Baxter Healthcare (Cade Clurman)

Please see attached comments.

October 19, 2026

Michelle Tarver, M.D., Ph.D.
Center Director
Center for Devices and Radiological Health
U.S. Food and Drug Administration

Re: Docket Number: FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled
Medical Devices: Discussion Paper and Request for Feedback

Dear Dr. Tarver,

Baxter appreciates the opportunity to submit comments in response to FDA’s discussion paper on
Considerations for the Regulation of Generative AI-Enabled Medical Devices. The discussion paper
envisions a modernized regulatory framework for generative AI-enabled medical devices (GenAI
devices), which includes a risk assessment, premarket evaluation, and postmarket monitoring, and
describes a potential risk-based, total product life cycle approach intended to support timely access to
safe and effective medical devices while applying least burdensome principles. Given the quickly
evolving, complex, and emerging AI landscape and AI focused policy priorities of this new
Administration, we applaud CDRH’s willingness to put forth these ideas in a discussion draft and allow
for public feedback to inform CDRH’s path forward.

About Baxter
Baxter International is a global medical technology company with more than 90 years of experience
developing products and services that support patients and caregivers across the continuum of care.
Following its acquisition of Hillrom in 2021, Baxter has expanded its reach as a leading healthcare
technology company, generating approximately $11.2 billion in global sales in 2025 and serving
patients in more than 100 countries. The company’s portfolio spans numerous therapeutic areas,
including anesthesia and critical care, biosurgery, drug delivery, infusion systems, IV access, nutrition,
renal therapies, and pharmacy workflow solutions.

A key area of innovation is Baxter’s RetinaVue platform, which integrates retinal imaging into primary
care settings to facilitate early disease detection and improve access to screening. RetinaVue enables
primary care physicians to perform retinal exams during routine wellness visits, particularly for
patients with diabetes, allowing for timely referral to ophthalmologists when needed. Through its
RetinaVue Network, Baxter has supported millions of diagnosed exams and is exploring artificial
intelligence algorithms for diabetic retinopathy, macular edema, age-related macular degeneration,
glaucoma, and cardiovascular disease risk assessment. This work reflects Baxter’s broader
commitment to expanding access to medically essential care and leveraging data-driven technologies
to improve patient outcomes.
General Comments
This discussion draft provides an important opportunity for FDA to continue its leadership in adapting
regulatory frameworks to advances in artificial intelligence and digital health technologies. Baxter
supports FDA's efforts to modernize AI-related policies and guidance, including through public
workshops, updated guidance, and a risk-proportionate, least burdensome regulatory approach.

AI-enabled imaging technologies, particularly in ophthalmology, have significant potential to improve
healthcare delivery and patient outcomes by expanding access to screening and diagnostic services,
reducing costs, alleviating workforce shortages, and supporting earlier disease detection. These
technologies can also help empower patients to play a more active role in managing their health
while advancing more efficient, patient-centered models of care.

Continued modernization of FDA's AI framework is consistent with broader federal efforts to promote
U.S. leadership in artificial intelligence, encourage the responsible use of high-quality datasets, and
leverage existing regulatory authorities to support innovation. It also aligns with FDA initiatives to
streamline regulatory pathways for AI, real-world evidence, and other emerging digital health
technologies.

Baxter appreciates the opportunity to share its experience with AI-enabled medical imaging
technologies and provide perspectives on the relevant questions posed in the discussion draft as FDA
considers updates to its regulatory framework for GenAI-enabled devices.

Section IV. Considerations for the Assessment of Risk for GenAI-Enabled
Devices
Question 3: 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?

