Predicting Stunting Risk in Toddlers Using Interpretable Machine Learning on Imbalanced Data with Random Forest

Abstract

Purpose: The prevalence of stunting in toddlers is still significant at over 20%, according to the Indonesian Nutritional Status Survey (SSGI). By creating machine learning based early prediction models especially made to handle unbalanced datasets, this project seeks to hasten the elimination of stunting.

Methods: Three hybrid data balancing methods SMOTE, SMOTE + Tomek Links, and SMOTE + ENN were used in conjunction with a Random Forest algorithm. An anthropometry dataset comprising 40.444 toddlers was used to train and evaluate the models. Data cleaning, labeling, and imbalance handling comprised preprocessing. Accuracy, precision, recall, F1-score, and specificity metrics produced from a confusion matrix were then used for evaluation.

Findings: Four key features significantly influence classification: ZSTB/U (0.698961), Height (0.116384), Weight (0.102773), and LiLA (0.035833). The Random Forest algorithm, paired with SMOTE-based techniques, achieved near-perfect performance across all metrics (approaching 1.0). This demonstrates excellent capability in accurately distinguishing the nutritional status of toddlers.

Originality: This study offers a scientific explanation of the main stunting variables as well as a high-performance classification methodology. It offers a strong framework for early stunting diagnosis in Indonesia by successfully correcting class imbalance through SMOTE, SMOTE-Tomek, and SMOTE-ENN.