The Two-stage PLS-SEM-LightGBM Hybrid for Causal-predictive Modeling of Stunting with Latent Feature Interaction Analysis

Stunting remains a problem in developing countries with low and middle incomes. Stunting often occurs in toddlers due to inadequate dietary patterns and portions, which disrupt their growth. This study proposes a hybrid two-stage modeling. This hybrid model integrates Partial Least Squares Structural Equation Modeling (PLS-SEM) with LightGBM, a type of gradient-boosting algorithm in machine learning. At the first process, the two-stage PLS-SEM–LightGBM hybrid employs PLS-SEM to examine latent variables. The latent variables include child growth status, dietary quality, family demographics, household environmental quality, maternal-child health behavior, parental education, and parental occupation. In the second process, the input to LightGBM consists of two components: the latent scores derived from PLS-SEM and feature engineering from the raw data. The LightGBM acts as a classifier to predict the multiclass outcome of stunting. The two-stage PLS-SEM–LightGBM hybrid enables improved prediction accuracy while maintaining interpretability by revealing causal connections between latent variables. The two-stage PLS-SEM–LightGBM hybrid has performance with an overall accuracy of 95%, a macro-averaged F1 Score of 0.91, and a macro-averaged ROC AUC of 0.99 when tested on a dataset of 1,687 children. The research results show that Height-for-age (HFA) is the most important predictor, followed by weight-for-age (WFA), height-for-age Z-score (HAZ), length, weight-for-height (WFH), age at measurement, weight, and other factors. The result of SHAP showed the impact of the features such as child growth status, household environmental quality, birth weight, and family factors. The findings demonstrate that the proposed two-stage PLS-SEM–LightGBM hybrid provides a high-performance tool for predicting stunting status.