A cardiologist notices something, for the third time this quarter, in a patient on a relatively new medication.
Nothing dramatic. Not the kind of reaction that lands someone in the ICU or triggers an automatic incident review. Just a pattern that is starting to feel less like coincidence: a specific, mild but unusual symptom, showing up in patients who share nothing obviously in common except this one drug.
He mentions it to a colleague in the hallway, who says she thinks she has seen something similar, though neither of them is entirely sure, and neither has counted carefully enough to be confident it is more than noise.
He opens the MedWatch form on his hospital's intranet, scans the fields it wants filled in, and closes it. It is not that the form is unreasonable. It is that what he has, honestly, is a hunch: three patients, a pattern he cannot yet articulate with confidence, and a nagging sense that filing a formal report for something this uncertain would make him look like he is chasing shadows.
So he does what the evidence says the overwhelming majority of clinicians in his position do. He says nothing further, to anyone outside that one hallway conversation, and moves on to the next patient.
He is very likely the first person in the country to notice a real pattern, and the only channel available to him requires a level of certainty he does not have yet, so the observation that could have started a signal investigation dies in a hallway instead.
The size of what goes unreported
Start with how large this gap actually is, because the number is more extreme than most people, including many clinicians, would guess.
A systematic review covering 37 studies found a median under-reporting rate of 94 percent for adverse drug reactions to spontaneous reporting systems, with an interquartile range of 82 to 98 percent. That means the vast majority of adverse reactions clinicians actually observe never reach the formal safety-surveillance system at all.
The number does not meaningfully improve for the reactions that matter most. Even for serious or severe reactions, the same review found under-reporting held at 85 percent.
This is not a niche methodological finding buried in a single study. It is the kind of number that, once you sit with it, reframes what "the FDA's adverse event database" actually represents: not a comprehensive picture of what is happening to patients on a given drug, but a thin, systematically incomplete sample of it, shaped as much by what clinicians feel confident enough to formally report as by what is actually occurring in practice.
The barriers are attitudinal, not logistical
It would be reasonable to assume under-reporting this severe reflects a broken, burdensome reporting form. The evidence points somewhere else.
A 2023 systematic review update, covering 65 studies published between 2007 and 2021, catalogued the specific reasons clinicians give for not reporting. The leading barrier, identified in 86.2 percent of the studies reviewed, was the belief that "only serious ADRs need to be reported." General lethargy or lack of interest followed at 84.6 percent. The belief that only safe drugs reach the market in the first place appeared in 46.2 percent of studies. Fear of appearing foolish for reporting a suspected reaction that turns out to be nothing showed up in 44.6 percent. Difficulty determining causation, simply not being sure the drug caused what was observed, appeared in 33.8 percent.
Every one of those barriers describes a confidence and trust problem, not a paperwork problem. A clinician who is unsure whether an observation is "serious enough," worried about looking foolish, or uncertain about causation is not deterred by form length. He is deterred by the binary nature of the choice in front of him: file a formal, permanent report, or say nothing.
There is no built middle option: no low-friction way to say "I think I might be seeing something, has anyone else" before committing to the formal step. That missing middle option is, on this evidence, most of the actual problem.
The 2023 review's authors made a point worth sitting with directly: it is an explicit update of a 2009 predecessor study, covering essentially the same terrain fourteen years later, and found "minimal improvement" in these attitudinal barriers over that period, despite them being identified as modifiable through education back in 2009. Fourteen years of awareness that the problem is attitudinal, and fourteen years of the same barriers persisting largely unchanged.
Why a hunch is exactly what matters most
Here is the part of this problem that is easy to miss if you think about signal detection only from the regulator's side rather than the clinician's.
A real safety signal's earliest statistical signature is not one clinician's confident, well-documented case. It is several independent clinicians, in different cities, each individually uncertain, each separately noticing something they cannot fully explain, each discounting their own single observation as probably nothing.
That is precisely the observation the current system is worst at capturing. The formal reporting pathway requires a clinician to decide, alone, that a specific observation clears the bar of "worth the friction of a formal report." That filter systematically screens out the tentative, "I'm not sure but" observations, which is exactly the category where convergent, independent noticing across multiple unconnected clinicians would be the real signal, if only it could be aggregated before each of them separately decided, alone, that it probably was not worth mentioning.
What the current workflow actually looks like
Trace the path an observation takes today, end to end.
A clinician notices something unusual. He either files a formal MedWatch or FAERS report, a structured, multi-field form that most clinicians experience as burdensome relative to how confident they feel about what they are reporting, given the 94 percent aggregate under-reporting rate found across the literature. Or he mentions it informally to a colleague, with no record kept anywhere. Or he does neither and simply moves on.
Pharmacovigilance teams, at the FDA and inside pharmaceutical companies, then work primarily from whatever sparse stream of formal reports results, supplemented by slower claims-based and registry-based signal-detection methods that lag real-world clinical observation, in the nature of how claims data accumulates and gets analyzed, by months or longer.
