American Academy of Ophthalmology (Michael X. Repka)

See attached file(s)

October 19, 2026

Division of Dockets Management (HFA- 305)
Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

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

The American Academy of Ophthalmology (the Academy) appreciates the opportunity to
comment on FDA’s Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback. The Academy is the largest association of eye
physicians and surgeons in the United States. A nationwide community of over 20,000 medical
doctors, we protect sight and empower lives by setting the standards for ophthalmic education,
supporting research, and advocating for our patients and the public. We support a risk-based
approach that promotes innovation while protecting patient safety, preserving clinical judgment,
and ensuring manufacturer accountability.

Question 1. Does the two-axis risk framework, organized around device activity and the
consequence of relying on an incorrect output, appropriately capture the dimensions most
relevant to the risk of a GenAI-enabled software function? If there are additional dimensions—
such as the reversibility of a resulting action, the availability of downstream safeguards, the
time pressure of the deployment setting, or the traceability of the output (i.e., to primary
source materials)—that should be represented in a risk framework, please describe and provide
examples of how they should be represented.

The Academy views the two-axis framework as a useful starting point but recommends explicitly
considering the expertise needed to evaluate an output, the quality and completeness of input
data, the confidence with which an output is presented, and the magnitude and clinical
consequences of an error. FDA could consider adding a third axis that conveys the quality and
completeness of input data in relation to output reliability.

The relevant expertise should reflect the intended user’s training and experience with the clinical
task, rather than specialty designation alone. Manufacturers should present information and
assessments in a manner that clinicians can understand and appropriately use for triage and
referral. These considerations should inform device evaluation and safeguards while preserving
physicians’ judgment about the limits of their expertise.

Traceability to verifiable primary source materials should also inform risk assessment. Where
relevant to the intended use, manufacturers should enable users to identify and review the
information supporting an output. FDA should also consider whether the resulting action is
reversible, whether timely consultation or other downstream safeguards are available, and
whether the setting permits meaningful verification before action.

The framework should distinguish among different types and degrees of error. Listing the correct
diagnosis second or third in a differential may have different consequences from omitting it
entirely. Similarly, a confidently incorrect recommendation may prompt different actions than an
output that appropriately communicates uncertainty. Neither uncertainty nor a broad differential
should automatically be presumed to reduce risk; the assessment should consider the resulting
clinical action and potential harm. False-positive and false-negative outputs should also be
evaluated separately because their consequences may differ substantially by clinical context.

Input quality and completeness are additional considerations. Where validated for the intended
use, a function should identify missing or unreliable information and suggest additional data that
could improve the reliability of its output. Such requests should not delay necessary escalation in
urgent situations. Demonstrated safeguards should be recognized as potential risk mitigations
and support evidence requirements proportionate to the remaining risk.

Question 3. CDRH seeks input on whether and when a GenAI-enabled function that results in
delivery of clinical information to patients, as opposed to HCPs, could present different or
higher risks, while also recognizing the potential benefits associated with improved patient
empowerment, engagement, and access to clinical information. What device characteristics,
output features, or safeguards might mitigate risks that could arise when a user lacks the
domain knowledge to independently evaluate an output, without unnecessarily
underestimating patient capability?

Risk should depend on the nature of the clinical information and its intended use, rather than an
assumption that all patient-facing functions are equally risky. Translating medical terminology in a clinical
note into plain language may improve patient understanding and engagement. A function that triages
symptoms or provides individualized clinical recommendations presents greater risk because its output
may influence whether and when a patient seeks medically necessary care.

Evaluation should assess whether the function appropriately challenges incorrect assumptions or instead
reinforces them to agree with the user. Testing should include situations in which a patient repeatedly
insists that erroneous information is correct, as well as situations in which an output could cause
inappropriate action, delay care, or misrepresent a clinician’s instructions. Evaluations should also
consider unsupported statements about the appropriateness of prior care that could undermine the
patient-clinician relationship.

Communicating uncertainty could be valuable, but a model’s expressed confidence should not be treated
as reliable without external validation. Manufacturers should demonstrate that any confidence or
uncertainty information meaningfully reflects input and output reliability.

Safeguards should include understandable language, clear descriptions of the function’s limits,
preservation of clinician-provided warning signs and return precautions, clear disclosure that
content is AI-generated, and accessible routes to human review. These safeguards would support
informed patient decision-making without assuming that patients cannot understand or evaluate
clinical information.

Question 4. CDRH seeks input on whether and how the distinction between generalist and
specialist physicians might inform the assessment of risk for HCP-facing GenAI-enabled
functions. Under what circumstances, if any, might risk be affected when an HCP who lacks the
relevant clinical specialist knowledge receives information that falls within a specialist area of
practice? What device characteristics or safeguards might mitigate such risks?

