Lightweight Hybrid Feature Engineering and Fusion with Variational Autoencoder and XGBoost for Nutritional Data Classification

Stunting remains a significant global health challenge, particularly among children under five years old. Early detection of nutritional status is crucial to prevent long-term impacts on both physical and cognitive development. This study proposes a lightweight approach based on hybrid feature engineering and fusion for stunting status classification without requiring complex data balancing strategies. The proposed method integrates basic anthropometric features with Rule-Based (RB) derived features (BMI, weight, and height changes since birth, growth rates, and z-score combinations), along with latent features generated by a Variational Autoencoder (VAE) that captures non-linear patterns in tabular data. All features are then concatenated and classified using XGBoost. The evaluation with 5-fold cross-validation demonstrated highly competitive performance, yielding an average F1-score of 0.9919 ± 0.0066 on the testing data. These findings highlight that a lightweight yet informative approach can serve as a practical solution to support early stunting detection, particularly in public health settings with limited resources.