Machine Learning for Predicting Critical Events Among Hospitalized Children
Sierra Strutz1, Huan Liang1, Kyle Carey2
1Department of Biostatistics & Medical Informatics, University of Wisconsin-Madison, Madison.
JAMA Network Open
|May 30, 2025
Summary
A new machine learning model effectively predicts critical events in hospitalized children, improving risk assessment across all hospital units. This unified approach enhances early detection and patient safety, reducing mortality and morbidity risks.
Area of Science:
- Pediatric critical care medicine
- Health informatics
- Machine learning in healthcare
Background:
- Unrecognized deterioration in hospitalized children leads to significant mortality and morbidity.
- Current pediatric risk stratification is fragmented, using different tools across hospital units (emergency, ward, intensive care).
Purpose of the Study:
- To develop a unified machine learning model for early detection of deterioration in pediatric patients.
- Enable continuous and consistent risk assessment throughout a child's hospital stay.
Main Methods:
- Retrospective cohort study of 135,621 pediatric admissions across 3 tertiary care academic hospitals.
- Developed and compared regression-based, extreme gradient-boosted machine (XGB), and deep learning models.
- Used 2 hospitals for model derivation and a third for external validation.
Main Results:
- The XGB model demonstrated superior discrimination (C statistic: 0.86) compared to existing ward-focused models (0.82 and 0.70).
- XGB required fewer alerts (6) at 80% sensitivity compared to ward models (9 and 11).
- The XGB model performed equivalently or better than unit-specific models.
Conclusions:
- A novel hospitalwide machine learning model was developed for continuous risk prediction of critical events in children.
- This model provides a unified framework for risk assessment in pediatric hospitals.
- The findings support the use of a single, adaptable model for improved pediatric patient safety.
