The Post-Market Surveillance Reckoning: Why the Old Playbook Is Running Out of Runway

For thirty years, "checking MAUDE" has been shorthand for post-market vigilance in the medical device industry. It is the database regulatory teams cite in literature searches, the tool clinical affairs pulls up before a submission, the first stop for anyone trying to understand how a device is behaving once it leaves the controlled world of a clinical trial and enters the far messier world of actual patients. By the end of May 2026, it will not exist in its current form. The FDA is folding the Manufacturer and User Facility Device Experience database into a new, unified Adverse Event Monitoring System, or AEMS, that consolidates seven previously separate adverse event databases — MAUDE among them — into a single platform.

On its face, this is a modernization story: fewer silos, better search, a system built for an agency that receives an extraordinary volume of reports. But look past the migration logistics and AEMS is really a symptom of something bigger. Post-market surveillance, as a discipline, has quietly stopped being a periodic compliance obligation and become something closer to a continuous intelligence problem — one that most quality and regulatory organizations are still resourced, staffed, and structured to handle the old way.

An agency drowning in its own data

The numbers explain why a rebuild was overdue. MAUDE has held more than 16 million medical device reports dating back to 1991, and it has been taking on roughly 2 million new reports every year. In its own account to the Government Accountability Office, the FDA said it receives and reviews two to four million adverse event reports annually — a volume that a ten-year search cap, a 500-record-per-query limit, and a fragmented set of legacy systems were never built to handle at scale.

The consequence of that strain shows up downstream, in how long it takes to actually close the loop on a safety signal. A GAO review found that Class I recalls — the FDA's most serious category, reserved for defects with a reasonable probability of causing serious injury or death — took an average of 263 business days from initiation to termination, with a range stretching as high as 882 days. Class II and Class III recalls weren't materially faster, averaging 269 and 248 days respectively. That is the better part of a year, on average, between identifying a serious problem and formally closing it out. And the volume of serious problems has been rising, not falling. Industry recall tracking puts total device recall events at 1,059 in 2024, a four-year high. Class I recalls specifically hit their highest level in fifteen years. Looking at the FDA's own recall and early-alert listings, the agency logged 111 of its most serious recalls and early alerts in 2025, and by the agency's own account, Class I recalls increased 232% between 2020 and 2025 — a trend an FDA official publicly called "concerning" at a RAPS conference session earlier this year, while also acknowledging the agency's device recall program is understaffed and mid-rebuild after 2025 workforce reductions.

None of this reads as a database migration story. It reads as a capacity story — and it raises an uncomfortable question for manufacturers: if the agency responsible for policing post-market safety is stretched this thin, how much of the actual surveillance burden has quietly shifted onto the companies themselves?

What's actually breaking

It helps to know what's driving the recalls in the first place, because the answer has shifted. For decades, manufacturing defects and process control failures were the leading causes of device recalls. That changed in 2024: device failure overtook them as the single largest recall category for the first time in more than five years, accounting for roughly one in nine recall events. Software problems are a meaningful contributor to that shift — devices that pass verification and validation in a controlled test environment but behave differently once exposed to the full variability of real clinical use, real interoperability with other systems, and real user behavior.

This is precisely the category of failure that traditional, periodic post-market surveillance is worst equipped to catch. A quarterly literature review or an annual trend report can identify a slow-building pattern of mechanical failure. It is much less suited to catching a software behavior that only manifests under a specific, uncommon combination of conditions — the kind of signal that tends to hide inside a handful of scattered complaint narratives until someone, or something, connects them.

The price tag, and who actually pays it

It's worth pausing on what a slow surveillance cycle actually costs, because the number is larger than most quality budgets account for. McKinsey has estimated that significant quality events — recalls, FDA enforcement actions, consent decrees, and related litigation — cost the global medical device industry $7 to $8.5 billion annually, roughly 2% of total sector sales. A separate industry estimate puts recall costs alone at up to $5 billion a year. In one widely cited study of major recalls, 52% of affected companies reported a total financial impact exceeding $10 million, and one in twenty saw the impact exceed $100 million.

Those figures also understate the real exposure, because they tend to capture only the direct costs — retrieval, disposal, customer reimbursement, regulatory response. Analysis from outside the device sector suggests direct costs typically represent a third or less of total recall impact, with business interruption, lost contract revenue, and litigation exposure running several times higher and persisting for a year or more after the recall is announced.

The gap between when a defect is detectable and when it is actually detected is where that cost accumulates. A recent recall involving reprocessed electrophysiology and ultrasound catheters illustrates the pattern: the manufacturer notified affected customers directly in July 2025, sent an expanded notification in August, and issued a more detailed customer letter in December describing residual particulate contamination in specific lots. The FDA did not publish the recall until March 2026 — and expanded its own listing again in June 2026 to cover nine device families and 134 lot numbers. Nothing about that sequence suggests bad faith on anyone's part; it reflects how much distance can open up between a manufacturer's internal signal and the point at which the full scope of a problem is understood, particularly across a wide product family. Every week inside that gap is a week where more of the already-manufactured, still-in-field product stays in active use.

