Local validation is a precondition, not a refinement.
AI value · Missed sepsis and alert burden
Verified
Missed 67% of sepsis while alerting on 18% of hospitalizations; AUROC .63
One health system; v1 only, not later v2.
Before
Monitor admitted patients and recognize sepsis through contemporary clinical practice. → Diagnoses and treats patients using ordinary EHR information.
After
Calculates a proprietary risk score every 15 minutes and generates alerts at the selected threshold. → Reviews or disregards the alert and retains all treatment decisions.
Human boundary
Clinicians own treatment; hospitals choose thresholds.
Why it matters
A deployed alert can add workload while missing most cases.
How the work changed
Before
How the work ran before the change.
Step 1 of 2
Hospital clinicians
Monitor admitted patients and recognize sepsis through contemporary clinical practice.
ControlTimely antibiotics are the study proxy for clinical recognition; no historical manual queue is reconstructed.
Step 2 of 2
Clinical team
Diagnoses and treats patients using ordinary EHR information.
ControlHuman diagnosis and treatment authority.
What changed
A deployed alert can add workload while missing most cases.
Decision rightHuman sets the threshold; AI decides each instance
After
How the same work runs now.
Step 1 of 2
Epic Sepsis Model v1
Calculates a proprietary risk score every 15 minutes and generates alerts at the selected threshold.
ControlHospital-selected threshold within the recommended range; model missed 67% and alerted on 18% of stays in the validation.
Step 2 of 2
Clinician
Reviews or disregards the alert and retains all treatment decisions.
ControlOrdinary surveillance remains the failure path for missed cases; alert fatigue is an explicit control concern.
Process model built from the published workflow evidence for Michigan Medicine / Epic Sepsis Model v1. Every step, actor, and control appears in full below.Every step, actor, and control
Exception path
Clinical judgment can override alerts; missed cases rely on ordinary surveillance.
Decision authority
Clinicians own treatment; hospitals choose thresholds.
Before
#
Actor
Action
Control
01
Hospital clinicians
Monitor admitted patients and recognize sepsis through contemporary clinical practice.
Timely antibiotics are the study proxy for clinical recognition; no historical manual queue is reconstructed.
02
Clinical team
Diagnoses and treats patients using ordinary EHR information.
Human diagnosis and treatment authority.
After
#
Actor
Action
Control
01
Epic Sepsis Model v1
Calculates a proprietary risk score every 15 minutes and generates alerts at the selected threshold.
Hospital-selected threshold within the recommended range; model missed 67% and alerted on 18% of stays in the validation.
02
Clinician
Reviews or disregards the alert and retains all treatment decisions.
Ordinary surveillance remains the failure path for missed cases; alert fatigue is an explicit control concern.
Work that left the path
No work removed; alert-review work added
Human role before
Hospital clinicians surveilled admitted patients through contemporary clinical practice and EHR review, then initiated diagnostic and treatment workflows when they recognized sepsis.
Human role after
Clinicians evaluate alerts and retain diagnosis and treatment authority.
AI role
Calculates a score every 15 minutes and generates threshold alerts.
Outcomes
Missed sepsis and alert burden
Verified
Contemporary practice→Missed 67% of sepsis while alerting on 18% of hospitalizations; AUROC .63
2018-12-06 to 2019-10-20 · 38,455 hospitalizations; 2,552 sepsis cases
One health system; v1 only, not later v2.
What leaders can reuse
Anti-pattern
Do not generalize v1 results to v2.
Questions
01Where is the operating threshold set and who can override it?
02What measured result would trigger rollback or retraining?
03Which residual decisions must remain human-owned?
Portability conditions
Local calibration
Version-specific testing
Alert monitoring
Reputation risk
medium
Evidence and authority
What the public record supports.
Current · updated
1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-09-06 · stable ID 998bcea29b46d14e