Risk assessment should consider whether intended users have the relevant training and
experience to evaluate a function’s outputs. However, generalist and specialist labels do not fully
capture clinical competence. Physicians exercise professional judgment in determining whether a
clinical question falls within their own expertise, and FDA’s evaluation of device safety and
effectiveness should preserve that judgment.

Manufacturers should clearly describe the function’s intended use, limitations, and any
knowledge or training needed for safe use. Outputs should be presented so that physicians can
understand the information and act appropriately within the intended clinical context,
particularly for triage, referral. A function supporting referral decisions should be evaluated
differently from one intended to provide subspecialty treatment recommendations.

Evaluation should assess whether intended users can interpret outputs, recognize important
limitations, and identify when consultation is appropriate. Safeguards should be proportionate to
the consequences of the clinical action rather than based solely on the user’s specialty
designation.

Traceability to supporting sources, clear limits on the function’s scope, and appropriate
communication of uncertainty can help clinicians assess recommendations outside their routine
experience. When available information cannot support a sufficiently reliable recommendation,
the function should identify that limitation and provide appropriate guidance for obtaining
additional information or escalating care. HCP-facing functions should likewise indicate when a
question falls outside their intended use or validated scope, rather than extrapolating beyond it.

Question 6. For GenAI-enabled care escalation functions, CDRH is considering how the
evaluation may account for both under-escalation and over-escalation. How could
manufacturers characterize and weigh these two directions of error, given that they may not be
commensurable and that acceptable trade-offs may vary by clinical context?

Care escalation is an integral part of clinical judgment. The Academy has developed numerous
Preferred Practice Pattern guidelines and Ophthalmic Technology Assessments that can inform
evaluation, benchmarking, and validation of escalation recommendations, decisions, and other
outputs. Where applicable, these resources should provide a verifiable basis for reviewing
recommendations, supporting explainability, and educating users.

When existing guidance is unavailable or does not adequately address the clinical situation,
evaluation should include structured review by clinicians with relevant experience, using cases
analogous to the assessment of clinical judgment in a board examination. We encourage
collaboration among CDRH, specialty societies, and sponsors to identify gaps and develop
appropriate reference criteria for emerging technologies. To advance these efforts, the Academy
has been working as part of the Collaborative Community for Ophthalmic Innovation to bring
stakeholders together to drive innovation within the field.

Manufacturers should report under-escalation and over-escalation separately and describe the
consequences of each. Under-escalation may delay necessary care, while over-escalation may
lead to unnecessary testing, referrals, costs, or pressure on limited specialty services. These
outcomes should not be treated as interchangeable.

Performance acceptance criteria should reflect the intended setting, urgency, and severity of
potential harm. For time-sensitive conditions that threaten vision, evaluation should place
particular emphasis on missed or delayed escalation. Evaluation should also examine whether
over-escalation could impede timely care for other patients. Manufacturers should justify these
trade-offs and test performance using representative cases, including ambiguous presentations
and incomplete information.

Question 11. CDRH is considering that clinical confirmation for a GenAI-enabled device might
not require a prospective clinical study in every case, and has described above a range of
approaches of increasing rigor and patient exposure. How might a sponsor select and justify a
confirmation approach tailored to a device’s intended use and proportionate to the device’s
risk profile? Are there device types or risk profiles for which one or more of these approaches
would be insufficient or inappropriate? How might the anticipated distribution of real-world
inputs be taken into consideration? Are there other methods of clinical confirmation that might
help inform the evaluation of GenAI-enabled devices?

The clinical confirmation approach should reflect the function’s intended use, its risks, and its
potential effect on patient care. When a function informs treatment, influences whether a patient
seeks medical care, or otherwise has the potential to affect clinical outcomes, the Academy
recommends clinical validation through appropriately designed studies, with the rigor and design
proportionate to the risk. For the highest-risk functions, this should generally include a
prospective clinical study.

Prospective clinical studies are particularly important when safety and effectiveness depend on
patient behavior, clinician reliance, or workflow effects that cannot be adequately evaluated
through retrospective data or simulated use. A sponsor proposing an alternative should explain
why the selected approach can establish safety and performance for the intended use, including
foreseeable ways that users may act on the output.

Functions limited to explaining existing information or supporting administrative tasks, such as
appointment scheduling or summarizing notes, may present lower risk and may not require a
prospective trial in every case. However, these functions still carry risk. For example, an incorrect
interpretation of return precautions could cause a patient not to report concerning symptoms.

Validation should reflect the anticipated distribution of real-world inputs, including variation in
language, health literacy, clinical complexity, and completeness of records. Sponsors should
justify the representativeness of their test data and evaluate performance across relevant patient
groups and deployment settings.

