Peer-reviewed article
Explainable Machine Learning for Current Stunting Status Classification in Toddlers: A Comparative Evaluation of SMOTEENN and SHAP-Based Interpretation
Siti Nurhasanah Nugraha, Muji Ernawati, Evita Fitri, Tri Wahyu Lestari ·
Summary
Written by AI, not yet AICS's view · claude-sonnet-5-5This study tests machine learning models that classify whether a toddler is currently stunted. It combines two secondary datasets and compares four algorithms trained on the original data and on data rebalanced with a method called SMOTEENN. Random Forest trained on the original data performed best. Rebalancing improved recall in every model but generally lowered overall performance in the tree-based models. SHAP analysis, used to explain the best model, pointed to height and age in months as the most influential features.
Key points
- Two secondary datasets were merged into 41,176 records, and 41,173 were kept after cleaning.
- Random Forest on the original data did best: accuracy 0.9600, precision 0.9584, F1-score 0.9505, ROC-AUC 0.9927.
- SMOTEENN consistently raised recall across all models but generally reduced overall performance of tree-based models.
- The largest gain was in Logistic Regression, where recall rose from 0.7058 to 0.8657.
- SHAP identified Height and Age (Month) as the most influential features.
Limitations
The abstract describes a classification of current stunting status within the two datasets analysed, so it cannot tell us how the models would perform on other populations or predict future stunting. It does not name the country or report any intervention or effect on children's outcomes; the Indonesia attribution comes from the metadata hint.
Abstract
Stunting remains a major nutritional problem that requires accurate and reliable assessment of children’s current growth status. Machine learning can support stunting classification; however, class imbalance and limited model interpretability remain important challenges. This study aims to evaluate the effect of SMOTEENN across classification algorithms with different learning characteristics and interpret the best-performing model using SHAP. Two secondary datasets were integrated, resulting in 41,176 records, of which 41,173 were retained after data cleaning and numerical conversion. Logistic Regression, Decision Tree, Random Forest, and XGBoost were evaluated using original and SMOTEENN-balanced training data. Hyperparameter optimization was performed using RandomizedSearchCV with stratified five-fold cross-validation and ROC-AUC as the optimization metric. The optimized configurations were fixed and applied to both experimental scenarios. Model performance was evaluated using Accuracy, Precision, Recall, F1-Score, and ROC-AUC on an independent test set. Random Forest trained on the original data achieved the best overall performance, with an Accuracy of 0.9600, Precision of 0.9584, F1-Score of 0.9505, and ROC-AUC of 0.9927. SMOTEENN consistently improved Recall across all models but generally reduced the overall performance of tree-based models. The largest improvement occurred in Logistic Regression, with Recall increasing from 0.7058 to 0.8657. SHAP analysis identified Height and Age (Month) as the most influential features. These findings show that the effect of SMOTEENN depends on dataset and algorithm characteristics, while SHAP provides transparent interpretation of current stunting-status classification within the analyzed datasets.