Machine learning for predicting prolonged length of stay among children with infectious mononucleosis

Infectious mononucleosis (IM) is usually self-limited in children but can require a prolonged inpatient course. Prediction of prolonged length of stay (PLOS) from information available at admission may inform the development of future early risk-stratification approaches; however, prediction models for pediatric IM remain limited. We therefore aimed to develop and temporally evaluate machine-learning models for PLOS in hospitalized children with IM. We conducted a retrospective single-center study of 499 children with confirmed IM admitted from April 2024 to March 2025. PLOS was defined as a hospital stay of more than 7 calendar days. Admissions from April to December 2024 formed the development cohort ( n = 420; 121 PLOS events), and admissions from January to March 2025 formed the locked temporal test cohort ( n = 79; 23 PLOS events). Missing-data handling, feature selection, hyperparameter tuning, and threshold selection were performed in development data only. Feature selection was performed using L1-penalized logistic regression, and seven machine-learning (ML) models were compared. Performance was assessed using discrimination, classification, and calibration metrics. Shapley Additive Explanations (SHAP) were used to describe model behavior. PLOS occurred in 144 of 499 children (28.9%). The L1-penalized logistic regression model retained peak fever before admission, AST, DBIL, percentage of CD3 + CD4+ T cells, and cystatin C. In the locked temporal test cohort, it achieved an AUC of 0.785 (95% CI 0.661–0.893), and the other six models had temporal-test AUCs ranging from 0.700 to 0.778. A parsimonious L1-penalized logistic regression model using five predictors available at admission in this hospital showed moderate discrimination for PLOS in a locked temporal test cohort. These findings provide proof of concept for future risk-stratification research in this setting. Independent external, preferably prospective multicenter, validation is required before clinical use. Not applicable.