Duke-Margolis Institute for Health Policy (Christina Silcox)

See attached file

October 19, 2026

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

To Whom It May Concern:

Established in January 2016, The Duke-Robert J. Margolis, MD Institute for Health Policy (“DukeMargolis”) is both an academic research center and a policy laboratory where stakeholders can come
together to analyze, propose, and evaluate ways to improve health in the United States and beyond. The
Institute’s mission is to improve health and health care value through practical, innovative, and
evidence-based policy solutions. By catalyzing Duke University’s leading capabilities, we conduct
research and convene activities focused on biomedical innovation and regulatory policy.
We appreciate the opportunity to respond to the U.S. Food and Drug Administration’s (FDA) carefully
considered discussion paper on generative artificial intelligence (GenAI)-enabled medical devices. We
believe the approaches described provide a reasonable starting point for advancing the safe and
effective use of these technologies. These efforts should proceed alongside regulatory science research
to determine how well the proposed methods predict clinical performance and identify emerging risks.
Pilots and other collaborative learning approaches can support implementation while generating
evidence about where methods require refinement or replacement. Given the pace of technological
change, FDA should establish explicit opportunities to reassess these practices and share findings
transparently, allowing the framework to evolve as understanding of GenAI-enabled devices and their
use improves. At the same time, generalizable, scalable, and reusable postmarket infrastructure could
help make monitoring for new risks or performance changes more efficient for these tools, whose
capabilities will continue to evolve quickly.

Overall Comments
Any framework for GenAI-enabled devices should account for the breadth of GenAI-enabled devices,
from single-function and bounded multifunctional tools to more open-ended agentic systems operating
across specialties (although within guardrails). More traditional task-specific assessment may be
appropriate for tools with bounded functions. Devices may also incorporate very different underlying
models, ranging from general-purpose large language models (LLMs) to foundation models of
physiological information, including radiological images and electrocardiographic signals. Tools may use
LLMs to fully execute functions or they may use LLMs to structure inputs but use rules or machine
learning algorithms to produce the clinical output. Assessment methods should therefore be selected
and tested with this diversity of functions and underlying models in mind. We encourage FDA to explore
multiple assessment approaches and share findings transparently so that regulatory practices can evolve
with the evidence rather than become entrenched before their utility is established.

Response to Questions 1–5, and 26
Risk Assessment and the Role of Users
The two-axis framework provides a useful initial organizing tool, but FDA should examine how the
additional risks identified in the discussion paper are represented within it. For example, the assessed
ability of the typical user to correctly recognize errors to a GenAI-enabled device and appropriately
respond should be considered as an aspect of the “consequences of the decision.” R

230 Science Drive 1201 Pennsylvania Ave, NW www.healthpolicy.duke.edu
Durham, NC 27708 Suite 500, Washington, DC 20004
Regarding this, we encourage FDA to consider users’ ability to recognize and appropriately respond to
incorrect outputs as more of a continuum, rather than relying on categories such as patient, generalist
health care professional, or specialist health care professional.

We also recommend that FDA prioritize regulatory science research on how users respond to device
outputs over time, including how differences in the directiveness of outputs may change behavior. The
research should support periodic reassessment of these risks as AI technologies become more widely
used.

Response to Questions 7, 9–11, 16—17
Evaluation Methods and Generalizability

We support exploring benchmarking followed by clinical confirmation as an evaluation approach.
Benchmarking offers the opportunity to examine a wider breadth of tasks and scenarios than may be
practicable through traditional clinical trials. However, research on these methods is still developing,
and the characteristics of an effective benchmark are not yet sufficiently understood.1 In particular,
performance on tested tasks does not establish how results generalize to untested tasks. Closed, third
party benchmarks can help reduce the risk of manufacturers “training to the test” but there are multiple
aspects of a quality benchmark test, and development by a third party does not ensure it is better than a
sponsor-developed benchmark test.

We therefore encourage FDA to establish monitoring tools and data infrastructure that can help
determine when benchmark results predict clinically meaningful performance across tasks, populations,
and settings. Such infrastructure would support systematic learning about the applicability and
limitations of different benchmarking approaches. Clinical confirmation methods should address
questions that remain unanswered after benchmarking, with prospective evaluation used when
retrospective assessments cannot adequately capture interactions with patients, users, or clinical
workflows.

Response to Questions 12–13
The Contribution of Synthetic Data

Synthetic data may contribute useful information to an evidence package, including the assessment of
uncommon hazards. However, important uncertainties warrant further investigation. A device may
detect features of synthetic inputs, infer that it is being tested, and respond differently than it would in
typical use. AI-based data generation may also fail to represent certain patient populations or patterns,
particularly rare patterns.2 These possibilities should be investigated rather than assumed to be resolved
by the apparent realism of generated data.

