Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Sepsis Watch deep-learning detection system

The nurse who gets the alert can't give the order: an authority-split detector

Duke University Hospital's Sepsis Watch alerts a rapid response nurse, who cannot order treatment. The alert becomes care only if the nurse persuades a physician.

See more

Sepsis Watch is deep-learning AI software that Duke built in house to detect sepsis. It scores every emergency department patient at Duke University Hospital every five minutes, using 86 variables from the electronic health record. It sends its alerts to rapid response nurses, not to physicians.

How it is used

Duke runs Sepsis Watch under a registered clinical trial, NCT03655626. Its alerts are not paged directly to physicians. Rapid response team nurses monitor the stream of alerts instead.

The nurses also track whether each patient's sepsis treatment bundle, a set of treatment steps, is completed on three-hour and six-hour timers.

A 2020 implementation study by Sendak and colleagues documented the system's successful integration into routine clinical care.

Who decides

The nurse who receives the alert cannot order treatment. Treatment authority belongs to the physician.

So the treatment the alert is meant to start depends on one person persuading another. The nurse has to persuade the physician across a professional hierarchy that does not automatically defer to nurses.

The hospital chose that arrangement. It sent the alerts to nurses to evaluate, while physicians keep treatment authority.

What an independent study found

An independent ethnography by Elish and Watkins, published by the research institute Data & Society in 2020, studied the deployment. An ethnography is a study that observes people at work. It named what made the system work: the nurses' hidden repair work.

The nurses mediated the professional hierarchy between nurses and physicians. They did the emotional labor of communicating a risk score upward. They judged 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. It was largely invisible in 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 it actually cost, and who paid.

Where to look

This network is modeled on Sepsis Watch. Watch the pathway from the nurse to the physician. A formal diagram of the system tends to leave it out.

It runs between two people with unequal authority. That is where the system's real work happens, and where its hidden cost falls.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other.

This case's budget of 13 units covers four tools: Understand the system, Assign a challenger, Check with a second model, and Escalate checks. They fund the nurses' hidden repair work, make the nurse's challenge routine, add an independent read of the model, and fund the attention a five-minute cadence demands.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets can be met. The cheapest combination of tools that meets them costs 7 of the 13 budget units.

Under Service and Safety Targets and under All Governance Targets, this case is not fully addressable with the available tools. Both levels ask you to close every failure pathway, among other targets. Every combination of tools was checked, and none meets the targets at either level.

Cost is not what stops it. Within the budget, you can stop mistakes from building on one another across the network.

No combination closes the two failure pathways named below. Every tool offered here, used at once, costs 18 units at these levels and still leaves them open. Every tool at its strongest setting costs 28 units and leaves the same two open.

One is the nurse's review of each Sepsis Watch alert, the step that putting a nurse first exists for. The other is Sepsis Watch reading the patient record. Closing it would stop the scoring itself.

That is a finding about the deployment, not a flaw in your choices.

Stylized model of a documented deploymentClinical decision support & deterioration alerting

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Sepsis-Watch-class alert with an authority split network: 5 components and 10 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 5 assumed · 1 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

Show all 6 assumptions
  • assumed

    The network assumes the treatment the alert is meant to start depends on the nurse persuading the physician. The nurse who receives the alert cannot order treatment. It also assumes the work is heavier than the staff time set aside for it. An independent ethnography, a study that observes people at work, found the system ran on hidden repair work by the nurses. They mediated the professional hierarchy and did the emotional labor of communicating a risk score upward. That labor was structurally necessary, largely invisible in the deployment's formal description, and undervalued. The network also assumes Sepsis Watch writes to the record continuously. It scores every patient every five minutes, over 86 variables.

  • assumed

    Kaiser Permanente's Advance Alert Monitor, another network in this Lab, also puts nurses ahead of the clinician who acts. The record separates the two deployments. Kaiser's regional program sends every alert to one screening team, staffed around the clock, and has a published evaluation. So that network assumes enough staff for the work. Sepsis Watch runs in a single emergency department, scoring every five minutes, on nurse labor the deployment did not account for. So this network assumes more work than staff, with the authority split as its fault line. The difference comes from documented staffing and routing.