The intended user and clinical context may affect the risk associated with a GenAI-enabled
informational function. As CDRH recognizes, healthcare professionals may be better positioned than
patients to interpret an output in context, recognize its limitations, and integrate it with other clinical
information. At the same time, patient-facing functions may improve patient empowerment,
engagement, and access to clinical information. Accordingly, patient-facing delivery should not, by
itself, determine the risk classification of a GenAI-enabled function. Risk should instead be assessed
based on the function’s intended use, the nature and directiveness of its outputs, the consequences
of relying on an incorrect output, and the safeguards incorporated into the device. Relevant
safeguards may include clear communication of the function’s intended scope and limitations;
communication appropriate to the intended user, clinical context, risk level, and health literacy level;
as well as clinically appropriate referral when timely, clear, and actionable escalation is warranted.
CDRH should apply these considerations in a manner that does not unnecessarily underestimate
patients’ ability to understand and use clinical information. The evaluation should focus on whether
intended users can understand the output, its limitations, and when escalation is needed, rather than
presume that all patient-facing functions present the same degree of risk. For example, where a
function serves solely to generate referral recommendations, FDA should account for the riskmitigating benefits associated with increased access to care. Such technologies can facilitate earlier
identification of potential disease and appropriate referral to specialists, particularly in settings with
limited health care resources.

Question 6: How could manufacturers characterize and weigh these two [Under-Escalation and
Over-Escalation] directions of error, given that they may not be commensurable and that
acceptable trade-offs may vary by clinical context?

Both under-escalation and over-escalation are relevant to evaluating a risk proportionate framework.
Under-escalation may delay needed treatment, while over-escalation may produce patient anxiety,
unnecessary tests or procedures, and erosion of patient trust. However, the relative significance of
under-escalation and over-escalation should be evaluated according to the device’s intended use,
clinical context, and the consequences of relying on an incorrect output and acceptance criteria
should be tailored to the relevant clinical setting and risk profile. For example, for AI functions that
provide referral recommendations rather than definitive diagnoses or treatment decisions, FDA
should consider the extent to which the technology reduces public health risk by improving access to
screening, enabling earlier clinical intervention, and supporting timely referral to appropriate
healthcare providers.

Section V: A Competency-Based Approach for Premarket Evaluation of
GenAI-Enabled Devices
Question 7. Is the competency-based approach described above, i.e., device benchmarking
followed by clinical confirmation, a useful and appropriate framework for evaluating GenAI-enabled
devices?

The competency-based approach of device benchmarking followed by clinical confirmation
appropriately emphasizes objective assessment of device performance while allowing clinical
confirmation to demonstrate that the technology performs safely and effectively in real-world use.
This is particularly important for AI-enabled medical imaging technologies, where robust
benchmarking against appropriate reference standards can establish analytical performance, and
clinical confirmation can demonstrate utility in intended care settings. For imaging-based AI systems,
for example, FDA should provide flexibility to leverage high-quality real-world evidence datasets for
benchmarking and performance validation, provided sponsors can demonstrate substantial
equivalence and meet applicable sensitivity and specificity expectations. Reference standards and
validation methodologies should be modernized to reflect current clinical practice and the technology
or services the AI is intended to supplement or replace. Consistent with a competency-based
approach, the level of clinical confirmation should be tailored to the device's intended use, risk
profile, and degree of autonomy.
Question 8: How might the two-axis risk framework described in Section IV be considered within
the competency-based approach described above to help determine the level of evidence needed
for a premarket submission?

Baxter supports CDRH’s consideration of a competency-based approach that is tailored to a device’s
intended use and proportionate to its risk. The nature, rigor, and amount of evidence should reflect
the device function, intended use, degree and independence of device activity, and consequences of
relying on an incorrect output. When a device has a well-understood intended use, established clinical
endpoints, and an appropriate predicate device, FDA should permit greater reliance on retrospective
real-world evidence, benchmarking, and other appropriate forms of clinical confirmation rather than
requiring prospective studies by default.

For moderate-risk imaging devices reviewed through the 510(k) pathway, FDA should consider the
availability of large real-world datasets, existing clinical experience with similar technologies,
demonstrated performance against accepted reference standards, and robust postmarket monitoring
capabilities. These factors may support a totality of evidence approach while maintaining confidence
in device performance. Another benefit is that this approach could be more cost effective and more
expeditious by utilizing existing resources, rather than starting from scratch.

