The Silent Sepsis Alert
A 340-bed community hospital deployed a sepsis screening alert in the EHR six weeks ago. The alert fires when two or more SIRS criteria plus a suspected-infection flag are present. Compliance with the alert (clinicians clicking 'acknowledge and act') is 91%, but the sepsis core measure bundle completion rate has fallen from 78% to 64%, and nurses report 'alert fatigue' on the medical-surgical units. Chart review of 40 missed bundles shows the alert frequently fires at 2 a.m. on patients later found to have simple viral illness, while three true sepsis cases on the night shift were never triggered because the lactate order was placed after the alert window closed. The CMIO wants to widen the criteria to catch more cases. The nursing informatics lead has data showing the alert fires 47 times per day hospital-wide, with a positive predictive value of 0.18.
1. The CMIO proposes widening the alert criteria. As the informatics lead, how should you respond, and what analysis would you bring to the discussion?
Read the worked answer
Widening the criteria would increase sensitivity but lower an already poor positive predictive value (0.18), worsening alert fatigue and likely further depressing bundle completion. The right response is to reframe the problem: the alert is not the bottleneck; the workflow around it is.
Bring these analyses:
- Signal-to-noise decomposition. Break the 47 daily firings into true positives, false positives, and the three missed night-shift cases. Show that the dominant failure mode is false positives, not missed sensitivity.
- Workflow timing analysis. The three missed cases failed because the alert window closed before the lactate order. That is a logic/timing defect, not a sensitivity defect.
- Bundle-failure attribution. Of the 40 missed bundles, stratify by whether the alert fired, whether it was acknowledged, and whether the delay was in ordering, collection, or result review.
- Cost of false positives. Estimate clinician time per dismissal and correlate with the drop from 78% to 64%.
Recommend a tiered response: tighten the rule to reduce false positives (e.g., require a sustained criterion or add a risk score), fix the alert-window timing so late lactate orders still trigger, and route low-risk night-shift firings to a quieter notification channel. Propose a pilot with pre-specified success metrics (PPV, bundle completion, dismissal rate) before any hospital-wide change. Widening criteria without fixing the noise and timing defects would be a predictable regression.
2. Nursing leadership asks you to design a rapid evaluation to determine whether the alert should be paused. What would you measure, over what interval, and what would trigger a pause?
Read the worked answer
Design a short-interval, multi-metric evaluation rather than a single-metric judgment.
Measures (over 2–4 weeks, stratified by shift and unit):
- Process: alert firing rate per 1,000 patient-days, acknowledgement rate, time-to-acknowledgement, dismissal-without-action rate.
- Outcome: sepsis bundle completion (all-or-none), time-to-first antibiotic, time-to-lactate, ICU transfer within 24h of trigger, mortality among triggered patients.
- Balancing: clinician override reasons (free-text coded), nurse-reported workload, and any documented near-misses on the night shift.
- Safety: missed true cases identified by concurrent chart review (not just triggered patients).
Pause triggers (pre-specified):
- Bundle completion continues to decline beyond a defined threshold (e.g., another 5 percentage points) despite no offsetting gain in case detection.
- Any confirmed sepsis-related harm attributable to alert-driven delay or distraction.
- Dismissal rate exceeding a defined ceiling (e.g., >85%) with PPV below a floor.
Governance: Present to the clinical decision support committee with a recommendation. A pause should be paired with a fallback (e.g., a passive dashboard or a nurse-driven screen) so detection does not go to zero. Document the decision and rationale for the change-control record.