Predicting Pediatric Mortality Across Five Intensive Care Units: Toward an Early Warning Using Machine Learning

Area of Science:

• Pediatric critical care medicine
• Machine learning applications in healthcare
• Biomedical data science

Background:

• Early identification of high-mortality risk in pediatric intensive care units (ICUs) remains a challenge.
• Heterogeneous ICU settings complicate risk stratification efforts.
• Existing methods may not fully leverage complex patient data for timely intervention.

Purpose of the Study:

• To develop and evaluate unit-specific machine learning (ML) models for early mortality risk prediction in pediatric ICUs.
• To assess the feasibility of data-driven early-warning systems tailored to different ICU environments.
• To identify key clinical and laboratory predictors of mortality in pediatric critical care.

Main Methods:

• Trained Random Forest (RF) classifiers separately for surgical (SICU), cardiac (CICU), general, neonatal (NICU), and pediatric (PICU) wards.
• Utilized admission data including demographics, diagnoses, medications, and laboratory features.
• Evaluated model performance using Area Under the Curve (AUC) metrics.

Main Results:

• Achieved strong discriminative performance with AUC values ranging from 0.86 to 0.97 across different units.
• Identified key predictors of mortality, including Lactate (LAC), red-cell distribution width (RDW), platelets (PLT), hemoglobin (Hb), and creatinine (Cr).
• Demonstrated the effectiveness of unit-specific ML models in diverse pediatric ICU settings.

Conclusions:

• Unit-specific machine learning models are feasible and effective for early mortality risk prediction in pediatric ICUs.
• Data-driven early-warning systems can enhance clinical decision-making and patient management.
• Key laboratory markers play a significant role in identifying high-risk pediatric patients.