Question 10: How should a sponsor establish that performance on a given benchmark predicts safe
and effective real-world behavior for the device’s intended use? What evidence should support the
construct validity of a benchmark used to gate device evaluation, and what role should sponsordeveloped benchmarks play given potential concerns around independence and optimization to the
test?

Sponsor-developed benchmarks should not be disfavored solely because they are sponsor-developed.
Appropriate safeguards may include prespecified methods and acceptance criteria, validation using
independent subsets of real-world data, and qualified independent adjudication where warranted.
Sponsors with access to large, clinically relevant datasets could be permitted to develop fit-forpurpose benchmarks that reflect the device’s intended use, patient population, care setting, and
conditions of use. For AI-enabled imaging technologies, datasets derived from routine clinical practice
may provide relevant evidence of performance under anticipated deployment conditions.

To support the predictive validity of a benchmark, sponsors should define the scope of testing
according to intended use and risk; prespecify evaluation methods and acceptance criteria; justify
scoring rubrics; identify qualified expert adjudicators when applicable; compare outputs against
clinically meaningful reference standards; and assess performance using data representative of
expected deployment contexts. For imaging algorithms, benchmark performance may be evaluated
using sensitivity and specificity metrics comparable to the requirement for prospective clinical studies
supporting relevant 510(k)s. Sponsors should not be required to satisfy materially different
performance thresholds solely because evidence is derived from real-world datasets rather than
prospective studies.
Question 11: 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?

Baxter supports CDRH’s recognition that clinical confirmation should be proportionate to risk and
need not require a prospective clinical study in every case. Sponsors should have flexibility to use
retrospective evaluation and other approaches based on the device’s intended use, degree of clinical
autonomy, risk profile, consequences of an incorrect output, and available evidence.

For AI-enabled imaging devices with well-characterized inputs, established reference standards, and
lower-risk or clinician-mediated uses, high-quality retrospective evidence may provide the most
appropriate and least burdensome means of clinical confirmation. A prospective study should not be
required when a representative real-world dataset, prospectively defined methods and acceptance
criteria, and comparison against an established reference standard or qualified expert adjudication
can demonstrate robust performance and generalizability. More intensive confirmation methods may
be appropriate when errors could result in serious harm, the device operates autonomously, the
intended-use population or setting is not adequately represented, or retrospective evidence cannot
address important uncertainties.

The RetinaVue experience illustrates the potential value of clinically relevant real-world datasets
derived from images collected primarily in primary care settings and interpreted through a nationwide
network of retinal specialists. FDA should permit sponsors to use such evidence when it reflects the
anticipated diversity of patients, care settings, image quality, and clinical conditions.

This approach would align the level of evidence with the device’s actual risk rather than applying a
uniform clinical study requirement. It would also advance FDA’s least burdensome principles, reduce
unnecessary development costs, and accelerate patient access without lowering applicable safety and
effectiveness standards. FDA should similarly allow sponsors to justify sensitivity, specificity, and other
performance endpoints based on intended use, risk, real-world evidence, and relevant predicatedevice standards.

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?

Baxter supports a flexible, fit-for-purpose approach to selecting performance comparators and
acceptance criteria. As CDRH recognizes, open-ended GenAI outputs may not have a single correct
response, and multiple responses may be acceptable. Depending on the device and intended use,
appropriate comparators may include a panel of qualified clinicians, a median clinician in practice, an
applicable standard of care, a traditional clinical reference standard, or the performance of a cliniciandevice team.
For imaging-based AI algorithms, including retinal image analysis systems, reference standards and
validation methodologies should reflect current clinical practice and the technology that the AI system
is intended to supplement or replace. FDA should update and modernize expectations for validation
methodologies and reference standards to facilitate the use of real-world evidence and
contemporary clinical datasets where appropriate. Reference standards should evolve alongside
clinical practice and should be aligned with the real-world clinical workflows, technologies, and
services that the AI system is designed to augment, streamline, or replace. FDA should allow sponsors
to use expert clinical adjudication, consensus review by qualified specialists, reference standards
derived from established clinical workflows, or ground truth established through validated clinical
outcomes, where appropriate.

