AI Healthcare Coalition (Cybil Roehrenbeck)

See attached file(s)

October 19, 2026

Kyle Diamantas
Acting Commissioner
U.S. Food & Drug Administration
10903 New Hampshire Avenue
Silver Spring, MD 20993

Re: Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback [Docket No. FDA-2026-N-7874]

Dear Acting Commissioner Diamantas,

On behalf of the AI Healthcare Coalition (AIHC), thank you for the opportunity to comment on the
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback (“GenAI Discussion Paper”). We appreciate the efforts of the U.S. Food & Drug
Administration (FDA) to engage stakeholders on the opportunities and challenges associated with
Generative AI (GenAI) enabled medical devices and ways to advance regulatory approaches for
these devices.

AIHC was founded in 2019 and convenes healthcare AI innovators to advocate for patient access to
safe, ethically developed AI medical services. Our members have created, or are in the process of
developing, AI medical services (or devices); some of which require FDA market authorization and
oversight, and our founding membership includes the first-ever autonomous AI innovator to obtain
FDA authorization.

We commend FDA for proactively engaging stakeholders as GenAI technologies continue to evolve
and become increasingly integrated into healthcare delivery. AIHC believes that FDA can effectively
regulate GenAI-enabled medical devices by building upon its existing frameworks, focusing on
intended use and clinical impact, maintaining clear regulatory pathways, and coordinating with
other relevant federal agencies.

AIHC Recommendations

 The GenAI Discussion Paper’s proposed risk axis will not enable the FDA to accurately
assess risk and could result in the misclassification of AI technologies.

AIHC is concerned that the proposed risk axis may place disproportionate emphasis on technical
characteristics of a model rather than on the clinical context in which the technology is deployed.
The same underlying GenAI technology may present dramatically different levels of patient risk
depending upon its intended use, user population, degree of clinical oversight, and role in medical
decision-making.

For example, a GenAI function that provides a clinician with a referenced summary of treatment
options presents fundamentally different risks than one that autonomously initiates a treatment
order, even if both rely on similar underlying foundation models.

If FDA retains the two-axis heuristic, AIHC recommends adding traceability and transparency of the
output as an explicit mitigating dimension. A function that shows the user the basis for its output, for
example by citing the source records it relied on, marking the region of an image that drove a
measurement, or disclosing what information was unavailable, allows the user to independently
evaluate that output. That ability should move a function lower on the consequences axis.

This is consistent with the independent-review criterion in section 520(o)(1)(E) of the Federal Food,
Drug & Cosmetic Act and with the GenAI Discussion Paper’s own observation that measurement
functions are higher risk when the user cannot evaluate the basis for the output. The converse should
also hold: where visual verification or source traceability is available, measurement and signal
processing functions should not be presumed higher risk.

AIHC also cautions against treating use by generalist clinicians as a risk-elevating factor. Extending
specialist-level information to emergency, primary care, and other non-specialist clinicians is a
principal benefit of AI-enabled devices. Any associated risk should be addressed through labeling,
human factors, and output transparency rather than through a higher presumptive risk tier.

Sponsors also need predictable criteria for where an informational function becomes actiondirecting. AIHC agrees that disclaimers alone should not decide the question and asks FDA to
publish concrete, example-based criteria, to characterize intended use through labeling and design
controls rather than emergent conversational trajectories, and to confirm that the Discussion Paper
does not alter FDA’s interpretation of the clinical decision support software exclusion under section
520(o)(1)(E).

Additionally, an overly rigid classification system may quickly become outdated as the technology
evolves. Given the rapid pace of GenAI development, FDA should avoid establishing risk categories
that could inadvertently lock the Agency into assumptions that are no longer valid within a short
period of time. Instead, FDA should preserve flexibility by evaluating products using existing deviceand use-case-specific risk principles that can adapt alongside technological advances.
  The FDA’s current regulatory framework is sufficiently flexible to allow appropriate
review and authorize GenAI devices. To the extent that the FDA believes changes are
required, the FDA should modify its existing methodologies rather than creating new
systems specific to GenAI.

GenAI innovators who seek FDA authorization require a predictable and certain pathway for review
and approval. Without such assurance, the essential capital and resources for GenAI innovators to
conduct safety studies and collect real-world evidence may be lacking, creating an insurmountable
hurdle for GenAI innovation.

Further, the FDA's existing methodologies for review, which include device-specific and medical
specialty-specific analyses, are well-suited to evaluating the clinical risks and benefits associated
with AI-enabled technologies. The relevant regulatory question should remain focused on a device's
intended use, clinical performance, risk profile, and impact on patient care, rather than on whether
a particular model architecture includes generative capabilities.