Why nobody owns the missing middle layer
Walk through the institutions touching this problem, and the reason none of them has built the low-friction, pre-formal layer this evidence points toward.
The FDA's FAERS and MedWatch system is a statutory compliance repository. It was built and is maintained to collect formal reports in a structured, legally defensible format, not to capture tentative, low-confidence observations before a clinician has decided whether they rise to that level. Redesigning it to also serve as an informal pre-reporting layer would be a fundamentally different mandate than the one it currently has.
Pharmaceutical company pharmacovigilance departments are structurally motivated to manage, not maximize, informal signal surfacing about their own products. This is not a claim of bad faith so much as an observation about incentives: a company's safety department exists within a business that has an obvious interest in signals about its own drug being handled carefully and deliberately, not amplified before they are well understood.
EHR vendors capture adverse-event data only when a clinician deliberately codes a structured field for it, which happens, by every indication in the reporting literature, even less often than formal MedWatch submission. The EHR is built around documentation for billing and clinical continuity, not around capturing a clinician's tentative, half-formed suspicion.
No commercial actor's incentives currently align with building a fast, low-friction, trusted channel for exactly the kind of observation this problem needs captured: the "I think I might be seeing something" moment, before a clinician has decided it is worth a formal report.
What already exists elsewhere, as a comparison point
The UK's Yellow Card scheme and the WHO's Uppsala Monitoring Centre operate broadly analogous spontaneous-reporting systems internationally, and under-reporting is a well-documented, persistent challenge for pharmacovigilance systems generally, not a uniquely American failure. This article did not independently verify precise comparative under-reporting rates across these systems and does not claim the US figure is uniquely severe; the point is that the underlying mechanism, a formal-only reporting pathway with no lower-friction pre-step, appears to be common to spontaneous-reporting systems as a category, not a single country's design flaw.
What would actually work
A low-friction flagging mechanism, distinct from formal reporting. The core design has to create space for a genuinely tentative observation, "has anyone else seen this," that requires far less commitment than filling out a MedWatch form, precisely because the evidence shows confidence, not effort, is the binding constraint.
Specialty and drug-class routing. An observation about a cardiology drug needs to reach other cardiologists prescribing that same drug, not a generic pool of clinicians, for convergent noticing to be recognized quickly rather than lost in noise.
Opt-in aggregation, visible to the reporting clinician. Showing a clinician "three other verified clinicians flagged something similar this month" directly addresses the fear-of-appearing-foolish barrier (44.6 percent of studies) by converting an isolated hunch into a validated, shared observation before anyone has to commit to a formal claim alone.
A clear, one-click bridge to formal reporting once a signal gains peer corroboration. The tentative layer has to feed into, not compete with or delay, the formal MedWatch system once a pattern looks real; it is a funnel toward better formal reporting, not a substitute for it.
Absolute clarity that this never substitutes for mandatory reporting. Certain settings carry legal mandatory ADR reporting obligations, and any informal layer has to be explicit, structurally, that it does not satisfy or delay those obligations where they apply.
No patient identifiers, ever. The conversation is about the clinical pattern and drug-class question, never chart-level detail, consistent with how de-identified peer consultation has to work everywhere else in medicine to be legally and ethically viable.
Strict firewalls against pharmaceutical company influence over what surfaces. For this to be trusted by the clinicians who would need to use it, it has to be structurally independent of the companies whose products are being discussed, funded in a way that does not create even the appearance that a manufacturer can suppress or shape which signals get attention.
What you can do now
If you are a prescribing clinician
Say the tentative thing out loud to a colleague, and count it. The evidence shows this already happens informally and constantly; the missing step is treating that hallway observation as data worth tracking, even just for yourself, rather than something to mention once and forget.
Report anyway, even when unsure. The 33.8 percent of studies citing difficulty determining causation as a barrier describes a real uncertainty, but MedWatch and FAERS are explicitly designed to accept suspected, not confirmed, associations; a report does not require you to be certain, only to have noticed something.
Ask your department whether anyone tracks near-miss or pattern observations informally. If three colleagues have each separately noticed something and never compared notes, a five-minute conversation at a department meeting can surface a pattern none of you could see alone.
If you lead pharmacovigilance or patient safety
Treat the 94 percent under-reporting figure as a design problem, not a compliance problem. Reminder campaigns and simplified forms address the wrong barrier if the dominant issue, per the literature, is confidence and fear of looking foolish rather than form length.
Build an internal, low-stakes way to surface tentative observations before they become formal reports. Even an informal departmental practice, a standing five-minute agenda item for "anything odd you've noticed lately," starts capturing exactly the signal that formal channels are currently missing.
Give feedback after every report, formal or informal. Lack of feedback after reporting was itself cited as a barrier in the 2023 review; clinicians who never learn what happened to a report they filed have less reason to file the next one.
If you build systems
Design the tentative-observation layer to feel nothing like a compliance form. The entire value proposition depends on this being lower-friction and lower-commitment than MedWatch; if it inherits MedWatch's structure, it will inherit MedWatch's 94 percent non-participation problem too.
