Prediction of child nutritional status using deep neural networks: A cross-sectional study of Egypt DHS data (2005–2014)

Abstract

Child malnutrition, particularly wasting and undernutrition, remains a pressing public health issue in low- and middle-income countries. Timely prediction of nutritional status using accessible demographic and health indicators can support early interventions. This study explores the use of a deep neural network (DNN) model to predict child nutritional status using nationally representative survey data from Egypt. To develop and evaluate the performance of a DNN model for classifying child nutritional status and to compare its performance against traditional machine learning models, including decision tree and Random Forest classifiers. This is a retrospective cross-sectional study based on pooled data from the Egypt Demographic and Health Surveys conducted in 2005, 2008, and 2014. A total of 36, 313 children under the age of 5 with complete anthropometric and demographic data were included. A deep neural network was trained on the Egypt Demographic and Health Surveys dataset using class weighting and synthetic minority oversampling technique to address class imbalance. Performance was assessed using accuracy, precision, recall, harmonic mean of precision and recall (F1-score), and receiver operating characteristic (ROC)–area under the curve (AUC) and compared against decision tree and random forest classifiers. The DNN outperformed all baseline models, achieving an accuracy of 89%, recall of 91%, and an ROC–AUC of 0.95. SHAP analysis revealed that maternal body mass index, child age, birth weight, and household wealth index were the most influential features. The random forest achieved a ROC–AUC of 0.95 but showed lower recall and F1-scores compared to the DNN. The DNN model demonstrated high performance in predicting child nutritional status and shows promise as a public health screening tool. Future work should focus on external validation across different populations and enhanced interpretability for clinical deployment.