Johns Hopkins Medicine (Sara Murray)

Please see comment from Johns Hopkins Medicine attached.

Johns Hopkins Medicine
720 Rutland Ave.
Baltimore, MD 21250
410-955-5000

October 19, 2026

Dockets Management Staff
U.S. Food and Drug Administration
5630 Fishers Lane, Room 1061
Rockville, Maryland 20852

To the Center for Devices and Radiological Health:

RE: Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper
and Request for Feedback, Docket No. FDA-2026-N-7874-0001

Johns Hopkins Medicine (JHM) appreciates the opportunity to comment on the Center for Devices and
Radiological Health’s (CDRH) Discussion Paper. We commend CDRH for seeking a regulatory
approach that protects patients while enabling responsible innovation and implementation. Regulatory
ambiguity for developers becomes de facto delegation of responsibility and associated liability to
deployers, and consequently a deterrent to adoption. Clear, proportionate regulation can give developers
and healthcare organizations confidence in the standards governing these tools and allow beneficial
technologies to reach clinical practice more quickly. JHM would welcome the opportunity to serve as a
partner to CDRH as this framework develops.

We believe the proposed two-axis framework offers a strong basis for proportionate regulation because
it considers the nature and independence of a device’s activity and the consequences of an incorrect
output. In that context, we recommend that CDRH treat healthcare professional supervision as a risk
modifier only when the intended user can reasonably identify a clinically significant error and intervene
before harm occurs. A clinician’s presence in a workflow should not, by itself, lower a device’s risk
classification. Conversely, CDRH should not require human review when a device has already
undergone rigorous evaluation for autonomous use.

We also commend CDRH’s emphasis on competency-based evaluation through benchmarking followed
by clinical confirmation, and its focus on assessing the final device in the configuration in which it will
be deployed. That approach should improve the relevance and reliability of premarket review. Likewise,
we encourage adoption of CDRH’s total-product-lifecycle approach, including ongoing monitoring,
reassessment after material changes, and continued sponsor responsibility throughout the product
lifecycle. To make this framework effective, CDRH should define the device’s clinical role and operating
boundaries clearly enough to support meaningful evaluation and monitoring. At the same time, CDRH
should recognize that those boundaries will differ across use cases, such as a narrowly defined
prescribing function and a broader patient-facing conversational tool. A consistent regulatory
framework will also help health systems adopt useful technologies without having to recreate validation
and monitoring infrastructure independently.

Our principal recommendation is that CDRH more clearly assign responsibility to the parties that
control each risk and are best able to prevent harm. Sponsors should remain responsible for the product
and for any claims about its performance. Healthcare organizations should remain responsible for
implementation decisions within their control. When both product-related and deployment-related
failures contribute to harm, the regulatory framework should reflect shared responsibility. Although
FDA may not determine civil liability, clear regulatory responsibilities would help support more
equitable downstream responsibility determinations. We therefore recommend that CDRH consider a
Lifecycle Responsibility and Accountability Plan to define these roles throughout the product lifecycle
and clarify that healthcare organizations can rely on premarket clearance or approval and associated
performance to deploy a device within its authorized use. We also recommend that CDRH pair greater
premarket uncertainty with stronger sponsor obligations for postmarket monitoring, disclosure,
correction, and remediation.

Finally, we recommend that CDRH address sponsor accountability for post-deployment monitoring
alongside the resources health systems need for clinical oversight and ongoing monitoring. Safe
adoption of these tools should not depend on the size of the organization. Without practical support for
the time, expertise, and financial resources that oversight requires, smaller and regional healthcare
providers may be unable to implement these technologies effectively, limiting access for the patients
they serve.

The detailed responses below expand on these recommendations, including protections for reasonable
clinical reliance and good-faith safety reporting.

Detailed Responses to the Discussion Paper
The comments below are organized according to the principal sections of CDRH’s Discussion Paper.
JHM focuses on the issues where its experience as both a developer and deployer of healthcare AI may
be most useful to CDRH.

Assessment of Risk

JHM supports the two-axis framework under consideration as a useful basis for proportionate
regulation. The nature and independence of the device activity and the consequences of relying on an
incorrect output are central to risk. Detectability of an error, reversibility of the resulting action, time
available for review, and downstream safeguards should inform the assessment as modifiers rather than
additional rigid axes. In addition, probability of consequences should be considered – overall harm
should be based on frequency in addition to degree of consequence.

