Melinda Rubinstein

I am submitting the attached public comment in response to FDA’s discussion paper, “Considerations for the Regulation of Generative AI-Enabled Medical Devices” (Docket FDA-2026-N-7874). Please consider the attached document, including its supporting references, as my complete comment.

I offer these recommendations in my individual capacity as a Johns Hopkins School of Nursing MSN candidate and Policy Honors Program participant, following more than 20 years working in technology implementation and organizational transformation.

I support the FDA’s exploration of competency-based evaluation, but encourage the Agency to assess clinical competence beyond benchmark accuracy. A generative AI-enabled device must demonstrate safe, equitable, understandable, and accountable performance in the settings where patients and healthcare professionals will actually use it.

My attached comment addresses four priorities:

1. PATIENT AGENCY AND ACCESS (Questions 3, 5, 9). Evaluate whether patients can understand, question, verify, and appropriately act on AI-generated clinical information. Assess uncertainty communication, multi-turn interactions, and appropriate escalation to professional care. Patient access should not itself be treated as a risk; safeguards should be proportionate to a device’s intended function and potential harm.

2. HEALTH EQUITY (Questions 9, 10, 12, 13). Require risk-proportionate evaluation across clinically relevant populations and settings, including subgroup performance and transparent disclosure of evidence gaps. Representative real-world data are important; synthetic data can supplement evaluation but should not be assumed to establish representativeness without validation.

3. NURSES IN AI GOVERNANCE (Questions 7, 14, 16, 21). Include nursing expertise in defining competencies, selecting clinical comparators, evaluating workflows, conducting clinical confirmation, and monitoring postmarket safety when devices materially affect nursing practice. Assess the performance of the actual human–AI care team, not only the model or a physician comparator.

4. DATA STEWARDSHIP AND INTEROPERABILITY (Questions 1, 9, 19, 24, 25). Evaluate how devices respond to missing, outdated, conflicting, or fragmented clinical information. Establish appropriate expectations for source traceability, data provenance, model-version documentation, and accountability for safety-relevant changes to third-party foundation models. Voluntary Foundation Model Device Master Files may help support regulatory transparency without displacing the finished-device manufacturer’s responsibility for safety and effectiveness.

Across these priorities, I recommend a risk-proportionate approach that recognizes implementation costs, privacy and proprietary constraints, limits of subgroup sample sizes, and the burdens of clinical reporting. These considerations should shape feasible safeguards rather than eliminate them.

The attached comment provides the full rationale, supporting research, and specific recommendations. Its central proposition is simple: technical capability is not the same as clinical competence.

Thank you for considering this feedback.

Public Comment on Considerations for the
Regulation of Generative AI-Enabled
Medical Devices
Docket No. FDA-2026-N-7874

Introduction
I appreciate the Food and Drug Administration's invitation to provide feedback on its
discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical
Devices. I support the Agency's exploration of a competency-based approach to these
technologies and encourage the FDA to consider a broader definition of clinical
competence. Specifically, the definition should account not only for technical accuracy
but also for the patients, clinicians, healthcare systems and information environments in
which these devices will operate.

I offer this perspective as a former technology leader with more than 20 years of
experience in technology implementation organizational transformation and helping
people adopt complex digital systems. I am now pursuing a Master of Science in Nursing at
Johns Hopkins University, where I participate in the School of Nursing's Policy Honors
Program. My transition from technology to nursing reflects a conviction that technological
innovation must ultimately serve human needs, particularly in healthcare, where the
consequences of poorly designed systems are profoundly personal.

Generative AI presents an extraordinary opportunity to improve access to clinical
information, support healthcare professionals and make complex health systems more
responsive to patients. But the effectiveness of these technologies cannot be established
through model performance alone. A system may answer medical questions accurately in
a controlled benchmark yet fail to communicate clearly with patients, perform consistently
across populations, integrate safely into clinical workflows or recognize that the
information upon which its conclusions depend is incomplete.

I encourage the FDA to consider four interconnected priorities in its developing regulatory
approach:
 1. Patient agency and meaningful access to clinical information.
2. Health equity and representative evaluation.
3. Nursing participation in AI design, evaluation and governance.
4. Patient data stewardship, interoperability and accountability for the information
upon which AI systems depend.

These priorities are not peripheral ethical considerations. They are integral to determining
whether a medical device can be used safely and effectively for its intended purpose.

1. Patient Agency and Access to Clinical Information
Relevant FDA Questions: 3, 5 and 9

The FDA's discussion paper appropriately recognizes the potential benefits of
democratizing access to clinical information, alongside the risks associated with patientfacing generative AI. I encourage the Agency to distinguish between the risks inherent in a
device's clinical function and assumptions about a patient's ability to use medical
information.

