Supervised machine learning for classification and prediction of stunting among under-five Egyptian children.

INTRODUCTION

Stunting, a significant form of chronic undernutrition, affects millions of children under five worldwide and poses substantial challenges to physical, cognitive, and socioeconomic development—particularly in low- and middle-income countries like Egypt.

AIMS

This study aims to apply and compare the performance of various supervised machine learning (ML) algorithms to classify and predict stunting among Egyptian children under five years old. It also aims to identify key risk factors that contribute to stunting.

METHODS

Data from the Egypt Demographic and Health Surveys (DHS) conducted in 2005, 2008, and 2014 were used. After extensive data cleaning and preprocessing—including handling missing values and addressing class imbalance—five ML classifiers (XGBoost, Logistic Regression, Random Forest, Gradient Boosting, and K-Nearest Neighbors) were trained and evaluated using 10-fold stratified cross-validation, performance metrics included accuracy, precision, recall, F1 score, and ROC-AUC.

RESULTS

Gradient Boosting and Random Forest achieved the highest predictive performance, with accuracy scores exceeding 90% and ROC-AUC values above 0.96. Logistic Regression also performed robustly, while K-Nearest Neighbors showed relatively lower performance due to sensitivity to noise and high-dimensional data Significant predictors of stunting included the child’s nutritional status, maternal education, birth size, wealth index, and rural residence.

CONCLUSION

The application of supervised machine learning, especially with the Gradient Boosting and Random Forest techniques, showed excellent accuracy in predicting stunting in children under five years of age in Egypt. The results of this study highlight the utility of machine learning in identifying vulnerable groups for targeted public health interventions. Further studies are encouraged to utilize more recent data and focus on multi-level feature selection and hyperparameter optimization to improve prediction precision further.