GBAS: A Hybrid GRU–BiLSTM–Attention–SVM Framework for Multiclass Stunting-Status Classification in Riau Province, Indonesia

Stunting-status classification using structured anthropometric data presents several challenges, including class imbalance, heterogeneous predictor characteristics, limited sample size, and the risk of target-derived information artificially inflating model performance. This study proposes GRU–BiLSTM–Attention–SVM (GBAS), a hybrid framework that combines recurrent deep representation learning, bidirectional processing, attention-based representation weighting, and margin-based classification for multiclass stunting-status classification. The study used 569 cross-sectional child-growth records from Riau Province, Indonesia, comprising Normal, Stunted, and Severely Stunted classes. Before model development, data-quality and leakage controls were applied, including measurement-unit standardization, exclusion of variables with substantial unavailable information, control of repeated feature combinations during partitioning, and removal of target-derived nutritional indicators. The remaining predictors were numerically encoded and standardized using parameters estimated from the training data, while class weighting was applied to address class imbalance without synthetic oversampling. In GBAS, GRU and BiLSTM operate as representation-learning components over a consistently ordered structured-feature representation, attention weights the resulting latent representations, and SVM performs final classification. On the held-out test set of 114 observations, GBAS correctly classified 113 observations, achieving 99.12% accuracy, 99.13% weighted precision, 99.12% weighted recall, and 99.12% weighted F1-score. Under the same experimental setting, GBAS outperformed the evaluated SVM, GRU, and BiLSTM baselines, which achieved accuracies of 89.46%, 95.78%, and 97.30%, respectively. These findings support the feasibility of the integrated GBAS framework for structured stunting-status classification. However, the results remain specific to the Riau dataset, feature ordering, and experimental configuration and require further validation through larger datasets, stronger tabular baselines, feature-order sensitivity analysis, component-level ablation, and external evaluation