RaySearch Laboratories AB (publ)

See attached file(s)

RAYSEARCH LABORATORIES AB’S FEEDBACK
RaySearch Laboratories respectfully submits the following comments on the discussion paper
"Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback" (Docket No. FDA-2026-N-7874).
Comments are organized according to the questions that were proposed in the Discussion Paper.

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 GenAIenabled 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 two proposed axes appear to provide a useful framework for characterizing the inherent clinical
risk of a GenAI-enabled function. Other factors, such as human oversight, traceability, explainability,
reversibility, and downstream safeguards, may be better considered as risk control measures that
influence residual risk and the associated evidentiary requirements, rather than as separate primary
dimensions of risk.

However, certain factors may be sufficiently fundamental to the likelihood or consequences of harm to
warrant explicit consideration as modifiers within the risk framework. For example, time pressure may
affect the opportunity for effective human intervention, while the ability of the user to independently
verify an output may affect the likelihood that an erroneous output results in harm. These factors
should therefore not necessarily be treated solely as post hoc mitigations.

While the proposed two-axis approach provides a valuable starting point, the framework does not
explicitly address how risk controls may reduce the risk associated with a GenAI-enabled function.
CDRH could consider defining categories or examples of risk controls that, when their effectiveness is
adequately demonstrated, may modify the risk characterization or corresponding evidentiary
expectations. For example, reversibility, independent downstream checks, and traceability may reduce
the likelihood or consequences of harm. Similarly, time pressure could either be introduced as an
explicit modifier or incorporated into the assessment of consequences and the opportunity for
intervention.

In particular, CDRH should consider the presence and effectiveness of deterministic, rule-based
controls when determining the type and extent of evidence appropriate for a GenAI-enabled device.
Such controls can constrain the consequences of uncertain or incorrect model outputs independently
of the probabilistic behavior of the underlying model. Examples include requiring user approval before
consequential actions are executed; enabling actions to be reversed, where appropriate; providing
links to source information to facilitate independent verification; restricting access to tools and actions
based on user roles; maintaining audit trails of user and system actions; and preventing operation with
model versions outside the manufacturer's validated configuration.

Where such controls demonstrably reduce the likelihood that an erroneous GenAI output will result in
harm, CDRH should consider them as part of the overall risk-control strategy when determining the
evidence necessary to support the intended use. Accordingly, evidentiary expectations should reflect
not only the intrinsic performance and uncertainty of the GenAI model, but also the extent to which the

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
overall system architecture constrains potentially harmful actions, enables detection of erroneous
outputs, supports verification, or allows effective human intervention before harm can occur.

2. CDRH seeks input on how to account for the spectrum of GenAI-enabled informational functions
that vary in the degree to which they direct a user to a particular action (i.e., between “non-directive”
and “action-directing”). What characteristics of an output—such as its wording, specificity,
personalization, or context—could be considered as modifiers of the risk associated with an
informational function, after accounting for the device’s overall functionality and intended use? What
additional information would provide manufacturers with sufficient clarity and predictability around risk
assessment for informational functions while recognizing that directiveness may exist along a
continuum rather than as a binary distinction?

Directiveness should not be assessed primarily based on the language used to present an output.
Rather, CDRH should consider a multidimensional assessment that includes four components:

• Personalization: whether the output is generic or individualized to a specific patient or user.
• Specificity: whether the output provides general concepts or recommends specific actions.
• Certainty: whether the output communicates possibilities or presents conclusions with a high
degree of confidence.
• Actionability: whether the output is informational or can be directly and immediately acted
upon.

Together, these factors may provide a more meaningful characterization of directiveness and its
potential impact on clinical decision-making than linguistic formulation alone.

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?

One potential benefit of GenAI-enabled devices is their ability to support users with different levels or
areas of training and expertise in performing tasks that might otherwise require a more specialized
healthcare professional. Accordingly, the intended user profile should be considered as part of the risk
classification, as the user's training, expertise, and ability to independently evaluate or verify GenAIgenerated outputs may influence the likelihood of harm.

In addition, intended users may include healthcare professionals and other clinical or technical roles
with different responsibilities and competencies, such as physicians, medical physicists, nurses, and
scheduling personnel. CDRH should therefore consider whether intended use and associated risk
should be characterized by user role, particularly where different roles are permitted to perform
different GenAI-enabled functions. Role-specific access controls could be used as a risk control
measure to restrict functions and actions to appropriately authorized users.

