Development of a machine learning model to diagnose pediatric lower respiratory tract infections

This retrospective cohort study aimed to develop and validate a two-step machine learning (ML) approach for diagnosing pediatric community-acquired lower respiratory tract infections (CA-LRTIs) and distinguishing bacterial from viral pathogens to aid decision-making regarding antibiotic use. We employed five ML algorithms for model development and two testing sets for model evaluation. Patients < 19 years presenting with symptoms suggestive of CA-LRTIs between January 2005 and November 2023 were enrolled, excluding patients with tuberculosis, coronavirus-19 disease, or nosocomial infections. Initially, 9,329 patients were enrolled in the study; after applying 1:1 undersampling during data preprocessing, the final analytic cohort comprised 8,583 patients. Among the five ML algorithms, the random forest model demonstrated the highest diagnostic performance, with an area under the receiver operating characteristic curve (AUROC) of 0.953 (95% confidence interval [CI]: 0.944-0.962) in the initial testing set and 0.961 (95% CI: 0.946-0.975) in the temporal testing set. For pathogen classification, the random forest model achieved an AUROC of 0.918 (95% CI: 0.874-0.958). If applied clinically, this model is predicted to reduce antibiotic use from 64.8% to 48.7%, primarily by decreasing inappropriate prescriptions. The ML models demonstrated high performance in diagnosing CA-LRTIs and reducing inappropriate antibiotic use, and will be particularly valuable in resource-limited settings and for addressing antibiotic resistance.