CDRH should also distinguish between a generative AI function that serves as an interface to an
established deterministic clinical protocol and one that generates clinical reasoning or actions through
non-deterministic model behavior. A function may warrant lower presumptive risk when the validated
protocol controls the clinical decision, the generative layer cannot select or alter the protocol or
populate safety-critical variables without validated controls, and the interface faithfully communicates
the protocol output. Describing a system as protocol based should not reduce its risk if the generative
component can depart from the protocol or introduce clinically meaningful omissions or errors.
Healthcare professional (HCP) supervision should reduce a device’s presumptive risk only when that
supervision is demonstrably effective. The presence of an HCP in the workflow does not establish that
the intended user has the information, time, expertise, or practical ability to identify an incorrect output.
This is particularly relevant to summarization and synthesis, where using the tool may reduce sourcelevel review and make omissions or inaccuracies difficult to detect. For all devices, the relevant inquiry is
whether the intended user can evaluate the output under actual conditions of use – and labeling should
reflect this intended use. For a subset of devices, labeled intended use may be specific to users with
specialty clinical training. In these cases, a generalist HCP without such training would be “off label,”
and presumptively higher risk.

When HCP supervision is relied upon to moderate risk or reduce evidentiary requirements, the sponsor
should demonstrate that clinically significant errors are reasonably detectable under actual conditions of
use. The sponsor should make the intended use, known limitations, and uncertainty clear at the point of
use and provide results for clinically relevant subgroups, including those underrepresented in training
data. Where appropriate, the user should be able to trace the output to supporting information. The user
must also be able to override or stop the device when intervention is needed. HCP supervision should
not transfer responsibility for latent product failures that a reasonable user could not detect. Absent the
information described above, reliance on HCP supervision as a risk moderator would shift patient safety
risk to clinician deployers and could chill adoption of these devices.

JHM also encourages CDRH to identify circumstances in which evidence can support autonomous use
rather than making human review a permanent requirement, including clinical-adjacent but immediately
actionable areas such as language access and translation. Machine translation of written clinical content
is one area where further interagency coordination would be valuable. These tools may provide
substantial benefit when qualified language services are not immediately available, yet current
Department of Health and Human Services (HHS) language-access rules require qualified human review
of machine-translated written material in certain high-stakes circumstances (45 C.F.R. § 92.201(c)(3)).
Because a translation function may not itself fall within FDA’s device jurisdiction, CDRH cannot
resolve this issue alone. FDA should work across HHS, including with the Office of Civil Rights (OCR)
and the Centers for Medicare and Medicaid Services (CMS), where financing or implementation is
implicated, to develop an evidence-based pathway for autonomous medical translation. For functions
within FDA’s device jurisdiction, that pathway should address performance by language and clinical
context, high-risk communications, uncertainty and escalation, and postmarket monitoring.

Risk assessment should reflect the output’s reasonably foreseeable influence on care rather than rely on
a binary distinction between informational and action-directing language. A summary can influence the
next clinical decision as strongly as an explicit recommendation. Patient-facing functions may create
different risks because a patient may act without clinical intermediation, but they should not be
presumed higher risk in every case. A function intended to diagnose disease, prescribe treatment, or take
another clinical action should be presumed to warrant review within the medical-device framework
unless a statutory exclusion applies. Such a function should generally be evaluated within a defined
condition, population, and scope of activity. A broader conversational tool may not permit the same
narrow scope, but it should still have a clear clinical role, prohibited actions, escalation requirements,
and safety-critical behaviors.
Multi-turn devices should be evaluated across complete conversations for compounding errors and
movement from information to recommendation or action. We suggest multi-turn devices be evaluated
against their broadest possible function, including unbounded inputs and outputs that test guardrails,
scope of practice, and relevant clinical conditions. Care-escalation testing should separately measure
under-escalation, which may delay care or create false reassurance, and over-escalation, which may
interrupt care plans, increase utilization and cost, worsen access, or fragment care. Sponsors should
prespecify measures and acceptance thresholds for both.