The FDA should evaluate whether a device supports patients in understanding,
questioning, verifying and appropriately acting on the information it provides. This
distinction is especially important because patients already navigate complex medical
information, often without adequate support. Generative AI may help make that
information more accessible. While there is evidence supporting that AI can help bridge
gaps in patient understanding (Kumar et al., 2025), research evaluating AI-generated
discharge instructions has identified potentially harmful inaccuracies, including
hallucinated information and medication-related errors (Stanceski et al., 2024). Improved
accessibility does not eliminate the need for rigorous clinical validation. These findings
support a regulatory approach that evaluates both the benefits of patient understanding
and the consequences of incorrect or misleading information.

For patient-facing devices, the FDA should consider evaluating:

• Whether patients can accurately understand the information presented, including
its uncertainty and limitations.
• Whether patients can ask clarifying questions, challenge outputs and identify the
sources or clinical basis for recommendations where appropriate.
• Whether a device appropriately identifies circumstances requiring professional
evaluation or emergency care.
 • Whether multi-turn interactions remain safe as patients provide new information,
challenge earlier conclusions or describe changing symptoms.

The degree of validation should be proportionate to the device's intended function and
potential consequences. A device translating a clinical note into plain language presents
different risks from one directing medication changes or advising a patient whether
emergency care is necessary. There are legitimate concerns about automation bias,
inappropriate reliance and the possibility that patients will act on incorrect information.
Those concerns warrant meaningful safeguards. They do not necessarily justify restricting
access to useful clinical information. I encourage the FDA to evaluate patient agency as a
component of safe device performance, rather than framing patient access and patient
safety as competing objectives.

2. Health Equity and Representative Evaluation
Relevant FDA Questions: 9, 10, 12 and 13

Aggregate accuracy of generative AI systems can obscure clinically consequential
differences in performance across patient populations. A device may perform well overall
while producing less accurate, less understandable or less appropriate recommendations
for particular populations.

Research has already demonstrated that clinical AI systems can reproduce inequities
embedded in healthcare data and evaluation methods. Obermeyer and colleagues (2019),
for example, demonstrated how a population-health algorithm using healthcare
expenditures as a proxy for health needs systematically underestimated the needs of
Black patients. Although this study evaluated a predictive algorithm rather than generative
AI, it illustrates a fundamental challenge: apparently objective performance measures can
conceal structural inequities.

More recent research examining large language models in simulated clinical decisionmaking has identified differences in medical recommendations associated with
sociodemographic characteristics, even when clinical details are held constant (Omar et
al., 2025). These findings reinforce the need to evaluate not only whether a model performs
accurately, but for whom, under what circumstances and according to which measures.

I encourage the FDA to consider the following expectations for risk-proportionate
evaluation:
 • Representative evaluation: Manufacturers should justify how evaluation datasets
and clinical scenarios reflect the intended patient population, including relevant
demographic, linguistic and clinical variation.
• Subgroup performance: Safety-critical performance should be evaluated across
clinically relevant subgroups, with transparent reporting of meaningful differences
and statistical uncertainty.
• Evidence gaps: Manufacturers should disclose when available data are insufficient
to establish performance for particular populations.
• Synthetic data: Synthetic cases may expand testing capacity but should not
automatically be treated as equivalent to representative real-world data. Their
validity for the intended evaluation purpose should be demonstrated.
• Ongoing evaluation: Postmarket monitoring should address performance
disparities that emerge as devices encounter populations and settings insufficiently
represented during development.

These recommendations must also acknowledge practical constraints. Rare conditions,
small patient populations, privacy protections and limited data availability can make
comprehensive subgroup testing difficult. The appropriate response is not necessarily to
require identical evidence for every device or population. Rather, the FDA should consider
a proportionate approach that requires manufacturers to justify their evaluation strategies,
disclose limitations and address residual uncertainty through additional validation or
postmarket monitoring.

Equity should be treated as a measurable dimension of clinical safety and
effectiveness, not simply as an aspiration expressed in product documentation.

3. Nurses in AI Design, Testing and Governance
Relevant FDA Questions: 7, 14, 16 and 21

The FDA's exploration of competency-based evaluation raises an important question
about whose clinical competence should serve as the benchmark. The discussion paper
considers comparisons with generalist and specialist physicians, as well as the
performance of human-AI teams. These are valuable approaches, but clinical care is not
delivered by physicians alone.

