A Comparative Methodological Study of Automated Machine Learning for Multiclass Stunting Prediction Using Anthropometric Data

Abstract

Stunting remains a major public health challenge among children under five years old and requires reliable early screening to support timely nutritional interventions, particularly in resource-limited healthcare settings. However, many existing stunting prediction studies rely on complex socio-economic variables and manually selected machine learning models, which limits reproducibility and practical deployment. This study proposes an automated machine learning (AutoML)–based framework for multiclass stunting prediction using routinely collected anthropometric data. The prediction task is formulated as a multiclass classification problem encompassing normal growth, stunted, severely stunted, and above-normal nutritional status. The proposed framework integrates standardized preprocessing, systematic model comparison, stratified 10-fold cross-validation, and controlled hyperparameter optimization, evaluated under SMOTE and non-SMOTE preprocessing scenarios. Experimental results demonstrate that reliable multiclass prediction can be achieved without socio-economic variables. Under SMOTE preprocessing, the optimized k-Nearest Neighbors model improves minority-class sensitivity, increasing accuracy from 0.9806 to 0.9820 with an MCC of 0.9688, while under non-SMOTE conditions, Random Forest achieves robust performance with an accuracy of 0.9985 and an MCC of 0.9975 without resampling. Confusion matrix, ROC, and learning curve analyses confirm strong discriminative capability and stable generalization for both models. Overall, the findings indicate that the proposed AutoML-based framework provides a practical, scalable, and reproducible solution for early multiclass stunting screening using anthropometric data alone.

GS Cited Analysis