PROGNOSTIC ACCURACY OF MACHINE LEARNING MODELS FOR IN-HOSPITAL MORTALITY AMONG CHILDREN WITH PHOENIX SEPSIS ADMITTED TO THE PEDIATRIC INTENSIVE CARE UNIT.

abstractObjective: The Phoenix sepsis criteria define sepsis in children with suspected or confirmed infection who have ≥2 in the Phoenix Sepsis Score. The adoption of the Phoenix sepsis criteria eliminated the Systemic Inflammatory Response Syndrome criteria from the definition of pediatric sepsis. The objective of this study is to derive and validate machine learning models predicting in-hospital mortality for children with suspected or confirmed infection or who met the Phoenix sepsis criteria for sepsis and septic shock. Materials and Methods: Retrospective cohort analysis of 63,824 patients with suspected or confirmed infection admission diagnosis in two pediatric intensive care units (PICUs) in Atlanta, Georgia, from January 1, 2010, through May 10, 2022. The Phoenix Sepsis Score criteria were applied to data collected within 24 h of PICU admission. The primary outcome was in-hospital mortality. The composite secondary outcome was in-hospital mortality or PICU length of stay (LOS) ≥ 72 h. Model-based score performance measures were the area under the precision-recall curve (AUPRC) and the area under the receiver operating characteristic curve (AUROC). Results: Among 18,389/63,824 (29%) children with suspected infection (median age [25th - 75th interquartile range [IQR]): 3.9 [1.1,10.9]; female, 45%, a total of 5,355 met Phoenix sepsis criteria within 24 h of PICU admission. Of the children with Phoenix sepsis, a total of 514 (9.6%) died in the hospital, and 2,848 (53.2%) died or had a PICU stay of ≥72 h. Children with Phoenix septic shock had an in-hospital mortality of 386 (16.4%) and 1,294 (54.9%) had in-hospital mortality or PICU stay of ≥72 h. For children with Phoenix sepsis and Phoenix septic shock, the multivariable logistic regression, light gradient boosting machine, random forest, eXtreme Gradient Boosting, support vector machine, multilayer perceptron, and decision tree models predicting in-hospital mortality had AUPRCs of 0.48-0.65 (95% CI range: 0.42-0.66), 0.50-0.70 (95% CI range: 0.44-0.70), 0.52-0.70 (95% CI range: 0.47-0.71), 0.50-0.70 (95% CI range: 0.44-0.70), 0.49-0.67 (95% CI range: 0.43-0.68), 0.49-0.66 (95% CI range: 0.45-0.67), and 0.30-0.38 (95% CI range: 0.28-0.40) and AUROCs of 0.82-0.88 (95% CI range: 0.82-0.90), 0.84-0.88 (95% CI range: 0.84-0.90), 0.81-0.88 (95% CI range: 0.81-0.90), 0.84-0.88 (95% CI range: 0.83-0.90), 0.82-0.87 (95% CI range: 0.82-0.90), 0.80-0.86 (95% CI range: 0.79-0.89), and 0.76-0.82 (95% CI range: 0.75-0.85), respectively. Conclusion: Among children with Phoenix sepsis admitted to a PICU, the random forest model had the best AUPRC for in-hospital mortality compared to the light gradient boosting machine, eXtreme Gradient Boosting, logistic regression, multilayer perceptron, support vector machine, and decision tree models or a Phoenix Sepsis Score ≥ 2. These findings suggest that machine learning methods to predict in-hospital mortality in children with suspected infection predict mortality in a PICU setting with more accuracy than application of the Phoenix sepsis criteria.