5. For multi-turn conversational GenAI-enabled devices that may migrate from providing “nondirective” information to “action-directing” information over the course of an exchange, how should risk
be assessed across realistic conversational trajectories? How could the intended use of such a device
be characterized when its behavior is emergent across a conversation?

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
The intended use should establish a defined boundary that the manufacturer enforces and verifies
through testing. Several approaches may be considered when determining the appropriate risk
classification of a conversational GenAI-enabled function:

• Risk could be assessed based on the highest-risk behavior that the system can reasonably
exhibit.
• Risk could be based primarily on the intended use, provided that the manufacturer
demonstrates that effective guardrails maintain system behavior within the defined scope.
• Risk could be assessed at the clinical workflow or task level, rather than at the level of the
GenAI model as a whole.
• Evaluation should consider representative conversational trajectories, rather than isolated
prompts, to assess how system behavior may evolve over a multi-turn interaction.
• The level of risk may change as a conversation evolves. The system could therefore apply
risk-proportionate safeguards dynamically when predefined boundaries or higher-risk
behaviors are detected.

A combination of these approaches may be appropriate. Risk classification should primarily reflect the
intended clinical workflow and claimed functionality, while manufacturers should demonstrate, through
testing of representative conversational trajectories, that the system remains within its defined scope
and appropriately handles attempts or circumstances that could move the interaction beyond that
scope.

For example, consider a conversational system intended to provide informational support that reliably
refuses to provide treatment recommendations but, under rare circumstances, generates actiondirecting advice. Several approaches could be considered for determining the appropriate risk
classification: classification based on the intended use, classification based on typical system
behavior, or classification based on the rare but highest-risk behavior the system can reasonably
exhibit. CDRH should clarify how such low-frequency but potentially higher-risk behavior should
influence risk classification and the corresponding evidentiary expectations.

Manufacturers could also implement a separate boundary-detection or supervisory mechanism to
identify when an interaction is transitioning from informational content toward higher-risk or actiondirecting behavior. Crossing a predefined boundary could trigger risk-proportionate controls, such as
refusal, escalation to an appropriately authorized user, additional verification, or application of stronger
safeguards. Such mechanisms could help maintain system behavior within the validated intended use
while addressing the dynamic nature of conversational GenAI systems.

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?

Manufacturers should characterize and assess both directions of error independently, as their clinical
consequences may differ. The assessment should consider relevant factors such as sensitivity,
specificity, and the benefit-risk balance associated with each type of error.

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
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?

Traditional validation approaches may not readily scale to the breadth and variability of GenAI-enabled
device outputs. The proposed competency-based approach may provide a suitable alternative or
complementary framework. However, the concept of “competency” should be more clearly defined,
including how competencies should be identified, measured, and demonstrated to support adequate
validation.

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?

A risk-based approach could be used to determine the type and extent of evidence necessary to
support a GenAI-enabled device, with evidentiary expectations increasing in proportion to the clinical
risk associated with the intended use. For example:

• Lower-risk functions: more limited benchmarking and clinical confirmation may be sufficient,
where appropriately justified.
• Higher-risk functions: more extensive benchmarking, greater clinical evidence, and more
robust postmarket monitoring may be warranted.

Such an approach would allow the depth and rigor of validation and postmarket monitoring to be
proportionate to the risk associated with the GenAI-enabled function, rather than applying uniform
evidentiary expectations across all applications.

9. Would a benchmarking structure such as the one described above be likely to provide adequate
evidence of clinical knowledge, analytic capabilities, safety behavior, communication, and
generalizability to support a reasonable assurance of safety and effectiveness? Are there elements
that are missing, redundant, or inappropriately categorized? Are there externally developed standards
that could be leveraged?

The proposed benchmarking framework, and the discussion paper more broadly, could more explicitly
address several factors that may influence the safety and effectiveness of GenAI-enabled devices:

• Explainability: whether the user can understand the basis for, and appropriately assess, the
generated output.
• Human factors: whether intended users can correctly interpret the output and take appropriate
action based on it.
• Workflow integration: how the GenAI-enabled function performs and is used within the
intended clinical workflow.

