Manual post-market surveillance breaks when report volume outgrows the attention available to read it.
Nearly 2.9 million device adverse-event reports reached the FDA in 2025, about 7,900 a day. A single manufacturer sees a fraction of that, but the fraction grows with every product code added and every similar device the procedure covers. The usual responses are to read faster, hire another analyst or quietly narrow what counts as in scope. Each delays the failure without changing the method.
What one signal takes
An analyst searches MAUDE for a product code, exports the results into a spreadsheet, then opens each report and reads the narrative. The narratives are dense and inconsistent. One describes a lead dislodgement in three clauses. Another buries a patient death in a paragraph of device-history boilerplate.
For each report the analyst decides the failure mode, classifies severity, notes the patient outcome and flags anything that might escalate. Then they repeat it a few hundred times a month, across products that don’t fail in convenient batches.
The judgment is skilled work. The copying and re-reading around it are not.
Where manual review fails
The first failure is the missed event. A record that changes the pattern can read much like a routine one, and attention thins as the pile grows.
The second is latency. A quarterly review can leave a new trend unnoticed for up to three months. A competitor’s recall reaches you through a news alert, though the reports behind it were available earlier.
The third is the crunch before each Periodic Safety Update Report (PSUR). The data exists, scattered across spreadsheets, PDFs and database exports. Each cycle the source trail is rebuilt from fragments before qualified review can begin.
A good team with a manual method still hits all three. The method assumes people can keep pace with volume that has no ceiling.
The case for a person on every report
Adverse-event narratives are ambiguous and severity is a judgment call. A classifier that is confidently wrong is worse than a slow, correct reader, because it writes the mistake into a record.
I think that argument is right. It is a reason to keep the person and take away the parts of the job that are not judgment.
What can be automated
Ingestion scales cleanly. A new product can receive its MAUDE history within minutes, then synchronize against openFDA’s weekly refresh.
Extraction scales too. A model can read a narrative and pull the failure mode, severity, patient outcome and FDA codes in seconds, with a confidence score and an explicit flag when the data is thin. I walked through how in how AI reads adverse event reports.
The review stays human. Each summary lands in a queue where someone acknowledges, flags or rejects it, with a signed decision when the procedure requires one. The analyst reads a structured summary beside the source narrative and spends attention on the calls that need it.
The bottleneck moves from how many narratives one person can read to how many decisions need a person. The second number is much smaller.
The bet a manual program makes
A manual program bets that volume won’t outpace the team. More products, more competitors and more reports arrive each quarter against the same number of hours.
DeviceWatch is built around that split. Reading, classifying and assembling run automatically, and the decisions stay with your team.