Domain Atlas / Clinical decision support & deterioration alerting
Sepsis Watch deep-learning detection system
Explore this deployment in the PAN Lab ↗
Sepsis Watch is a deep-learning sepsis-detection system scoring every emergency-department patient every five minutes over 86 variables, deployed at an academic hospital under a registered clinical trial, with alerts fronted by rapid-response-team nurses who track treatment-bundle completion on three- and six-hour timers. Its structural fault line is an authority split: the operator who receives the alert (the nurse) is not the operator empowered to act on it (the physician who holds treatment authority), so the correction runs through a peer-persuasion edge. An independent ethnography found the system worked because nurses performed hidden repair work — mediating the professional hierarchy and doing the emotional labor of communicating a risk score upward — labor that was structurally necessary, largely invisible to the deployment's formal description, and undervalued.[2]
What happened
Sepsis Watch is a deep-learning sepsis-detection system that scores every emergency-department patient every five minutes across 86 variables, deployed at an academic hospital under a registered clinical trial. Its alerts are not paged directly to physicians. They are fronted by rapid-response-team (RRT) nurses, who monitor the alert stream and track treatment-bundle completion against three- and six-hour timers. An implementation study documented the system's successful integration into routine clinical care.
The structural feature this deployment exposes is an authority split. The operator who receives the alert — the RRT nurse — is not the operator empowered to act on it. Treatment authority belongs to the physician. So the correction the alert is supposed to produce does not run along the model-to-operator edge; it runs along a peer edge between two operator classes with different authority, and that edge carries its own friction. The nurse who sees the alert has to persuade the physician who can order the treatment, across a professional hierarchy that does not automatically defer to them.
An independent ethnography of the deployment named what made this work. The nurses performed hidden "repair work": mediating the nurse-physician professional hierarchy, doing the emotional labor of communicating a risk score upward, and calibrating when and how hard to push a physician who might not want to hear it. This labor was structurally necessary — the system did not function without it — yet it was largely invisible to the deployment's formal description and undervalued in how the work was accounted for. The implementation study documents that the integration succeeded; the ethnography documents what the integration actually cost and who paid it. This is the domain's cleanest case of a check whose real cost sits on a peer edge between people with unequal authority, not on the model itself.
The sociotechnical reading
Sepsis Watch is the case that shows where the cost of a working AI alert actually lands: not on the model-to-operator edge everyone looks at, but on a peer edge between two people with unequal authority. The person who receives the alert cannot act on it. A nurse gets the score; a physician holds the treatment authority. So the alert only becomes care if the nurse successfully moves the physician — across a professional hierarchy, under time pressure, with a risk number as the only leverage. The deployment's formal description calls this "the nurse fronts the alert." The ethnography calls it what it is: repair work — mediating the hierarchy, managing the emotional register of paging someone more senior, judging when and how hard to push. The system worked because that labor was done, and the labor was largely invisible and undervalued.
The governable surfaces are all on that peer edge, and they are the ones a formal system diagram tends to omit. The first is recognition: the repair work is real work, and a deployment that does not name it, resource it, or account for it is running on unpaid labor that will quietly fail when the nurse doing it burns out or moves on. The second is the authority split itself: the correction pathway depends on persuasion succeeding, so the same alert produces care in one nurse-physician pairing and nothing in another, and no amount of model accuracy closes that gap — it is a gap in who can act, not in what the model knows. The third, shared with the rest of this domain, is independence: the implementation study was developer-led, and it took an outside ethnography to surface the labor the formal account left out — which is exactly the kind of thing developer-produced evidence does not see about itself. The honest boundary throughout: no patient or sepsis outcome is computed on the Lab diagram. The patients scored every five minutes are boundary-only; the alerts, pages, and bundle timers are institutional signals, and the authority split and the repair work live in the case file, never on any network.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.