Kestra Medical Technologies, Inc. (Beverly Magrane)

Comments on Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback, docket number FDA-2026-N-7874 from Kestra Medical Technologies, Inc. Please see attached.

KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

October 19, 2026
Submitted electronically via Regulations.gov
Dockets Management Staff (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Room 1061
Rockville, MD 20852
Re: FDA-2026-N-7874, Comments on “Considerations for the Regulation of Generative AIEnabled Medical Devices: Discussion Paper and Request for Feedback”

Kestra Medical Technologies, Inc. (Kestra) appreciates the opportunity to comment on the
Center for Devices and Radiological Health discussion paper, “Considerations for the
Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for
Feedback” (Discussion Paper). Kestra supports FDA’s effort to develop a risk-based, total
product life cycle framework that protects patients while enabling timely access to beneficial
medical technology.
Kestra develops and manufactures medical devices used in the care of patients at risk of
serious cardiac events. From that manufacturer perspective, Kestra supports the Discussion
Paper’s focus on intended use, device activity, consequences of incorrect output,
competency-based evaluation, clinical confirmation, postmarket monitoring, and change
control. Additional clarity, however, is needed so manufacturers can design studies,
establish acceptance criteria, maintain quality system controls, and implement postmarket
monitoring in a predictable and technically feasible manner.
Kestra recommends that FDA: (1) define concrete traceability and auditability expectations
for GenAI-enabled device functions, including the ability to reconstruct safety-relevant
recommendations, communications, and downstream actions; (2) address patient- and
provider-facing agentic workflows separately from traditional decision support, including
outreach, care coordination, escalation, handoff, and documentation risks; (3) establish
practical governance expectations for third-party foundation model changes, including
notification, versioning, degradation disclosure, validation support, rollback, and
compensating controls; (4) explicitly incorporate AI-specific cybersecurity and model supplychain risks into risk management and postmarket monitoring; and (5) apply benchmarking,
clinical confirmation, PCCPs, and postmarket obligations through a risk-proportionate total
product life cycle framework.

I. Comments on the Proposed Risk Framework
A. The Two-Axis Framework Is a Useful Starting Point but Is Not Sufficient by Itself
The proposed axes of device activity and consequence of reliance on an incorrect output
address two important risk drivers. FDA should retain these concepts as the foundation of
the framework. However, devices that appear similar on those two axes may present
materially different risks depending on whether the resulting action is reversible, whether the
error is detectable before harm occurs, whether a qualified user can intervene, and whether
the output can be traced to reliable primary source information.

Kestra Medical Technologies, Inc. | 1
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

FDA should therefore treat the two-axis matrix as an organizing tool, not a complete
classification mechanism. At minimum, the framework should incorporate modifiers that
manufacturers can apply during design control, risk management, validation planning, and
submission preparation. These modifiers could include reversibility of the resulting action;
time available for intervention; detectability of an erroneous output; availability and
independence of downstream safeguards; user population and clinical expertise; source
traceability; degree of personalization; cumulative risk across multi-turn interactions; and the
ability to audit and reconstruct safety-relevant recommendations, communications, and
actions.
FDA could incorporate these modifiers by asking sponsors to first place the device function
within the two-axis matrix, then justify evidentiary and control expectations based on specific
modifiers. For example, a GenAI function that drafts a non-urgent patient-education
summary may remain in a lower-risk category when the output is reviewable, reversible,
traceable to source material, and reviewed by a qualified clinician before communication to
the patient. By contrast, a GenAI function that recommends urgent therapy escalation based
on patient-specific data may warrant a higher evidentiary tier when the output is timesensitive, highly personalized, difficult for the user to independently verify, or likely to be
relied on before error detection is possible.
Similarly, the same nominal device activity may warrant different requirements depending
on available safeguards. A system that summarizes electrocardiogram data with linked
source strips, clear uncertainty statements, and clinician confirmation where appropriate
presents a different risk profile than a system that provides an untraceable diagnostic
impression directly to a patient or automatically triggers a downstream clinical action. In the
first case, targeted verification, human factors validation, and documentation of traceability
controls may be sufficient. In the second case, more robust clinical confirmation, predefined
critical-error limits, cybersecurity and data-integrity controls, and postmarket monitoring
capable of detecting safety-critical failure modes may be warranted.
FDA could also use modifiers to define when a device function remains within a lower-risk
cell of the matrix versus when additional controls are needed without necessarily changing
the nominal cell. For example, limited personalization, source traceability, independent
clinician review, and sufficient time for intervention could support reliance on bench testing,
retrospective evaluation, and human factors validation. In contrast, high personalization,
limited source traceability, autonomous action, time-critical use, or the absence of
independent safeguards may increase the need for more rigorous clinical evidence, which
may include prospective evidence where warranted, stricter acceptance criteria, bounded
change control, and enhanced postmarket surveillance.
For high-consequence functions, lack of traceability should increase the level of evidence
required. An output supported by linked, verifiable source data is materially different from an
output whose basis cannot be reconstructed. For higher-risk functions, manufacturers
should be able to reconstruct the safety-relevant recommendation or action path, including
the user input, system configuration, model version, retrieval sources, retrieved content,
generated output, confidence or uncertainty signals, tool calls, human review, downstream
actions, exceptions, and decision-relevant intermediate artifacts or routing logic where
available. Traceability is relevant not only to premarket evaluation, but also to complaint