While GenAI introduces novel technical considerations, these considerations generally represent
extensions of issues that FDA already evaluates for software and AI-enabled devices, including
transparency, validation, human factors, cybersecurity, performance monitoring, and post-market
oversight. FDA should therefore build upon its established risk-based framework and existing
software and AI policies, leveraging targeted guidance where necessary rather than creating a
separate regulatory paradigm for GenAI-enabled devices.

 The FDA's competency-based approach should be applied flexibly and consistent with
least burdensome principles.

AIHC supports FDA's recognition that clinical confirmation need not require a prospective clinical
study in every case, and urges FDA to accept retrospective evaluation, shadow deployment, realworld data, OUS data, and synthetic data where scientifically justified. Sponsors should also be able
to rely on sponsor-developed benchmarks tailored to their device's intended use.

FDA should avoid requirements that add cost without commensurate safety benefit, such as
mandating adjudicators independent of both the sponsor and the foundation model developer, or
benchmarking across clinical domains outside the device's intended use.

Where a device addresses unmet need, performance should be assessed against the care likely to
occur in the device's absence (e.g., unaided clinical judgment or delayed specialist review), rather
than presumptively against a specialist panel.

The benchmarking principles FDA describes, including prespecified methods and justified rubrics
and acceptance criteria, align with existing design verification practice under 21 CFR Part 820 and
IEC 62304. They should be implemented through those requirements rather than as a GenAI-specific
evidentiary structure. Retrieval, prompt, and guardrail components should likewise be verified and
change controlled as software components under existing frameworks, and transparency of the
sources a device relied on should count as evidence toward the safety elements FDA describes.

For care escalation functions, AIHC asks FDA to recognize that under-escalation and overescalation are not symmetric harms. A missed stroke or pulmonary embolism is not
commensurable with an unnecessary consult. Sponsors should be able to justify error weighting by
clinical context rather than against a uniform standard.

AIHC supports the use of synthetic data and notes the Discussion Paper’s caution that synthetic
inputs generated by models of the same class may reproduce the gaps under evaluation. Sponsors
should be able to address this through documented safeguards, such as clinician-authored seed
cases, held-out real-world data, or generation by a different model family, rather than through
exclusion of synthetic data.

 Third party evaluators are likely to add complexity and burden to GenAI innovation and
may result in varied safety and quality assessment.

AIHC appreciates FDA's interest in exploring mechanisms to support efficient review of GenAIenabled devices. However, introducing third-party evaluators could create additional complexity,
increase development costs, and generate uncertainty regarding regulatory expectations.

Medical device sponsors ultimately remain responsible for the safety, effectiveness, and quality of
their products. Introducing additional entities into the review process may lead to inconsistent
interpretations of FDA standards, variation in assessment methodologies, and potentially
conflicting recommendations. Such outcomes could undermine predictability for developers and
delay patient access to beneficial technologies.

Moreover, because GenAI-enabled devices often require highly specialized clinical, technical, and
regulatory expertise, ensuring consistency across multiple third-party organizations may prove
challenging. FDA's direct oversight and review processes remain the most reliable mechanism for
maintaining uniform standards and accountability.

FDA's experience with existing third-party programs shows the limits of this approach. Both the
510(k) Third Party Review Program and the Accreditation Scheme for Conformity Assessment (ASCA)
program were intended to improve efficiency and reduce sponsor burden. In practice, they have not
consistently done so.

To the extent FDA identifies resource constraints or specialized technical review needs, AIHC
encourages the Agency to continue building internal expertise and utilizing existing advisory
mechanisms rather than establishing a new layer of third-party review requirements.
  The FDA’s existing postmarket review processes are sufficient for ongoing safety
monitoring of GenAI; GenAI does not require an alternate postmarket review process.

FDA's existing postmarket authorities provide a robust framework for monitoring the performance
and safety of GenAI-enabled medical devices once they enter the market. Through established
mechanisms, including Medical Device Reporting (MDR), complaint handling requirements,
corrections and removals reporting, quality system obligations, and postmarket surveillance
authorities where appropriate, FDA already possesses the tools necessary to identify, assess, and
address emerging safety concerns.

Importantly, postmarket oversight should remain focused on a device's real-world performance and
clinical impact rather than on the technological architecture used to develop the product. GenAIenabled devices are not unique in their need for ongoing monitoring; software-based medical
devices of all types may experience performance changes resulting from evolving clinical
environments, user interactions, data inputs, or cybersecurity threats. FDA's existing risk-based
approach is designed to address these circumstances and can be applied equally to GenAI-enabled
products.

