An explainable spatial machine learning approach for modeling and predicting spatial patterns of stunting in Central Java, Indonesia

This study proposes an explainable spatial machine-learning framework for modelling district/city stunting prevalence in Central Java, Indonesia. The analysis integrates spatial feature engineering, LASSO-based variable selection, and a focused comparison of XGBoost, CatBoost, and LightGBM under leave-one-out cross-validation. The candidate feature space comprised local public-health indicators and queen-contiguity spatial-lag predictors, reflecting the spatial nature of stunting and the high predictor-to-sample ratio. LASSO selected 10 predictors from 70 candidates. The best predictive performance was obtained by CatBoost (RMSE=2.82, MAE=2.22, R²=0.619), followed by XGBoost and LightGBM. Model interpretation using permutation importance, SHAP, and partial dependence indicates that neighboring vitamin A coverage, facility-based delivery, ANC K6, and low-birth-weight indicators are central to the spatial pattern of stunting. The findings show that explainable spatial machine learning can complement conventional spatial regression by producing validated predictions, residual spatial diagnostics, and policy-readable explanations for geographically targeted intervention.