Sepsis Watch

By Duke Health (Duke Institute for Health Innovation)
· United States
Deep learning early-warning tool that uses routine EHR data to flag emergency department patients at risk of sepsis, for review by a rapid response nurse team, as an alert aid.
Predictive
Critical Care
Production
Not Formally Regulated
Regulatory status
Not Formally Regulated
Deployment
Production
Clinical role
Adjunct (decision support)
Headquarters
United States
AI type
Predictive

Precise indication

Predicts the risk of sepsis in the next 12 hours from routine EHR data (vital signs, lab results, medications) for emergency department patients. A rapid response nurse reviews each alert and escalates to the treating team; it does not diagnose sepsis.

Regulatory status

Internal academic tool, not a commercially marked/cleared device.

Clinical evidence

Vendor-supported

Valan et al., npj Digit Med 2025 (Duke and Cohere-Med co-authors): test in 4 Ohio emergency departments, 205,005 encounters; AUROC 0.91-0.96, AUPRC 0.18-0.25.
Duke reports its 3-hour sepsis bundle compliance doubled after launch (self-reported, not peer-reviewed).

Independent

Deployed at Duke’s own emergency department and validated through peer-reviewed publications, including npj Digital Medicine and NEJM AI.

Real-world deployment

Deployed in Duke University Health System’s emergency department, with peer-reviewed validation.

Interoperability

European status

Editorial verdict

Limits and uncertainties

– No FDA clearance or CE marking found; not available as a certified medical device in Europe.
– No published peer-reviewed study of effect on mortality or other patient outcomes.
– Validation published by Duke and its licensee; no fully independent clinical evaluation found.
– Low AUPRC (0.18-0.25) at external sites: alert precision and alert burden need local assessment.
– Relies on a dedicated rapid response nurse team to review alerts; needs staff time and training.

Sources

Sepsis Watch™: the implementation of a Duke-specific early warning system for sepsis
Vendor — Cohere-Med, Sepsis Watch solution profile, MIT Solve, October 2023 https://solve.mit.edu/solutions/79311 Study (vendor co-authored) — Valan B et al., Evaluating Sepsis Watch generalizability through multisite external validation, NPJ Digit Med, June 2025 https://pubmed.ncbi.nlm.nih.gov/40500319/ Study (vendor co-authored) — Valan B et al., abstract on Duke Scholars, NPJ Digit Med, 2025 https://scholars.duke.edu/publication/1678008 Study (vendor) — Sendak MP et al., Real-world integration of a sepsis deep learning technology into routine clinical care, JMIR Med Inform, July 2020 (Duke Scholars abstract) https://scholars.duke.edu/publication/1452380 Report (independent) — Elish MC, Watkins EA, Repairing Innovation: A Study of Integrating AI in Clinical Care, Data & Society, September 2020 https://datasociety.net/library/repairing-innovation

This profile is not medical advice or an endorsement. Hospitals should perform their own due diligence (regulatory compliance, GDPR, cybersecurity, integration, local clinical validation). How we evaluate

At a glance
Vendor
Duke Health (Duke Institute for Health Innovation)
Headquarters
United States
AI type
Predictive
Specialty
Clinical role
Adjunct (decision support)
CE marking
Notified body
EU AI Act
Data hosting
Last verified
Reviewed by
How we evaluate

Every profile documents the precise indication, regulatory status, vendor-supported and independent evidence, real-world deployment, limits and dated sources.

A independent systematic review or prospective multicentre study · B at least one independent peer-reviewed study · C peer-reviewed, vendor-supported · D vendor claims only.

Read the methodology →

Spotted an error?

Vendors, authors and clinicians can request a correction. Corrections are dated on the profile.