Creating a separate postmarket review framework specifically for GenAI could result in duplicative
reporting obligations, increased compliance burdens, and regulatory uncertainty without producing
commensurate patient safety benefits. Such requirements may also divert resources away from
meaningful safety monitoring and product improvement efforts. FDA should instead leverage
existing postmarket processes and, where needed, issue targeted guidance regarding the types of
performance metrics, documentation, or monitoring activities that may be particularly relevant for
GenAI-enabled devices.

To the extent FDA identifies unique considerations associated with adaptive or continuously learning
systems, those concerns can be addressed through existing lifecycle management approaches,
including predetermined change control plans, software update procedures, and risk management
processes. These mechanisms allow FDA and innovators to proactively manage foreseeable
modifications while maintaining appropriate oversight and accountability. By relying on its
established postmarket authorities and continuing to apply a technology-neutral, risk-based
approach, FDA can effectively monitor the safety and effectiveness of GenAI-enabled medical
devices without creating a separate regulatory structure that could impede innovation and patient
access to beneficial technologies.

As FDA considers accepting greater premarket uncertainty in exchange for postmarket monitoring,
such arrangements should remain a sponsor-elected option rather than a default expectation. To be
workable, any such commitment should:
o be built on existing mechanisms, including predetermined change control plans, quality
system requirements, MDR, and section 522 authority, with no parallel GenAI reporting
structure;
 o be prespecified by the sponsor and bounded in metrics, cadence, duration, and exit criteria,
proportionate to the device’s risk;
o recognize manufacturer-conducted monitoring without mandating independent
adjudicators or third parties;
o keep accountability with the manufacturer and not transfer monitoring obligations to
deploying institutions, so that health systems are not deterred from adopting these
devices—institutional participation in monitoring, such as data sharing and clinician
feedback, should remain voluntary and contractual; and
o allow changes to third-party foundation models to be triaged through the sponsor’s quality
system, with re-benchmarking reserved for changes that affect safety-critical behavior.
Sponsor-set performance thresholds should trigger investigation under the quality system
rather than automatic regulatory consequences, and authorizations supported by such
commitments should not be labeled as provisional.

 The FDA should not venture into licensure, credentialing, or traditionally stateadministered systems for patient safety oversight.

AIHC respectfully urges FDA to remain focused on its core statutory mission: evaluating the safety
and effectiveness of medical devices. Professional licensure, provider credentialing, scope of
practice determinations, and healthcare facility oversight have historically been administered by
states, professional boards, and healthcare organizations.

Attempting to incorporate such considerations into FDA's regulation of GenAI-enabled devices
could create duplicative oversight, regulatory confusion, and potential conflicts with longstanding
state authorities. Questions regarding which healthcare professionals may utilize a technology, the
circumstances under which it may be deployed, and how providers are credentialed to use advanced
technologies are separate from FDA's determination of whether a device is safe and effective for its
intended use.

FDA can continue to promote patient safety through appropriate device labeling, indications for use,
warnings, training requirements, and post-market surveillance without expanding into areas that
have traditionally fallen outside the Agency's jurisdiction.

 The FDA and the Centers for Medicare & Medicaid Services (CMS) should coordinate and
align approaches in this emerging field.

As GenAI-enabled technologies become increasingly integrated into healthcare delivery, alignment
between FDA and CMS will become critical to ensuring patient access and encouraging innovation.
Patients derive little benefit from a technology that receives FDA authorization but faces significant
barriers to coverage, reimbursement, or clinical adoption.

Developers frequently must navigate separate regulatory, payment, and coverage frameworks that
are not always aligned in terms of evidence expectations or evaluation criteria. Greater coordination
between FDA and CMS could reduce unnecessary duplication, improve transparency regarding
evidentiary requirements, and facilitate more efficient development pathways.

We commend the agencies’ recent efforts through the FDA Technology-Enabled Meaningful Patient
Outcomes (TEMPO) for Digital Health Devices Pilot and CMS Advancing Chronic Care with Effective,
Scalable Solutions (ACCESS) programs to align approaches on these issues. Such coordination
helps ensure that innovative technologies can move efficiently from authorization to meaningful
patient access while preserving the high standards necessary to protect beneficiaries and promote
quality care.

Conclusion

Thank you for the opportunity to provide AIHC's feedback on the GenAI Discussion Paper. We
appreciate FDA's commitment to ensuring safe and effective access to AI medical services and look
forward to continued collaboration on these important issues. If you have any questions about our
comments, please contact me at cybil.roehrenbeck@ai-coalition.org.

Sincerely,

Cybil Roehrenbeck
Executive Director, AI Healthcare Coalition