Premarket Evaluation

JHM supports benchmarking followed by clinical confirmation as a reasonable response to the difficulty
of exhaustively testing devices with open-ended inputs and variable outputs. The approach appropriately
parallels clinical credentialing, in which structured assessment is followed by supervised practice before
greater independence is granted. Evaluation should focus on the final user-facing device in its intended
configuration, including its retrieval sources, prompts, guardrails, tools, interface, and workflow
integration, rather than on the foundation model in isolation. Competencies should be tied to the
device’s intended clinical role and operating boundaries rather than broad claims of clinician-like
capability.

JHM supports allowing sponsors to propose testing methods and acceptance criteria because they
understand their products and intended uses. The flexibility to define such methods and criteria should
be paired with independent demonstration of the methods’ clinical relevance and validity. Rigor should
scale with device autonomy and potential consequences, and higher-risk functions may warrant
sequestered datasets, independent adjudication, or testing by a qualified third party. Where a sponsor
defines the grounds on which its device is evaluated, the sponsor should bear increased responsibility,
including appropriate liability, if the device fails to perform as represented.

Clinical confirmation should reflect actual use. That includes the performance of the human-AI team
when the device is intended to support a clinician, multi-turn interactions for conversational devices,
and representative clinical workflows and populations. Prospective evaluation may be warranted when a
function acts autonomously, produces high-consequence outputs, or interacts with patients in ways that
are difficult to supervise; sponsors should justify when it is not used. The sponsor should explain why
the selected approach is sufficient and identify the uncertainty that remains for postmarket monitoring.

Performance comparators should match the task, condition, population, and intended use. Comparison
with a median clinician may provide useful evidence, but it should not by itself always establish
acceptability where current practice is variable or known to be suboptimal – especially where there is a
clearly defined standard of care. For a defined clinical function, consensus among qualified clinicians
applying the relevant standard of care may be a more appropriate comparator. Further, a device
intended to perform a defined function in a specialty, such as oncology, should be compared with
clinicians who have the expertise required for that function rather than with a generic generalist or
specialist category. An HCP trained only in internal medicine would not reasonably advise on
chemotherapy; GenAI-enabled devices should be compared to HCPs with relevant training and
expected scope of practice.
The care likely to occur without the device may also be relevant to benefit-risk assessment, particularly
where patients face long delays or lack access to specialists. This counterfactual is important but highly
dependent on local and regional conditions, including access, wait times, and available services, and it
may change as the technology itself becomes more widely available. It should not lower the minimum
safety standard for the product. If authorization or labeling relies on a counterfactual, the assumptions
should be explicit and supported by evidence, and CDRH should require reassessment when the
conditions underlying that determination materially change.

Finally, if FDA issues guidance, greater clarity on what types of products are excluded from such
regulation will also be helpful to accelerate adoption, including differentiating devices that fall under
CDS guidance as opposed to GenAI-enabled device guidance. Further, there has been differentiation at
the state level between AI devices developed only for internal health system use and those developed for
external distribution or commercialization. Clarity on whether future regulation intends to capture
internal AI tools, often developed for internally specific quality improvement initiatives, would be
helpful.

Postmarket Monitoring and Modifications

JHM supports an appropriate role for postmarket evidence combined with risk proportionate premarket
evaluation. Greater premarket uncertainty should be accepted only when the remaining uncertainty is
well characterized, and the product can be meaningfully monitored. The sponsor should have a
prespecified, adequately resourced plan to detect failures and act on them promptly, and should accept
correspondingly greater responsibility, including appropriate liability for product failures within its
control. This approach may be inappropriate for autonomous, difficult-to-supervise, or highconsequence functions whose failures are difficult to detect or reverse.

Any reduction in premarket evidence should be paired with greater sponsor responsibility for
postmarket monitoring, disclosure, correction, and remediation. Sponsors, who are most familiar with
the grounds on which the device was originally evaluated, are best positioned to perform product-wide
surveillance because they have access to the model architecture, relevant development information,
device and model versions, and performance signals across deployments. Robust postmarket monitoring
requires substantial infrastructure, technical expertise, and data access that most healthcare organizations
do not have. Further, sponsor-led monitoring is also necessary for consistency, especially in autonomous
and agentic devices. As GenAI-enabled devices become a meaningful part of the clinical workforce,
deployer-led monitoring would produce innumerable, inconsistent approaches, such that a risk detected
at one health system could go undetected at another despite using the same device in active clinical care.
Standardized monitoring anchored in the sponsor’s infrastructure and required as a condition of
marketing authorization would ensure consistency across deployments, allow monitoring to be updated
as new risks emerge, and support the rapid-cycle risk correction that increasingly autonomous devices
will require. Monitoring should include standardized clinical outcomes where appropriate, exposure
denominators, subgroup performance, repeated low-frequency failures, near misses, delayed care, and
over- and under-escalation. CDRH should also develop or designate a common mechanism for
reporting serious incidents and near misses so that signals can be aggregated across sites and acted upon
promptly.