Nurses represent an essential part of the clinical environment in which AI-enabled medical
devices may be used. They assess patients, recognize changes in condition, administer
treatments, coordinate care, provide patient education and frequently identify when
information or technology does not align with a patient's clinical presentation. When AIenabled devices affect these functions, nursing expertise should inform how clinical
competence is defined and evaluated.

Research examining nurses' experiences with machine-learning clinical decision support
has identified concerns related to clinical autonomy, usability, implementation, and the
importance of involving nurses in technology design and development (Wieben et al.,
2025). Similarly, research involving nursing informaticists highlights the importance of
preparing nurses to participate in AI implementation planning, education, and evaluation
(Turchioe et al., 2024). While these studies do not establish that any particular governance
structure will improve clinical outcomes, they do reinforce the importance of evaluating
technology with the professionals whose work it will affect.

My previous career in technology adoption informs this recommendation. Even technically
sophisticated systems can fail when evaluation focuses on what the technology can do
rather than how people will use it. In healthcare, those failures can have direct
consequences for patient safety.

I encourage the FDA to consider:

• Competency definition: Include relevant nursing expertise when defining
benchmarks for devices that affect nursing assessment, triage, patient education,
care coordination or other nursing functions.
• Clinical comparators: Select human comparators according to the device's
intended clinical task and users, rather than defaulting to physician performance
across all functions.
• Clinical confirmation: Include nurses in usability testing, workflow simulation,
clinical adjudication and prospective evaluation when nursing practice is materially
affected.
• Postmarket governance: Establish meaningful pathways for nurses and other
frontline clinicians to identify, report and help evaluate emerging safety concerns.

These expectations should be proportional to intended use. Not every device requires
nursing participation and multidisciplinary evaluation creates additional coordination and
resource demands. However, when a device affects nursing practice, excluding nursing
expertise risks overlooking important dimensions of clinical performance. Postmarket
monitoring also requires clear accountability. Frontline clinicians should be able to report
unsafe outputs and workflow failures without assuming responsibility for correcting
problems that remain within the manufacturer's control. Clinical competence should be
evaluated in the context of the healthcare team that will actually use the device.

4. Patient Data Stewardship, Interoperability and
Accountability
Relevant FDA Questions: 1, 9, 19, 24 and 25

The quality of generative AI outputs depends on the quality and availability of the
information upon which those outputs are based. Healthcare information remains
fragmented across organizations, systems and care settings. Patients frequently bear the
burden of reconciling incomplete records, transferring information between providers and
ensuring that clinicians have access to relevant medical histories. Generative AI may
improve our ability to interpret and connect clinical information, but it cannot reliably
reconstruct information that was never captured, is inaccessible or is inaccurately
represented. This challenge has direct implications for device safety.

A systematic review by Li and colleagues (2022) examined the relationship between
electronic health record interoperability and healthcare safety and quality. Although the
evidence was heterogeneous, the review identified potential benefits related to medication
safety, patient safety and care delivery. Research evaluating generative AI for clinical
summarization has also documented hallucinations, omissions and other clinically
relevant errors, reinforcing the importance of source fidelity and information quality (Asgari
et al., 2025). Together, these findings suggest that device evaluation should consider not
only whether a model can generate an appropriate response when provided with complete
information, but how it behaves when the clinical information available to it is incomplete,
inconsistent, outdated or contradictory.

I encourage the FDA to consider the following:

• Data quality and completeness: Evaluate device behavior when relevant clinical
information is missing, conflicting, outdated or unavailable.
• Source traceability: Assess whether safety-critical outputs can be traced to their
underlying clinical information when necessary for the device's intended use.
• Interoperability-related risks: Include realistic scenarios involving fragmented
records, conflicting medication lists and information exchanged across systems.
 • Foundation-model documentation: Consider expectations for documenting
relevant data provenance, known limitations, model versions, evaluation methods
and update procedures.
• Change management: Clarify responsibilities for detecting, evaluating,
communicating and responding to safety-relevant changes in third-party foundation
models and connected information sources.

The FDA's exploration of Foundation Model Device Master Files presents a particularly
valuable opportunity to improve transparency while recognizing that device manufacturers
may not control every aspect of the foundation models upon which their products depend.
There are legitimate concerns about proprietary information, patient privacy, the cost of
documentation and the practical limitations of data availability. A proportionate
framework could distinguish between information provided confidentially to regulators and
information necessary for clinicians and patients to use a device safely.

More broadly, I believe patients should have meaningful access to and agency over their
health information. Questions of patient data rights and interoperability extend beyond the
FDA's medical-device jurisdiction, but the safety consequences of fragmented or
unreliable clinical information fall squarely within the Agency's consideration of device
performance. A clinically capable AI system must be evaluated not only for how well it
processes information, but also for how safely it responds when the information
foundation is incomplete or unreliable.

