Machine learning model for daily prediction of pediatric sepsis using Phoenix criteria

Abstract

Background: Early sepsis diagnosis is essential for initiating prompt treatment, preventing the progression of organ failure, and improving the survival rate of critically ill children. The aim of this study was to develop and validate a machine learning sepsis prediction model for patients admitted to a pediatric intensive care unit (PICU) who met the Phoenix Sepsis Score Criteria using EMR data.

Methods: Data were obtained from two PICUs within the same healthcare system. Readily available variables were used to develop and validate machine learning models predicting the onset of sepsis in critically ill children.

Results: A total of 63,875 PICU encounters were included, of which there were 5248 who met the criteria for Phoenix Sepsis. We trained and tested 4 machine learning models using vital signs, laboratory tests, demographic data, medications, and organ dysfunction scores. The Categorical Boosting (CatBoost) model had the best performance with an AUROC of 0.98 (95% CI, 0.98-0.98), and an AUPRC of 0.83 (95% CI, 0.82-0.83).

Conclusions: The implementation of our model capable of predicting the onset of sepsis defined by the Phoenix Sepsis Score criteria may help clinicians recognize and manage children with sepsis more efficiently to reduce morbidity and mortality.

Impact: Sepsis is a life-threatening condition with high rates of morbidity and mortality in children, especially in pediatric critical care units. However, there is no validated model using readily available variables in the electronic medical record data to identify critically ill patients with sepsis. The use of machine learning and electronic medical health records data to develop a predictive model can automate the identification of patients at high risk for sepsis-related organ dysfunction. The implementation of this tool can improve recognition of sepsis and prevent the progression of sepsis-related organ dysfunction leading to death.