Postmarket Monitoring for Generative AI Medical Devices

Postmarket monitoring for generative AI medical devices should test whether the device continues to perform as intended after deployment.

That means tracking more than uptime and aggregate accuracy.

Teams need:

• to watch generated output
• human corrections
• clinical escalation
• subgroup performance
• connected-system failures
• changes to the model
• workflow

FDA's August 2026 discussion paper identifies postmarket monitoring as a central topic for generative AI-enabled devices. The paper is not guidance, but the operational lesson is already clear: a variable, updateable system cannot be governed through a one-time premarket snapshot alone.

Define the Monitoring Question Before Launch

Each measure should answer a decision question. Is the device producing unsupported content? Are clinicians correcting the same type of output repeatedly? Did an update change behavior? Is one population experiencing more errors? Is the tool creating new delay or review burden?

A dashboard full of numbers is not monitoring unless someone knows what action each number can trigger.

Monitor Generated Output

Useful output measures may include:

• unsupported statements
• missing clinical context
• incorrect citations
• omitted warnings
• inconsistent recommendations
• responses outside intended use

The exact measures depend on the device function.

Sampling should include ordinary cases, edge cases, high-risk cases, and cases where users disagreed with the system. Teams should preserve enough context to understand why the output failed without exposing more patient information than necessary.

Use Human Corrections as Safety Data

Corrections, overrides, rejected suggestions, and escalation requests are not noise. They are some of the best signals that the device or workflow needs review.

Monitoring should distinguish between harmless style edits and corrections that change clinical meaning. A rising correction rate may reflect model drift, a new user group, a workflow change, or a prompt update. The response depends on the cause.

Track Model and Dependency Changes

Generative AI devices may depend on foundation models, retrieval systems, terminology services, or external tools. Version control should cover those dependencies, not only the application release number visible to users.

Every meaningful change should have a record:

• what changed
• why it changed
• what testing was completed
• which users were notified
• whether local revalidation was required

Monitor Agentic Actions Separately

If the system can call tools or initiate actions, monitor the action sequence as well as the generated explanation.

Track:

• failed tool calls
• unauthorized attempts
• incomplete actions
• incorrect routing
• duplicate actions
• cases where a human stopped the sequence

Action logs should make it possible to reconstruct what the system did and which control allowed or blocked each step.

Subgroup and Setting Performance

Aggregate performance can hide differences by patient population, language, care setting, device type, or clinical service. Monitoring should reflect the groups and environments named in the intended use and local risk assessment.

Small samples need careful interpretation. A signal may justify review without proving a performance difference. The point is to detect credible concerns early enough to investigate them.

Operational Measures Matter

• response latency and failed-run rate
• user adoption and abandonment
• correction and override burden
• escalation volume and response time
• downtime and recovery behavior
• changes in documentation or ordering workflow
• support tickets and clinician complaints

Set Review and Stop Criteria

Monitoring needs thresholds and owners.

Some signals can trigger routine review.

Others may require:

• a limited pause
• rollback
• user warning
• vendor escalation
• full removal from use

Stop criteria should be defined before the device is under pressure. It is much harder to decide what counts as unacceptable after a safety event or widespread workflow failure.

A Practical Monitoring Record

• device, model, and dependency versions
• intended use and user groups
• clinical and operational measures
• sampling method and review cadence
• named owners for each signal
• thresholds, escalation, and stop rules
• change history and revalidation decisions
• incidents, corrective actions, and closure records

Related AI Medicine Now Coverage

• Regulated Generative AI Medical Devices
• Generative AI Medical Device Risk Assessment
• Monitoring and Observability
• Monitoring Clinical AI After Deployment
• Clinical AI Inventory

Reviewed: September 2, 2026. Next review: October 20, 2026.