Mortality prediction models for pediatric intensive care: comparison of overall and subgroup specific performance

Area of Science:

• Pediatric critical care medicine
• Clinical epidemiology
• Healthcare outcomes research

Background:

• Accurate mortality prediction is crucial for resource allocation and clinical decision-making in pediatric intensive care units (PICUs).
• Existing models like the Pediatric Index of Mortality (PIM) and Pediatric Risk of Mortality (PRISM) require validation in diverse patient populations and settings.

Purpose of the Study:

• To validate the performance of PIM and PRISM model variants for mortality prediction in a Dutch PICU population.
• To assess model performance across various subgroups, including neonates, different diagnoses, admission urgency, and length of stay (LoS).

Main Methods:

• Comparative analysis of PIM and PRISM model variants using calibration and discrimination metrics (Area Under the Curve - AUC).
• Evaluation of model performance in the overall cohort and predefined subgroups (diagnoses, age, urgency, LoS).
• Data from 12,040 admissions of patients <16 years in 8 Dutch PICUs (2006-2009) were analyzed, excluding referred or rapidly deceased patients.

Main Results:

• PIM2 variants demonstrated the best calibration. All models showed good discrimination (AUC > 0.83), including in neonates.
• PRISM3-24 exhibited the highest discrimination overall (AUC 0.90) and in most subgroups (13/14).
• Model performance, particularly discrimination (AUC < 0.73), declined for patients with LoS > 6 days.

Conclusions:

• PIM2 and PRISM3-24 (after recalibration) are suitable for individualized mortality risk prediction in Western European PICUs.
• Both models demonstrate good discrimination in most subgroups, including neonates.
• Predictive accuracy is limited for patients with extended PICU stays (>6 days).