Derivation and validation of clinical prediction models for viral etiologies of acute diarrhea in North American children presenting for emergency care

Background: Diarrheal illness in children leads to 3.5 million care visits and 200,000 hospitalizations annually in the US. Viruses are responsible for most pediatric diarrheal cases, yet limited guidance on distinguishing viral from bacterial etiologies complicates clinical decision-making, especially regarding empiric antibiotic use. Methods: We used clinical and qualitative molecular etiologic data from the Implementation of Molecular Diagnostics for Pediatric Acute Gastroenteritis (IMPACT) study to develop prediction models for viral etiology of diarrhea. We used conditional random forests to identify informative clinical and environmental predictors and evaluated model performance using logistic regression and random forests within a 5-fold cross-validation framework. We conducted external validation using the Alberta Provincial Pediatric Enteric Infection Team (APPETITE) dataset. Results: Variables predictive of viral etiology included younger age, non-bloody diarrhea, winter season, and presence of vomiting. External validation showed that an AUC of 0.82 can be achieved with a parsimonious 5-variable model, yielding a sensitivity of 0.92 and specificity of 0.55 Conclusion: Our results suggest that in North American healthcare settings, clinical prediction models can inform decision-making by identifying children with a high probability of viral diarrhea, improving diagnostic clarity, and reducing unnecessary testing and treatment.