Machine learning flags which children with mononucleosis face longer hospital stays

Infectious mononucleosis, the classic glandular fever driven by Epstein-Barr virus, is usually a self-limited illness in children, yet a meaningful fraction of young patients end up needing more than a week of inpatient care. For pediatricians, knowing at the moment of admission which child will follow the short, uneventful course and which will drift into a prolonged hospitalization has long been a matter of clinical intuition. A new retrospective study from the Children’s Hospital of Nanjing Medical University, published in BMC Infectious Diseases, now offers a data-driven alternative: a compact machine-learning model built from just five admission-time variables that can identify children at elevated risk of a lengthy stay with moderate accuracy.

The research team, Juxin Ye and Mingqi Peng, assembled a cohort of 499 children with confirmed infectious mononucleosis admitted between April 2024 and March 2025. Prolonged length of stay was defined as a hospital stay exceeding seven calendar days, a threshold that captures the inpatient course requiring extended supportive care, monitoring of complications, or management of organ involvement. Across the full cohort, prolonged stay occurred in 144 children, or 28.9 percent, confirming that nearly one in three hospitalized children with this viral illness faces a substantially longer admission than the typical self-limited course would suggest.

What distinguishes this study from many earlier prediction-model efforts is its temporal design. Rather than splitting the data randomly, the authors locked aside all admissions from January to March 2025 as a temporal test cohort of 79 children containing 23 prolonged-stay events. All the messy decisions of model construction, including missing-data handling, feature selection, hyperparameter tuning, and threshold selection, were performed exclusively on the development cohort of 420 admissions from April to December 2024. This locked temporal evaluation matters because models that look impressive on randomly partitioned data often degrade when applied to genuinely later patients, where seasonal patterns, circulating viral strains, and clinical practices may shift.

Feature selection relied on L1-penalized logistic regression, a technique that adds a penalty to the model’s coefficients so that weak predictors are driven to exactly zero, effectively pruning the variable set while fitting the data. The authors compared seven machine-learning approaches in total, spanning penalized logistic regression alongside ensemble methods such as extreme gradient boosting, light gradient boosting machine, and gradient boosting decision trees, as well as support vector machines. Synthetic minority over-sampling, or SMOTE, was used within the development data to address class imbalance, since prolonged-stay events made up less than a third of admissions and naive models would otherwise be biased toward predicting short stays.

When the dust settled, the parsimonious L1-penalized logistic regression model emerged with five retained predictors, all available at admission: peak fever recorded before admission, aspartate aminotransferase, direct bilirubin, the percentage of CD3-positive CD4-positive T cells, and cystatin C. These variables sketch a coherent clinical picture. Peak fever reflects the intensity of the viral illness at presentation. Aspartate aminotransferase and direct bilirubin are markers of hepatic involvement, a recognized complication of Epstein-Barr virus infection in children. The CD4 T-cell percentage captures the degree of immune perturbation, since infectious mononucleosis is fundamentally a disease of disordered T-cell and B-cell interaction. Cystatin C, a sensitive marker of glomerular filtration rate, hints at renal function and systemic inflammatory burden.

Performance was assessed across discrimination, classification, and calibration metrics. In the locked temporal test cohort, the penalized logistic regression model achieved an area under the receiver operating characteristic curve of 0.785, with a 95 percent confidence interval of 0.661 to 0.893. The six competing models produced temporal-test AUCs ranging from 0.700 to 0.778, meaning the simplest model performed as well as or better than its more complex counterparts. This outcome illustrates a recurring theme in clinical machine learning: gradient-boosted ensembles and support vector machines often deliver marginal gains on development data but fail to justify their added complexity, opacity, and overfitting risk when confronted with temporally separated patients.

To describe how the model behaves, the authors turned to Shapley Additive Explanations, a framework borrowed from cooperative game theory that assigns each predictor a contribution value for individual predictions. SHAP analysis allows clinicians to see not just that a model predicts prolonged stay but which variables push a given child’s risk upward or downward, addressing some of the black-box concerns that have slowed the adoption of machine learning in pediatric medicine. Decision-curve analysis, another evaluation tool, was used to gauge the clinical net benefit of acting on the model’s predictions at various risk thresholds, ensuring the model is not merely statistically significant but potentially useful for triage decisions.

The authors are careful to frame their findings as proof of concept rather than a deployable clinical tool. Independent external validation, preferably prospective and multicenter, is required before any clinical use, and the model’s calibration and thresholds were tuned to a single hospital’s population and practices. The retrospective single-center design, the modest size of the temporal test cohort, and the absence of external data all temper the enthusiasm with which the results should be interpreted. Still, the study provides a template for how pediatric infectious-disease research can move from intuition toward risk-stratification frameworks, using variables that are already collected during routine admission workups.

For clinicians on the wards, the practical message is that a handful of admission-time laboratory values, including liver enzymes, bilirubin, lymphocyte subsets, and cystatin C, together with the trajectory of fever before admission, carry real prognostic information about the likely inpatient course of infectious mononucleosis in children. For researchers, the study demonstrates that temporally locked evaluation, penalized feature selection, and explainable machine learning can be combined into a rigorous pipeline for prognosis research in pediatric virology. As machine-learning methods continue to permeate predictive medicine, studies like this one, which balance technical sophistication with clinical parsimony and honest temporal testing, offer a realistic path toward early risk stratification that could eventually help pediatricians plan discharge conversations, allocate beds, and counsel families from the very first day of admission.

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

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

Article References: Ye, J., & Peng, M. (2026). Machine learning for predicting prolonged length of stay among children with infectious mononucleosis. BMC Infectious Diseases. https://doi.org/10.1186/s12879-026-14408-x

Image Credits: AI Generated

DOI: 10.1186/s12879-026-14408-x

Keywords: Machine, learning, predicting, prolonged, length, stay, children, infectious, mononucleosis, scientific research