Integrating Environmental Data, Spatial Epidemiology, and Machine Learning to Characterize Childhood Stunting in Bangladesh

Childhood stunting remains a major public health challenge in Bangladesh. We integrated Bangladesh Demographic and Health Survey 2022 child data with geographic coordinates and environmental indicators to characterize spatial structure and evaluate predictive modeling. Stunting was defined as a height-for-age z-score < -2 SD among children aged 0-59 months. Spatial analyses included global and local autocorrelation, hotspot analysis, and multiscale spatial regression. Machine learning used nonoverlapping primary sampling unit groups for model development, calibration, and held-out evaluation, with survey-weighted fitting and alternative feature-selection pipelines. The Machine Learning analysis included 3,620 children from 666 primary sampling units. Survey-weighted stunting prevalence in the complete case sample was 22.84% (95% CI 21.43-24.60%). Spatial autocorrelation was weak and sensitive to neighborhood definition. Multiscale Geographically Weighted Regression achieved the highest in-sample R² and adjusted R² among the fitted global OLS, GWR, and MGWR specifications (R² = 0.1884; adjusted R² = 0.1621). Grouped backward feature selection with AdaBoost had the numerically highest internal held-out comparative ROC-AUC (0.6773; 95% CI 0.6276-0.7259, without demonstrable superiority over the pipelines. GeoShapley identified geographic and feature-by-geography predictive components within the reduced explanatory surrogate.