The blind spot nobody planned for

That gap becomes sharper still when the device in question is itself powered by AI or machine learning. A systematic review of FDA post-market surveillance data covering roughly 950 AI/ML-enabled devices approved between 2010 and 2023 concluded that the existing surveillance infrastructure is not adequate to properly evaluate the safety and effectiveness of this device category. That is a striking finding for an industry that has approved AI/ML devices at an accelerating pace over the past several years: the monitoring apparatus was largely built for static, hardware-centric devices, and it is being asked to also govern products that can behave differently after a model update than they did at the time of clearance.

Regulators outside the U.S. are wrestling with the same problem from a different angle. The European Society of Radiology recently published consensus recommendations noting that post-market surveillance responsibilities for AI-enabled imaging devices are effectively shared between manufacturers and the clinicians using them, in a regulatory environment that the ESR says still lacks imaging-specific guidance for how that shared responsibility should actually work in practice.

What "good" is starting to look like

The encouraging half of this story is that the same forces creating the problem — data volume, unstructured narrative text, the need for continuous rather than periodic review — are also what artificial intelligence is well suited to address, and there is now real evidence of it working. One published evaluation of an AI-assisted literature surveillance platform found it increased the rate of relevant safety-signal detection roughly tenfold compared with manual literature searching, while also completing the search faster and with fewer resources — a result the authors described as significant for a task that regulatory teams are required to repeat multiple times a year under tight deadlines.

Proposed frameworks in the field now describe AI-assisted post-market surveillance as a three-part workflow: automated triage and signal detection across structured adverse event data, natural language processing applied to the unstructured narrative text that traditionally required a human reader to parse line by line, and predictive analytics that flag an emerging pattern before it reaches recall-triggering severity — feeding directly into earlier, more targeted CAPA activity rather than a reactive scramble after the fact.

The FDA's own Section 522 postmarket surveillance program is a useful proxy for where the agency wants the industry to go. Manufacturers ordered into a 522 study increasingly rely on real-world evidence and registry data rather than bespoke prospective trials, and the FDA has publicly signaled it intends to issue further guidance specifically on using real-world evidence in postmarket surveillance. That is a regulator explicitly asking industry to lean more heavily on the kind of large-scale, real-world data analysis that manual, spreadsheet-driven surveillance processes were never designed to do well.

The strategic reframe

Put these threads together and a pattern emerges that goes beyond any single database migration. The agency is consolidating its systems because the old ones couldn't handle the volume. Recall timelines remain stubbornly long because closing a safety signal loop is fundamentally a data-synthesis problem, not a paperwork problem. The fastest-growing category of device failure is exactly the kind that periodic review is worst at catching. And the device category growing fastest — AI/ML-enabled products — is the one regulators openly admit the current surveillance model wasn't built to handle.

For manufacturers, there's a near-term checklist and a longer-term posture shift, and it's worth separating the two.

The near-term items are procedural but time-bound:

• Update post-market surveillance SOPs to reference AEMS as MAUDE's successor system, and maintain parallel references to both during the transition window rather than switching over all at once.

• Validate, before the legacy system goes dark at the end of May, that AEMS searches on your product codes return results equivalent to what MAUDE was producing — and flag discrepancies early rather than assuming continuity.

• Brief complaint-handling, MDR-reporting, and PMS staff explicitly on what is and isn't changing: the underlying reporting obligations under 21 CFR Part 803 are unchanged, only the system they're filed into is different.

• Revisit literature-surveillance and signal-detection workflows with an eye toward where manual review is the actual bottleneck, rather than assuming the constraint is data access.

  
  

The longer-term shift is the one underneath the systems migration. Post-market surveillance built around a quarterly or annual review cycle, run largely by hand against a handful of databases, is a model designed for a world that no longer exists. The organizations that treat surveillance as a continuously running intelligence function — one that reads every complaint narrative, tracks every emerging pattern across structured and unstructured data alike, and surfaces a signal in days rather than the better part of a year — are the ones that will spend less time explaining a recall after the fact, and more time preventing one.

Sources: U.S. Government Accountability Office; FDA Manufacturer and User Facility Device Experience (MAUDE) database and Adverse Event Monitoring System (AEMS) transition materials; Regulatory Affairs Professionals Society (RAPS); Sedgwick 2025 U.S. State of the Nation Recall Index; FDA Section 522 Postmarket Surveillance Studies Database; Regulatory Rapporteur/TOPRA; European Society of Radiology; McKinsey & Company.