These factors may be important when determining whether benchmarking results adequately reflect
the performance and risks of a GenAI-enabled device under its intended conditions of use.

10. CDRH recognizes that publicly available benchmarking assets may be subject to data
contamination, saturation, and limited real-world representativeness. How should a sponsor establish
that performance on a given benchmark predicts safe and effective real-world behavior for the

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
device’s intended use? What evidence should support the construct validity of a benchmark used to
gate device evaluation, and what role should sponsor-developed benchmarks play given potential
concerns around independence and optimization to the test?

The predictive value of a benchmark should be established by demonstrating that it is representative
of the device’s intended users, clinical workflows, input characteristics, permitted actions, and
reasonably foreseeable failure modes. Sponsor-developed benchmarks may be appropriate when
they are based on representative historical data collected across multiple healthcare institutions and
clinical workflows, provided that the benchmark’s construction, coverage, acceptance criteria, and
limitations are transparently documented.

Benchmark validity should be assessed for the complete device workflow and the specific model
version deployed, rather than solely at the level of the underlying foundation model. The breadth of the
device’s permitted input space should also be considered when determining the appropriate extent of
benchmark testing. For example, a device operating on structured or otherwise bounded inputs may
allow more comprehensive test coverage than a device accepting unrestricted conversational input.

Where appropriate, sponsor-developed benchmarks should be supplemented by independent clinical
review and/or postmarket evidence to mitigate the risk of overfitting or optimization to the sponsor’s
own benchmark and to provide additional assurance that performance is maintained under
representative conditions of use.

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?

Clinical confirmation is an important component of the evaluation of GenAI-enabled devices.
Approaches to clinical confirmation should support:

• Scalability, given the breadth and variability of potential GenAI outputs.
• Evaluation of rare but clinically relevant events.
• Stress testing, including challenging or boundary conditions.

Postmarket monitoring can complement premarket clinical confirmation by providing evidence
regarding device performance under real-world conditions and helping identify behaviors or failure
modes that may not be fully represented during premarket evaluation.

The type and extent of clinical confirmation should be proportionate to the risk associated with the
intended use, and prospective clinical studies may not be necessary in all circumstances. Relevant
considerations should include the intended use, risk profile, level of autonomy, potential
consequences of erroneous outputs, and the opportunity for meaningful human review or intervention.
CDRH should consider flexible approaches that enable evaluation using representative real-world data
while appropriately protecting patient data.

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
One potential approach would be to allow certain evaluations to be conducted within a healthcare
institution's local environment, with only aggregate performance results transferred outside the
institution. Clinical confirmation could also include protocol-defined sampling and review of actual
device outputs within the healthcare institution, performed by qualified clinical personnel or, where
appropriately validated for the intended purpose, by a machine-based reviewer operating locally. The
manufacturer should prospectively define the evaluation protocol, acceptance criteria, sampling
methodology, and relevant failure modes.

CDRH should also clarify the circumstances under which clinical evidence generated outside the
United States may supplement U.S. clinical evidence. For competencies that are not inherently
dependent on country-specific characteristics, pooling U.S. and international evidence may provide
greater volume and diversity of data. U.S.-specific evidence could then focus on factors that may
materially affect performance in the U.S. healthcare environment, such as language, units of
measurement, document formats, clinical workflows, and standards of care.

12. How should sponsors achieve statistically meaningful performance measurement for GenAIenabled devices? Where synthetically generated inputs supplement real patient data, how should
sponsors account for differences between the synthetic and real-world distributions when estimating
performance, and under what conditions, if any, is it appropriate to combine benchmarking evidence
and clinical confirmation evidence to support a single performance estimate?

Statistically meaningful performance measurement may require evidence from multiple healthcare
institutions and geographic regions, particularly for less common inputs, failure modes, and patient
subgroups. CDRH should clarify when U.S. and non-U.S. evidence may be pooled and what evidence
is necessary to demonstrate that relevant differences between populations and clinical environments
do not materially affect the competency being evaluated. Where country-specific characteristics may
affect performance, such as language, terminology, units of measurement, document formats, clinical
workflows, or standards of care, these factors should be explicitly represented and analyzed.

