Enhancing Nutritional Predictive Models: Addressing Class Imbalance with Machine Learning

Malnutrition remains a critical public health concern, particularly in low-resource settings where early detection is essential yet often constrained by limited infrastructure. While machine learning (ML) has emerged as a promising tool for nutritional risk prediction, many existing models fail to address class imbalance, resulting in biased outcomes and poor minority class detection. This study introduces an optimized ML framework that integrates imbalance-handling techniques—specifically SMOTE and Bagging—into the classification of stunting, wasting, and underweight among children in Banjarmasin, Indonesia. A curated dataset from 26 community health centers was used to train and evaluate five algorithms (Neural Network, Random Forest, Decision Tree, Logistic Regression, and XGBoost) across three treatment phases. Performance was assessed using 10-fold cross-validation and multi-method statistical validation, including ANOVA, Kruskal-Wallis, Dunn’s, and Friedman tests. XGBoost consistently outperformed other models, achieving the highest accuracy (90.7%) and F1 scores across all indicators. The integration of oversampling and ensemble methods yielded substantial improvements in minority class detection, with F1 score gains ranging from 1.15% to 419.42%. Spatial validation revealed regional disparities, underscoring the need for adaptive modeling strategies. These findings contribute to the development of scalable, equitable, and context-aware nutritional surveillance systems, offering actionable insights for targeted interventions and public health policy.