Sepsis Alert Firing on the Wrong Patients
A hospital deploys a sepsis early-warning alert in the EHR. The alert fires when 2 of 3 SIRS criteria plus a suspected-infection flag are present. After 6 weeks, the alert fires on 41% of adult inpatients, and the sepsis coordinator reports that fewer than 1 in 8 fired alerts correspond to a patient the clinical team agrees is at risk. Nurses describe 'alert fatigue' and have started dismissing without reading. Meanwhile, the hospital's sepsis mortality metric has not improved. The CDS build team wants to add more criteria to 'make it more specific.'
1. Diagnose why the alert is failing. Is 'add more criteria' the right fix?
Read the worked answer
The alert is failing on specificity and actionability, not on sensitivity.
- Firing on 41% of adult inpatients means it is essentially a background noise generator. SIRS criteria alone are notoriously non-specific - many inpatients meet 2 of 3 for benign reasons (post-op, minor infection, etc.).
- The 1-in-8 agreement rate confirms the clinical team does not find the alerts meaningful.
- Dismissal-without-reading is the predictable result of low positive predictive value. Once nurses stop reading, even a good alert gets ignored.
'Add more criteria' is the wrong reflex if done blindly. Adding criteria can improve specificity but also risks missing early sepsis, and it doesn't address the deeper problem: the alert lacks a recommended action and isn't tied to a workflow. A better alert says 'this patient meets criteria - here is the sepsis bundle step to initiate.'
The fix is a redesign, not a tweak: tighten the trigger, add actionability, and measure PPV before and after.
2. Propose a redesign of the alert logic and its presentation. What do you change and how do you measure it?
Read the worked answer
Logic changes:
- Require a higher-specificity combination - e.g., SIRS plus organ dysfunction indicator (lactate, hypotension, altered mentation) rather than suspected infection alone.
- Add an exclusion list (comfort-care patients, known chronic conditions that mimic SIRS).
- Consider a risk score (e.g., NEWS or qSOFA-style inputs already in the EHR) rather than a flat 2-of-3 rule.
Presentation changes:
- Make the alert action-oriented: display the recommended next step (e.g., 'Order lactate; consider sepsis bundle').
- Route to the right role - a sepsis alert to a nurse and to a rapid-response team may warrant different presentations.
- Suppress repeat firing for the same patient within a window unless the patient's condition changes.
Measurement:
- Track positive predictive value (clinician-agreed true positives / total fired), alert volume per 100 patient-days, dismissal rate, and time-to-bundle-initiation.
- Set targets: PPV > 30%, volume reduced by at least half, dismissal rate < 50%.
- Run a 4-week pilot on 2 units before hospital-wide rollout.
3. The CDS team is pushing to ship the redesign in 10 days. What governance and safety steps do you require before go-live?
Read the worked answer
Ten days is tight but workable if the safety steps are not skipped. Require:
- Clinical review and sign-off. The redesigned logic and exclusions must be reviewed by the sepsis committee or a designated clinical owner. Document the rationale.
- Silent-mode validation. Run the new logic in shadow mode for at least 1-2 weeks, comparing what it would have fired on versus the current alert, before it reaches clinicians. This catches logic errors and gives you a real PPV estimate.
- Exclusion-list review. Confirm exclusions don't inadvertently suppress alerts for vulnerable populations (e.g., immunocompromised, elderly).
- Rollback plan. Know how to revert to the prior logic within hours if the new alert misfires.
- Monitoring plan. Assign an owner to review alert performance weekly for the first month.
If the team cannot accommodate silent-mode validation in 10 days, recommend a phased rollout (pilot units first) rather than a hospital-wide launch. Patient safety outranks the deadline.