Pouyan Golshani

Docket FDA-2026-N-7874. Comment from Pouyan Golshani, MD, practicing interventional and diagnostic radiologist, submitted in a personal capacity. Disclosure: partner in a California medical group; founder of GigHz, which develops AI-assisted clinical decision-support and radiology reporting software.

Full comment attached. In brief, I support FDA’s stated goal of a nimble, least burdensome approach. The right benchmark for a generative AI tool is the care patients actually receive without it, which overnight and in understaffed settings is often delay, not perfection.

Question 15: Allow comparison against care without the device, including delayed specialist review.

Questions 11 and 18: Do not require a prospective study by default. For lower-risk and clinician-reviewed functions, accept more premarket uncertainty and rely on practical postmarket monitoring.

Question 1: Give real credit to downstream safeguards, especially clinician review, and do not add dimensions that raise the bar for every device.

Questions 2 and 3: Keep informational and patient-facing functions light-touch, with simple built-in failsafes against confabulation (source one click away, statements linked to their source, critical findings and uncertainty preserved); do not underestimate patients.

Question 16: Keep third-party benchmarking voluntary so it does not become a costly gatekeeper that favors large companies.

Questions 22 and 24: Let routine changes, including foundation model updates, be managed in the manufacturer’s quality system rather than new submissions.

Background: in our peer-reviewed analysis of 1,430 FDA AI device authorizations (Cureus 2026;18(7):e112583, https://doi.org/10.7759/cureus.112583), annual authorizations rose from 1.8 to 264 per year, and 76.5% were reviewed by the Radiology panel.

Division of Dockets Management, Food and Drug Administration

Docket No. FDA-2026-N-7874
Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback
Submitted by Pouyan Golshani, MD, interventional radiologist
Disclosure: I submit this comment in a personal capacity. It does not represent my medical group or any other
organization. I am a partner in a California medical group and the founder of GigHz, which develops AI-assisted clinical
decision-support and radiology reporting software. Those products could be affected by the approaches discussed.

Summary
I am a practicing, board-certified interventional and diagnostic radiologist and lead author of a peerreviewed analysis of all 1,430 FDA authorizations of AI/ML-enabled medical devices from 1995 through
2025. I support the paper’s stated goal of “a nimble regulatory approach that employs least burdensome
principles.” My central point is that the right benchmark for a generative AI tool is the care patients
actually receive without it, not perfection. Overnight, in rural areas, and in understaffed settings, the
realistic alternative is often delay. Evidence requirements set too high would keep useful tools away
from the patients who would benefit most. I respond to the questions where clinical experience is most
relevant:

• Question 15: compare devices against care without the device, including delayed specialist review.
• Questions 11 and 18: do not require a prospective study by default; accept more premarket
uncertainty for lower-risk functions and rely on postmarket monitoring.
• Question 1: give real credit to downstream safeguards, especially clinician review, when assessing
risk.
• Questions 2 and 3: keep informational functions light-touch, with simple built-in failsafes against
confabulation, and do not underestimate patients.
• Question 16: keep any third-party role voluntary so it does not become a new gatekeeper.
• Questions 22 and 24: let routine changes, including foundation model updates, be handled in the
manufacturer’s quality system.

Background
In our analysis of the FDA’s public list of AI/ML-enabled devices (Golshani P, Joseph MS. Cureus.
2026;18(7):e112583; https://doi.org/10.7759/cureus.112583), annual authorizations rose from an
average of 1.8 per year (1995–2014) to 264 per year (2023–2025). Devices reviewed by the Radiology
panel accounted for 1,094 of 1,430 authorizations (76.5%). Pathology had 9, Obstetrics and Gynecology
4, and no authorizations were recorded under a psychiatry or behavioral health panel. These figures
count authorization records by review panel, not clinical use. They show that FDA already has a large
body of experience with AI in imaging to draw on, and that many clinical areas have little or none. A
framework that is workable for smaller developers matters if those areas are to see useful tools.

Question 15: Compare against care without the device
Yes. For many functions the most meaningful comparator is what happens without the device: unaided
judgment, delayed specialist review, or no review at all. In imaging, many studies are read after hours,
and a clinician may act before a specialist sees the case. A tool that is imperfect but available
immediately can be better for the patient than a specialist read that arrives hours later. Holding devices
only to an ideal specialist standard would reject tools that improve on what patients actually receive
today. Sponsors should be able to justify this kind of comparator.

Questions 11 and 18: Proportionate premarket evidence
• A prospective clinical study should not be the default. Retrospective review of real cases, reader
studies, and simulated-use testing are adequate for many functions, especially those used with
clinician review.
• For lower-risk and clinician-reviewed functions, FDA should accept more premarket uncertainty and
rely on postmarket monitoring. A practical program is enough: logging, periodic sample review, and
a clear process for acting on problems. This gets tools to patients sooner and produces better realworld evidence than a longer premarket study would.

Question 1: Credit for downstream safeguards
The two-axis framework is reasonable. Of the additional dimensions listed, the availability of
downstream safeguards matters most. When a qualified clinician reviews an output before it affects
care, and the output links to its source so it can be checked quickly, the assessed risk should be lower,
with correspondingly lower evidence requirements. I would not add further dimensions that raise the
bar for every device.

Questions 2 and 3: Informational functions and patients
Functions that summarize, explain, or organize information without directing a specific action should
stay light-touch. Patient-facing tools that explain results in plain language can improve understanding
and engagement. Simple features are usually sufficient: a clear statement of what the tool does and
does not do, and a prompt to discuss results with a clinician. FDA should not assume patients cannot
handle clinical information.
Light-touch does not mean no safeguards. Summarizing tools can confabulate, adding a finding that is
not in the source or dropping one that is. A few simple, built-in failsafes address most of this risk without
slowing development: keep the original source (for example, the full radiology report) one click away
from the summary; link statements in the summary to where they came from; and preserve critical
findings and stated uncertainty rather than smoothing them away. These are design features a
manufacturer can show in its submission, not a reason for heavier review.

Question 16: Third parties
Independent benchmarking can be useful, but it should be one option among several, not a
requirement. A mandatory third-party step would add cost and delay, favor large companies that can
absorb it, and risk turning a small number of testing organizations into gatekeepers. Sponsor-developed
testing, reviewed by FDA, should remain acceptable.
Questions 22 and 24: Changes, including foundation model updates
• Most updates, such as prompt changes, retrieval updates, interface changes, and many model
version changes, can be managed within the manufacturer’s quality management system, with
records available to FDA, rather than through new submissions.
• For third-party foundation models, practical tools already exist: record the model version used, rerun the manufacturer’s own test set when the model changes, and use contractual notice or version
pinning where available. These are reasonable engineering practices to describe in a PCCP; they
should not trigger new premarket review.

Closing
Generative AI can extend specialist-level support to places and hours where it is not available today. I
encourage FDA to keep the framework proportionate, measure devices against the care patients
actually receive, and keep the path to market practical for smaller developers. I would be glad to provide
additional detail from clinical practice.
Pouyan Golshani, MD
pouyan.golshani@gighz.com