Machine Learning Models Show High Accuracy in Predicting Pediatric Sepsis Within 48 Hours of ED Visit

A new study published in JAMA Pediatrics reports the successful development and validation of machine‑learning models that estimate a child’s risk of developing sepsis—or septic shock—within 48 hours of arriving at the emergency department (ED).

Sepsis remains a leading cause of pediatric mortality, and early detection is critical. Despite decades of effort, traditional predictive tools have not reliably improved early diagnosis. In this work, researchers leveraged electronic health record (EHR) data from five health systems across the Pediatric Emergency Care Applied Research Network (PECARN) to derive and validate novel models using data from January 2016 to February 2020, and then temporally validate them on data from 2021 to 2022.

The study included more than 1.6 million eligible pediatric ED visits (aged 2 months to <18 years) in the training cohort and over 719,000 in the test cohort. Visits with trauma, existing sepsis, or ED disposition of death or transfer were excluded. Predictive features were drawn from patient demographics and physiologic data collected in the first 4 hours of ED care (for instance, emergency severity index, vital signs adjusted for age, and medical complexity).

The models were benchmarked using both logistic regression (with ridge regularization) and gradient tree boosting techniques. For predicting sepsis, the logistic regression model achieved an area under the receiver operating characteristic curve (AUROC) of 0.92 (95% CI 0.92–0.93), while the gradient boosting model reached an AUROC of 0.94 (95% CI 0.93–0.94). Models for septic shock performed similarly, with AUROCs of 0.92 or higher. The gradient boosting models also demonstrated robust positive likelihood ratios—for sepsis, approximately 4.7 to 6.2; for septic shock, approximately 4.2 to 5.8—suggesting clinically meaningful discrimination.

In fairness analyses, model performance was largely consistent across demographics, except by insurance status: the AUROC was actually higher in patients with Medicaid compared to those with commercial insurance.

The authors conclude that their approach—combining EHR data and machine learning across a large, multicenter cohort—can predict the risk of pediatric sepsis with high accuracy, days in advance. They emphasize that these models should be integrated with clinical judgment in future studies to assess whether they can meaningfully improve outcomes in real-world ED settings.