412,000 Alerts and a Petition to Turn Them Off
You are the clinical informatics pharmacist for a 500-bed academic medical center. Last year the EHR fired 412,000 medication-related alerts and the overall override rate was 91%. Drug-drug interaction alerts were overridden 96% of the time and duplicate-therapy alerts 94%; allergy alerts 88%; renal dose alerts 71%. A pharmacy resident reviewed 20 overridden renal alerts and found four genuine dosing errors that reached patients. Physicians have petitioned to disable drug-drug interaction alerts system-wide. The vendor states the alert logic is validated content and that any customization requires a $75,000 change order. Nursing is not at the table. The pharmacy and therapeutics committee meets monthly and owns medication policy.
1. How do you decide which alerts to keep, modify, and remove?
Read the worked answer
Override rate is not the decision variable. It tells you that clinicians clicked a button, not whether the alert was correct. The decisions should rest on positive predictive value and actionability, segmented by alert class.
Method:
- Segment the data by alert type, severity tier, drug pair, ordering specialty, and patient population. A 96% override rate on class-level interaction alerts is a different problem from a 71% override rate on renal dosing, and averaging them hides both.
- Sample and adjudicate. Pull a defensible sample of overridden alerts per segment and review the chart against the intended action. The resident's 4-of-20 renal finding is the right kind of signal, but 20 charts is a pilot, not a conclusion - expand it before you act on it.
- Triangulate with sources that do not depend on alerts firing: ADE reports, incident reports, pharmacist interventions, and trigger tools.
- Use published benchmarks as context, not policy.
Decision rule: keep alerts with high severity and demonstrated actionability even at low volume; modify alerts whose content is right but whose targeting is wrong (wrong population, wrong workflow step, wrong user); remove alerts that are non-actionable by design or redundant with another control.
2. Physicians want drug-drug interaction alerts disabled system-wide. How do you respond, and what is the governance path?
Read the worked answer
Answer plainly: no blanket disablement. But do not answer with 'the vendor says so' - that argument loses the room.
To the physicians: agree with the real complaint - volume and irrelevance, not the existence of interaction checking - and commit to a data-driven decision within 60 days. Offer them a seat in the review so the eventual change is theirs as much as yours.
Governance: the P&T committee owns medication policy. The proposal goes there with segmented data, chart-review findings, a written recommendation, and a defined evaluation period. Include pharmacy and nursing; a physician-only vote on alert design has predictable failure modes, and nursing currently has no seat.
Substantive options that may not require the $75,000 change order:
- Suppress non-actionable and class-duplicate interactions using institution-level rules layered on top of vendor content, which often does not require modifying the vendor library itself. Confirm this in writing with the vendor before promising it.
- Move selected alerts from order entry to pharmacist verification on high-risk units.
- Require structured override reasons instead of free text, so the next analysis is possible.
- Tailor by specialty where the evidence supports it.
Document the decision, the rationale, and the scheduled review. A reversible, monitored change is defensible; an unmonitored blanket disablement is not.
3. How will you evaluate the change at six months?
Read the worked answer
- Alert volume per 1,000 orders, by class - the burden measure.
- Override rate and override-reason distribution together. A falling override rate accompanied by rising free-text 'OK' entries is not success; it is a quieter version of the same problem.
- Positive predictive value from a repeat chart-review sample using the same method as baseline.
- Safety counterweight: confirmed ADEs, pharmacist interventions, and specifically any event that a removed alert would have caught. Someone must be assigned to look for this rather than waiting for it to surface.
- Workflow proxy: time-to-order for high-volume medications.
- Clinician perception survey using the identical instrument as baseline.
- Alert migration check: new or previously rare alerts firing more often because logic upstream changed.
Put it on a single dashboard reviewed monthly at P&T. The point of the exercise is a standing feedback loop, not a one-time cleanup.