Deep learning outperforms traditional machine learning methods in predicting childhood malnutrition: evidence from survey data

Childhood malnutrition remains a major public health concern in Nepal and other low-resource settings, while conventional case-finding approaches are labor-intensive and frequently unavailable in remote areas. This study provides the first comprehensive assessment of machine learning and deep learning methodologies for identifying malnutrition among children under five years of age in Nepal. We systematically compared 16 algorithms spanning deep learning, gradient boosting, and traditional machine learning families, using data from the Nepal Multiple Indicator Cluster Survey (MICS) 2019. A composite malnutrition indicator was constructed by integrating stunting, wasting, and underweight status, and model performance was evaluated using ten metrics, with emphasis on F1-score and recall to account for substantial class imbalance and the high cost of failing to detect malnourished children. Among all models, TabNet demonstrated the best performance, likely attributable to its attention-based architecture, and outperformed both support vector machine and AdaBoost classifiers. A consensus feature importance analysis identified maternal education, household wealth index, and child age as the primary predictors of malnutrition, followed by geographic characteristics, vaccination status, and meal frequency. Collectively, these results demonstrate a scalable, survey-based screening framework for identifying children at elevated risk of malnutrition and for guiding targeted nutritional interventions. The proposed approach supports Nepal’s progress toward the Sustainable Development Goals and offers a transferable methodological template for similar low-resource settings globally.

💡 Research Summary

This study evaluates the predictive performance of a wide range of machine‑learning and deep‑learning algorithms for identifying malnutrition among children under five in Nepal, using the 2019 Nepal Multiple Indicator Cluster Survey (MICS). After cleaning, 6,416 children were retained, with 42 % classified as malnourished based on a composite indicator that combines stunting, wasting, and underweight status according to WHO z‑score thresholds. Twenty‑three candidate predictors spanning child care, health, demographic, maternal, household, and geographic domains were initially selected; a rigorous feature‑selection pipeline (mutual information, chi‑square, ANOVA, Pearson correlation, variance threshold, recursive feature elimination, L1‑regularized logistic regression, and the Boruta algorithm) reduced this set to 16 variables, including maternal education, household wealth index, child age, vaccination status, meal frequency, and provincial residence.

Sixteen models were trained and evaluated under a unified cross‑validation framework: four deep‑learning architectures (standard DNN, Wide & Deep, ResNet, and TabNet), five gradient‑boosting methods (AdaBoost, CatBoost, XGBoost, LightGBM, histogram gradient boosting), and seven traditional classifiers (SVM, LDA, Random Forest, Extra Trees, Decision Tree, K‑Nearest Neighbors, Logistic Regression). Ten performance metrics were computed, with particular emphasis on F1‑score and recall to mitigate the consequences of class imbalance and to prioritize the detection of malnourished children.

TabNet emerged as the top performer, achieving an F1‑score of 0.62, precision of 0.63, recall of 0.68, and balanced accuracy of 0.62, modestly surpassing the best gradient‑boosting (AdaBoost, F1 = 0.61) and traditional (SVM, F1 = 0.61) models. Other deep‑learning models (DNN, Wide & Deep, ResNet) and gradient‑boosting variants yielded F1‑scores in the 0.55–0.59 range, while conventional classifiers generally lagged behind (0.54–0.56).

Feature‑importance analysis, integrating TabNet’s internal attention weights with importance scores from Random Forest, XGBoost, and L1‑regularized logistic regression, consistently highlighted maternal education, household wealth, and child age as the strongest predictors. Geographic factors (province, urban/rural residence), vaccination status, and meal frequency also contributed meaningfully. These findings align with existing literature on malnutrition determinants and demonstrate that attention‑based deep models can capture complex, non‑linear interactions inherent in survey data.

The authors conclude that attention‑driven deep learning, exemplified by TabNet, provides a scalable, interpretable, and highly accurate framework for survey‑based malnutrition screening in low‑resource settings. The methodology—encompassing data preprocessing that respects survey skip patterns, robust feature selection, and multi‑metric evaluation—offers a transferable template for other countries seeking to accelerate progress toward Sustainable Development Goal 2 (Zero Hunger) through data‑driven targeting of nutritional interventions.