Performance should be evaluated using sufficiently large and representative real-world datasets,
where appropriate. Longitudinal historical data generated through the same or closely related clinical
workflows may be particularly valuable because they can reflect the types of inputs, workflow
variations, and errors encountered in clinical practice. Such data may provide important evidence of
real-world representativeness. However, usability and human factors studies remain important for
evaluating how intended users interact with, interpret, and act upon GenAI-generated outputs.

Synthetic data may supplement real-world data, but sponsors should characterize relevant differences
between synthetic and real-world data distributions and assess their potential effect on performance
estimates. Sponsors should demonstrate that synthetic data are fit for the specific evaluation purpose,
for example by characterizing distributional similarity, evaluating whether performance trends are
comparable between synthetic and real-world datasets, and validating the synthetic data generation
process, as appropriate. Even where synthetic data cannot be justified for inclusion in an overall
performance estimate, they may still provide value for stress testing, robustness testing, and
evaluation of rare or challenging scenarios.

Benchmarking, clinical confirmation, synthetic data, and real-world evidence should only be combined
into an overall performance estimate when the populations, endpoints, metrics, and data distributions

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
are sufficiently comparable to support such pooling. Otherwise, results should be reported separately
to preserve interpretability and allow meaningful assessment of potential performance differences.

For rare scenarios in which statistically conclusive testing is impractical, targeted benchmarking,
transparent characterization of uncertainty, appropriate risk controls, and expert clinical review may
provide complementary evidence. CDRH should consider providing guidance on how such
complementary evidence can be used when conventional statistical approaches are not feasible due
to the rarity of the relevant scenario.

Finally, the use of trained patient actors, as discussed in the paper, may present limitations that
warrant further consideration. Actor-based evaluations may be susceptible to bias and may not
adequately represent the breadth and variability of the real-world input space. CDRH should therefore
clarify the role of trained patient actors within the overall evidence framework and whether additional
evidence may be necessary to demonstrate that performance observed in actor-based evaluations is
representative of performance under real-world conditions.

13. For which clinical domains, device functions, or subpopulations is synthetic data particularly wellsuited, or particularly inadequate, as a supplement to real world evidence? What safeguards would
mitigate the risk that synthetic data generated by models of the same class as the device under
evaluation reproduces the very performance gaps the evaluation is intended to detect, particularly for
underrepresented subgroups?

Synthetic data may be particularly valuable for evaluating edge cases and rare but clinically relevant
scenarios, as a supplement to real-world evidence. However, CDRH should clarify the documentation
and evidence necessary to demonstrate that synthetic data are fit for the intended evaluation purpose.

Where synthetic data are intended to represent real-world inputs, sponsors should characterize the
data-generation methodology, assess comparability with relevant real-world distributions, demonstrate
coverage of clinically relevant characteristics and subpopulations, and identify known limitations.
Particular caution may be warranted when synthetic data are generated using models that share
characteristics, training sources, or potential failure modes with the GenAI-enabled device under
evaluation. In such cases, the synthetic data could reproduce, mask, or fail to expose the same
performance limitations that the evaluation is intended to detect. Independent real-world data therefore
remain important for identifying potential performance gaps, particularly for underrepresented
subgroups and clinically consequential scenarios.

Synthetic data may also be particularly useful for deliberate challenge and stress testing. Test
datasets could incorporate intentionally introduced errors, inconsistencies, rare events, and boundary
conditions, such as incorrect patient-document associations, implausible laboratory values, conflicting
information, or incomplete inputs. Such datasets would not necessarily need to reproduce the realworld prevalence of these conditions, because their purpose is to systematically evaluate whether the
device can detect, appropriately respond to, or safely handle reasonably foreseeable abnormal
conditions.

Accordingly, evidence derived from representative synthetic data, deliberate challenge datasets, and
real-world clinical data should be clearly distinguished, as each serves a different evaluation purpose
and provides different evidence regarding the performance, robustness, and safety of a GenAIenabled device.

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
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?

Where feasible, individual output elements should be objectively verifiable against an authoritative
source, without necessarily requiring expert adjudication or comparison with clinician performance.
Examples may include dose levels, medication names, patient weight, or other information that can be
directly verified against source data. Expert adjudication may instead be appropriate for more openended aspects of performance, such as completeness, relevance, and degree of directiveness.

