A CMIO who has sat through even one root-cause review knows the pattern before the chart is pulled up. A serious drug interaction fired correctly. The clinician clicked through it in under a second, the same way they click through the other 40 interruptions that day. The alert worked exactly as designed and still failed, because the system around it had trained the clinician to stop reading. That is the paradox at the center of clinical decision support (CDS): a tool built to catch errors can become background noise that hides the one warning that mattered.

Alert fatigue is not a training problem or a discipline problem. It is a design and governance problem, and it has been studied long enough that the field has real data on what works. This piece lays out the mechanics of building CDS that clinicians actually use, drawing on the CDS Five Rights framework, published override-rate research, and the governance structures that separate health systems with usable alerting from those drowning in it.

What Actually Causes Alert Fatigue?

Alert fatigue is the desensitization that occurs when clinicians are exposed to a high volume of low-value interruptions, to the point that they begin dismissing alerts reflexively rather than evaluating them. The Agency for Healthcare Research and Quality’s PSNet primer on alert fatigue traces the problem to two compounding factors: sheer alert volume and low alert specificity. AHRQ cites an ICU physiologic monitoring study that logged more than two million alerts in a single month, roughly 187 per patient per day, and notes that Veterans Affairs primary care clinicians have received upward of 100 alerts daily. When most of that volume is clinically inconsequential, the signal-to-noise ratio collapses and clinicians rationally, if dangerously, learn to override on reflex.

This is why override-rate data is so consistent across health systems and vendors. A widely cited 2013 study in the Journal of the American Medical Informatics Association found that clinicians overrode 96.5% of drug-allergy alerts and roughly 91% of drug-drug interaction alerts, and a separate analysis using 2014 data reported drug-drug interaction override rates around 95%, compared to about 91% for drug-allergy alerts. A follow-up look at the Meaningful Use era found no material improvement in override rates despite years of certification-driven CDS deployment. A systematic review published in JMIR Medical Informatics found average override rates across studies ranging from roughly 46% to 96%, with appropriateness of those overrides varying just as widely. None of this means clinicians are careless. It means the alerts, as designed, were not worth stopping for most of the time they fired — and clinicians correctly learned that pattern.

Why Volume Alone Isn’t the Whole Story

It is tempting to treat alert fatigue purely as a volume problem, but burden varies enormously even among comparable institutions. A six-site pediatric CDS study found alert burden per clinician-day at the highest-burden site was 43.8 times higher than at the lowest-burden site, despite all sites using similar EHR platforms. That gap is a governance and configuration story, not a clinical-population story. It means the variable most within a health system’s control — how alerts are built, tiered, and maintained — is also the biggest lever for reducing fatigue.

What Is the CDS Five Rights Framework?

The most durable model for designing individual CDS interventions is the Five Rights framework, first articulated by Osheroff and colleagues in 2007 and still referenced by AHRQ, HealthIT.gov, and EHR vendors as the baseline design checklist. The framework holds that an effective CDS intervention delivers:

  • The right information — recommendations grounded in current evidence, not stale or generic rule sets
  • The right person — targeted to the clinician, nurse, pharmacist, or care team member who can actually act on it
  • The right format — presented as an interruptive alert, a passive reminder, an order set, a dashboard indicator, or an infobutton, matched to how consequential and time-sensitive the information is
  • The right channel — delivered inside the EHR workflow, a mobile task list, or another medium the user is already working in, rather than a side system nobody checks
  • The right time in the workflow — surfaced at the moment a decision is actually being made, not before the relevant data exists or after the order has already been placed

The framework’s value is that it forces a format and channel decision for every rule before it ships. A drug-allergy contraindication that could cause serious harm justifies an interruptive, hard-stop alert. A guideline reminder about screening intervals does not, and forcing it into the same interruptive channel is precisely how low-value alerts erode trust in high-value ones.

