A language model can turn an adverse event narrative into structured data in seconds. That changes which part of post-market surveillance needs a person, and it does not remove the person.
Nearly 2.9 million device adverse-event reports reached the FDA in 2025. The narratives are free text written by different reporters, with inconsistent terms, abbreviations and gaps. Some reports have no published narrative and others get one later. Reading and classifying them is where the hours go.
What a model extracts from a narrative
Take one sentence: “The catheter tip fractured during insertion, migrating to the right atrium, requiring surgical retrieval.”
A model can pull out the failure mode (fracture), the component (catheter tip), the complication (migration), the location (right atrium) and the intervention (surgical retrieval). Applied across every report, that supports four jobs:
Entity extraction. The device components, anatomy, procedures and outcomes mentioned in each narrative.
Severity classification. MAUDE’s event type field separates death, injury and malfunction. The narrative can separate an injury that needed emergency surgery from one that needed extra monitoring.
Failure mode grouping. Fifteen reports that describe the same failure in different words become one pattern.
Duplicate detection. A manufacturer and a user facility often report the same event, and MAUDE does not deduplicate them.
How signal detection uses the result
The simplest method counts. It flags a product code when this period’s reports exceed the historical average by a set margin. That catches obvious jumps and is easily fooled by market growth or a change in reporting habits.
Disproportionality methods are harder to fool. A proportional reporting ratio compares how often one device-event combination is reported against the background rate for similar devices. Bayesian methods such as the Multi-item Gamma Poisson Shrinker, which the FDA uses for pharmacovigilance, do the same with a correction for small numbers.
Severity classification adds a signal counting cannot see. If total reports hold steady while the share of high-severity events climbs, the count looks calm and the pattern is not.
Why severity needs more than MAUDE’s fields
“Serious injury” covers a wide range. A model reading the narrative can weigh how invasive the corrective intervention was, how long the harm lasted, whether the event was life-threatening and whether the patient recovered.
That turns a flat list of serious injuries into a graded dataset. Risk-benefit analysis, trend reporting and PSURs all depend on that grading.
Where the reviewer sits
A model’s output is a draft. EU MDR Article 83 puts the duty to gather and analyse post-market data on the manufacturer, and a classifier that is confidently wrong writes its mistake into the record unless someone checks it.
The sound design is a pipeline. The model ingests, extracts, classifies and flags trends. A qualified reviewer sees the source narrative beside the model’s summary, confirms or overrides it and decides on escalation, CAPA or reporting.
The same design produces the audit trail. Each summary is timestamped and versioned, and each one links to the decision a named person made about it.
How DeviceWatch applies this
DeviceWatch ingests new MAUDE reports from the openFDA API each week. It uses Claude to analyse each narrative into a structured summary with a severity classification, then puts the summary in a review queue where a regulatory professional validates and acknowledges it. The workflow keeps an audit trail.
The case for automating this work at all is in manual MAUDE surveillance doesn’t scale.