Kestra Medical Technologies, Inc. | 2
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

investigation, medical device reporting, corrective and preventive action, cybersecurity
investigation, postmarket change assessment, and continued assurance of safety and
effectiveness.

B. Directiveness Should Be Evaluated Functionally, Not by Wording Alone
Kestra agrees that directiveness exists on a continuum. FDA should avoid classifying an
output based primarily on whether it uses words such as “recommend,” “should,” or
“consider.” Prompt design or superficial wording changes should not change regulatory
treatment when the foreseeable reliance and risk remain the same.
Directiveness should be treated as a cross-cutting modifier within the two-axis framework.
After a device function is initially placed based on device activity and consequence of
reliance on an incorrect output, FDA should assess whether the output’s specificity,
presentation, workflow placement, personalization, and foreseeable reliance increase or
decrease the likelihood that the user will act on the output without independent confirmation.
That assessment should inform the evidentiary burden, risk controls, user information,
human factors validation, and postmarket monitoring expectations.
For informational functions, FDA should provide factors that allow manufacturers to
distinguish general education, assistive interpretation, and individualized clinical direction in
a predictable way. Relevant factors include whether the output identifies or prioritizes a
preferred clinical action, whether alternative actions are presented, whether uncertainty and
source limitations are visible, and whether the user can independently verify the basis for
the output before acting. FDA should provide examples showing how similar content may
present different levels of directiveness depending on workflow placement, user population,
presentation, personalization, and foreseeable reliance.
For example, a general explanation of treatment options that presents balanced information,
source limitations, and uncertainty without identifying a preferred action may be considered
lower directiveness. A patient-specific summary that identifies relevant clinical
considerations but requires clinician interpretation and does not rank or select an action
may present moderate directiveness. A patient-specific output that identifies a preferred
therapy, timing, or escalation step should be viewed as higher directiveness, particularly
when presented prominently in the workflow or under time pressure. An output that initiates,
automates, gates, or materially conditions a clinical action should be treated as highly
directive even if the wording does not use directive terms.
Disclaimers such as “consult a healthcare professional” should not, by themselves, reduce
the risk classification when the balance of the communication directs a specific clinical
action or creates foreseeable reliance on the output.

C. Patient-Facing Clinical Information Can Present Different Risks and Requires
Tailored Safeguards
Patient-facing GenAI functions may present different risks than HCP-facing functions
because patients may have less clinical context and a more limited ability to recognize
incomplete, misleading, or care-plan-inconsistent outputs. FDA should not, however, treat
patient use as inherently high risk. Properly bounded patient-facing functions can improve

Kestra Medical Technologies, Inc. | 3
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