CDRH should preserve flexibility for a sponsor to select a generalist or specialist comparator based on
the device’s intended use, intended user, clinical task, and deployment setting. In particular,
evaluation standards should reflect the clinical context in which the device will be used and the level
of expertise that would ordinarily perform the relevant clinical task in practice.

Question 15: Are there ways in which performance might be assessed relative to the care,
technology, or course of action likely to occur in the absence of the device, rather than to the
comparators described in this section? What approaches might be used to identify and justify a
comparator such as unaided clinical judgment, delayed specialist review, or no intervention?

Baxter supports assessing performance relative to the care, technology, or course of action likely to
occur in the absence of the device when that comparator is appropriate to the device’s intended use
and deployment setting. For AI-enabled imaging technologies, the comparator should reflect realworld clinical practice and the technology or service that the AI function is intended to supplement or
replace. Depending on the context, appropriate comparators may include established clinical
workflows, unaided clinical judgment, delayed specialist review, expert adjudication, consensus
review by qualified specialists, or ground truth established through validated clinical outcomes, such
as a diagnosis confirmed by biopsy or expert consensus.

The sponsor should identify and justify the comparator based on the device’s intended use, intended
user, clinical task, and deployment setting. Evaluation should reflect how the device will be used in
practice. For clinician-support tools, assessment of the clinician-device team may provide the most
clinically relevant measure of performance, whereas evaluation of the device operating independently
may be appropriate for autonomous functions. FDA should preserve flexibility to select fit-for-purpose
comparators rather than requiring a single comparator across device types and care settings. This
approach would support clinically meaningful evaluations while maintaining rigorous, prospectively
defined performance criteria.
Question 16: Should independent third parties be involved in some or all aspects of a competencybased approach, including device benchmarking and clinical confirmation? If so, in what ways might
qualified, independent third-party participation contribute to this approach, and what
qualifications and independence criteria should apply? Are there aspects of the assessment for
which third-party involvement would be impractical or inadvisable? What safeguards or program
design features would be critical to prevent third-party participation from limiting competition or
preventing innovation?

There is value in utilizing qualified independent third-party participation in device benchmarking and
clinical confirmation. Independent third parties could develop and maintain benchmark datasets,
maintain sequestered evaluation datasets, establish consensus validation methodologies, serve as
independent expert adjudicators, and support development of standards relevant to AI-enabled
imaging technologies. Third-party participation should be tailored to the nature of the device and the
assessment. Qualification and independence criteria should address relevant technical and clinical
qualifications and structural independence from the device sponsor and, where applicable, the
foundation model developer. Any third-party framework should incorporate appropriate safeguards to
avoid unnecessarily restricting competition, innovation, or the adoption of beneficial technologies.
Transparency and broad stakeholder engagement are important components of such safeguards,
including opportunities for public input, stakeholder education, and access to relevant guidance,
validation methodologies, and performance benchmarks.

To support these objectives, FDA should consider convening a public-private workshop involving
industry, regulators, clinicians, academic researchers, and patient stakeholders to inform future
guidance on the use of real-world evidence (RWE) for the development, training, validation, and
evaluation of AI-enabled medical imaging systems. Such a forum could help establish more consistent
expectations, promote scientifically rigorous and clinically relevant evaluation approaches, and
support the modernization of validation methodologies and reference standards to better reflect realworld clinical practice and evolving technologies.

Section VI: Postmarket Monitoring for GenAI-Enabled Devices
Question 18: CDRH is considering whether it may be appropriate to accept greater premarket
uncertainty regarding a GenAI-enabled device’s benefit-risk profile through greater reliance on
postmarket monitoring. Under what conditions might such an approach be appropriate, and what
characteristics of a monitoring program would need to be in place to justify reduced premarket
evidence? Are there device types or risk profiles for which this approach would not be appropriate?