More generally, the concept of “qualification level” should be more clearly defined when selecting
performance comparators and establishing acceptance criteria. One practical approach would be to
consider the qualifications, training, and expertise of the professional who would ordinarily perform the
relevant task in the absence of the GenAI-enabled function.

The appropriate performance comparator should also reflect the intended purpose and claimed benefit
of the GenAI-enabled function. A fundamental distinction is whether the function is intended primarily
to improve efficiency, for example by enabling the device or a less specialized user to perform a task
that would otherwise require a more specialized professional, or to improve the quality of performance
through augmentation, for example by enabling a physician working with the GenAI-enabled device to
achieve better performance than the physician alone.

Accordingly, CDRH should consider the intended user, clinical workflow, degree of autonomy, and
claimed benefit when determining appropriate performance comparators and acceptance criteria,
including whether performance should be evaluated for the device independently, for the human-AI
team, or relative to the professional who would ordinarily perform the task.

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?

Regardless of the evaluation approach used, there is a need to establish objective and clinically
meaningful performance thresholds for GenAI-enabled medical devices.

The performance comparator should reflect the care or clinical workflow that would realistically occur
in the absence of the device. Where a GenAI-enabled function automates or supports an existing
manual activity, historical human-performance data, including the frequency and clinical significance of
errors occurring within the existing workflow, may provide an appropriate baseline. CDRH should
clarify when demonstration of non-inferiority or improvement relative to the existing standard of care or
workflow may be sufficient, rather than comparison against an error-free theoretical standard.

For rare workflows, inputs, or failure modes where statistically conclusive premarket evaluation is
impractical, a risk-based approach to residual uncertainty may be appropriate. Sponsors could provide

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
targeted premarket evidence, characterize the remaining uncertainty and its potential clinical
significance, implement appropriate risk controls, and use postmarket monitoring to generate
additional evidence regarding performance under real-world conditions.

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?

A risk-based approach to the balance between premarket evidence and postmarket monitoring is
appropriate. Lower-risk GenAI-enabled functions may reasonably permit greater residual uncertainty
at the time of market authorization, whereas higher-risk functions should generally require stronger
premarket evidence.

Greater reliance on postmarket monitoring may also be appropriate for rare but clinically relevant
workflows or scenarios where sufficient premarket data cannot reasonably be accumulated to
generate statistically conclusive performance estimates. Such an approach should be considered
where the principal failure modes are adequately characterized, appropriate risk controls are
implemented, initial performance is acceptable relative to the existing standard of care or relevant
human-performance baseline, and the manufacturer can obtain sufficiently timely and reliable
postmarket data to detect, characterize, and investigate performance issues.

Conversely, reduced premarket evidence should generally not be appropriate where failures could
result in severe or irreversible harm, where the device operates autonomously without meaningful
human review or intervention, or where reliable postmarket performance signals cannot be obtained in
a timely manner.

CDRH should consider providing further clarity on the circumstances under which postmarket
evidence may appropriately address residual premarket uncertainty, including the factors that should
determine the acceptable balance between premarket evidence, risk controls, and postmarket
monitoring.

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?

An event-based approach to reassessment may be preferable to reassessment at fixed intervals,
provided that the triggering events are comprehensive, clearly defined, and capable of being reliably
detected. Relevant triggers could be categorized as follows:

Planned change triggers:
• Introduction of a new model version.
• Expansion or modification of the intended use or intended user population.
• Introduction or replacement of an underlying foundation or frontier model.

Performance-related triggers:

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
 • Detection of performance or data drift.
• A meaningful increase or trend in complaints or identified errors.
• Identification of an adverse event or other safety-related signal.

Environmental or use-related triggers:
• Deployment in a new or materially different patient population.
• Introduction of new input modalities or other material changes to the operating or clinical
environment.

The effectiveness of an event-based reassessment approach depends on the manufacturer's ability to
detect relevant changes and performance signals. GenAI-enabled devices should therefore support
the structured capture of appropriate information regarding, for example, sessions, model versions,
clinical workflows, output review, corrections, and identified errors, subject to applicable privacy and
data-protection requirements. Expectations regarding access to privacy-protected performance and
feedback data when such information is necessary for manufacturers to perform effective postmarket
monitoring and determine whether reassessment is warranted could be better clarified.