health literacy, engagement, adherence, access to information, and shared decisionmaking.
FDA should treat patient-facing use as a modifier within the two-axis framework, not as an
automatic basis for higher classification. Risk should turn on whether the intended use,
output content, workflow, and safeguards reduce foreseeable misunderstanding, overreliance, or action inconsistent with appropriate care. Risk may be increased by patientspecific outputs, time-sensitive use, preferred-action language, limited source traceability,
limited clinician review, urgent use contexts, or outputs that could lead a patient to delay
care, change therapy, or disregard clinician instructions.
Risk can be mitigated by bounding the function to education, clarification, adherence
support, or preparation for clinician discussion rather than diagnosis, treatment selection, or
therapy modification. FDA should identify expectations for a defined intended use,
prohibited uses, escalation criteria, and limits on personalization. The system should avoid
instructions that conflict with the known care plan, device labeling, or clinician instructions,
recognize out-of-scope requests, and direct the patient to an HCP or emergency resource
when the content involves urgent symptoms, device malfunction, therapy changes, or
potential clinical deterioration.
Output features should support appropriate reliance. FDA should encourage plain language,
visible uncertainty, clear distinction between education and medical advice, balanced
options where appropriate, source traceability, explanation of when clinician judgment is
needed, and avoidance of false precision. For patient-specific outputs, the system should
identify the information used, key limitations, and assumptions that materially affect the
response.
Appropriate safeguards may include human factors validation with representative patient
users, health-literacy testing, multilingual validation where applicable, usability testing under
foreseeable stress or time pressure, source-linked responses, conservative escalation rules,
refusal behavior for out-of-scope requests, audit trails, complaint and safety-signal
monitoring, and mechanisms to update patient-facing content when clinical instructions or
labeling change. These safeguards should be risk-proportionate and should support
informed patient use without unnecessarily restricting access to understandable clinical
information.
FDA should also distinguish patient-facing informational functions from patient-outreach,
provider-outreach, and care-coordination agents that initiate communications, coordinate
follow-up, monitor adherence, or execute care-management workflow steps. These systems
can create distinct risks even when the underlying clinical content is accurate, including
delayed or failed escalation, contact with the wrong party, unclear communication,
incomplete handoff, missed documentation, provider-notification failure, or execution of an
action outside the intended workflow. FDA should evaluate communication quality, routing
accuracy, timing, handoff reliability, documentation, and escalation completion separately
from clinical content accuracy for agents that communicate with patients, caregivers,
providers, or care teams.

Kestra Medical Technologies, Inc. | 4
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

D. Generalist and Specialist Use Should Be Considered as Part of the Intended User
and Workflow
FDA should treat the generalist/specialist distinction as a user-population modifier within the
two-axis framework. After initial placement based on device activity and consequence of
reliance on an incorrect output, FDA should assess whether the intended HCP user can
understand, verify, and appropriately act on the output. The distinction should not
automatically drive higher classification. Generalists can appropriately use specialistinformed outputs when the function is bounded, transparent, and supports clinical judgment,
referral, or escalation rather than replacing specialist expertise.
Risk may increase when an output addresses a specialist domain, is patient-specific,
identifies a preferred diagnosis or treatment pathway, uses unexplained specialist
terminology, appears at a time-sensitive decision point, or could lead the HCP to defer
referral, delay escalation, or act outside the clinician’s training or scope. Risk may also
increase when the output lacks traceability or presents uncertainty in a way a non-specialist
may not understand.
Risk can be mitigated by defining the intended users, specialty context, and authorized use.
For generalist-facing functions, FDA should identify expectations for plain-language
specialist concepts, source traceability, visible uncertainty, referral or escalation prompts,
red-flag handling, and limits on outputs that imply specialist-level diagnosis or treatment
selection. The device should identify when specialist consultation is appropriate and avoid
foreseeable reliance beyond the intended user’s expertise.
Safeguards may include representative-user validation, human factors testing of
terminology, uncertainty, and escalation prompts, workflow-based performance evaluation,
documented referral logic or consultation triggers, and postmarket monitoring for overreliance or delayed escalation. These controls should be risk-proportionate and should
support access to specialist-informed information without assuming that only specialists can
safely use it.

E. Multi-Turn Conversational Functions Should Be Assessed Across Realistic
Trajectories
FDA should treat multi-turn conversational behavior as a trajectory-based modifier within the
two-axis framework. After initial placement based on device activity and consequence of
reliance on an incorrect output, FDA should assess whether risk increases as the exchange
evolves. Informational content may become directive as context accumulates, follow-up
questions narrow options, or the system applies patient-specific facts.
Intended use should define the authorized conversational scope, intended users, clinical
context, permitted and prohibited outputs, escalation points, and maximum level of
directiveness or autonomy. For emergent behavior, sponsors should define the function’s
outer bounds based on what the system is designed, trained, constrained, and reasonably
expected to do across foreseeable interactions.
Controls should address both content and trajectory, including conversation-state
monitoring, limits on cumulative personalization, refusal or redirection for prohibited
requests, escalation triggers, source traceability across turns, contextual audit logs, and
Kestra Medical Technologies, Inc. | 5
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

representative testing of adversarial, ambiguous, and high-pressure exchanges.
Acceptance criteria should confirm that the system remains within authorized scope across
the conversation, not only that individual responses are accurate.

