Read Part 1 of the Sepsis Watch story to learn more about development of Sepsis Watch at Duke and Part 2 on the problem with scaling internally built models, as learned through the commercialization of Sepsis Watch to a point solution company.

  1. What is Sepsis Watch™? Sepsis Watch™ is a machine learning model that uses real-time physiological and medication data to predict the onset of sepsis 12 hours in advance. It launched at Duke University Hospital in 2018 as the first deep learning model integrated into routine clinical care in the U.S.
  2. Why was it needed? Sepsis itself is the leading cause of mortality in U.S. hospitals, with related inpatient stays rising nearly 40% between 2016 and 2021 across the country. Before Sepsis Watch™, Duke clinicians relied on an EHR alert that generated so many false alerts that 86% were dismissed.
  3. Has it worked? Yes. Duke University Hospital increased Sep-1 bundle compliance by 89%, decreased sepsis mortality by 29%, and the Sepsis Watch solution isestimated to save about 8 lives per month by predicting sepsis a median of 5 hours before clinical presentation.
  4. Where is it in use today? It operates continuously across all three Duke Health hospitals — Duke University Hospital, Duke Raleigh Hospital, and Duke Regional Hospital — making real-time hourly predictions for adult patients.
  5. Who receives the alerts? Alerts are routed to the Rapid Response Team (RRT), a clinically experienced nursing team with the expertise to act on the model's outputs, building trust between RRT nurses and ED physicians rather than overburdening already-stretched physicians.
  6. Does it support treatment as well as detection? Yes. Beyond flagging sepsis risk, Sepsis Watch™ provides real-time nudges when further treatment steps are needed and tracks compliance with documentation and treatment standards for reporting to payors.
  7. Why hasn't it spread to more hospitals? Scaling an internally built AI model proved far more complex than expected. In 2018 there was no standard playbook for commercializing healthcare AI, and few understood that the actual “product” was both the model and its workflow. This made it much more difficult to replicate at another hospital.
  8. What happened when it was deployed at another hospital? At Summa Health in Ohio, the model performed well clinically, but change management proved a bigger obstacle. Summa had fewer dedicated staff and different practice standards than Duke, making it difficult for clinicians to adopt a tool they perceived as adding work rather than reducing it.
  9. What is the core lesson from trying to scale it? There is no such thing as a “turnkey” implementation. A model without a clinical workflow and dedicated implementation support has no future. Successful AI adoption in healthcare requires ongoing collaboration between technology teams and frontline staff.
  10. What does effective implementation look like? Effective implementation is alignment across clinical roles, strong operational champions, IT buy-in, and time spent with frontline staff to understand their specific workflows and pain points before ANY technology is implemented. Peer credibility from nurses and clinicians who have lived the implementation also accelerates adoption more than vendor messaging alone.

Sepsis Watch is available on the Vega Health Marketplace.