Baxter supports a risk-proportionate approach under which robust postmarket monitoring may, in
appropriate circumstances, complement premarket evidence. The appropriateness of accepting
greater premarket uncertainty should depend on the device’s intended use, risk profile, degree and
independence of device activity, consequences of relying on an incorrect output, and the ability of the
monitoring program to detect relevant changes in performance.
Reduced premarket evidence may be more appropriate when the device has a well-understood
intended use, established clinical endpoints, an appropriate predicate device, demonstrated
performance against accepted reference standards, relevant real-world evidence, and robust
postmarket monitoring capabilities. These factors may support a more flexible evidentiary approach
for moderate-risk imaging devices reviewed through the 510(k) pathway while maintaining confidence
in device performance.

A monitoring program supporting such an approach should use prespecified analyses and thresholds
and should be proportionate to the device’s risk profile. Relevant components may include periodic
reassessment against premarket benchmarking thresholds, sample-based review of real-world inputs
and outputs by qualified independent clinicians, and monitoring for performance degradation
resulting from changes in the input population, data environment, or underlying model components.

Question 19: Please comment on the potential approaches to postmarket performance evaluation,
including periodic re-benchmarking, sample-based clinician review, and performance degradation
monitoring. What additional approaches should CDRH consider, and how should the cadence and
triggering events for reassessment be determined?

CDRH’s proposed postmarket monitoring approaches, including periodic reassessment of
benchmarks, periodic sample-based clinician review, and performance degradation monitoring are
consistent with a total product life cycle framework in which the scope, frequency, and level of
postmarket evidence are proportionate to the device’s intended use and risk profile. Periodic
reassessments of benchmarks may allow a sponsor to reassess the device against prespecified
thresholds established during premarket evaluation. The initial competency-based assessment may
serve as a baseline for evaluating subsequent performance and determining whether a modification
has affected device safety or effectiveness.

Since sponsors are the most knowledgeable about their devices, they should be permitted to specify
the analyses and thresholds used to detect performance degradation. The cadence of routine
reassessment and the triggering events for additional evaluation should be informed by the device’s
intended use, risk profile, deployment environment, and relevant changes to the underlying model or
other components of the deployment architecture. The framework should remain sufficiently flexible
to accommodate differences among devices while maintaining prespecified methods, thresholds, and
criteria for identifying changes that may affect safety or effectiveness.

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

Clinicians, healthcare institutions, professional societies, standards-setting bodies, public-private
consortia, patients, payers, and government authorities may have roles in the deployment, postmarket monitoring, reporting, and ongoing evaluation of GenAI-enabled devices. For example,
clinicians and health care institutions may contribute real-world inputs, outputs, and clinical
experience relevant to periodic performance evaluation. Qualified independent clinicians may also
review samples of real-world inputs and outputs against prospectively defined criteria. Professional
societies and standards-setting bodies may contribute to consensus validation methodologies and the
development of standards that promote consistency across testing and review. Independent third
parties or public-private consortia may maintain benchmark or sequestered evaluation datasets and
serve as independent expert adjudicators.

These roles should be clearly structured so that stakeholders provide information or expertise
relevant to monitoring while the manufacturer remains responsible for evaluating device
performance. As previously mentioned, CDRH should also consider a public-private workshop
involving industry, regulators, clinicians, academic researchers, and patient stakeholders to support
consistent expectations and scientifically sound approaches for the use of real-world evidence in
training and validating AI-enabled medical imaging systems.

Question 23: How might [Predetermined Change Control Plan] PCCP concepts or other changecontrol approaches be adapted for GenAI-enabled devices when the nature or scope of future
modifications cannot be fully prespecified?

