Physician response to drug-drug interaction alerts — severity tiering, interruptive vs. passive design, override rate, and alert fatigue in the EHR
The earliest drug-drug interaction (DDI) checking modules in electronic health records made a simple, well-intentioned design choice: fire an alert for every interaction present in the underlying drug knowledge base, regardless of clinical severity. The result, documented across decades of clinical informatics literature, is a torrent of interruptions that clinicians learn to dismiss reflexively — undermining the very safety goal the alerts were built to serve.
Commercial drug knowledge bases (First Databank, Multum, Medi-Span) encode tens of thousands of documented pairwise interactions, spanning everything from theoretical pharmacokinetic curiosities to combinations that are fatal within minutes. Early computerized physician order entry (CPOE) systems treated this list as flat: any match between two active medications on a patient's list triggered the same interruptive modal, with the same visual weight, the same required click-through, and the same generic language ("A significant interaction has been detected").
This design ignores a foundational finding of clinical decision support (CDS) science: alert value is a function of specificity, not just sensitivity. A system tuned to catch every possible interaction necessarily catches enormous numbers of interactions that carry negligible clinical risk in the actual patient context — an already-tolerated combination, a short course, a dose too low to matter, an interaction the ordering physician already compensated for. When 85–90% of fired alerts are judged irrelevant by the clinician who receives them, the alert stream stops functioning as a warning system and starts functioning as background noise that must be cleared to keep working.
"Alert fatigue" describes the well-documented process by which repeated exposure to low-value interruptions desensitizes a clinician's response — not only to the alerts being dismissed, but to all alerts of that class, including genuinely dangerous ones. It is not a metaphor; it is measurable in override-rate curves that climb over a shift and in response-time data showing modal dismissal accelerating from seconds to a practiced, unread reflex.
The landmark systematic review by van der Sijs, Aarts, Vulto, and Berg (JAMIA 2006) pooled override-rate data from dozens of CPOE deployments and found rates for drug safety alerts ranging from 49% to 96%, with DDI-specific alerts clustering toward the high end of that range at many sites. A widely cited follow-up study by Isaac and colleagues (Archives of Internal Medicine, 2009) found override rates near 90% for drug interaction alerts in ambulatory CPOE — including a nontrivial share of alerts flagging combinations with genuine, serious clinical risk.
This is the central paradox of unfiltered alerting: a system designed to never miss a true positive, by presenting every true positive with the same urgency as every false positive, ends up training clinicians to miss true positives anyway.
Beyond the direct patient-safety risk of desensitized clinicians, an unfiltered alert flood carries measurable operational costs: added seconds per order multiplied across hundreds of daily orders per clinician translate into meaningful lost time across a shift; documented "alert burden" correlates with physician burnout metrics in multiple health-system surveys; and workaround behaviors emerge, including batch order entry specifically to minimize alert exposure, pre-emptive override rationales copy-pasted across encounters, and clinician requests to disable entire alert categories at the system level — the opposite of the intended safety posture.
This stage of the simulation models that unfiltered state: every generated interaction, minor to contraindicated, fires as an interruptive modal with identical visual treatment. Watch the override pile grow almost as fast as the accepted pile, and the fatigue gauge climb quickly even though only a small fraction of these alerts represent real danger.
The first structural fix to alert fatigue is not reducing alert volume directly — it is classifying what already fires. A severity classification engine maps every interaction pair to a discrete tier (minor, moderate, major, contraindicated) sourced from a maintained drug knowledge base, so that downstream alerting logic has something meaningful to key off of.
First Databank (FDB), Cerner Multum, and Medi-Span — the three drug knowledge bases underlying the large majority of U.S. EHR CDS — each maintain clinical pharmacist-curated interaction monographs. Each documented pair is scored along axes that typically include: severity of the potential clinical outcome (from asymptomatic lab-value shift to life-threatening), quality and quantity of supporting evidence (case report vs. controlled trial vs. established pharmacologic mechanism), and onset (rapid vs. delayed). These axes are collapsed into a small number of discrete severity tiers that CDS logic can act on — commonly Minor, Moderate, Major, and a distinct Contraindicated tier reserved for combinations with a well-established, severe, and largely unavoidable risk (e.g., two drugs that together reliably precipitate a fatal arrhythmia).
Critically, the contraindicated tier is deliberately narrow: across major knowledge bases, well under 2% of all catalogued interaction pairs meet the bar. This scarcity is what makes a contraindicated-tier hard stop tolerable to clinical workflow — if 40% of alerts were "contraindicated," a hard stop on that tier would simply recreate the original flood problem under a new label.
Tagging severity alone does not fix alert fatigue if every tier is still displayed identically. What severity tagging enables is everything downstream: it lets a health system's CDS governance committee make deliberate, auditable policy decisions — "Minor and Moderate interactions display as a passive, non-blocking indicator; Major interactions interrupt with a dismissable modal; Contraindicated interactions require an attending-level override reason or block the order outright."
