P2.066: A Machine Learning Approach to Classifying Phoenix Sepsis in Pediatric Emergency Encounters: Phase One Results and Next Steps

Background: Sepsis is a major contributor to pediatric morbidity and mortality, requiring timely and accurate identification to improve outcomes. While the Phoenix Sepsis (PS) criteria offer a refined diagnostic framework, it was not designed for early screening of children at risk for developing sepsis. This study aims to develop and evaluate a machine learning model for identifying PS in pediatric emergency department (ED) encounters using clinical and laboratory data.

Methods: This retrospective analysis utilized two datasets derived from ED encounters of children < 18 years old at a single institution:

A) Mixed Cohort (MC) dataset: a random sample of encounters between 8/15/2023 and 8/20/2023 was initially extracted. Due to the relative infrequency of sepsis, this was supplemented with additional encounters identified via ICD diagnostic codes for sepsis between 1/1/2017 and 12/31/2024, yielding a combined dataset of 780 encounters, including both ICD-coded sepsis and non-sepsis cases.

B) ICD-sepsis Cohort (IC) dataset: A subset of encounters identified exclusively by ICD diagnostic codes for sepsis within the same 2017–2024 date range, supplemented by any PS-positive cases identified in MC.

In Phase 1, logistic regression models were trained to classify encounters as PS-positive or PS-negative. Input variables included all components of the PS criteria as well as patient age. Both datasets were randomly split into 80% training and 20% testing sets, and model training employed 5-fold cross-validation to address class imbalance. Two models were developed: one using MC and the other using IC.

Results: Of the 780 ED encounters, 104 encounters had an ICD code for sepsis and 23 had sepsis by PS criteria. The MC test set consisted of 8 (5.1%) PS cases out of a total of 156 encounters. The MC trained model yielded a sensitivity of 87.5% and specificity of 99.3% for PS. The IC test set consisted of 5 (23.8%) PS cases out of 21 encounters. The IC-trained model had a sensitivity of 80.0% and a specificity of 93.8% for PS, respectively. Variables in the neurological category, GCS and pupil reactivity, emerged as the most significant features with weight of 0.36 and 0.14, respectively.

Conclusion: Despite significant data imbalance, the Phase 1 models performed well, with only one false positive and one false negative – highlighting the potential of machine learning for PS prediction. Neurologic variables may have appeared more predictive due to fewer missing values compared to other PS components.

In Phase 2, we will incorporate features beyond PS components to enhance model performance. Phase 3 will focus on optimizing early detection and risk stratification, with an emphasis on improving sensitivity. These refinements aim to support more accurate and timely sepsis screening, ultimately advancing pediatric emergency care.