Postmarket changes may include intentional and discrete modifications, passive or incremental model
evolution, and unplanned changes arising from updates to an underlying third-party foundation
model. Because the nature or scope of all future modifications may not be fully prespecified, a
change-control approach should focus on the types of changes anticipated, the methods used to
evaluate those changes, and the evidence needed to determine whether safety or effectiveness has
been affected. The manufacturer’s premarket competency-based assessment could serve as the
baseline for evaluating later modifications. A device benchmarked against a defined set of capabilities
before authorization could be reassessed against those capabilities following a modification. The
extent of reassessment should be proportionate to the nature and expected impact of the change.

The change-control framework should allow for a range of regulatory oversight that is least
burdensome. Depending on the nature and potential effect of a modification, oversight may range
from documentation within the manufacturer’s quality management system to FDA review and
authorization before implementation. A PCCP may provide a mechanism for facilitating certain
changes without requiring a new premarket submission. For devices incorporating third-party
foundation models, the framework should also account for changes initiated by the foundation model
developer. However, the device manufacturer should remain responsible for demonstrating the safety
and effectiveness of the final device incorporating the foundation model.

Conclusion
Baxter appreciates CDRH’s proactive engagement with stakeholders as the Agency evaluates future
regulatory approaches for generative AI-enabled medical devices. As AI technologies continue to
advance and become more deeply integrated into healthcare delivery, FDA has an important
opportunity to establish a modern, risk-based regulatory framework that promotes innovation while
maintaining appropriate assurances of safety and effectiveness.

Baxter encourages CDRH to adopt a risk-proportionate framework that bases evidentiary expectations
on the intended use and clinical impact of a device function rather than on whether the technology is
patient-facing. The level of evidence required should be commensurate with the device’s actual risk,
role in clinical decision-making, and potential impact on patient care. Such an approach would
support innovation while ensuring that regulatory oversight remains appropriately tailored to the
intended use of the technology.

Consistent with this principle, Baxter supports a framework that: (1) scales evidence requirements
according to the risk and clinical significance of the device function; (2) provides flexibility to use
retrospective real-world evidence as a principal source of clinical confirmation where scientifically
appropriate; (3) permits sponsor-developed, fit-for-purpose benchmarks and clinical competency
standards that reflect intended-use populations, real-world deployment environments, and
contemporary clinical practice; (4) modernizes comparator and reference-standard expectations to
reflect the technology or clinical services a device is intended to supplement or replace; and (5)
leverages qualified independent third parties, where appropriate, to support consistent, transparent,
and scientifically rigorous evaluation.

Baxter further encourages CDRH to recognize that performance expectations for AI-enabled medical
imaging technologies should be aligned with existing regulatory standards for comparable devices. In
particular, sensitivity and specificity thresholds used to assess AI-enabled functions should be
consistent with those applied in prospective clinical studies supporting relevant 510(k) submissions,
rather than establishing novel performance expectations solely because AI methodologies are
involved.

In addition, Baxter recommends that CDRH convene a public-private workshop involving regulators,
industry, clinicians, academic researchers, and patient stakeholders to consider strong use cases of
leveraging RWE to support AI, such as retinal scans, to get into technical details that the guidance may
not be able to address. This workshop also would offer an opportunity to advance shared
understanding of best practices for the development, training, validation, and evaluation of AIenabled medical imaging technologies. Such a forum would help establish consistent expectations
regarding the appropriate use of real-world evidence, benchmarking approaches, clinical competency
standards, and performance evaluation methodologies, while fostering scientific consensus on
emerging technical and regulatory issues.

Collectively, these actions would advance CDRH’s goals of applying least burdensome principles,
supporting innovation, and promoting timely patient access to safe and effective medical devices. By
continuing to engage stakeholders and modernize its regulatory framework for AI-enabled
technologies, FDA can foster regulatory predictability, encourage responsible innovation, and ensure
that patients benefit from transformative advances in digital health and medical imaging.

Sincerely,

Cade Clurman
Associate Vice President of Government Affairs
Baxter Healthcare