The loss or absence of sufficient postmarket feedback, logging, traceability, or model-version
information could itself constitute a reassessment trigger, as insufficient information may prevent the
manufacturer from adequately evaluating whether the device continues to perform safely and
effectively within its intended use.

20. Please comment on whether the proposed approaches to postmarket monitoring can be facilitated
by machine-based supervisory agents. What considerations, including the evaluation and reliability of
the supervisory agent itself, should CDRH take into account for such an approach?

Machine-based supervisory agents could facilitate postmarket monitoring of GenAI-enabled medical
devices, particularly by providing an additional mechanism for detecting outputs or behaviors that fall
outside predefined intended-use or operating boundaries. Such agents may be particularly valuable in
supporting an event-based postmarket monitoring approach, as they could enable continuous
surveillance for predefined trigger conditions rather than reliance solely on periodic reassessment.

However, the use of a supervisory agent should not merely transfer the reliability challenge from the
medical device to another AI system. The intended use, predefined operating boundaries, and known
or reasonably foreseeable failure modes of the GenAI-enabled device should inform what the
supervisory agent is expected to monitor and the criteria against which its effectiveness is evaluated.

For example, an independent supervisory agent operating with a separate context from the primary
model could evaluate selected device outputs against predefined acceptance criteria without relying
on the same conversational context or reasoning process used to generate the original output. This
approach could provide an additional layer of oversight and facilitate identification of potential
boundary violations, anomalous outputs, or other predefined performance signals.

However, the independence, reliability, and effectiveness of the supervisory mechanism should be
demonstrated rather than assumed, particularly where it is relied upon as a risk control or as a source
of postmarket safety and performance signals. CDRH should consider clarifying the evidence
necessary to support reliance on machine-based supervisory agents for these purposes.

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
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?

The allocation of roles and responsibilities could generally remain consistent with existing medical
device oversight, with relevant stakeholders contributing information and evidence while the
manufacturer retains responsibility for the safety, effectiveness, and regulatory compliance of the
device.

Potential stakeholder roles could include:

Clinicians
• Report device-related issues and potential safety or performance concerns.
• Participate in clinical review and evaluation activities.
• Contribute to the generation of real-world performance evidence.

Professional societies

• Update relevant clinical practice guidelines as knowledge and experience with GenAI-enabled
technologies evolve.
• Identify changes in clinical practice and standards of care that may affect device performance
or benefit-risk considerations.

Healthcare institutions

• Collect relevant device-use and performance information.
• Identify workflow-related issues and changes in the clinical environment that may affect device
performance.
• Facilitate appropriate feedback regarding real-world device performance.

Manufacturers

• Conduct ongoing safety and performance monitoring.
• Investigate identified issues and implement corrective and preventive actions, as appropriate.
• Fulfill applicable regulatory reporting obligations.
• Maintain and reassess the device's benefit-risk profile throughout its lifecycle.

CDRH could consider clarifying expectations for information sharing among these stakeholders,
particularly where access to real-world performance information is necessary for manufacturers to
fulfill their postmarket monitoring responsibilities.

22. CDRH envisions that a manufacturer’s premarket competency-based assessment might serve as
a baseline against which post-deployment modifications could be re-evaluated. How might the extent
of re-benchmarking or other evidence be scaled to the nature and expected impact of a given
modification? For example, are there categories of change that might not significantly affect safety or
effectiveness of a GenAI-enabled device, or might be appropriately managed within a sponsor’s
quality management system versus requiring FDA premarket review and authorization before
implementation? Are there categories of changes that might be appropriate for inclusion in a PCCP?

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
Different categories of changes or modifications to GenAI-enabled medical devices may warrant
different regulatory responses during the postmarket phase. CDRH should consider a risk-based,
tiered approach based on the nature of the change and its potential impact on device safety,
effectiveness, and intended use. For example:

• Minimal-impact changes: Changes with no meaningful impact on safety, effectiveness, or
intended use could generally be managed within the manufacturer's quality management
system without additional FDA review.
• Moderate-impact changes: Changes with a potentially meaningful but bounded impact on
device performance could potentially be managed within an authorized Predetermined
Change Control Plan (PCCP), where applicable, with appropriate re-evaluation or
benchmarking against the established performance baseline.
• Major-impact changes: Changes that could materially affect the device's intended use, risk
profile, or degree of autonomy may warrant additional regulatory review or a new marketing
submission, as applicable. Examples may include a new or significantly modified intended
use, replacement of the underlying foundation model, or introduction of new autonomous
functionality.