Applying the Five Rights to a Single Rule

Consider a sepsis screening rule. The right information is a validated early-warning score, not a generic vital-signs threshold. The right person is the bedside nurse and covering physician, not every clinician who opens the chart. The right format is a passive dashboard flag for borderline scores and an interruptive alert only above a validated threshold. The right channel is the nursing workflow view where vitals are already being charted. The right time is continuous reassessment, not a single point-in-time check at admission. Each of the five decisions independently affects whether the alert gets used or ignored, which is why treating them as a single design step, rather than five separate ones, is a common source of failure.

Interruptive vs. Non-Interruptive Alerts: How Should Tiering Work?

The single highest-leverage governance decision in CDS design is which alerts get to interrupt the clinician’s workflow at all. Interruptive alerts (hard stops or modal pop-ups requiring an action before proceeding) should be reserved for a narrow set of high-severity, high-confidence scenarios: contraindicated drug combinations with serious harm potential, critical lab values, or hard patient-safety stops like wrong-site or wrong-patient triggers. Everything else belongs in non-interruptive channels — passive flags, sidebar reminders, order-set defaults, or infobuttons the clinician can consult on demand.

A 2022 review of CDS stewardship practices, published in JAMIA Open and summarized on PubMed Central, frames this as a severity-tiering exercise: classify alerts into a small number of tiers (for example, severe, moderate, and minor), then reserve interruptive presentation for the top tier only, with lower tiers delivered non-interruptively. The same review found that acceptance rates for vendor-supplied interruptive alerts were often in the single digits to low double digits — in the 4% to 11% range in the systems studied — which is a strong signal that most vendor-default interruptive rules are miscalibrated for a given institution’s population and workflow and need local tuning rather than blanket adoption.

Practical Tiering Criteria

Health systems that have successfully reduced alert burden typically apply a small set of criteria before allowing any new rule to be interruptive:

  • Does the underlying evidence support a clinically significant, not just theoretically possible, harm?
  • Is the false-positive rate low enough that most firings represent true positives?
  • Has a comparable, currently active alert already covered this risk?
  • Can the same information be delivered non-interruptively (a persistent flag, an order-set default) without materially increasing risk?

A risk-prioritization method borrowed from patient-safety engineering, such as Healthcare Failure Mode and Effects Analysis, gives committees a structured way to answer these questions instead of relying on individual clinician requests or vendor defaults, which tend to over-include alerts out of liability caution rather than clinical necessity.

Who Should Govern CDS, and How?

Alert fatigue is rarely fixed by a single build change; it is fixed by an ongoing governance process that treats every active rule as something to be maintained, measured, and retired, not just deployed once. The CDS stewardship literature recommends a standing alert governance committee with three structural elements.

An Approval and Maintenance Model

New rules and modifications should route through a defined approval body with representation from bedside clinicians, pharmacy, nursing informatics, clinical informatics/IT, quality and patient safety, and administration. Some organizations centralize this in a single committee; others use a federated model with domain-specific subcommittees (medication safety, sepsis, radiology) reporting into a central CDS governance body. Either structure works as long as no rule reaches production without clinical sign-off and a documented rationale.

Justification and Design Standards

Every proposed alert should document the evidence base, the target population, the expected firing frequency, and why the chosen format (interruptive vs. passive) matches the severity. Design standards — consistent color coding for severity, a clear statement of the clinical consequence, and an actionable next step rather than a vague warning — reduce the cognitive load per alert even when volume can’t be cut further.

A Retirement and Re-Tuning Process

Rules should be reviewed on a fixed cadence, not left active indefinitely. If an alert’s override rate exceeds a defined threshold with no improvement after specificity tuning, governance should have explicit authority to retire, narrow, or reclassify it. Without a sunset mechanism, alert inventories only grow, since it is organizationally much easier to add a new rule than to remove an old one that a single stakeholder still wants.

How Should Health Systems Measure Whether CDS Is Working?