F. Care Escalation Functions Should Separately Evaluate Under- and Over-Escalation
FDA should treat care escalation behavior as a function-specific modifier within the two-axis
framework. After initial placement based on device activity and consequence of reliance on
an incorrect output, FDA should separately assess under-escalation and over-escalation as
distinct error types, not interchangeable false negatives and false positives. Underescalation may delay necessary care. Over-escalation may increase burden, alarm fatigue,
unnecessary utilization, anxiety, and downstream intervention risk. These harms are not
always commensurable, and acceptable trade-offs will vary by clinical context, severity,
reversibility, available alternatives, and time to intervention.
Sponsors should prespecify the clinical rationale for weighting each direction of error. The
rationale should address the target condition, intended population, care setting, escalation
pathway, consequences of missed and unnecessary escalation, and whether harm can be
detected and corrected before patient impact. For high-acuity or irreversible conditions,
tolerance for under-escalation should generally be lower. In low-acuity or resourceconstrained settings, excessive over-escalation may also create meaningful risk.
Performance evaluation should include separate acceptance criteria for under-escalation
and over-escalation, clinically meaningful error categories, subgroup analyses, and
threshold stress testing. Aggregate accuracy or a single net-benefit score should not mask
unacceptable performance in either direction. Threshold selection should be justified based
on clinical risk, intended workflow, user population, and available mitigations.
Controls may include conservative escalation defaults for high-severity conditions, tiered
escalation categories, human confirmation for higher-risk actions, clear rationale and
uncertainty information, audit trails, monitoring for alert fatigue or delayed escalation, and
threshold adjustment when real-world performance shows an unacceptable imbalance
between missed and unnecessary escalations. FDA should allow different trade-offs when
justified by clinical context, but the justification should be explicit, prespecified, and
supported by evidence.

II. Comments on Competency-Based Premarket Evaluation
A. Benchmarking and Clinical Confirmation Are Appropriate Components
Kestra supports evaluating the final user-facing device in its deployed or representative
configuration, rather than evaluating the foundation model in isolation. Kestra agrees that
device benchmarking followed by clinical confirmation is a useful and appropriate framework
for evaluating GenAI-enabled devices, provided both steps are tied to the device’s intended
use, risk profile, user population, and clinical workflow. The proposed competency domains
of safety, clinical proficiency, generalizability, and agentic capability provide a useful
structure. Benchmarking, however, should not substitute for device-specific verification,
validation, human factors engineering, cybersecurity evaluation, risk management, or

Kestra Medical Technologies, Inc. | 6
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

clinical evidence when those elements are necessary to provide reasonable assurance of
safety and effectiveness.
FDA should use the two-axis framework to help manufacturers determine the level of
evidence needed in a premarket submission. After initial risk placement based on device
activity and consequence of reliance on an incorrect output, modifiers such as autonomy,
directiveness, reversibility, traceability, intended user, and use environment should inform
the rigor of benchmarking, clinical confirmation, human factors validation, and postmarket
controls.
FDA should clarify that benchmark selection should be tied to intended use, patient
population, operating environment, user population, foreseeable misuse, and identified
hazardous situations. A benchmark should provide evidence of clinical knowledge, analytic
capability, safety behavior, communication quality, and generalizability only when its
construct validity is justified for the specific device function. Public benchmarks may be
useful for broad characterization, but they may be insufficient when contamination,
saturation, limited representativeness, or optimization to the benchmark undermines realworld relevance.
Sponsors should explain why benchmark performance predicts safe and effective real-world
behavior for the intended use. That justification should address data provenance,
independence from training and tuning data, task fidelity, input distribution, output scoring,
clinical relevance of errors, subgroup coverage, and evaluator independence. Sponsordeveloped benchmarks may be appropriate when public assets are inadequate, but they
should include safeguards against optimization to the test, such as locked test sets, blinded
evaluation, prespecified scoring, independent review, and separation from model
development.