CDRH could provide clear criteria for distinguishing among these categories, including the extent of
revalidation, re-benchmarking, and regulatory review expected for each type of modification.

23. How might PCCP concepts or other change-control approaches be adapted for GenAI-enabled
devices when the nature or scope of future modifications cannot be fully prespecified?

Rather than focusing exclusively on the pre-authorization of specific individual modifications, CDRH
could consider whether a process-based approach to change control may be appropriate for certain
GenAI-enabled medical devices. Under such an approach, modifications that remain within predefined
competency and risk boundaries and successfully meet established performance criteria could
potentially be implemented under an authorized Predetermined Change Control Plan (PCCP), where
applicable.

A PCCP could prospectively define, without necessarily identifying every individual future modification:

• Categories and boundaries of permitted changes.
• Evaluation and validation procedures.
• Predefined acceptance criteria.
• Benchmark suites and other evidence used to demonstrate continued acceptable
performance.

An additional consideration is model version identification and traceability. Validation evidence for a
GenAI-enabled device function, including benchmarking, clinical confirmation, and postmarket
performance evidence, is associated with the model configuration evaluated by the manufacturer. A
challenge may arise where an external model provider modifies model behavior without changing the
model identifier available to the medical device manufacturer. Such changes could affect the
continued applicability of the manufacturer's validation without being readily detectable.

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
CDRH should therefore consider establishing expectations for unambiguous identification of
foundation-model versions used in medical devices. For example:

• Each model response could provide a unique identifier corresponding to the model
configuration being served. Possibility of introducing version metadata together with each
response from the GenAI should be taken into consideration.
• An intentional change that could materially affect model behavior, including changes to model
weights or numerical representation, provider-controlled input or context, decoding methods,
or output post-processing, could require a new identifier.

Minor numerical differences between hardware platforms would not be covered because the models
themselves are already non-deterministic.

Such an approach would not necessarily require disclosure of proprietary training data or model
internals. Instead, it would provide manufacturers with a means to determine whether the model
configuration being served remains within the configuration they have validated. Manufacturers could
implement controls to prevent use of model configurations outside the validated set and perform
appropriate re-evaluation before enabling a new configuration, including through an authorized PCCP
where applicable.

Under this framework, the model provider would be responsible for accurate and reliable modelversion identification, while the medical device manufacturer would retain responsibility for the safety
and effectiveness of the medical device using that model.

More broadly, CDRH could consider establishing requirements applicable to foundation models
incorporated into or relied upon by medical devices, rather than directly regulating foundation-model
providers solely because they supply AI models. A standardized model-identification and changenotification framework could provide a scalable mechanism for supporting medical device
manufacturers' validation and change-control obligations and may reduce reliance on manufacturerspecific contractual arrangements with individual model providers.

24. For devices built on third-party foundation models, changes to the underlying model may be
initiated by the third-party model developer rather than the device manufacturer. How can a
manufacturer detect, evaluate, and respond to such changes in a timely manner? What
mechanisms—for example, contractual, technical, or through a PCCP—could provide reasonable
assurance that third-party developer-initiated changes do not compromise the safety or effectiveness
of the device?

Technical controls, model-version stability, change detection, and continuous monitoring should be
considered essential elements of postmarket monitoring when a GenAI-enabled medical device relies
on a third-party foundation model. As discussed in the response to Question 23, CDRH should also
consider establishing clear expectations for model identification and change control for foundation
models used in medical devices.

In addition to providing a stable and technically verifiable model-version identifier, foundation model
providers should provide machine-readable notification of relevant model changes, deprecations, and
planned withdrawal of model versions used in validated medical device configurations. A model

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
identifier should not remain unchanged when provider-controlled modifications could materially affect
the behavior or performance of the medical device.

Manufacturers should have mechanisms to detect material changes to the foundation model before, or
at the latest when, the modified model becomes available for use by the device. This would enable
manufacturers to determine whether the modified model remains within the validated configuration
and whether additional evaluation, risk assessment, or other change-control activities are necessary
before continued use.

