A comparative analysis of ensemble and deep learning approaches for predicting viral gastroenteritis in paediatric patients

Viral gastroenteritis, primarily caused by rotavirus and adenovirus, remains a major contributor to paediatric morbidity and mortality worldwide. Early and accurate identification of viral etiology is essential for appropriate clinical management and antimicrobial stewardship. However, laboratory confirmation is often inaccessible in resource-limited settings. This study aimed to develop and validate a machine learning model to predict rotavirus/adenovirus-positive acute gastroenteritis in children using observable clinical features, including symptoms, demographic factors, and vaccination status. A cross-sectional analysis was conducted on 320 paediatric patients with acute gastroenteritis, of whom 39.38% tested positive for rotavirus and/or adenovirus by rapid antigen testing. Twelve original clinical variables were expanded to 24 engineered features, incorporating symptom severity scores, interaction terms, and composite risk indicators. Class imbalance was managed using SMOTE. Eleven predictive models were evaluated, including eight traditional machine learning algorithms, two ensemble approaches, and one deep neural network. Performance was assessed using ROC-AUC, accuracy, precision, recall, and F1-score with five-fold stratified cross-validation. Seven predictors showed statistical significance ( p < 0.05), including stool type, dehydration, fever, vomiting, vaccination status, weakness, and vomiting frequency. The Extra Trees classifier achieved the best performance (ROC-AUC = 0.8795; accuracy = 84.38%), with strong cross-validation results (mean ROC-AUC = 0.8886 ± 0.0409). Engineered interaction features contributed substantially to model performance. The study developed a non-laboratory-based predictive model for rotavirus/adenovirus-positive gastroenteritis that, pending external validation, shows potential to support triage-level clinical decision-making and rational antibiotic use in resource-constrained settings.