We recommend that results from synthetic and real-patient data be analyzed and reported separately.
As with phantom testing and clinical testing of imaging devices, both may contribute useful evidence
while providing different information about performance. Combining results into a single estimate could
obscure differences that are important for interpreting the evidence. For tools spanning specialties or

1 Large Language Models Often Know When They Are Being Evaluated
2 Margolis Synthetic Data Generation

230 Science Drive 1201 Pennsylvania Ave, NW www.healthpolicy.duke.edu
Durham, NC 27708 Suite 500, Washington, DC 20004
supporting longitudinal care, assessments should examine whether integrated capabilities improve care,
beyond performance on discrete tasks modeled on individual areas of clinical expertise.

This is another area that needs significant regulatory science research, specifically on whether patterns
reproduced by synthetic data-generating models introduce bias into device evaluations or inadequately
represent relevant populations, and whether detection of artificial testing conditions changes device
behavior. The role of synthetic evidence should be adjusted as findings accumulate, including through
comparisons with real-world clinical results.

Response to Questions 14–15
Comparators and Clinically Meaningful Performance

Performance comparators and acceptance criteria should align with the device’s labeling, intended role,
and claimed benefit. Typical care pathways, including realistic assessments of current patient population
access to specialized expertise, may be appropriate comparators for tools intended to expand access.
“No care” should not be the default comparator where care options exist, but exceptional individual
physician performance should not become a universal threshold for access-expanding tools. However,
tools claiming to improve on gold-standard care should be assessed against comparators that
meaningfully test those claims. Evaluation should address the human–AI system in its intended
workflow, ideally through clinical outcomes.

Response to Questions 18–20
Lifecycle Evidence and Postmarket Monitoring

We encourage FDA to allocate premarket and postmarket evidence requirements around the
information each can provide. Additional premarket testing may offer limited value for questions best
answered during deployment. In those circumstances, postmarket evidence may provide more useful
information relative to the burden of generating additional premarket evidence. Sufficient risk-based
premarket evidence of safety and effectiveness nevertheless remains necessary.

Where an evidence strategy relies on postmarket monitoring, the manufacturer’s submission should
include a rigorous, feasible plan addressing data access, detection capability, resources, responsibilities,
and timely action. The plan should establish how the information needed for monitoring will be
obtained and how findings will support an effective response.

We encourage FDA to develop and test initial risk-based postmarket assessment practices that connect
failure modes identified during premarket evaluation, including through adversarial testing, to
monitoring priorities, evaluation of guardrails, and predefined thresholds for intervention. These
practices should address when changes in device capabilities, workflows, or underlying models warrant
reassessment because earlier evidence may no longer adequately characterize performance. Depending
on the identified risk, responses could include restriction of particular functions, reversion to a
previously evaluated version, or suspension of operation, with provisions for continuity of care. FDA
should evaluate the effectiveness of these regulatory practices and refine them as evidence
accumulates.

A priority for regulatory science should be identifying characteristics of AI-enabled tools that predict
clinically meaningful performance deterioration across deployment settings, geographic locations, and
time. These findings should inform the tailoring of monitoring requirements. FDA should also examine

230 Science Drive 1201 Pennsylvania Ave, NW www.healthpolicy.duke.edu
Durham, NC 27708 Suite 500, Washington, DC 20004
how complementary monitoring methods detect different failures. Fixed benchmarks alone cannot
establish performance on changing real-world inputs, and changing user behavior may affect outcomes
even when device performance remains stable.

Closing

We encourage FDA to advance these regulatory science research recommendations collaboratively
through pilots, sandboxes, postmarket monitoring, learning communities, and shared findings involving
health systems, manufacturers, professional societies, and academic and FDA researchers. Explicit
opportunities to reassess whether evaluation methods predict clinical performance, identify meaningful
deterioration, and support effective intervention would help FDA revise approaches that do not meet
those objectives. We appreciate FDA’s leadership in initiating this discussion and encourage continued
progress accompanied by deliberate testing and transparent learning.

Sincerely,

Christina Silcox, Ph.D.

Research Director, Digital Health

Adjunct Assistant Professor

Duke-Margolis Institute for Health Policy

230 Science Drive 1201 Pennsylvania Ave, NW www.healthpolicy.duke.edu
Durham, NC 27708 Suite 500, Washington, DC 20004