<![CDATA[Accurate prediction of stunting among toddlers plays a vital role in enabling early intervention and supporting data-driven public health planning. Nonetheless, constructing robust predictive models remains a challenge due to significant class imbalance and heterogeneous data characteristics derived from multiple sources. This research proposes an enhanced Deep Neural Network (DNN) model for multiclass stunting classification by integrating advanced data balancing (SMOTE, SMOTE-NC, and SMOTE-ENN). The model was trained using anthropometric data obtained from both primary and secondary datasets. Experimental evaluations indicate that the combination of the DNN and SMOTE-ENN model yielded optimal performance, obtaining an accuracy of 98.57%, precision of 98.59%, F1-Score of 98.58%, and ROC-AUC of 99.97%. The confusion matrix analysis showed a similar misclassification pattern, predominantly between the Severely Stunted and Stunted classes and between the Normal class and the adjacent growth-status classes, which suggests a considerable overlap of features between adjacent classes. Furthermore, the smooth training and validation loss curves show that the proposed model can effectively avoid the underfitting and overfitting problems, guaranteeing the reliability and robustness of the prediction. Overall, the combination of DNN+SMOTE-ENN was the best approach to deal with class imbalance, enhance class separation, and boost the performance of multiclass stunting classification.]]>
