Development and External Validation of a Machine Learning Model for Prediction of Potential Transfer to the PICU

Area of Science:

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

Background:

• Unrecognized clinical deterioration in hospitalized children poses significant mortality and morbidity risks.
• Early identification of children at risk for critical events is crucial for timely intervention.

Purpose of the Study:

• To develop and externally validate machine learning algorithms for predicting intensive care unit (ICU) transfer in pediatric patients within 12 hours.
• To improve the accuracy of identifying children at risk of clinical deterioration.

Main Methods:

• An observational cohort study was conducted using electronic health records from two urban, tertiary-care, academic hospitals.
• Machine learning models (logistic regression, random forest, gradient boosted machine) were developed using patient data including age, vital signs, and laboratory results.
• Model performance was evaluated against a modified Bedside Pediatric Early Warning Score (PEWS) using discrimination (C-statistic), sensitivity, and specificity.

Main Results:

• The study included over 139,000 pediatric admissions across two sites.
• The gradient boosted machine model demonstrated superior accuracy in predicting ICU transfer compared to the modified PEWS.
• The gradient boosted machine achieved higher discrimination (C-statistic: 0.84 vs 0.71 at site 1; 0.80 vs 0.74 at site 2) and improved sensitivity and specificity.

Conclusions:

• A novel machine learning model was successfully developed and validated for early detection of ICU transfers in hospitalized children.
• The developed model significantly outperforms existing tools like PEWS in identifying children at risk of deterioration.
• This advancement offers potential for earlier interventions and improved clinical outcomes in pediatric care.