Hyperparameter Optimization in Machine Learning Models for Automated Nutritional Status Assessment in Ecuadorian Children

Abstract

Child malnutrition remains a critical public health challenge in Ecuador, where chronic child malnutrition affects 19.3% of children under two years of age according to the ENDI 2023–2024 national survey. Access to WHO growth reference tables is often limited in primary care settings, which motivates the systematic hyperparameter optimization of machine learning models presented in this study for the automated computation of WHO z-scores (HAZ, WAZ, WHZ) and the classification of nutritional risk. The analysis draws on the ENDI 2022 and 2023-2024 datasets, comprising 27,511 children aged 24-59 months. To this end, a three-phase pipeline was developed, integrating multi-output regression, WHO threshold-based rules, and supervised binary classification. Random Forest and XGBoost were evaluated across n_estimators (100-500) and max_depth (4-12), with 10 repeated trials per scenario. The optimal configuration (n_estimators=500, max_depth=8) achieved $R^{2}=0.9977-0.9979$ in regression and Accuracy=0.9898, F1=0.9695, AUC=0.9996 in classification, supporting integration as a deployable clinical decision support tool for national nutrition surveillance in Ecuador.