Conclusion
I appreciate the FDA's willingness to explore regulatory approaches that account for the
distinctive capabilities and risks of generative AI-enabled medical devices. A competencybased framework offers an opportunity to move beyond static measures of technical
performance and consider whether these technologies function safely and effectively in
the clinical environments for which they are intended.

I encourage the Agency to incorporate four considerations into its ongoing work:

1. Evaluate patient agency, comprehension and appropriate use as dimensions of
patient-facing device safety.
2. Require risk-proportionate evidence of equitable performance, including
representative evaluation and transparent reporting of subgroup limitations.
3. Incorporate relevant nursing and interdisciplinary expertise into competency
development, clinical confirmation and postmarket monitoring.
 4. Address data quality, provenance, source traceability, interoperability-related risks
and accountability for changes to foundation models.

Generative AI has the potential to make healthcare information more accessible, clinical
workflows more effective and care more responsive to the needs of patients. Realizing that
potential requires a regulatory framework that recognizes a fundamental distinction that
technical capability is not the same as clinical competence. Clinical competence must
ultimately be demonstrated through safe, equitable, understandable and accountable
performance in the real environments where patients receive care.

Thank you for considering these comments.

Respectfully,

M. Rubinstein

MSN Candidate, Johns Hopkins School of Nursing

Submitted in an individual capacity

References
Asgari, E., Montaña-Brown, N., Dubois, M., Khalil, S., Balloch, J., Au Yeung, J., & Pimenta,
D. (2025). A framework to assess clinical safety and hallucination rates of LLMs for
medical text summarisation. npj Digital Medicine, 8, Article 274.
https://doi.org/10.1038/s41746-025-01670-7

Kumar, A., Wang, H., Muir, K. W., Mishra, V., & Engelhard, M. (2025). A cross-sectional
study of GPT-4–based plain language translation of clinical notes to improve patient
comprehension of disease course and management. NEJM AI, 2(2).
https://doi.org/10.1056/AIoa2400402

Li, E., Clarke, J., Ashrafian, H., Darzi, A., & Neves, A. L. (2022). The impact of electronic
health record interoperability on safety and quality of care in high-income countries:
Systematic review. Journal of Medical Internet Research, 24(9), e38144.
https://doi.org/10.2196/38144
Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in
an algorithm used to manage the health of populations. Science, 366(6464), 447–453.
https://doi.org/10.1126/science.aax2342

Omar, M., Soffer, S., Agbareia, R., Bragazzi, N. L., Apakama, D. U., Horowitz, C. R.,
Charney, A. W., Freeman, R., Kummer, B., Glicksberg, B. S., Nadkarni, G. N., & Klang, E.
(2025). Sociodemographic biases in medical decision making by large language models.
Nature Medicine, 31, 1873–1881. https://doi.org/10.1038/s41591-025-03626-6

Stanceski, K., Zhong, S., Zhang, X., Khadra, S., Tracy, M., Koria, L., Lo, S., Naganathan, V.,
Kim, J., Dunn, A. G., & Ayre, J. (2024). The quality and safety of using generative AI to
produce patient-centred discharge instructions. npj Digital Medicine, 7, Article 329.
https://doi.org/10.1038/s41746-024-01336-w

Turchioe, M. R., Pepingco, C., Lytle, K., & Austin, R. (2024). Nurses' roles in artificial
intelligence implementation: Results from a mixed-methods study. Studies in Health
Technology and Informatics, 315, 223–227. https://doi.org/10.3233/SHTI240140

Wieben, A. M., Alreshidi, B. G., Douthit, B. J., Sileo, M., Vyas, P., Steege, L., & GilmoreBykovskyi, A. (2025). Nurses' perceptions of the design, implementation, and adoption of
machine learning clinical decision support: A descriptive qualitative study. Journal of
Nursing Scholarship, 57(1), 82–94. https://doi.org/10.1111/jnu.13001

U.S. Food and Drug Administration. (2026). Considerations for the regulation of generative
AI-enabled medical devices: Discussion paper and request for feedback.
https://www.fda.gov/medical-devices/digital-health-center-excellence/considerationsregulation-generative-ai-enabled-medical-devices-discussion-paper-and-request

U.S. Food and Drug Administration, Health Canada, & Medicines and Healthcare products
Regulatory Agency. (2021). Good machine learning practice for medical device
development: Guiding principles. https://www.fda.gov/media/153486