Sepsis kills children. It is a leading cause of pediatric death, with an in-hospital mortality of 5-20% in the United States [1, 2]. Early recognition and treatment improve outcomes, but predicting which emergency department (ED) patients will develop sepsis remains a major clinical challenge. Current sepsis screening tools are designed primarily to identify children who currently have sepsis or to predict its development within hours among patients already suspected of having infection [3, 4]. Can we predict sepsis earlier—up to 48 hours before organ dysfunction appears—in the general ED population, when early intervention could make the biggest difference?
A new PECARN study “Derivation and Validation of Predictive Models for Early Pediatric Sepsis” published in JAMA Pediatrics suggests the answer may be yes [5]. Using machine learning models applied to routine electronic health record (EHR) data, researchers successfully predicted which children would develop sepsis within 48 hours of ED arrival—even when those children showed no signs of sepsis during their first 4 hours of care. This extended prediction window in an undifferentiated ED population represents a uniquely different approach from existing tools.
Study Design
This retrospective multi-center study analyzed ED visits from 5 health systems in the PECARN network between 2016-2022. The study:
- Inclusion criteria: All ED visits aged 2 months to 18 years
- Exclusion criteria:
- ED disposition of death or transfer to a facility not within the database
- Patients presenting with trauma based on ICD codes
- Sepsis criteria met during the monitoring timeframe (4 hours or ED disposition)
- Sample size: 1,604,422 visits in the training cohort; 719,298 visits in the test cohort
- Outcome measured: Development of sepsis, defined using the Phoenix Sepsis Criteria [1, 2]
Methods
Researchers developed machine learning models using data from over 1.6 million visits (training cohort) and validated them on over 700,000 visits from a later time period (2021-2022). They compared two machine learning approaches—logistic regression and gradient tree boosting—to demonstrate the robustness of the predictive signal across different algorithms.
Results
Among over 2.3 million ED visits analyzed, the prevalence of sepsis occurring within 48 hours was 0.35% in the training cohort and 0.37% in the validation cohort
Both machine learning models achieved excellent predictive performance for sepsis:
- Logistic regression: AUROC of 0.923
- Gradient tree boosting: AUROC of 0.936
(AUROC = Area Under the Receiver Operating Characteristic curve—a measure of how well the model distinguishes between patients who will and won’t develop sepsis, where 1.0 is perfect and 0.5 is no better than chance)
The researchers evaluated model performance at a 90% sensitivity threshold (meaning the model would identify 9 out of 10 children who develop sepsis). At this threshold, the gradient tree boosting model achieved:
- Specificity: 80.7%
- Number needed to evaluate: 59 (meaning for every 59 alerts, one represents a child who will develop sepsis)
Important predictive features included:
- Markers of medical complexity (complex chronic conditions, recent ED visits)
- Emergency Severity Index category
- Age-adjusted vital signs and shock index
- Oxygen saturation measurements
Why This Matters
The 48-hour prediction window is what makes this study different. While some existing sepsis screening tools can predict sepsis several hours before it develops, most work over shorter timeframes and focus on patients already suspected of having infection. This model predicts future risk up to 48 hours in advance in an undifferentiated ED population, using only routine EHR data from the first 4 hours of ED care—creating a potential opportunity for earlier identification before organ dysfunction develops.
These tools are best suited to augment clinician judgment rather than replace it. The models could help prioritize which patients need closer reassessment, monitoring, or escalation of care, rather than triggering automatic treatment protocols. Given the low prevalence of sepsis (less than 0.4% of ED visits), any screening approach must carefully balance sensitivity with alert burden—even with strong predictive performance, 58 children would be flagged for every one who develops sepsis at the 90% sensitivity threshold.
However, this is a prediction model derivation and validation study, not an implementation study. Several important questions remain before this could be used in clinical practice:
- Alert design and integration: How should alerts be delivered to minimize alarm fatigue while ensuring timely action?
- Clinical workflow: What specific actions should clinicians take when the model identifies a high-risk patient?
Caution
These findings are promising but require prospective validation and real-world implementation studies before informing routine clinical practice.
Take Home Message
This study demonstrates that machine learning can successfully predict pediatric sepsis up to 48 hours before organ dysfunction develops, using only routine ED data from the first 4 hours of care. This is a critical proof-of-concept: we can identify high-risk children much earlier before organ dysfunction develops with the potential to support earlier recognition and more targeted evaluation.
Why is this important for patients and caregivers
- While sepsis is associated with significant morbidity and mortality, it can be very difficult to predict which patients are at risk for developing this condition. This study created machine learning predictive models for sepsis to identify children who have not yet developed it.
- Using a dataset of over 2.3 million pediatric ED visits, researchers created and then validated a predictive model that demonstrated excellent accuracy in predicting sepsis within 48 hours of arrival to the ED.
- While additional study is needed, these findings suggest that in the future, with AI complementing clinical judgment, sepsis can be recognized earlier in patients, leading to earlier treatment and potentially better outcomes.
References
- Tan B, Wong JJ, Sultana R, et al. Global Case-Fatality Rates in Pediatric Severe Sepsis and Septic Shock: A Systematic Review and Meta-analysis. JAMA Pediatr. 2019;173(4):352-362. doi:10.1001/jamapediatrics.2018.4839
- Weiss SL, Balamuth F, Hensley J, et al. The Epidemiology of Hospital Death Following Pediatric Severe Sepsis: When, Why, and How Children With Sepsis Die. Pediatr Crit Care Med. 2017;18(9):823-830. doi:10.1097/PCC.0000000000001222Alpern ER, Scott HF, Balamuth F, et al. Derivation and Validation of Predictive Models for Early Pediatric Sepsis. JAMA Pediatr. 2025;179(12):1318-1325. doi:10.1001/jamapediatrics.2025.3892
- Schlapbach LJ, Watson RS, Sorce LR, et al. International Consensus Criteria for Pediatric Sepsis and Septic Shock. JAMA. 2024;331(8):665-674. doi:10.1001/jama.2024.0179
- Sanchez-Pinto LN, Bennett TD, DeWitt PE, et al. Development and Validation of the Phoenix Criteria for Pediatric Sepsis and Septic Shock. JAMA. 2024;331(8):675-686. doi:10.1001/jama.2024.0196
- Alpern ER, Scott HF, Balamuth F, et al. Derivation and Validation of Predictive Models for Early Pediatric Sepsis. JAMA Pediatr. 2025;179(12):1318-1325. doi:10.1001/jamapediatrics.2025.3892
