Greater New York Hospital Association (Kenneth E. Raske)

See attached file(s)

October
Nineteen
2026

Mr. Kyle Diamantas
Acting Commissioner of Food and Drugs
Office of the Commissioner
U.S. Food and Drug Administration
10903 New Hampshire Avenue
Silver Spring, MD 20993

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

Dear Commissioner Diamantas:
Greater New York Hospital Association (GNYHA) appreciates the opportunity to submit
comments in response to FDA’s Request for Feedback on “Considerations for the Regulation of
Generative AI-Enabled Medical Devices,” published in August 2026 (discussion paper).
GNYHA is a trade association that represents more than 200 hospitals and health systems in
New York, New Jersey, Connecticut, and Rhode Island. Since 2023, GNYHA has been in
discussion with experts from our member hospitals and health systems to examine the benefits
and challenges associated with AI use in health care.
Broadly speaking, our members are especially interested in receiving clearer guidance from FDA
on 1) FDA’s enforcement posture on GenAI software and enabled devices 2) the basis and
premarket evidence for FDA authorization, and 3) the role for hospitals in postmarket
monitoring.
We appreciate that the discussion paper raises many of the most salient questions facing policy
makers and stakeholders. We are providing the attached comments from the perspective of
GNYHA members as purchasers, operators, and users of AI-enabled medical devices, focusing
on the questions that are pertinent to these roles.
Thank you for considering our input.
Sincerely,
/s/
Kenneth E. Raske
President
 Detailed Comments
The discussion paper1 uses a risk-based, total product lifecycle approach and explores a
competency-based premarket framework composed of benchmarking followed by clinical
confirmation, with the rigor of confirmation scaled to risk. Like FDA’s January 2025 draft
guidance on lifecycle management and marketing submission recommendations for AI-enabled
device software functions,2 the discussion paper raises the need to account for real-world inputs,
relevant subgroups, workflow and user characteristics, limitations of validation data, and
performance criteria that reflect the device’s intended clinical use.
GNYHA supports this approach. It is consistent with other frameworks3 and generally consistent
with the approaches that many hospitals and health systems are proactively adopting for the
governance and oversight of AI- and GenAI-enabled devices.
The key issues we hear from our members concern 1) the need for greater transparency in
labeling and 2) the relative roles and responsibilities of manufacturers vis-à-vis operators/users in
postmarket activities. There are a host of open questions that FDA can help answer with further
guidance.
As a threshold matter, GNYHA urges FDA to try, to the greatest extent possible, to leverage
existing frameworks and processes rather than develop entirely new ones for AI- and GenAIenabled devices. We also strongly urge FDA to clarify its enforcement posture on GenAI clinical
decision support (CDS) tools. While FDA has set forth guidance on CDS,4 stating that where a
CDS tool generates a single recommendation and meets other non-device criteria as set forth
under Section 3060 of the 21st Century Cures Act, it may be subject to FDA enforcement
discretion.5 GNYHA urges FDA to formalize this position by committing that it will not exercise
enforcement authority over such tools. FDA should build upon this kind of balanced approach as
it develops its regulatory framework for GenAI technology in general, whose outputs may vary
depending on the clinical context and may range from informational support to specific
diagnostic or treatment recommendations.
Section IV—Considerations for the Assessment of Risk for GenAI-Enabled Devices
The following comments are intended to be generally applicable to Section IV of the discussion
paper and specifically responsive to portions of Questions 1, 2, 5, and 6.
The discussion paper’s two-axis framework offers a useful but incomplete way of evaluating the
risk of Gen-AI medical devices. GNYHA recommends that additional factors be incorporated
into the framework, including the following as examples: whether the device consists of multiple
U.S. Food and Drug Administration, Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback (August 2026), Docket No. FDA-2026-N-7874.
Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission
Recommendations, Draft Guidance for Industry and Food and Drug Administration Staff (January 7, 2025).
National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0),
NIST AI 100-1, (January 2023), https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf;
U.S. Food and Drug Administration, Clinical Decision Support Software: Guidance for Industry and Food and Drug
Administration Staff (January 2026), Docket No. FDA-2017-D-6569.
U.S. Food and Drug Administration, Clinical Decision Support Software: Guidance for Industry and Food and Drug
Administration Staff (January 2026), Docket No. FDA-2017-D-6569.
AI functions or agents; who the intended users are and the intended clinical environment;
characteristics of the output; level of autonomy; probability of harmful consequences; and
opportunities for human review and verification of outputs.
We also suggest FDA consider reframing the activity types within the two-axis framework.
Between “Informational: Action-Directing” and “Action-Taking: HCP-Supervised” there may be
increments of risk that are not reflected and do not necessarily increase in a stepwise fashion
consistent with those categories. For example, just because a device contemplates or requires
health care provider supervision does not mean it is inherently less risky than an activity that is
informational in nature. The totality of the device’s features, how they work together, and the
intended use should be considered in assessing the level and probability of risk.
With respect to “multi-turn” devices, we believe FDA must balance evaluation of the final
intended configuration with also determining whether interactions between functions change
