Dynamic Mortality Risk Predictions for Children in ICUs: Development and Validation of Machine Learning Models

Area of Science:

• Pediatric critical care medicine
• Machine learning in healthcare
• Health informatics

Background:

• Accurate, real-time assessment of mortality risk is crucial for pediatric intensive care unit (PICU) patients.
• Existing methods may not adequately capture dynamic changes in patient status during ICU stays.

Purpose of the Study:

• To assess a machine learning method for serially updated mortality risk estimation in critically ill children.
• To develop and validate a model that dynamically tracks mortality risk throughout an ICU admission.

Main Methods:

• Retrospective analysis of a national database (Health Facts) including 27,354 pediatric ICU admissions (2009-2018).
• Development of the Criticality Index-Mortality (CI-M) model using physiology, therapy, and care intensity data, updated every 6 hours up to 180 hours.
• Model performance evaluated using discrimination (area under the ROC curve) and calibration (Hosmer-Lemeshow test).

Main Results:

• The CI-M model demonstrated strong discrimination with an overall AUC of 0.852.
• The model showed good calibration across 29 time periods, indicating reliable risk estimates.
• Clinical validity was confirmed through analysis of patient trajectories and greater volatility observed in non-survivors.

Conclusions:

• Machine learning models integrating patient physiology, therapy, and care intensity can effectively monitor dynamic changes in hospital mortality risk in the ICU.
• The CI-M framework offers a promising approach for real-time monitoring of clinical improvement and deterioration in critically ill children.