Among pediatric patients presenting to the emergency department (ED) with high fevers, pinpointing those at risk for developing sepsis “is akin to looking for a needle in the haystack,” said Elizabeth Alpern, MD, MSCE, Division Head of Emergency Medicine at Ann & Robert H. Lurie Children’s Hospital of Chicago.

“We see a lot of children who have fever, cold symptoms, or other infectious disease symptoms — tens or hundreds of thousands of cases nationally,” Alpern explained. “The question is: How do we identify that one child in that mass of otherwise well children who will progress to develop sepsis and have a life-threatening infection?” 

For suspect cases, the urgency with which ED clinicians must act, and the intensive workup that follows, can heighten anxiety for families. But a new artificial intelligence (AI) model may better specify which children are at high risk for sepsis before organ dysfunction develops, while also sparing lower-risk patients from unnecessary rapid intervention.

In a study led by Alpern, researchers used routine electronic health record (EHR) data collected during the first 4 hours of an ED visit to develop and validate machine-learning models that predicted sepsis within 48 hours with high accuracy. The study included more than 1.6 million ED visits across five health systems in the Pediatric Emergency Care Applied Research Network. Results are published in JAMA Pediatrics.

Sepsis is a leading cause of death in children worldwide, and although early recognition improves outcomes, existing predictive tools have not reliably identified children with evolving infection before clinical deterioration, Alpern said. The model represents the first application of AI-based prediction using the new Phoenix Sepsis Criteria, which were developed through an international consensus effort to standardize pediatric sepsis definitions. 

Study Design and Model Performance

The team examined ED visits from January 2016 through February 2020 for model derivation and from 2021 through December 2022 for temporal validation. Eligible patients were aged 2 months to < 18 years and did not have sepsis during the initial 4-hour feature window. Data analysis followed the TRIPOD-AI reporting guidelines.

Researchers used logistic ridge regression and gradient tree boosting to predict sepsis as defined by suspected infection plus a Phoenix Sepsis Criteria score ≥ 2 or death within 48 hours. Predictive variables included

• Triage severity,
• Age-adjusted vital signs, and
• Markers of medical complexity.

Model performance was high across cohorts. The gradient tree boosting model achieved an area under the receiver operating characteristic (AUROC) curve of 0.94 (95% CI, 0.93-0.94) for predicting sepsis and demonstrated positive likelihood ratios ranging from 4.67 to 6.18. Performance for septic shock prediction was similarly strong, with AUROCs ≥ 0.92. Fairness assessments showed comparable results across demographic groups.

Clinical Implications

Because the models rely only on data already captured routinely in the ED, the researchers envision an alert-based system that integrates EHR inputs with clinician judgment. The models narrowed the number of children who would need evaluation to identify one case of sepsis; the data show approximately 45-68 cases for each one case of sepsis, she noted.

Future work will focus on refining risk thresholds, improving predictive value, and evaluating the model’s performance when embedded in ED workflows, which will require thoughtful implementation, Alpern said.

“This is not an off-the-shelf tool that can simply plug into an EHR,” Alpern said. A provider-facing alert that integrates the model’s output with clinical judgment “may help us move toward preemptively identifying children with sepsis.”

Alpern reported having no relevant financial relationships. The research was funded by the NIH.