clinical meaning or risk, as set forth in earlier guidance.6 Hospitals need clear information about
which functions have undergone FDA review, which have not, and how potential interactions
among functions have been evaluated for safety and effectiveness. This information is essential
to support informed procurement, implementation, clinical oversight, and ongoing monitoring.
Section V—A Competency-Based Approach for Premarket Evaluation of GenAI-Enabled
Devices
The following comments are intended to be generally applicable to Section V of the discussion
paper and specifically responsive to portions of Questions 7, 11, 12, 13, 15, and 16.
GNYHA supports the use of a competency-based approach with validated benchmarks and
clinical confirmation to assess GenAI-enabled medical devices before they are introduced more
broadly into clinical care. Benchmarks should be evaluated for representativeness, subgroup
performance, and overlap with training data that could inflate apparent performance. FDA should
be cautious about the use of synthetic data, which can increase the risk of bias.
We believe it is appropriate for sponsors to benchmark certain devices against real-world
outcomes not using GenAI. Much has been written about the relative merits of AI- versus human
intervention, and our members tend to think there is no single answer to whether GenAI models
or devices perform better than human judgment or human-operated devices. FDA should
therefore require, in appropriate cases, specific benchmarking against such comparators.
Finally, the prevalence of GenAI-enabled devices and their associated risks dictate that a better
resource be developed to help with information sharing and disclosure of premarket evidence.
For higher-risk generative AI-enabled medical devices, FDA should require a standardized,
publicly accessible summary describing how the device was evaluated, the evidence supporting
its authorization, key performance results, and any significant limitations. Hospitals need to
understand not only whether a device has received FDA authorization, but also the nature and
strength of the supporting evidence, the methods used to evaluate its performance, the clinical
settings and patient populations represented, and any remaining limitations or uncertainties. This
information should be reported consistently across applicable regulatory pathways, regardless of
See, Multiple Function Device Products: Policy and Considerations. FDA Guidance, July 2020; sections IV–VI.
https://www.fda.gov/media/112671/download.
whether the evaluation included a prospective clinical study and should be updated as clinically
meaningful postmarket performance data become available. This would enable hospitals to make
informed decisions about the continued use of these technologies, including the need for
additional guardrails in certain clinical scenarios. Existing mechanisms, such as summaries of
safety and effectiveness data and 510(k) summaries, could serve as models for this approach but
do not, in themselves, meet the need.
Independent third parties may have an important role to play in benchmarking, postmarket
monitoring, and clinical confirmation. Given their resource constraints, it is important that
hospitals and health systems have a reasonable number of reliable options for seeking such
information, including not-for-profit organizations.
It may be useful for FDA to consider, as a potential model for a competency-based approach for
premarket evaluation, the National Institutes of Health's scientific peer review process, in which
independent experts evaluate proposed study designs and methodologies. FDA could establish a
similar mechanism, drawing on clinical, methodological, patient safety, and technical expertise,
to assess benchmarking and other premarket evidence and, importantly, whether manufacturers'
proposed postmarket monitoring plans are sufficiently rigorous to identify clinically meaningful
performance limitations, emerging safety concerns, and changes in performance over time. Such
an approach could provide a pathway for evaluating certain GenAI-enabled devices while
maintaining appropriate manufacturer accountability and independent oversight.
Section VI—Postmarket Monitoring for GenAI-Enabled Devices
The following comments are intended to be generally applicable to Section VI of the discussion
paper and specifically responsive to portions of Questions 18, 19, and 21.
Before addressing postmarket monitoring, GNYHA would like to underscore the need for robust
labeling information on GenAI-enabled devices. Comprehensive labeling information could
support post-deployment monitoring and site-specific clinical validation. FDA needs to move
beyond traditional labeling with static disclosures about intended use, warnings, and limitations
to more detailed, actionable information that hospitals and clinicians need to understand how a
GenAI device performs in particular clinical settings, when it can be relied upon, and where
human oversight is necessary. Specific types of information include the identity and version of
the underlying model; which inputs or features are most influential in producing outputs;
intended and expected performance thresholds; applicable restrictions; and—perhaps most
importantly for certain devices—instructions needed for ongoing performance monitoring and
management.
Hospitals also need to know which functions and combinations FDA has reviewed, the
conditions of that review, and how updates to a model, prompt, interface, or connected system
may affect the authorization and hospital implementation responsibilities. In addition to being
important to hospitals and their clinician users, such labeling information could promote
continued innovation and advancement of these technologies in the health care space.
With respect to postmarket monitoring, GNYHA encourages FDA to consider a risk-based
approach that places greater emphasis on structured postmarket evaluation of GenAI-enabled
medical devices. For certain higher-risk devices, FDA should require manufacturers to submit a
prospective postmarket evaluation plan for review before deployment, with the scope and rigor
