Dave Kunstman

See attached file(s)

PUBLIC COMMENT TO THE U.S. FOOD
AND DRUG ADMINISTRATION
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback
Docket: FDA-2026-N-7874
Submitted by: Dave Kunstman, MD
Professional background: Family Physician and Associate Chief Medical Information Officer
Organization: Essentia Health
Capacity: Individual comment. The views expressed are my own and do not necessarily represent
those of Essentia Health.
Questions addressed: 1, 11, 18-21, and 26

To the Food and Drug Administration:
Thank you for the opportunity to comment. I am a family physician and Associate Chief Medical
Information Officer at Essentia Health. My work includes evaluating, implementing, governing, and
monitoring AI used in clinical care.

These comments are my own. They are based largely on the practical challenges we have seen as AI
moves from evaluation into routine clinical care.

I support FDA’s overall direction. I would also encourage alignment with principles developed by the
Coalition for Health AI (CHAI), where they fit. A common approach to evaluating and monitoring AI
would be helpful to both developers and healthcare organizations.

Question 1: Does the two-axis framework appropriately capture risk?
Yes, as a starting point.
Activity and consequence are useful primary dimensions. I would keep the framework simple, but
also consider whether an error is likely to be caught before it causes harm.
A wrong recommendation that is obvious to a clinician is different from a wrong answer that looks
reasonable and quietly moves through the workflow.
Human oversight only reduces risk if the human can realistically recognize the problem and has an
opportunity to intervene.
Question 11: How should clinical confirmation be approached?
I support testing in real or clinically representative settings.
The workflow matters as much as the model. Did the system get the right information? Did the right
person see the output? Could that person recognize a bad recommendation? What happened next?
For tools designed to support clinicians, the most useful comparison may often be clinician plus AI
versus usual care, rather than AI versus clinician. That better reflects how these tools will actually be
used.
Questions 18-21: Postmarket monitoring
This may be the hardest part to get right.
Re-benchmarking and monitoring for model degradation are useful, but they may not tell us how a
tool is actually performing in clinical care.
The useful signals may be much simpler: clinician corrections, overrides, failed actions, inappropriate
escalation, safety reports, or other unexpected behavior. The challenge is capturing those signals
reliably without creating another large burden for clinicians and health systems.
Manufacturers and healthcare organizations will both have a role. FDA could help by being clear
about what each should monitor and how important safety signals should be shared. Manufacturer
responsibility for product safety should remain clear.
I would also be cautious about using AI to monitor AI. It may help, but the monitoring system needs
to be evaluated too.
Question 26: Agentic AI
Agentic AI deserves additional attention because it can take actions, not simply make
recommendations.
For these systems, I would evaluate the path as well as the result.
Did it act on the right patient? Did it use the right information and tools? Did it stop when something
went wrong? Did it involve a person when it was supposed to?
A correct final result does not make an unsafe series of actions acceptable.
As these systems become more autonomous, we need to be able to see what they did and intervene
when needed.

Closing
I support FDA’s move toward risk-based evaluation and stronger postmarket monitoring.
My main concern is making sure the expectations are practical. Healthcare organizations are still
learning how to monitor these systems well. We should not assume that because something can be
measured in a controlled evaluation, it can be routinely monitored in clinical practice.

CHAI principles can provide useful structure where applicable. FDA can also help by setting clear
expectations for what manufacturers and healthcare organizations should each be able to see and
measure after deployment.

That will become even more important as AI moves from giving advice to taking action.

Respectfully submitted,
Dave Kunstman, MD
Family Physician
Associate Chief Medical Information Officer
Essentia Health