This is the pattern recommended by CDS design consensus work, notably Phansalkar and colleagues' JAMIA 2013 publication proposing a specific, limited list of DDIs that should be non-interruptive to reduce alert fatigue, paired with companion work identifying a short list of interactions serious enough to warrant hard, order-blocking stops. The core insight: interruption is a scarce resource. Spending it on every tier bankrupts its value for the tier that actually needs it.
The severity threshold control in this simulator approximates a site-level CDS configuration choice: should the system surface minor interactions at all, or only escalate moderate-and-above, major-and-above, or contraindicated-only? Raising the threshold sharply cuts total alert volume (fewer pairs qualify to fire anything) at the cost of some sensitivity to genuinely low-risk-but-non-zero interactions. Most CDS governance committees converge on "moderate and above" or a customized allowlist as the practical default — full suppression of minor interactions, with moderate/major/contraindicated retained and differentiated by interruption style, which is the subject of the next stage.
Once severity is known, the second lever is interruption style: does the alert require the clinician to stop, read, and act before proceeding (interruptive), or does it surface as a persistent but non-blocking indicator the clinician can review at their own pace (passive)? This stage differentiates alert presentation by tier — the design pattern most consistently associated with reduced override rates for high-severity alerts in the CDS literature.
CDS alert presentation spans a spectrum rather than a binary. At the passive end, a small sidebar or chart-header icon indicates an interaction exists, viewable on demand — no workflow interruption, no forced acknowledgment, appropriate for well-documented but low-urgency findings. Moving inward: a non-modal banner that appears in the order-entry view but does not block navigation; a dismissable modal that must be clicked through but accepts a single click with no required justification; a modal requiring a structured override reason (e.g., "monitoring in place," "benefit outweighs risk," "prescriber aware") before dismissal; and, at the interruptive extreme, a hard stop that blocks order submission entirely absent a higher-privilege override, typically reserved for the contraindicated tier.
Each step toward the interruptive end increases the probability the clinician actually engages with the content of the alert — but also increases workflow friction, time cost, and (if overused) the incentive to develop systematic workarounds.
The core design principle validated across multiple health-system CDS optimization studies: interruption should be reserved for alerts where (a) the potential harm is severe, (b) the evidence is strong, and (c) an action by the clinician meaningfully changes the outcome. Interactions meeting all three criteria are rare — which is precisely why the contraindicated tier is small. Applying hard stops only to that narrow tier keeps the total number of forced interruptions low enough that clinicians do not habituate to overriding them, while passive presentation for lower tiers still preserves the information (visible in the chart, reviewable, auditable) without imposing a workflow tax on every order.
Sites that adopt this differentiated approach consistently report the same pattern: total interruptive alert volume drops by an order of magnitude, override rates for the alerts that remain interruptive fall substantially (because they are now rare enough to be taken seriously), and — critically — the true-positive catch rate for clinically significant interactions is preserved or improved, because clinician attention is no longer diluted across thousands of low-value interruptions.
Hard stops are a powerful but narrow tool, and CDS governance literature is consistent in warning against overuse. When hard stops are applied broadly — for example, blocking orders for any "major" tier interaction rather than only the narrower "contraindicated" tier — clinicians facing a clinically appropriate but technically-interacting order (common in complex, polypharmacy patients) are forced into workarounds: canceling and re-entering orders under a different route, contacting pharmacy to manually override, or documenting a boilerplate justification that adds documentation burden without adding clinical scrutiny. Each of these workarounds erodes both efficiency and the audit trail's value. The interruption style slider in this simulator makes that trade-off visible: pushing toward maximal interruptiveness for every tier reproduces flood-like fatigue even after severity tagging is in place.
Alert design does not operate on a passive recipient — clinicians adapt their behavior in response to the alerting environment, and that adaptation is itself a major determinant of whether a CDS system achieves its safety goal. This stage examines the behavioral feedback loop between alert volume, interruption style, and the clinician's evolving decision heuristics across a shift.
Cognitive science offers a direct mechanism for the override curves observed in CDS override-rate studies: habituation, the well-established decline in behavioral response to a repeated, non-reinforced stimulus. Each low-value alert dismissed without consequence very slightly lowers the threshold of attention paid to the next one. Because most fired alerts in an unfiltered or poorly tiered system are low-value, this habituation generalizes: by the middle of a shift, a clinician's "System 1" (fast, automatic) response to any interruptive alert is dismissal, not evaluation — regardless of the alert's actual severity tier, unless the visual and interaction design forcibly interrupts that automaticity.
