Back to Bigger Picture

BIGGER PICTURE Better care

The hospital records that can raise a sepsis alarm

Hospital records contain many clues, spread across time and screens. TREWS tried to turn those clues into a timely clinical review before a dangerous infection response was missed.

Suchi Saria points at waveforms on a wall-sized clinical-data display while another person watches.Original photograph
Computer scientist Suchi Saria discusses clinical data on a large display. Her team developed the TREWS early-warning approach for sepsis.Will Kirk / Johns Hopkins UniversityEditorial reuse with attribution
Study hospitals
Five
Flagged sepsis cases analysed
6,877
Adjusted mortality difference
3.3 percentage points

How the idea works

  1. Combine clinical clues

    Records supply evolving measurements and context.

  2. Request a sepsis review

    The model flags a concerning pattern.

  3. Confirm and treat

    Clinicians assess the patient and choose appropriate action.

The signs do not arrive as one tidy diagnosis

Sepsis can begin with findings that also occur in other conditions. TREWS combines information from the electronic health record to identify a pattern that merits attention. Its purpose is to get a clinician to assess the patient sooner, not to let a risk score independently decide that an infection exists.

Sources: [2]

A warning inside the clinical workflow

The alert appears where care teams already work with patient information. Clinicians can review and confirm the concern, then make treatment decisions. This differs from an algorithm tested silently in the background: the intervention includes how people encounter, interpret and respond to its prediction. A good prediction cannot help a patient if nobody responds.

Sources: [1][2]

The patients behind the mortality result

The 2022 analysis examined 6,877 sepsis patients flagged before antibiotics. When clinicians confirmed the alert within three hours, adjusted in-hospital mortality was 3.3 percentage points lower than in the comparison group. The often-cited relative reduction was 18.7%. The first expresses the absolute difference; the second expresses the change relative to the comparison risk.

Sources: [1]

Why response capacity belongs in the design

Patients were not randomly assigned to prompt or delayed confirmation. Differences in teams, patients and attention could contribute despite statistical adjustment. TREWS illustrates actionable prediction with encouraging clinical observations; it does not establish that installing an alert alone causes the reported mortality reduction. Response capacity belongs in the design from the beginning.

Sources: [1]

The original sources

  1. Nature MedicineProspective, multi-site study of patient outcomes after implementation of TREWS
  2. Johns Hopkins Malone CenterSepsis detection platform prevents thousands of deaths

What to explore next

The invisible drawing work before cancer treatmentWhy the crop-diagnosis app asks for six leaves
Explore all 99 applications