Predicting Intensive Care Readmission Among Hospitalized Children

Area of Science:

• Pediatric critical care medicine
• Machine learning applications in healthcare
• Patient safety and quality improvement

Background:

• Pediatric intensive care unit (PICU) readmissions are linked to worse patient outcomes.
• Predicting PICU readmissions can enable timely interventions to improve care.
• Developing accurate predictive models is crucial for patient safety.

Purpose of the Study:

• To develop and validate machine learning models for predicting PICU readmission risk at the time of patient transfer.
• To assess the generalizability of these models across different clinical sites.

Main Methods:

• Retrospective observational cohort study of 35,601 children admitted to three quaternary care PICUs (2012-2019).
• Developed and validated four models (logistic regression, elastic net, random forest, XGBoost) using vital signs, patient characteristics, and lab results.
• Primary outcome: unplanned PICU readmission within 48 hours of transfer.

Main Results:

• Internal validation showed consistent model performance (AUC 0.70-0.73) across sites.
• External validation revealed a significant decrease in performance (AUC 0.60-0.69).
• Key predictive variables varied by site, indicating limited generalizability.

Conclusions:

• Machine learning models for PICU readmission prediction exhibit limited generalizability.
• Locally derived models showed modest performance and may aid decision-making if prospectively validated.
• Models developed externally are unlikely to perform well in predicting PICU readmissions.