Derivation and Validation of Predictive Models for Early Pediatric Sepsis

Area of Science:

• Pediatric Emergency Medicine
• Clinical Informatics
• Artificial Intelligence in Healthcare

Background:

• Sepsis is a critical cause of mortality in children, necessitating early detection.
• Existing predictive models have not significantly improved early sepsis diagnosis.
• Machine learning offers potential for enhanced sepsis prediction.

Purpose of the Study:

• To develop and validate machine learning models for predicting pediatric sepsis within 48 hours.
• To compare the performance of different machine learning algorithms in sepsis prediction.
• To identify key predictive features from electronic health records (EHR).

Main Methods:

• Multisite registry study using EHR data (2016-2022).
• Model derivation and validation using logistic regression (ridge) and gradient tree boosting.
• Inclusion criteria: pediatric ED visits (2 months to <18 years), excluding specific conditions.

Main Results:

• Gradient tree boosting achieved high predictive performance (AUROC 0.94 for sepsis, >=0.92 for shock).
• Models demonstrated positive likelihood ratios for predicting sepsis and septic shock.
• Key predictors included emergency severity index, vital signs, and medical complexity.

Conclusions:

• Validated machine learning models effectively predict pediatric sepsis and septic shock using EHR data.
• These models offer a promising tool for improving early sepsis detection in emergency departments.
• Future research should integrate these models with clinical judgment for optimal prediction.