Build the escalation path to formal reporting as a first-class feature, not an afterthought. The system's legitimacy depends on visibly strengthening, not competing with, the formal pharmacovigilance infrastructure that already exists and that no serious regulator will accept being bypassed.
Frequently asked questions
Why do doctors underreport adverse drug reactions? Primarily due to attitudinal barriers rather than logistical ones. A 2023 systematic review of 65 studies found the leading reasons were the belief that only serious reactions need reporting (86.2 percent of studies), general lack of engagement (84.6 percent), belief that only safe drugs reach market (46.2 percent), fear of appearing foolish (44.6 percent), and difficulty determining causation (33.8 percent) (Garcia-Abeijon et al., Drug Safety, 2023).
What percentage of adverse drug reactions go unreported? A systematic review of 37 studies found a median under-reporting rate of 94 percent, with an interquartile range of 82 to 98 percent. Even serious or severe adverse reactions were under-reported at a rate of 85 percent (Hazell and Shakir, Drug Safety, 2006).
How does the FDA detect new drug safety signals? Primarily through its FAERS spontaneous reporting system, supplemented by claims-based and registry-based active surveillance methods such as the FDA's Sentinel Initiative. Because spontaneous reporting captures only a small fraction of actual adverse reactions, per the 94 percent median under-reporting finding, these systems work from a structurally incomplete picture of real-world drug safety experience.
What are the main barriers physicians cite for not reporting side effects? The most commonly cited barriers across 65 studies reviewed in 2023 were the belief that only serious reactions warrant reporting, general disengagement from the reporting process, the assumption that marketed drugs are inherently safe, fear of appearing foolish over an uncertain observation, and difficulty establishing that the drug actually caused the reaction (Garcia-Abeijon et al., Drug Safety, 2023).
Would a peer network actually help doctors report drug problems faster? This is a plausible, evidence-consistent hypothesis rather than a proven outcome: because the dominant barriers are attitudinal (confidence, fear of judgment, causal uncertainty) rather than logistical, a lower-friction way to compare notes with trusted peers before committing to a formal report addresses the barriers the literature actually documents. No pilot data confirming this specific mechanism works was available for this article.
Has underreporting of adverse drug reactions improved over time? Not meaningfully, according to the available evidence. A 2023 review that explicitly updated a 2009 predecessor study on the same question found "minimal improvement" in the attitudinal barriers driving underreporting over that fourteen-year period, despite those barriers having been identified as modifiable through education back in 2009 (Garcia-Abeijon et al., Drug Safety, 2023).
The bottom line
The cardiologist in the hallway is not being careless, and he is not an outlier. He is doing exactly what the literature says the overwhelming majority of clinicians do with an uncertain observation: he notices, he mentions it once to someone nearby, and then, absent from a way to make that observation count for anything, he lets it go.
A median 94 percent of adverse drug reactions never reach the formal reporting system that exists to catch them. Even serious reactions go unreported 85 percent of the time. And fourteen years after researchers first identified the attitudinal barriers behind that gap, a follow-up review found almost no improvement, despite those same barriers being flagged as fixable the whole time.
The fix the evidence points toward is not a better form. It is a different kind of channel entirely: one built for the tentative, uncertain, "has anyone else seen this" observation that the current binary choice, file a formal report or say nothing, was never designed to hold. Nothing about that channel needs to compete with or delay mandatory formal reporting; done right, it should feed more, better-formed reports into that system, not fewer.
Until it exists, the earliest, most valuable signature of a real safety signal, several independent clinicians each half-noticing the same thing, keeps dying the same way it died in that hallway conversation: real, shared between two people, and then gone.
Part of a series on the missing professional infrastructure of healthcare. Previously: The Invisible Contribution Economy
Evidence note: the under-reporting figures (median 94 percent across 37 studies, interquartile range 82 to 98 percent, 85 percent even for serious reactions) are from Hazell L, Shakir SAW, Drug Safety, 2006. The attitudinal barrier figures (86.2 percent, 84.6 percent, 46.2 percent, 44.6 percent, 33.8 percent, and the "minimal improvement" finding relative to a 2009 predecessor review) are from Garcia-Abeijon P, Costa C, Taracido M, Herdeiro MT, Torre C, Figueiras A, Drug Safety, 2023, a review of 65 studies published 2007 to 2021. This article did not locate a verified, sourced historical drug-safety episode with a reconstructible timeline between earliest informal clinical suspicion and formal regulatory action, and deliberately does not name or date any specific past drug-safety case; any general reference to historical signal-detection delay and associated cost should be read as a plausible pattern consistent with the pharmacovigilance literature, not a documented specific example. Comparative claims about the UK Yellow Card scheme and WHO's Uppsala Monitoring Centre are limited to noting that under-reporting is a documented challenge across spontaneous-reporting systems generally; this article did not independently verify or compare precise under-reporting rates across countries. Nothing in this article is clinical or regulatory guidance for reporting any specific adverse event.
