Sepsis Alert Fatigue in a Community ED
A 220-bed community hospital turns on a commercial EHR sepsis best-practice advisory. It fires when 2 of 3 SIRS criteria plus a suspected-infection flag are present. In month one it fires 1,140 times across the ED and floors. Chart review of a random 100 alerts shows 71 were dismissed without action; of those, 9 patients later met Sepsis-3 criteria within 12 hours. The ED medical director wants the alert disabled entirely, citing alarm fatigue. Nursing leadership wants it kept because two of the missed cases went to the ICU. Baseline SEP-1 bundle compliance is 58%, and the informatics team has 6 weeks before a Joint Commission survey. The vendor offers no native threshold tuning without a change request costing 8 weeks.
1. As the clinical informatics lead, how do you quantify whether this alert is worth keeping, and what intermediate option do you propose given the 6-week constraint?
Read the worked answer
Start with the operating characteristics rather than the raw fire count. Positive predictive value from the sample is roughly (29 actionable)/(100 reviewed) ≈ 29% overall, but the clinically relevant number is the miss rate: 9 of 71 dismissed alerts (≈13%) had a downstream Sepsis-3 event, so blanket dismissal is unsafe. Frame the decision as a harm trade-off: alerts per patient-day vs. preventable deterioration. Since vendor tuning is 8 weeks out, propose a locally configurable interim: (1) restrict firing to ED and high-acuity units where base rate is higher, raising PPV; (2) suppress re-firing for 24h after a dismissal by the same clinician unless new objective data (lactate, hypotension) arrives; (3) add a structured dismissal reason to generate data for the eventual change request. Keep the alert live but reduce volume, and commit to a 30-day re-measure using the same sampling method so the medical director and nursing see comparable data. Document the risk decision with the patient safety committee.
2. What evaluation design would you use to show the alert's effect on SEP-1 compliance and patient outcomes without a randomized trial, and what confounder most threatens it?
Read the worked answer
Use an interrupted time series (or a difference-in-differences if a comparable non-adopting unit exists) with at least 12 pre- and 12 post-intervention monthly points on SEP-1 bundle compliance and a balancing measure like antibiotic time-to-first-dose. ITS handles secular trends better than a simple pre/post comparison. The dominant confounder is co-intervention: the same period likely includes education campaigns, order-set changes, or survey-driven attention, so bundle gains cannot be attributed to the alert alone. Mitigate with a control series (e.g., a parallel hospital or non-targeted unit), segment the analysis, and pre-specify the primary outcome. Report both intended effects (compliance) and unintended effects (alert burden, override rates, time-to-antibiotic in dismissed cases).
3. The ED director asks you to simply turn the alert off. What is your governance response?
Read the worked answer
Decline unilateral disabling and route it through the clinical decision support governance committee with a written risk assessment. Present the data: volume, PPV, miss rate, and the two ICU escalations. Offer a time-boxed modification (scoped firing plus suppression) with defined success metrics and a rollback plan if volume does not fall and miss rate rises. Record the decision, owners, and review date. This protects patients, distributes accountability across clinical and informatics leadership, and creates an audit trail for the upcoming survey.