AI models predict sepsis in children (news.northwestern.edu)

Researchers at Northwestern University and Ann & Robert H. Lurie Children’s Hospital developed and validated machine-learning models that predict which children in the emergency department will develop sepsis within 48 hours, using routine EHR data collected during the first four hours of care. Published in JAMA Pediatrics and the first to use the new Phoenix Sepsis Criteria for pediatrics, the multi-center study drew on PECARN data across five health systems. By excluding patients already septic on arrival, the models focus on true early prediction so clinicians can initiate proven, time-sensitive therapies before organ dysfunction appears—potentially reducing mortality while avoiding unnecessary aggressive treatment for low-risk children. Technically, the models were discovered on retrospective ED visits from 2016–2020 and temporally validated on 2021–2022 data, demonstrating balanced identification of at-risk children without excessive false positives. Key predictive inputs included triage acuity, heart and respiratory rates, and comorbidities such as cancer, all elements routinely available in EHRs—supporting feasibility for real-time deployment. The team evaluated bias and emphasizes combining AI outputs with clinician judgment; next steps include prospective implementation studies and workflow integration to confirm impact on outcomes and equity. The work was supported by NICHD grant R01HD087363.