Imbalance-aware explainable machine learning for concurrent child undernutrition modeling

Abstract

Abstract Child undernutrition remains a significant public-health issue in Ethiopia, shaped by socioeconomic, environmental, and health-related factors. This study develops a machine-learning (ML) classification model that accounts for concurrent nutritional outcomes in male and female children. Using baseline data from the Young Lives Cohort Study across five Ethiopian regions, we applied Random Forest (RF) for feature selection and trained multiple ML algorithms with k-fold cross-validation. To address class imbalance, we used the Synthetic Minority Oversampling Technique (SMOTE). Performance was evaluated with accuracy, sensitivity (recall), specificity, and F1-score. RF achieved the strongest cross-validation performance (mean accuracy 94.9%; Macro-F1 0.932), with hold-out test accuracy of 85.7% and Macro-F1 of 0.780, followed by Gradient Boosting (CV accuracy 87.8%; hold-out 82.0%). The preprocessing pipeline notably improved classification of underrepresented concurrent categories, such as US (underweight + stunted) and UW (underweight + wasted). Confusion-matrix analyses showed high accuracy overall, with correct classifications predominating in normal (N) and wasted (W) categories. However, 6 males and 10 females classified as Normal (N) were misclassified as undernourished, and 2 instances of stunting (S) were correctly identified for both males and females, highlighting residual challenges in borderline cases. The proposed models provide a reliable framework for predicting child nutritional outcomes from socioeconomic and health-related features. In particular, the RF-based system offers a robust tool for classifying undernutrition, including concurrent conditions, in Ethiopia. High performance supports early identification of at-risk children and can inform targeted nutritional interventions and community outreach.