Explainable machine learning integrating environmental chemical biomarkers and maternal clinical factors for prediction of hypertensive disorder of pregnancy.

Hypertensive disorders of pregnancy (HDP) represent a leading cause of maternal morbidity, yet environmental chemical contributors to HDP risk remain poorly characterized within predictive frameworks. We developed and evaluated an explainable machine-learning model integrating environmental biomarkers with maternal clinical characteristics for HDP prediction using data from 4,260 pregnant women enrolled in the Korean Children's Environmental Health Study (Ko-CHENS) prospective birth cohort. An Extreme Gradient Boosting (XGBoost) classifier incorporating 18 environmental biomarkers-including blood lead, cadmium, mercury, and endocrine-disrupting chemicals-alongside maternal clinical covariates was trained using stratified 60/20/20 data splits with Bayesian hyperparameter optimization. The optimized model achieved a cross-validated ROC-AUC of 0.787 and an independent test ROC-AUC of 0.748 (95% CI: 0.587-0.887), with a negative predictive value of 0.994 under severe class imbalance (HDP prevalence: 1.7%). SHapley Additive exPlanations (SHAP) analysis identified late-pregnancy blood lead concentration and pre-pregnancy body mass index as the dominant predictors, each exhibiting non-linear risk gradients; formal SHAP interaction values (1.6% of combined attribution) and an independent logistic-regression interaction term (β = 0.418, 95% CI - 0.278-1.115, p = 0.239) indicated additive rather than synergistic BMI-lead pathways. These findings demonstrate that environmental biomarkers contribute independent, non-linear predictive value beyond established clinical risk factors, supporting integration of environmental exposure surveillance into precision obstetric risk stratification.