B. FDA Should Provide Principles for Comparators and Acceptance Criteria
A significant implementation gap is the lack of predictable principles for establishing passing
performance. Sponsors need a defensible method to prespecify clinically meaningful
acceptance criteria for open-ended and variable outputs. FDA should provide guidance on
how to define critical errors, weighing different error types, addressing repeated-run
variability, establishing subgroup performance criteria, and determine when aggregate
performance may mask unacceptable safety-critical failures.
Clinician comparison should not be the default when an objective reference standard is
available. Appropriate reference standards may include adjudicated clinical outcomes,
pathology, laboratory or physiological measurements, validated signal annotations, or
expert consensus. When clinician comparison is appropriate, the comparator should reflect
the intended user and use environment. Specialist performance may be relevant when the
function substitutes for specialist judgment, while generalist or human-AI team performance
may be more relevant when the device supports a generalist in actual practice.
For many assistive functions, the clinically meaningful comparison is the performance of the
human-device team against current practice, not the isolated performance of the device
against a specialist. FDA should permit sponsors to justify the comparator that best reflects
the intended workflow and course of care.

Kestra Medical Technologies, Inc. | 7
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

C. Clinical Confirmation Should Be Risk-Proportionate
Kestra agrees that a prospective clinical study should not be required for every GenAIenabled device. Retrospective evaluation, shadow deployment, standardized patient
interactions, clinician adjudication, and prospective studies may each be appropriate
depending on intended use and risk.
FDA should identify factors that govern the choice of confirmation method, including severity
and reversibility of harm, degree of autonomy, novelty of the function, ability to detect error
before action, representativeness of available data, dependence on user interaction, and
feasibility of real-world monitoring. For autonomous treatment, medication dosing, therapy
control, life-supporting or life-sustaining functions, or other high-consequence uses,
postmarket monitoring should not be used as a substitute for premarket evidence needed to
provide reasonable assurance of safety and effectiveness.
Sponsors should also prespecify statistically meaningful performance measurement
methods. The statistical plan should define the analysis population, sampling strategy,
sample size rationale, confidence intervals, repeated-run handling, subgroup analyses,
missing or indeterminate outputs, and criteria for clinically unacceptable errors. Performance
estimates should be reported separately for safety-critical outputs and clinically important
subgroups rather than relying only on aggregate metrics.
Synthetic data may supplement real patient data when its generation process, fidelity,
coverage, independence, and limitations are justified. Sponsors should account for
differences between synthetic and real-world distributions. Benchmarking and clinical
confirmation evidence should not be combined into a single performance estimate unless
the data sources, populations, endpoints, and uncertainty are compatible and the pooling
method is prespecified.
Synthetic data is better suited to rare but well-characterized scenarios, stress testing,
workflow edge cases, and robustness evaluation. It is less appropriate as primary evidence
for novel clinical presentations, poorly characterized subgroups, or safety-critical
performance where real-world distributional assumptions are uncertain.
Where synthetic data is generated using models of the same class as the device under
evaluation, sponsors should address the risk of correlated bias. Safeguards may include
independent data sources, independent generation and evaluation methods, comparison to
real patient cases, subgroup-specific performance checks, prespecified limits on syntheticdata contribution, and sensitivity analyses showing whether conclusions depend on
synthetic inputs. Synthetic data should not substitute for adequate evaluation of clinically
important subgroups.

D. Third-Party Participation May Strengthen Benchmarking and Clinical
Confirmation
Independent third-party participation may strengthen a competency-based approach when it
improves objectivity, benchmark integrity, clinical adjudication, or statistical review. Thirdparty involvement may be most useful for locked benchmark maintenance, blinded scoring,
comparator panel management, subgroup analysis review, and verification that evaluation
assets are independent from model training and tuning data.
Kestra Medical Technologies, Inc. | 8
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

FDA should define qualifications and independence criteria for any third party, including
clinical expertise, statistical competence, data-governance controls, conflict-of-interest
safeguards, confidentiality protections, and separation from model development. Third-party
review should not be required where it would be impractical, duplicative, or likely to delay
access without improving the evidence. FDA should also avoid program designs that create
exclusive benchmark gatekeepers or barriers to entry for smaller manufacturers.

