Early prediction of antibiotic need and bacteremia risk in non-immunocompromised pediatric emergency patients using machine learning

Tom Velez, Computer Technology Associates, Cardiff, CA, USA.
Oluwakemi Badaki-Makun, Division of Pediatric Emergency Medicine, Department of Pediatrics, Johns Hopkins University School of Medicine, Baltimore, MD, USA.
Danielle Hirsch, Division of Pediatric Emergency Medicine, Department of Pediatrics, Johns Hopkins All Children's Hospital, St. Petersburg, FL, Germany.
Danielle Claire Mercurio, Division of Pediatric Emergency Medicine, Department of Pediatrics, Johns Hopkins All Children's Hospital, St. Petersburg, FL, Germany.
Holly Depinet, Division of Emergency Medicine, Cincinnati Children's Hospital Medical Center and Department of Pediatrics, University of Cincinnati College of Medicine, Cincinnati, OH, USA.
Maya Dewan, Department of Pediatrics, Division of Critical Care Medicine, Cincinnati Children's Hospital Medical Center, University of Cincinnati College of Medicine, Cincinnati, OH, USA.
Rishikesan Kamaleswaran, Department of Surgery, Duke University School of Medicine, Durham, NC, USA.
Jocelyn Grunwell, Department of Pediatrics, Division of Critical Care Medicine, Emory University School of Medicine, Atlanta, GA, USA.
Maria Triantafyllou, Children's National Research Institute, Washington, DC, USA.
Fehima Abdelrahman, Children's National Research Institute, Washington, DC, USA.
Charles Macias, Division of Pediatric Emergency Medicine, University Hospitals Rainbow Babies and Children's Hospital, and Case Western Reserve University School of Medicine, Cleveland, OH, USA.
Ioannis Koutroulis, Children's National Research Institute, Washington, DC, USA. ikoutroulis@gwu.edu.

Document Type

Journal Article

Publication Date

12-12-2025

Journal

Pediatric research

DOI

10.1038/s41390-025-04656-z

Abstract

BACKGROUND: Timely identification of serious bacterial infections in children presenting to emergency departments is critical, especially among non-immunocompromised children, where early symptoms can be nonspecific. Although many children receive empiric antibiotic treatment based on clinical suspicion, true bloodstream infection is relatively uncommon, and unnecessary antibiotics can contribute to adverse effects and antimicrobial resistance. METHODS: To support individualized decision-making, we developed and evaluated a two-part machine learning framework using retrospective electronic health record data from 5706 pediatric patients aged 3 months to 17 years across six emergency departments. The first model predicted clinical deterioration-defined as admission to intensive care, use of vasopressors, mechanical ventilation, or in-hospital death-among children in whom antibiotics were initially withheld. The second model predicted the likelihood of bacteremia among those who received early empiric antibiotics. Both models were built using XGBoost and evaluated through cross-validation. RESULTS: Performance was strong, with high area under the curve values and negative predictive values above 96%. Predictive features included supplemental oxygen use, fever, low oxygen saturation, age, and abnormal laboratory values. CONCLUSIONS: This dual-model framework offers interpretable, evidence-based support for early treatment decisions and could improve both patient safety and antibiotic stewardship in pediatric emergency care. IMPACT: This study introduces a dual machine learning framework that informs early antibiotic decisions in non-immunocompromised pediatric emergency patients. It adds a novel two-model approach: one to predict deterioration when antibiotics are initially withheld, and another to predict bacteremia in those treated. Unlike prior tools, it uses harmonized multi-center EHR data and SHAP-based explain ability to support bedside clinical use. The impact lies in enhancing antibiotic stewardship and patient safety by identifying who may benefit from early antibiotics and who may safely avoid them.

Department

Pediatrics

Link to Full Text

Share

COinS