Using explainable machine learning to better understand the environmental and socioeconomic contributions to preterm birth

Preterm birth (PTB) remains a leading cause of child mortality, yet the role of ambient air pollution remains disputed. In a cohort of 52,642 singleton births in Southwest China (2020–2023), we combined an automated machine-learning (AutoML) method with SHapley Additive exPlanations (SHAP) to rank and quantify 12 environmental, clinical, and sociodemographic predictors of PTB. Environmental factors collectively explained 48.5 % of the importance of the model features, with residential ambient PM_2.5 (20.7 %), elevation (17.3 %), and the normalized difference vegetation index (NDVI, 10.5 %) emerging as the top three contributors. The exposure–response curve demonstrated a marked increase in PTB risk above a PM_2.5 threshold of 50 μg/m³, with a mean SHAP value of 1.81 (CI: 1.75–1.87). The adverse effects of PM_2.5 were amplified among mothers with low educational levels (mean SHAP value 1.85 vs. 1.78 in the high education group) and varied by infant sex, with female infants exhibiting greater susceptibility when PM_2.5 concentrations were in the highest exposure bin (50–80 μg/m³). This study introduces a reproducible AutoML–SHAP framework for comprehensive PTB risk quantification and highlights that stringent air-quality control, coupled with targeted interventions, could substantially reduce prematurity.