Derivation and validation of clinical prediction models for Cryptosporidium-attributed acute diarrhea in African children.

Abstract

Clinical decision-support tools have the potential to guide empiric treatment and prioritize diagnostic testing at the individual-level, and screening strategies for vaccine and therapeutic trials at the population-level. Such tools for the parasitic pathogen Cryptosporidium, a leading cause of morbidity and mortality among children in low- and middle-income countries (LMICs), are lacking. We aimed to develop clinical prediction models to identify Cryptosporidium-attributed acute diarrhea in young children. For model derivation, we used clinical and demographic data from the Global Enteric Multicenter Study (GEMS) and integrated site-specific weather data from the NOAA database. Random forest (RF) and logistic regression (LR) models were trained and evaluated using 5-fold cross-validation. Using only the top three predictors, prediction models achieved a mean area under the curve (AUC) of 0.77 for RF and 0.75 for LR models. RF classification identified mean rainfall and mean temperature in the 30 days preceding enrollment, along with patient age, as the most important predictors of Cryptosporidium-attributed diarrhea. For external validation, models were trained on the full GEMS dataset and tested on the Vaccine Impact on Diarrhea in Africa (VIDA) dataset, with the LR models having better calibration and net benefit. In conclusion, we derived and externally validated clinical prediction models for identifying Cryptosporidium infections among LMIC children with acute diarrhea. Our findings highlight the value of incorporating weather variables into clinical prediction models for diarrheal diseases, and their potential use for real-time decision support in the context of climate change and extreme weather events.