CLaGAtt: A Hybrid CNN-LSTM-GRU-Attention Model for Stunting Classification Based on Anthropometric Sequences

Abstract

Stunting is a chronic nutritional problem that requires accurate early identification because it affects child growth, cognitive development, and long-term human capital. This study adapts the CLaGAtt model, a hybrid CNN–LSTM–GRU–Attention architecture, for stunting classification using anthropometric sequence data. Rather than proposing a new deep learning architecture, the main contribution of this study lies in adapting the existing CLaGAtt framework through an integrated preprocessing pipeline, sequence construction strategy, class balancing using SMOTE, and an evaluation protocol specifically designed for stunting prediction. The preprocessing pipeline included irrelevant-column removal, data transformation, label encoding, standard scaling, class balancing using SMOTE, and sequence generation with a time step of five and a step of one. Three train–test split scenarios were evaluated, namely 90:10, 80:20, and 70:30. Experimental results showed that the 90:10 split produced the best performance, with 91.42% accuracy, 91.50% precision, 91.50% recall, and 91.43% F1-score. The 80:20 and 70:30 scenarios achieved 87.14% and 85.71% accuracy, respectively, indicating that larger training proportions improved model generalization in the available dataset. These findings suggest that the adapted CLaGAtt framework can effectively integrate convolutional feature extraction, sequential learning, and temporal attention for stunting classification from structured anthropometric data. Future work should validate the model on external datasets and integrate regional visualization to support priority intervention mapping.