aimed to develop a trimester-specific, interpretable machine learning (ML) framework to predict expanded fetal-growth vulnerability phenotype using Kriging-derived ambient air pollution estimates and to compare predictive feature importance with adjusted epidemiologic associations. We analyzed 4274 mother–infant pairs from the Korean CHildren’s ENvironmental health Study (Ko-CHENS). The primary outcome was expanded SGA phenotype, defined as birth weight below the 20th percentile for gestational age and sex; sensitivity analyses used the conventional < 10th percentile definition. Trimester-specific and full-pregnancy datasets were constructed using Kriging-based estimates of sulfur dioxide (SO₂), carbon monoxide (CO), ozone (O₃), nitrogen dioxide (NO₂), particulate matter ≤ 2.5 μm (PM₂.₅), and particulate matter ≤ 10 μm (PM₁₀). Seven ML classifiers evaluated using receiver operating characteristic and precision–recall curves, calibration analysis, and lift curves. Model interpretability was assessed using Shapley Additive Explanations (SHAP), partial dependence plots, interaction analyses, and local explanation plots. Pollutant-specific SHAP rankings were compared with adjusted logistic regression estimates. Extreme Gradient Boosting showed the highest predictive performance. The first trimester achieved the highest discrimination (AUC 91.97%), followed by the third trimester (90.11%), full pregnancy (89.73%), and second trimester (88.27%). Average precision was also highest in the first trimester (78.14%). Calibration and lift analyses favored early pregnancy. SHAP analyses identified O₃, PM₂.₅, and PM₁₀ as influential features, with the strongest alignment between ML importance and regression estimates observed in the third trimester. Expanded fetal-growth vulnerability phenotype prediction was gestationally structured, with early pregnancy providing the strongest predictive signal and late pregnancy showing the greatest agreement between predictive and epidemiologic inference.