Design: Retrospective multicentre diagnostic accuracy cohort
Population: 166,943 paediatric visits to 8 emergency departments in New York
Primary outcome: Sensitivity and predictive values for sepsis within 24 hours
Effect: Sensitivity 36% (29 to 43); PPV 3%; under-fives 20%

An independent evaluation tested a commercial electronic health record sepsis alert across eight emergency departments in New York, covering 166,943 paediatric visits and 189 sepsis cases defined by Phoenix criteria within 24 hours.

The alert identified 36% of children with sepsis, and only 30% before sepsis criteria were met. Its positive predictive value was 3%, so most alerts were false. Sensitivity was lower in children under five (20%), in Black children (22%) and outside tertiary children's hospitals (25%).

An alert that misses most cases and fires mostly on children without sepsis can breed both false reassurance and alert fatigue. Clinical recognition remains the safeguard.

Do not rely on an electronic alert to flag paediatric sepsis; it missed about two-thirds of cases here
Reassess any unwell child with fever for perfusion, mental state and work of breathing
Be most alert in children under five, where the tool performed worst
Ask parents whether this illness feels different from previous ones

Why it matters

Software sold to detect sepsis may be performing far worse than clinicians assume.

The statistics, in plain English

A sensitivity of 36% means it missed 64 of every 100 children with sepsis. A positive predictive value of 3% means 97 of every 100 alerts were for children who did not have sepsis.