  • baseline

    This network models the authority-split alert pattern documented in the Sepsis Watch case file. It does not rebuild the actual model. Its defining feature is that the rapid response nurse who receives the alert is not the physician who can act on it. So the treatment the alert is meant to start depends on one person persuading another who holds more authority.

  • assumed

    The network assumes nurses do persuade physicians, and the cost of the system's benefit falls on that persuasion. An independent ethnography found the system worked because nurses did hidden repair work. They mediated the professional hierarchy and did the emotional labor of communicating a risk score upward. That labor was structurally necessary, largely invisible in the formal description, and undervalued. So the same alert produces care with one nurse and physician, and nothing with another. The gap is in who can act, not in what the model knows.

  • assumed

    The Independent validation check stands for the independence the record lacks. The implementation study of record was led by the system's developers. It took an outside ethnography to surface the labor the formal account left out. That is exactly what evidence produced by a system's own developers tends not to see about itself. The study did not publish the share of alerts that were true sepsis cases. The network reads putting a nurse first as the deployment's own answer to a dense stream of alerts.

  • assumed

    The network models no patient or sepsis outcome. It follows how mistakes pass between Sepsis Watch, the staff, and the record inside the hospital. The patients scored every five minutes stay outside the network. The authority split and the repair work come from the case file, and nothing in this network computes them.

What this example does not show

Show all 2 limitations
  • This example models no patient or sepsis outcome. The Lab follows how mistakes pass between the model, the staff, and the records inside the hospital. The patients scored every five minutes stay outside it. The authority split and the repair work come from the case file, not from anything computed here.
  • The finding that Sepsis Watch was integrated successfully comes from an implementation study led by its developers. That study did not publish the share of alerts that were true sepsis cases. The finding about hidden repair work comes from a separate, independent ethnography of the deployment. It is not the result of a randomized comparison.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • 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.

    empirical
    • Peer-reviewed Sendak, M.P., et al. (2020). Real-World Integration of a Sepsis Deep Learning Technology Into Routine Clinical Care: Implementation Study. JMIR Medical Informatics, 8(7), e15182. https://doi.org/10.2196/15182 https://medinform.jmir.org/2020/7/e15182/
    • Advocacy Elish, M.C., & Watkins, E.A. (2020). Repairing Innovation: A Study of Integrating AI in Clinical Care. Data & Society Research Institute. https://datasociety.net/library/repairing-innovation/

  • The Advance Alert Monitor is an in-hospital deterioration model running around the clock across 21 hospitals of an integrated health system, scoring inpatients hourly and firing roughly twelve hours before predicted deterioration; a 2020 New England Journal of Medicine evaluation associated its alert-driven rapid-response workflow with lower mortality. Its defining feature is where the alert goes: not to the bedside, but to a dedicated regional tier of critical-care virtual quality nurse consultants who screen every alert around the clock, work up the chart, and only then escalate to the on-site rapid-response team — so the measured benefit is priced against the whole two-tier staffing topology, not the model alone.

    empirical
    • Academic Escobar, G.J., Liu, V.X., Schuler, A., Lawson, B., Greene, J.D., & Kipnis, P. (2020). Automated Identification of Adults at Risk for In-Hospital Clinical Deterioration. New England Journal of Medicine, 383(20), 1951-1960. https://doi.org/10.1056/NEJMsa2001090 https://www.nejm.org/doi/full/10.1056/NEJMsa2001090
    • Academic The Kaiser Permanente Northern California Advance Alert Monitor Program: An Automated Early Warning System for Adults at Risk for In-Hospital Clinical Deterioration (2022). Joint Commission Journal on Quality and Patient Safety. https://www.jointcommissionjournal.com/article/S1553-7250(22)00110-6/fulltext

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
  • Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.

All of them in context on the Clinical decision support & deterioration alerting domain page.

Levers available here and the patterns behind them

Documented case histories