Pediatric sepsis is difficult to diagnose until it’s already dangerous.
Children who eventually develop organ dysfunction frequently arrive at the emergency department looking stable, sometimes well enough to be admitted to the floor or even discharged home.
By the time sepsis declares itself, hours may have passed, and the opportunity for early intervention narrowed.
Machine learning — an AI process that enables computer systems to identify patterns in data to make predictions — may help clinicians diagnose sepsis much earlier, a new multicenter study suggests.
In a JAMA Pediatrics study spanning more than 1.6 million pediatric emergency department visits, researchers developed and validated predictive models to identify those likely to develop sepsis or septic shock within the next 48 hours.
“We wanted to know whether there were signals early in the emergency department visit that could tell us a child was on a trajectory toward organ dysfunction,” said Elizabeth R. Alpern, MD, a pediatric emergency physician and clinical epidemiologist at Ann & Robert H. Lurie Children’s Hospital of Chicago.
Read next: A New Paradigm in Pediatric Sepsis Care
Despite major improvements in sepsis care, most existing tools focus on identifying children who already meet sepsis criteria.
“There were no studies in predicting sepsis in undifferentiated patients in the emergency department, and those are crucial moments for sepsis prevention,” Alpern said. “It’s very difficult to predict something that can have so many different causes and can develop in so many different ways.”
Building a predictive window
To address that gap, investigators turned to the PED Screen Registry, a harmonized electronic health record database developed across five pediatric health systems. The dataset included visits from both quaternary children’s hospitals and affiliated community emergency departments using Epic and Cerner platforms.
The study focused on children ages 2 months to 17 years who did not meet sepsis criteria during their first four hours of emergency department care. Researchers then analyzed whether those children went on to meet Phoenix Sepsis Criteria or develop septic shock within the next 48 hours.
Using machine learning approaches — including logistic regression and extreme gradient boosting — the team trained models on data from 2016 to early 2020 and validated them on a separate post-pandemic cohort from 2021 to 2022.
Key predictors included age-adjusted vital signs, oxygen saturation, shock index, emergency severity index triage category, and markers of chronic medical complexity.
How well did the models perform?
By traditional performance standards, the AI models were strong, with areas under the receiver operating characteristic curve (AUROC) exceeding 0.92. But the researchers emphasized metrics more relevant to clinical reality, especially given how rare pediatric sepsis is.
At high-sensitivity thresholds, the models identified one child who would go on to develop sepsis for roughly every 45 to 68 evaluations triggered.
“That’s a good start, but emergency department teams would need to consider workload, alarm fatigue, and how this information would be used in practice,” Alpen said.
Septic shock, as expected, was rarer and required higher numbers needed to evaluate.
Crucial caveats
The authors are careful to point out that these models are not ready to function as stand-alone sepsis alerts.
Some predictive features may reflect clinician judgment rather than physiology alone. Others highlight the inherent difficulty of predicting rare outcomes. And Phoenix Sepsis Criteria identify a smaller, higher-acuity group of children than some earlier sepsis definitions.
Most importantly, the goal is not automatic escalation of care.
“This isn’t about pushing a button and launching a full sepsis response,” Alpern said. “It’s about augmenting clinician judgment — helping teams pause and rethink disposition, monitoring, or diagnostic decisions for a child who might otherwise seem low risk.”
A flagged patient, for example, may warrant closer observation, delayed discharge, earlier antibiotics, or clearer communication with inpatient teams.
Next steps
The machine learning models will require prospective testing and thoughtful integration into existing workflows at children’s hospitals looking to implement this work. Alarm fatigue, explainability, and clinician trust will be central challenges, particularly in emergency departments already navigating multiple screening tools.
Still, the findings mark a meaningful step toward shifting sepsis recognition earlier in the care trajectory.
“Sepsis is still a needle in a haystack,” Alpern said. “What this work shows is that the haystack may be giving us clues earlier than we realized. The challenge now is learning how to use those clues wisely.”
Like this story? Get proven strategies for recognizing and treating pediatric sepsis based on eight years of learning and data from more than 100,000 sepsis cases.