Measurement is where many CDS programs stop short. It is common to track how many alerts fire, but far less common to track whether firing changes anything clinically. The CDS stewardship literature distinguishes two measurement tiers that a mature program should track together.

Proximal Measures

These capture immediate user response: acceptance rate, override rate, and time-to-action. They are easy to pull directly from EHR logs and are useful for spotting acutely miscalibrated rules, but a low override rate alone doesn’t prove clinical value — a rule can be widely accepted and still be clinically trivial.

Distal Measures

These capture downstream effect: did the targeted adverse event rate actually fall, did guideline-concordant ordering increase, did length of stay or readmission for the targeted condition change? Distal measures are harder to attribute cleanly to a single alert, but they are the only way to confirm that accepted alerts are producing the outcome the rule was built for.

Alongside acceptance and outcome measures, burden metrics matter for population-level fatigue tracking: alerts per clinician-day, alerts per encounter, and alerts per 100 orders are the four denominators most commonly used in the published literature, and tracking them by service line or by individual rule is what surfaces the outlier alerts responsible for most of the burden, since a small number of high-frequency rules typically account for a disproportionate share of total firings.

What Should a CDS Optimization Cycle Look Like?

Bringing these pieces together, a workable optimization cycle looks less like a project with an end date and more like a standing operational function: new rule proposals are evaluated against Five Rights criteria and routed through governance before build; every live rule is tiered by severity with interruptive status reserved for the top tier; burden and override metrics are reviewed on a recurring schedule at both the aggregate and individual-rule level; and any rule that shows a persistently high override rate without a credible clinical explanation is flagged for re-tuning or retirement. HealthIT.gov’s own guidance frames CDS optimization the same way, pointing to a 2016 National Academy of Medicine collaborative that treated CDS as a continuous improvement discipline rather than a one-time implementation milestone — a framing that has held up even as the underlying EHR technology has changed considerably since.

The organizations that keep alert burden manageable are not the ones with the fewest rules in absolute terms. They are the ones that never let a rule go unmeasured, and that treat “we built the alert” as the start of the work rather than the finish line.

Frequently Asked Questions

What is considered a high override rate for a clinical alert?

There is no single regulatory threshold, but published studies commonly report drug-drug interaction override rates near 90 to 95%, and general CDS override rates ranging from roughly 46% to 96% depending on alert type. Most stewardship programs treat sustained override rates above 90% without a credible clinical rationale as a signal the rule needs re-tuning or retirement.

What is the difference between interruptive and non-interruptive CDS alerts?

Interruptive alerts halt the clinical workflow, typically as a modal pop-up requiring an action before the user can proceed, and should be reserved for high-severity, high-confidence risks. Non-interruptive alerts, such as passive dashboard flags or order-set defaults, convey lower-severity information without blocking workflow, preserving interruptive channels for genuine emergencies.

Who should sit on a clinical decision support governance committee?

Effective committees combine bedside clinical representation (physicians, nurses, pharmacists), clinical informatics and IT staff who understand build constraints, quality and patient safety leadership, and administrative sponsorship. This mix ensures new rules are clinically justified, technically sound, and aligned with organizational safety priorities before deployment.

How often should existing CDS alerts be reviewed?

Published stewardship frameworks recommend a fixed, recurring review cadence rather than one-time deployment, since alert relevance and burden shift as populations, formularies, and workflows change. Reviews should examine override rates, alert burden per clinician, and whether the rule still reflects current evidence, with clear authority to retire or reclassify rules that no longer perform.

Does reducing alert volume compromise patient safety?

Evidence suggests the opposite: indiscriminately high alert volume degrades clinician attention to the point that even critical warnings get overridden reflexively. Tiering alerts by severity and reserving interruptive formats for genuinely high-risk scenarios, as recommended in the CDS Five Rights framework, is intended to protect attention for the alerts that carry real clinical consequence.