Electronic Sepsis Alarm for Newborns Flags Infection Hours Before Blood Tests

Late-onset sepsis remains one of the most feared complications in neonatal intensive care. Babies who spend weeks in a NICU, often born many weeks premature and dependent on central lines and ventilators, can deteriorate from subtle early warning signs to septic shock within hours. Because the earliest symptoms are notoriously vague, clinicians have long struggled to distinguish the beginning of a bloodstream infection from ordinary fluctuations in a fragile infant’s condition. A new case-control study published in BMC Pediatrics by Kelsey A. Simek and colleagues at the University of Utah and Primary Children’s Hospital offers a promising answer: an electronic health record based screening algorithm that continuously scans routine clinical data and flags infants at risk of culture-positive late-onset sepsis before anyone draws the decisive blood culture.

The research team had previously built a paper-based neonatal sepsis screening tool, or NSST, designed for use in a level IV referral NICU, the highest acuity category of neonatal care. That version was scored intermittently, at scheduled care times, by nurses working through a checklist of physiologic parameters. The new study asked a critical question for the digital era: what happens to the tool’s performance when it is embedded in the electronic health record and scored continuously, drawing on every available data point rather than a handful of spot checks during shifts? Continuous scoring could theoretically sharpen detection by capturing deterioration the moment it appears, but it could also flood the system with false positives if transient abnormalities trigger the algorithm unnecessarily.

To answer that question, the investigators conducted a retrospective case-control study at a single level IV NICU. They searched the electronic health record for infants who experienced their first episode of blood culture positive late-onset sepsis during 2021, designating each of these as a case. Each case was then matched with up to two control infants of comparable baseline status who remained culture negative. The final analysis paired 31 sepsis cases with 60 controls. The NSST algorithm was applied retrospectively to each case’s data from the 24 hours preceding collection of the positive blood culture, and to each control’s data from the corresponding days of life, creating a direct comparison of how the algorithm would have behaved in infected versus uninfected infants.

The results were striking. A positive NSST score occurred in 25 of the 31 sepsis cases, or 80.6 percent, compared with only 13 of the 60 controls, or 21.7 percent. Translated into standard test characteristics, the tool achieved a sensitivity of 0.81 and a specificity of 0.80. Perhaps most compelling was the odds ratio: an infant with a positive NSST screen had 16.6 times the odds of having a positive blood culture compared with a screened-negative peer. In a population where every unnecessary antibiotic carries risks ranging from altered gut microbiome development to later metabolic disease, and every missed infection can be fatal, a screening test that balances both error types near 80 percent represents a meaningful advance.

Just as informative as the headline statistics was the anatomy of the alerts. Heart rate was the most common contributing variable driving a positive screen, followed by respiratory status and urine output. This hierarchy makes physiologic sense. Neonates respond to bloodstream infection with compensatory tachycardia long before fever or hypothermia becomes evident, and rising oxygen or ventilator requirements often accompany the inflammatory cascade. Declining urine output signals falling renal perfusion, an early marker of circulatory compromise. Notably, abnormal capillary refill time, a bedside measure of peripheral perfusion that was routinely documented in all infants in the unit, was observed exclusively among the sepsis cases, suggesting that this simple physical examination finding carries genuine discriminative weight when captured systematically.

The timing of the alerts may matter even more than their accuracy. Among the culture-positive cases, NSST alerts fired between 0 and 20 hours before blood cultures were collected. In practical terms, the algorithm frequently raised the alarm while clinicians were still deciding whether an infant looked sick enough to warrant a sepsis workup. In neonatal care, where antibiotic administration within the first hour of recognized deterioration is associated with improved outcomes, a lead time measured in hours rather than minutes could translate directly into earlier treatment, earlier source control such as line removal, and potentially lower mortality. The tool essentially functions as a continuous, tireless second observer watching every monitored parameter in the unit.

The technical shift from intermittent to continuous scoring deserves emphasis. Paper checklists are vulnerable to sampling bias: an infant whose heart rate spikes between care times may go unnoticed until the next scheduled assessment. By contrast, an EHR-based algorithm ingests every heart rate measurement, every ventilator setting change, every recorded urine output, and every blood pressure entry as they are documented. The Utah team’s finding that performance held up, with specificity remaining at 0.80 despite the far greater volume of data being evaluated, suggests that the algorithm’s thresholds were robust to the denser data stream rather than being overwhelmed by noise. This is a nontrivial validation, because many screening tools that perform well at scheduled assessments degrade badly when applied continuously.

The study’s design also carries important caveats that the authors acknowledge candidly. With only 31 cases, the confidence intervals around the sensitivity, specificity, and odds ratio estimates are necessarily wide, and the precision of the findings is limited. The retrospective, single-center design raises questions of generalizability: a level IV referral NICU cares for an unusually fragile population, and the case-control matching, while methodologically sound for evaluating test characteristics, cannot capture how the tool would behave in a prospective implementation where clinicians act on alerts in real time. Alert fatigue, workflow disruption, and the downstream consequences of responding to positive screens all remain unmeasured. The University of Utah Institutional Review Board approved the study with a waiver of consent, and the authors report no competing interests, but they are explicit that larger, multi-center studies are needed before the tool can be recommended for routine deployment.

Those caveats notwithstanding, the study lands at a moment of intense interest in pediatric clinical informatics. Sepsis screening tools embedded in electronic health records have been associated with decreased mortality in adult and pediatric intensive care, yet the ideal algorithm for neonates has remained uncertain precisely because infant physiology differs so profoundly from that of older patients. Preterm infants have baseline heart rates that would signal tachycardia in an adult, immature thermoregulation, and respiratory patterns that fluctuate for reasons unrelated to infection. A neonatal-specific algorithm, validated in the population where it will be used, addresses a genuine gap rather than simply porting adult logic to the crib.

The broader lesson of the work may be about how clinical prediction tools evolve. The Utah NSST began as a bedside paper instrument, was translated into an electronic algorithm, and was then re-evaluated under the very different conditions that digital integration creates. That iterative cycle, from paper to pixel to retrospective validation against gold-standard blood cultures, is exactly the pathway that responsible clinical artificial intelligence and decision support should follow. If multi-center prospective studies confirm these results, continuously scored EHR screening could become standard practice in NICUs, converting the torrent of monitoring data that already surrounds every fragile newborn into an early warning system that speaks up hours before sepsis declares itself. For the smallest and most vulnerable patients, those hours can be the difference between recovery and catastrophe.

Subject of Research: Validation of an electronic health record based screening algorithm for early detection of late-onset sepsis in neonatal intensive care

Article Title: Case control evaluation of a novel late-onset neonatal sepsis screening tool

Article References: Case control evaluation of a novel late-onset neonatal sepsis screening tool. (n.d.). https://doi.org/10.1186/s12887-026-07821-3

Image Credits: AI Generated

DOI: 10.1186/s12887-026-07821-3

Keywords: neonatal sepsis, late-onset sepsis, NICU, screening tool, electronic health record, early detection, case-control study, clinical decision support, neonatology, predictive markers, blood culture, pediatric informatics