E. Model Architecture Should Inform, but Not Replace, Device-Specific Evaluation
The competency-based approach should apply across model architectures, but the
evaluation should be tailored to the device function and architecture-specific hazards.
Multimodal and vision-language models may require evaluation of image quality, signal
fidelity, modality alignment, missing or conflicting inputs, temporal context, and whether text
outputs remain grounded in the source data. Generative or predictive world models may
require additional assessment of simulation validity, causal assumptions, rare-event
behavior, and whether generated trajectories reflect clinically plausible real-world states.
FDA should not evaluate architecture labels in isolation. The premarket evidence should
address how the deployed device uses the model, what inputs and outputs are safetycritical, how errors propagate through the workflow, and what controls prevent unsupported
clinical conclusions. For architectures that integrate multiple modalities, predict future states,
or generate simulated scenarios, sponsors should justify the relevance of benchmarks and
clinical confirmation methods to the device’s intended use and real-world operating
conditions.

III. Postmarket Monitoring and Change Control
A. Reduced Premarket Evidence Should Be Limited to Defined Circumstances
FDA should accept greater premarket uncertainty only when the remaining uncertainty is
limited, the anticipated benefits justify that uncertainty, clinically meaningful performance
can be monitored, safety signals can be detected promptly, and effective mitigation,
rollback, or suspension mechanisms are available. This approach should generally not be
used when failures may cause immediate, serious, or irreversible harm before detection or
intervention.
A manufacturer's postmarket monitoring plan should define performance metrics, sampling
strategy, subgroup analyses, data-quality controls, review cadence, triggering events, alert
thresholds, escalation pathways, corrective actions, and reporting responsibilities.
Monitoring should include periodic re-benchmarking, sample-based clinician review,
performance degradation monitoring, complaint and safety-signal review, and targeted
review of safety-critical outputs.
Cadence and triggers should be risk-proportionate and should account for device criticality,
deployment scale, observed drift, complaint trends, model or data-source changes,
cybersecurity events, changes in clinical practice, unexpected use patterns, new failure
modes, increased override or escalation rates, subgroup performance concerns, or
upstream model changes.

Kestra Medical Technologies, Inc. | 9
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

Machine-based supervisory agents may support postmarket monitoring by screening
outputs, detecting drift, flagging out-of-scope use, monitoring escalation behavior, and
identifying potential safety signals. FDA should identify validation expectations for the
supervisory agent’s intended monitoring function, including sensitivity to safety-critical
failures, false alert burden, robustness to distribution shift, cybersecurity controls,
auditability, and independence from the device behavior it monitors. Supervisory agents
should augment, not replace, manufacturer review and accountability.
Clinicians, healthcare institutions, professional societies, standards-setting bodies, and
other stakeholders may contribute to postmarket monitoring by reporting safety signals,
supporting adjudication of clinical significance, developing consensus metrics, identifying
emerging standards of care, and informing subgroup or workflow-specific concerns. These
roles should be structured through defined reporting pathways, data-use controls, conflictof-interest safeguards, and clear escalation criteria. Manufacturer responsibility for
complaint evaluation, risk management, corrective action, reporting, and continued
assurance of safety and effectiveness should remain clear and non-delegable.

B. Re-Benchmarking Should Be Scaled to the Change and Hazard Impact
FDA should align re-evaluation with the nature of the modification and its potential effect on
safety and effectiveness. Changes to prompts, retrieval sources, guardrails, orchestration
logic, model version, user interface, output formatting, tool permissions, and deployment
infrastructure can each affect device behavior and should be addressed in the change
impact assessment.
Changes that do not affect intended use, safety-critical behavior, clinical logic, model
performance, or risk controls may be managed within the quality system when supported by
documented verification and risk assessment. Changes that affect autonomous authority,
clinical scope, safety-critical performance, or the probability or severity of harm may require
evaluation under an authorized predetermined change control plan or a new premarket
submission, as applicable.
For PCCPs, FDA should allow manufacturers to prespecify bounded categories of
modifications, validation protocols, acceptance criteria, monitoring commitments, and
implementation controls even when every future parameter cannot be identified. The
authorized envelope should be defined by the type and clinical effect of change, not solely
by a list of exact technical modifications.