of monitoring proportionate to the device's potential clinical risks. Higher-risk applications, more
autonomous technologies, and devices whose errors may be difficult for clinicians to identify
should be subject to more rigorous clinical confirmation. Manufacturers will have to engage with
a variety of customers (e.g., major academic medical centers and community hospitals) to ensure
the collection of adequate information about how the device performs in different settings and
with different populations. Many hospitals have governance structures in place to support
oversight and monitor performance, and these structures can be leveraged to support postmarket
monitoring given sufficient guidance and support from manufacturers.
Importantly, manufacturers should retain primary responsibility for ensuring these evaluations
are performed and reporting associated findings to FDA. Hospitals should not be expected to
independently design and conduct extensive postmarket studies to compensate for gaps in
manufacturer evaluation. A recent article published in NEJM AI highlights the potential
consequences of easing regulations governing AI-enabled health products, characterized as
“signaling a potential strategic shift from a premarket approval framework toward postmarket
surveillance for clinical AI technologies.”7 Although that approach could reduce regulatory
burdens for sponsors and speed products to market, the authors caution that it “risks transferring
the regulatory burden traditionally assumed by the FDA onto health systems, which would now
serve as the primary entities responsible for postmarket surveillance.”8
Hospitals are devoting increasing operational and financial resources to adopting and overseeing
emerging technologies. Nevertheless, the Centers for Medicare & Medicaid Services (CMS)
recently finalized the repeal of the alternative New Technology Add-On Payment (NTAP)
pathway, including for FDA-designated Breakthrough Devices, beginning with applications for
fiscal year 2028, subject to transitional provisions for certain devices.9 GNYHA opposes such a
policy, which may create additional financial barriers to hospital adoption of innovative
technologies and delay patient access to these products. We mention the CMS proposal here
because it underscores the need for a coordinated Federal approach that accounts for the
cumulative operational and financial implications of policies affecting the adoption and oversight
of emerging technologies, including GenAI-enabled medical devices.
We also must note that the very significant question of liability for AI and GenAI-related health
care is unsettled. Clinicians, health care institutions, and others have an important role to play in
postmarket monitoring, but they should not be expected to compensate for reduced manufacturer
premarket evidence through resource-intensive local monitoring. While it is possible to envision
scenarios where a GenAI-enabled device holds such promise that some will argue it is
inappropriate to delay entry into the market pending collection of premarket evidence, it is
crucial that manufacturers’ obligation to demonstrate baseline effectiveness and safety not be
outsourced to hospitals and clinicians.
Manufacturers should be required to provide the necessary tools, instructions, technical support,
and escalation procedures to carry out any site-specific checks or monitoring required for safe
use. Monitoring plans should define actionable performance thresholds, alert recipients, response
timelines, and interim safeguards, with hospital tasks proportionate to risk and operationally
Wong A, Adler-Milstein J, Price WN II, Singh K. How repeal of the NTAP alternative pathway will impact U.S.
clinical AI innovation. NEJM AI. 2026;3(10). doi:10.1056/AIpc2600711
Id.
91 Fed. Reg. 49679–49680 (August 4, 2026)
feasible. Hospitals should receive clear notice of updates, whether the changes fall within the
existing authorization or an authorized predetermined change control plan, and whether local
implementation checks must be repeated. Manufacturers should provide their customers with
ongoing performance information, investigate safety signals, notify hospitals of material changes
or emerging risks, and take appropriate corrective action when performance concerns are
identified.
To this end, FDA should also consider standardized reporting requirements to ensure that
hospitals receive timely, actionable information about device performance. There should be a
more general and standardized system for performance data collection for GenAI-enabled
devices so that information is not siloed or only in the hands of manufacturers. Existing adverseevent reporting systems10 were not designed for the routine exchange of performance
information or for complex devices such as those that are GenAI-enabled, whose performance
may vary across clinical settings, workflows, and patient populations. The current disclosure
system is too fragmented for hospitals to develop a clear view of product performance.11 We
believe it is crucial for FDA to address this gap in order to promote the safety and effectiveness
of GenAI-enabled medical devices.
Conclusion
GNYHA thanks FDA for soliciting comments on this important topic. We urge FDA to consider
the impacts on and role of hospitals, health systems, and their clinicians in developing guidance
on GenAI-enabled medical devices.
Examples include FDA’s Medical Device Reporting (MDR) framework under 21 C.F.R. Part 803, which requires
specified reports of device-related deaths, serious injuries, and certain malfunctions, and MedWatch, through which
health care professionals may voluntarily report adverse events and product problems. These general device-safety
reporting mechanisms can receive reports involving AI-enabled devices, but are not AI- or GenAI-specific channels
for routine exchange of local results. See U.S. Food and Drug Administration, Medical Device Reporting (MDR):
How to Report Medical Device Problems (March 27, 2025).
Certain systems, e.g., TPLC, MAUDE, MedSun, contain various types of information but there is no comprehensive
source for device performance.