Duke Health’s Sepsis Watch

Sepsis, the body’s extreme response to an infection, is one of the most common causes of death in hospital, and it is hard to spot early. Speed matters, which makes it an obvious target for prediction. Yet as the team behind Duke Health’s Sepsis Watch wrote, “successful integrations of machine learning into routine clinical care are exceedingly rare.”¹

Sepsis Watch is one of the exceptions. It has been running in Duke University Hospital’s emergency department since November 2018.⁵ What makes it worth studying is not the algorithm. It is the deliberate, often invisible work that went into making the algorithm useful.

What was built

In 2016, Duke formed a multidisciplinary team of statisticians, data scientists, data engineers and clinicians to build the system.¹ The model is a recurrent neural network trained on Duke’s own electronic health record data from October 2014 to December 2015.⁵ Its development and evaluation drew on more than 32 million data points, and in use it produces a risk score for each patient every hour.³

The crucial design choice was where the alerts would go. Rather than sending them straight to busy emergency doctors, Duke routed them to a centralised team of Rapid Response Team nurses, who reviewed each high-risk patient using a custom risk tool and then contacted the doctors responsible.² As Madeleine Clare Elish and Elizabeth Anne Watkins of Data & Society put it, Sepsis Watch “adds a skilled nurse between the system and the clinician receiving the new information.”³

The work nobody designed

Elish and Watkins studied the system’s six-month pilot. They found that the nurses did a great deal of what they call “repair work” to make the system fit into the hospital.³ The nurses worked on a different floor from the emergency department, so they had to phone doctors they could not see. They learned to time calls around shift changes, to group calls by area of the department, to open with questions rather than instructions, and to review patient records so they could back up the risk score with evidence.

In other words, the nurses negotiated professional hierarchies and built trust one call at a time. That labour was essential to the system’s success, and it appeared in no technical specification. Their conclusion applies well beyond healthcare: “Innovation happens not only when an AI system is being developed, but also when the system is being integrated.”⁴

What the evidence shows, and what it does not

Duke reports that it has doubled its compliance with the three-hour SEP-1 sepsis care bundle, a national quality measure reported to the US Centers for Medicare & Medicaid Services.² That is a meaningful improvement in how care is delivered.

However, we have not found published evidence from Duke showing that Sepsis Watch reduced deaths. A 2025 study validated the model on more than 205,000 emergency department encounters at four sites run by another health system, Summa Health in Ohio, with strong predictive performance (an AUROC of 0.906 to 0.960). Its authors were clear, though, that “Sepsis Watch will need to be prospectively tested to measure impact.”⁵

This is worth stating plainly, because sepsis AI is often described as “saving lives” without the evidence to match. Patient-outcome evidence of the kind leaders should look for does exist for other systems: a before-and-after study of a different deep learning model, COMPOSER, at UC San Diego Health reported a 17% relative reduction in in-hospital sepsis deaths.⁶ Even that study was not a randomised trial.

Lessons for leaders

1. Budget for integration, not just development. The model was the visible part of Sepsis Watch. The workflow, the new nursing role and the daily repair work were what made it function. Plan and fund that work explicitly, and recognise the people who do it.
2. Decide deliberately where the human sits. Duke chose to place skilled nurses between the algorithm and the doctors. That intermediary role filtered alerts, added clinical judgement and built trust. Every AI deployment involves a choice about who receives its output and what they are equipped to do with it.
3. Measure outcomes, not just adoption. Running continuously for years and improving process compliance are real achievements. They are not the same as proving that patients are better off. Be clear about which kind of evidence you have.

Questions to ask in your organisation

1. For our most important AI system, who does the day-to-day “repair work” that keeps it useful, and is that work recognised, resourced and written down?
2. Who receives our AI’s outputs, and are they the right people, with the right authority and skills, to act on them?
3. What would count as evidence that our AI has improved outcomes, not just activity, and have we set out to collect it?