25. Would voluntary Foundation Model MAFs be practical for foundation model developers to provide
and to keep current, and what content might they include? Given that participation would be voluntary
and that model developers may have limited incentive to disclose safety-relevant information, what
would make such a program sufficiently useful for premarket review? Are there alternative
mechanisms CDRH should consider for obtaining information about underlying foundation models?

Up-to-date, safety-relevant information regarding foundation models used in medical devices should
be made available to manufacturers where such information is necessary to support premarket
evaluation, risk management, and regulatory submissions. To be relied upon as part of a medical
device submission, this information should be sufficiently reliable, current, and maintained throughout
the relevant model lifecycle.

A purely voluntary (and not mandatory) approach may not provide manufacturers with sufficient
assurance that the necessary information will be consistently available or updated. CDRH should
therefore consider establishing clear requirements or expectations regarding the safety-relevant
information that should accompany foundation models intended for use in medical devices, including
expectations for maintaining and communicating updates to that information.

26. Are there additional considerations that inform the premarket and postmarket evaluation of agentic
GenAI-enabled devices, beyond those applicable to non agentic GenAI-enabled devices? How should
the elevated risk associated with autonomous multi-step action, tool use, and reduced opportunity for
human review be reflected in acceptance criteria and oversight?

Agentic GenAI-enabled devices may warrant additional considerations because autonomy introduces
risks not only from individual generated outputs, but also from the sequence of actions that an agent is
authorized to initiate and execute independently.

A key distinction should be made between actions that modify a work product, such as adjusting
objective weights during an optimization process, and actions that modify the clinical intent or
acceptance criteria against which that work product is evaluated, such as changing clinical goals,
regions of interest (ROIs), prescriptions, constraints, or other clinically defined parameters. An agent
should not be permitted to autonomously modify the criteria against which its own output will
subsequently be evaluated without appropriate authorization and oversight. These boundaries should
be explicitly defined, controlled, and verified.

Premarket evaluation should therefore consider the agent's scope of authority, permitted tools and
actions, ability to modify governing parameters, potential for failures to propagate across multiple
steps, and opportunities for meaningful human review or intervention. The evidence, acceptance
criteria, and oversight mechanisms should be commensurate with the degree of autonomy and the

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.
potential consequences of an incorrect action. Activities that affect clinical intent or acceptance criteria
may warrant stronger controls and explicit human approval than actions performed within predefined
and validated boundaries.

A distinction should also be made between agentic functionality operating directly within a clinical
workflow and agent-assisted development performed in a non-clinical environment (test environments
not used to treat patients). Agent-assisted development may represent a different risk profile where
the agent supports the generation or modification of a work product, but that output remains subject to
the manufacturer's established verification and validation processes and must independently satisfy
predefined acceptance criteria before clinical use.

Postmarket oversight should additionally evaluate whether the agent continues to operate within its
authorized boundaries and whether changes to its behavior, available tools, operating environment, or
underlying models could affect the continued validity of previously established risk controls, validation
evidence, or acceptance criteria.

Finally, an agent performing a bounded task within a validated workflow should not necessarily be
evaluated in the same manner as an open-ended conversational agent. The evaluation should reflect
the defined input space, permitted actions, objective acceptance criteria, downstream verification,
opportunities for human intervention, and the agent's ability to modify clinical intent or governing
parameters. However, the fact that an agent operates internally and is not directly visible to the user
should not, by itself, reduce the evidence necessary to address risks arising from its actions.

More broadly, agentic GenAI-enabled devices may introduce additional challenges related to reduced
predictability when systems are capable of generating novel solutions or reasoning paths, incorrect or
hallucinated outputs, cybersecurity threats involving manipulation of AI models or training data,
increased complexity of verification and validation, and allocation of responsibilities when generalpurpose AI models are incorporated into medical devices. These considerations may become
increasingly important as the degree of autonomy, number of available tools, and potential
consequences of autonomous actions increase.

At the same time, these risks might be addressed through a risk-based and proportionate regulatory
approach rather than through restrictions on agentic AI as a technology class.

RaySearch Laboratories AB (publ)
Reg. no. 556322-6157 Visitors: Eugeniavägen 18 C The registered office
P.O. Box 45169 info@raysearchlabs.com of the Board of Directors
SE-104 30 Stockholm www.raysearchlabs.com is in the Municipality
Sweden Phone +46 (0)8 510 530 00 of Stockholm.