This is precisely why differentiated interruption style (Stage 3) matters causally, not just cosmetically: a contraindicated-tier hard stop that a clinician encounters perhaps once or twice per shift does not benefit from — or suffer from — the same habituation curve as a moderate-tier alert encountered forty times per shift.
A frequently under-discussed driver of alert-flood persistence is medicolegal: health systems and EHR vendors have historically erred toward firing more alerts, not fewer, out of concern that suppressing an alert — even a low-value one — could be characterized in a malpractice claim as the system failing to warn. This produces a "defensive alerting" equilibrium where liability-driven caution generates alert volumes that clinical usability evidence says are actively harmful to safety, because they drive the exact override-habituation behavior that lets truly dangerous alerts slip through unread.
The resolution favored by CDS governance and legal-informatics scholarship is not fewer safeguards but better-documented, evidence-based tiering: a structured, defensible, professionally reviewed severity and interruption policy (of the kind produced by pharmacy and therapeutics committees using published consensus lists) is a stronger liability position than an undifferentiated flood, precisely because it demonstrates deliberate, clinically reasoned design rather than reflexive alert generation.
Multiple health-system CDS optimization projects report the same counter-intuitive result: cutting total DDI alert volume by 50–90% through tiering and passive redesign did not increase adverse drug event rates — and in several published deployments, was associated with measurable decreases, because the alerts clinicians still saw were more likely to be read.
Requiring a structured override reason — a short pick-list ("Patient already tolerating combination," "Will monitor labs," "Benefit outweighs risk," "Alternative unavailable") rather than a single dismiss click — has a measurable effect independent of the underlying alert content: it forces a brief cognitive checkpoint before dismissal, and it generates auditable data that CDS governance teams use to identify which alerts are being routinely overridden for legitimate clinical reasons (candidates for re-tiering to passive) versus which are being overridden without adequate justification (candidates for tighter interruption or targeted clinician education). This turns the override log itself into a continuous improvement signal for the alerting system — the mechanism underlying the optimization achieved in the final stage.
The endpoint of severity classification, interruption differentiation, and override-log-driven refinement is a tiered alerting system that fires far fewer interruptions overall while directing the interruptions it does fire toward the small set of combinations where clinician action genuinely changes outcomes. This is not a reduction in safety rigor — the underlying interaction knowledge base is unchanged — it is a redesign of how that knowledge is surfaced.
An optimized tiered DDI alerting configuration typically combines several concrete design decisions validated in the CDS literature: minor interactions suppressed from interruptive display entirely (visible passively in the medication list or chart, not surfaced as an alert); moderate interactions consolidated into a single non-blocking summary rather than one alert per pair; major interactions presented as a dismissable interruptive modal requiring a structured override reason; and a short, evidence-vetted contraindicated list — often drawn from or adapted from published consensus lists such as Phansalkar et al.'s high-priority interruptive DDI set — triggering a genuine hard stop that requires pharmacist or attending-level override to bypass.
The practical effect: a clinician who previously saw 60–100 DDI alerts per shift, nearly all dismissed without reading, now sees perhaps 5–15 interruptive alerts, each one meaningfully more likely to represent a combination worth stopping for — and, because the volume is low enough to avoid habituation, meaningfully more likely to actually be read.
It is intuitive to assume that firing fewer alerts must mean catching fewer true positives — but this assumes catch rate is purely a function of what is displayed, ignoring that catch rate is a function of what is displayed times the probability the clinician actually attends to it. Sensitivity (interactions detected by the underlying knowledge base) is preserved unchanged in a tiered redesign; what changes is specificity of the interruptive channel and, as a direct consequence, attentional yield. Reducing false-positive interruptive burden by 80–90% while keeping true-positive detection logic fully intact is precisely the mechanism by which several published health-system deployments report stable or improved detection of clinically significant interactions alongside dramatically lower override rates for the alerts that remain interruptive.
Tiered alerting is not a static configuration — it requires a standing pharmacy and therapeutics (P&T) or CDS governance committee that periodically reviews override logs (per Stage 4's structured-reason data), incorporates knowledge base updates (monthly FDB/Multum/Medi-Span releases can add or re-tier interactions), and adjusts thresholds as new evidence emerges. Systems that treat tiering as a one-time launch activity tend to drift back toward flood-like states as new drugs and new interaction entries accumulate untiered at the default (often maximally interruptive) setting. The healthiest deployments couple this simulator's two control axes — severity threshold and interruption style — to a recurring, data-driven review cycle rather than a fixed initial configuration.
Van der Sijs and colleagues' foundational 2006 review found published DDI alert override rates ranging from 49% to 96%. That wide range is not primarily a difference in clinician diligence across institutions — it is very largely a difference in alerting system design: the same clinician who dismisses 95% of an unfiltered flood will act on the great majority of a well-tiered, low-volume, high-specificity contraindicated-tier hard stop.