JHM supports the shared monitoring ecosystem described in the Discussion Paper and urges CDRH to
clearly assign responsibilities. Sponsors should remain responsible for design, validation, performance
representations, known or reasonably discoverable limitations, product-wide surveillance, and safetyrelevant changes. Healthcare organizations should remain responsible for local implementation, material
configuration, user training, use outside authorized conditions, and response to safety notices. For an
autonomous function, a health system’s local credentialing role should include confining the device to its
authorized scope, limiting access to the systems necessary for that scope, and integrating it into an
appropriate clinical workflow. When product and deployment failures jointly cause harm, responsibility
should reflect each party’s contribution and capacity to prevent it.

CDRH should require a Lifecycle Responsibility and Accountability Plan that defines the device’s
intended clinical role and operating boundaries and assigns responsibility for monitoring, reporting,
investigation, modification, notification, suspension, and remediation. The plan should make clear which
regulatory responsibilities remain with the sponsor regardless of labeling or contractual arrangements.
Although FDA may not determine the ultimate allocation of civil liability, clear regulatory
responsibilities would support fairer allocation. FDA should also make clear that clearance or
authorization provides a reasonable basis for HCPs to rely on the device’s disclosed performance within
its authorized use. FDA and related agencies should work with Congress and other agencies on
additional protections for reasonable clinical reliance and mechanisms for allocating responsibility when
developers, manufacturers, deployers, and users jointly contribute to an injury.

Healthcare organizations conducting good-faith monitoring and reporting should receive appropriate
protection without excusing negligent implementation. To enable health systems to offer patient-facing
functions responsibly, HHS, including FDA and CMS as appropriate, should align regulatory and
reimbursement policy with control, protect health systems from primary responsibility for latent product
failures they could not reasonably detect or prevent, and provide sustainable support for the clinical
oversight, monitoring, and infrastructure required for safe deployment.

Reassessment after a modification should depend on the expected clinical effect of the change rather
than the size of the software change. Changes affecting clinical role, autonomy, data sources, retrieval
methods, guardrails, intended population, or the ability of users to review outputs may require
substantial re-evaluation. The device sponsor should remain responsible for the safety and effectiveness
of the marketed device when a change originates with a third-party foundation-model provider and
should maintain advance notice, version traceability, regression testing, rollback, and suspension
capabilities.

Foundation Models and Agentic Systems

Voluntary Foundation Model Device Master Files could improve consistency and reduce duplicative
review if they remain current and contain information relevant to healthcare use, including version
history, known limitations, failure modes, subgroup results, guardrails, audit capabilities, and changenotification commitments. A Master File should not constitute authorization of the foundation model
or reduce the device sponsor’s responsibility to evaluate the final configured device. Where a Master File
is unavailable, the sponsor should obtain equivalent safety-relevant information and change-notification
commitments through contractual or technical mechanisms.

Agentic systems require clear limits on the actions they may plan and execute, the tools and data they
may access, and the circumstances requiring confirmation or escalation. Evaluation should address
complete action sequences, accumulation of small errors, tool failures, auditability, and the ability to
interrupt or reverse an action. A system that diagnoses, prescribes, places orders, or controls another
device should be evaluated for each action and for the safety of the overall sequence. Reduced
opportunities for meaningful human review should result in stronger evidence, monitoring, and sponsor
accountability rather than nominal oversight checkpoints.

Thank you for considering our feedback, and JHM looks forward to continued engagement with CDRH
and other stakeholders on practical approaches that can support safe adoption across different care
settings.

Sincerely,

Sara Murray, MD, MAS
Vice President, Enterprise Chief Medical Informatics & AI Officer
Johns Hopkins Medicine

Deanna Hanisch, MS
Vice President and Chief Information Officer
The Johns Hopkins Health System Corporation

Brian Hasselfeld, MD
Executive Medical Director, Digital Health and Innovation
Johns Hopkins Medicine