Question 14. For open-ended device outputs, how should performance comparators and
acceptance criteria be selected? When a panel of qualified clinicians serves as the comparator,
how should the applicable standard (for example, the standard of care versus the performance
of a median clinician in practice) be defined and justified? Should generalist or specialist
physicians be used as a performance standard? When should human-AI team performance,
rather than the device operating alone, serve as the basis for evaluation?

Performance comparators should reflect the function’s intended clinical task and claims.
Generalist physicians may be appropriate comparators when the function’s main purpose is
patient triage. For functions involving specialized medical care, specialist physicians should
generally be the default comparator for validation. Comparator qualifications should reflect
relevant training and demonstrated clinical experience.

The reference standard used to establish the correctness of an output should be distinguished
from the clinician comparator used to assess relative performance. Where applicable, Academy
Preferred Practice Pattern guidelines, Ophthalmic Technology Assessments, and established
clinical reference standards should inform evaluation and benchmarking. These resources can
also support review of recommendations, explainability, and user education. When clinician
panels are used, sponsors should justify panel composition, adjudication methods, and the
reliability of the resulting reference standard.

For open-ended outputs, acceptance criteria should address clinical correctness, completeness,
potentially harmful omissions, appropriate escalation, and appropriate abstention when the
function lacks sufficient information, or a question falls outside its intended use. A medianperformance benchmark should be accompanied by explicit safety criteria so that acceptable
average performance does not obscure serious errors. The CCOI publication, Foundational
Considerations for Artificial Intelligence Using Ophthalmic Images, provides a useful framework
for considering reference-standard quality separately from minimum acceptable performance.

Rigorous studies should evaluate clinicians alone, the device alone, and clinicians working with an
assistive or augmentative device, as appropriate to the intended use and risk. These comparisons
should establish whether the intended configuration improves clinical performance and patient
care.

When the intended use requires clinician review or action in response to an output, evaluation
should include human-AI team performance under realistic conditions. Device-only testing
remains useful for identifying errors but cannot, by itself, establish whether the intended
workflow improves care.

Question 21. What roles might clinicians, healthcare institutions, professional societies,
standards-setting bodies, and other stakeholders appropriately play in postmarket monitoring,
and how can these roles be structured without diffusing manufacturer accountability?

Manufacturers should retain ultimate accountability to users, patients, and other stakeholders for
the safety, effectiveness, accuracy, and quality of their devices. Clinicians, healthcare institutions,
professional societies, and standards-setting bodies should contribute to postmarket monitoring,
but their participation should not transfer manufacturer responsibility.

FDA should require manufacturers of GenAI-enabled medical devices to establish ongoing
monitoring proportionate to the function’s risks. Monitoring should assess whether performance
remains acceptable in clinical use, including after updates and across relevant patient populations
and deployment settings. Manufacturers should remain responsible for collecting and
investigating reports, identifying performance changes, implementing corrective actions, and
communicating relevant findings to FDA and users.

Aggregate findings relevant to safety and effectiveness should be publicly available, with
appropriate protection of patient privacy and proprietary information. FDA should consider a
public repository of standardized manufacturer performance reports that identifies the device
version, intended use, evaluation methods, and relevant limitations.

Clinicians and patients should retain accessible reporting channels. Healthcare institutions can
help identify workflow problems and recurring safety concerns, while professional societies and
standards-setting bodies can support development and maintenance of clinically meaningful
performance measures and consistent reporting practices.

FDA and manufacturers should also address how participating stakeholders will be compensated
for the time, expertise, and infrastructure required for monitoring. Clearly defined roles,
reporting pathways, and resources should support meaningful participation while preserving
identifiable manufacturer responsibility for monitoring product effectiveness and addressing
safety concerns.
Closing Remarks

The Academy sincerely appreciates the opportunity to share our feedback with FDA on
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback. Ophthalmology has been at the forefront of clinical AI, including FDA’s
2018 authorization of the first device to provide autonomous AI-based screening decisions
without clinician interpretation of the images or results. Although that technology differs from
generative AI, its use highlights the importance of appropriate validation and ongoing monitoring.
As FDA develops its approach to generative AI-enabled medical devices, we encourage careful
consideration of user expertise, the consequences of incorrect outputs, evidence requirements
proportionate to risk, and manufacturer accountability. The Academy welcomes continued
collaboration with FDA to advance safe and effective technologies that improve patient care.

Should you have questions about any of our comments or seek additional input, please feel free
to contact me or Scott Haber, Director of Public Health Advocacy, at shaber@aao.org or via
phone at 202-737-6662.

Sincerely,

Michael X. Repka, MD, MBA
Medical Director, Governmental Affairs
American Academy of Ophthalmology