A 13-year cohort study using clinical machine learning to differentiate bacterial and viral infections in young infants in a dengue hyperendemic region.

Abstract

Differentiating bacterial from viral infections in febrile young infants is challenging, particularly in dengue-hyperendemic regions. We developed and internally validated a clinical machine-learning model to enhance diagnostic accuracy in this risk population in Colombia. We retrospectively analyzed a pediatric infectious admission cohort (<18 years) at a reference hospital in southern Colombia from 2007 to 2019. 4671 admissions (2251 bacterial and 2420 viral) were included. Nine clinical and laboratory variables were used to train an eXtreme Gradient Boosting (XGBoost) classifier. We divided the data into development (70%) and test (30%) sets, with Youden's J statistics defining the optimal threshold. Penalized logistic regression (LR) and single-marker rules [leukocytosis, C-reactive protein (CRP)] served as comparators. The young-infant XGBoost achieved an area under the receiver-operating characteristic curve (AUC) of 0.896, outperforming LR (0.790) and single markers (0.746-0.706). Sensitivity was 93.5%, specificity 76.1%, positive predictive value 87.9%, and negative predictive value 86.4% in the temporal validation cohort. Discrimination was highest in children aged 6-10 years (AUC 0.967). CRP positivity, leukocytosis >16 × 10³ µl-1, and thrombocytopenia <150 × 10³ µl-1 were the most informative features. A nine-variable XGBoost model using routine clinical and hematologic variables accurately differentiated bacterial from viral infections in children from a low-resource dengue-endemic setting. Performance remained stable during temporal validation. Improved specificity with preserved sensitivity supports earlier targeted therapy and antibiotic stewardship. Multicenter studies and exploration of clinical challenges are the next steps for this kind of tool.