C. Third-Party Foundation Models Require Contractual and Technical Controls
For devices that depend on third-party foundation models, the device manufacturer must
retain responsibility for the safety and effectiveness of the finished device. That
responsibility must be supported by practical access to information needed to detect,
assess, validate, document, and respond to upstream changes within the manufacturer’s
quality system.
FDA should encourage practical sponsor-vendor governance provisions that enable
manufacturer accountability, including defined notification windows for safety-relevant model
changes; version persistence or controlled migration where feasible; access to release
notes, model-change history, and safety-relevant technical information; disclosure of
Kestra Medical Technologies, Inc. | 10
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

degradation events, cybersecurity incidents, or other safety-relevant issues; validation
support; rollback capability or compensating controls; and continued support or controlled
transition planning for authorized versions. Technical controls may include version locking
where feasible, automated change detection, regression testing, canary or staged
deployment, performance monitoring, independent impact verification, and the ability to
suspend affected functionality.
FDA should also clarify expectations when a foundation model provider will not provide
sufficient transparency or control. In such cases, the manufacturer should demonstrate
compensating controls or select an alternative architecture that permits adequate design
control and postmarket accountability.

IV. Foundation Model Master Files
Kestra supports a voluntary Foundation Model Device Master File program, if FDA
establishes such a program, provided it reduces duplicative review while preserving sponsor
accountability. A Foundation Model Master File could include model architecture and
versioning information; training-data provenance at an appropriate level of detail; known
limitations and failure modes; healthcare-relevant evaluation results; subgroup performance;
safety constraints and guardrails; update and incident notification commitments;
cybersecurity information; audit-log capabilities; and change history.
The program should not be interpreted as FDA authorization of a foundation model for
medical use, nor should a sponsor’s right of reference replace device-specific evidence.
FDA should establish minimum content, maintenance, version control, and notification
expectations so that a referenced file remains current and useful. To encourage
participation, FDA should provide a clear review pathway, protect confidential commercial
information, and explain how reliance on a current Master File may streamline sponsor
submissions.
Voluntary Master Files may be practical when the content is standardized, submission
burden is reasonable, confidential commercial information is protected, and update
obligations are limited to changes that could affect safety, effectiveness, cybersecurity, or
device performance. Because foundation model developers may have limited incentive to
disclose safety-relevant information, the program will be useful only if FDA identifies what
information is expected, how it will be reviewed, how sponsors may rely on it, and what
change notifications are needed to keep the file current.
FDA should also consider alternative mechanisms when a voluntary Master File is
unavailable or incomplete. These may include sponsor-held technical files, contractual
disclosure obligations, standardized model cards, confidential FDA review packages,
independent audits or attestations, third-party conformity assessments, provider changenotification commitments, PCCP-linked update controls, and sponsor validation packages
demonstrating that limited model information is offset by adequate technical, contractual,
and postmarket controls.

Kestra Medical Technologies, Inc. | 11
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

V. Agentic GenAI-Enabled Devices
Agentic systems warrant additional controls because risk can arise from planning,
sequencing, tool selection, external data retrieval, execution, and interaction among
individually acceptable steps. Evaluation should assess the complete action chain, including
cumulative and emergent hazards.
For high-consequence agentic device functions, FDA should identify expectations for a
defined scope of authority, least-privilege tool and data access, human authorization before
irreversible or high-consequence actions, interruptibility and safe-state behavior, reviewable
audit trails sufficient to reconstruct safety-relevant actions, verification or validation checks
for safety-relevant external tool outputs, prompt-injection resistance, recovery from partial
execution or tool failure, and prevention of unauthorized clinical-scope expansion.
Oversight and acceptance criteria should increase with autonomy, clinical consequence,
tool access, and reduced opportunity for human review. For higher-risk agentic functions,
evaluation should include final outputs, tool calls, data retrieval accuracy, sequencing,
authorization checks, failure recovery, prohibited actions, unsafe tool use, unverified data
reliance, boundary failures, and failure to obtain required human approval. Postmarket
monitoring may include, as appropriate, action logs, tool-use patterns, near misses,
overrides, and attempted out-of-scope actions.
FDA should explicitly incorporate AI-specific cybersecurity and model supply-chain risks into
expectations for GenAI-enabled and agentic devices. In addition to prompt-injection
resistance and conventional device cybersecurity controls, sponsors should address data
poisoning, retrieval poisoning, model and dependency provenance, compromise of retrieval
pipelines, unsafe tool orchestration, multi-agent attack chains, abnormal tool-use patterns,
and degradation or compromise of third-party model components. Security testing should
evaluate prompts, retrieval sources, orchestration logic, tool integrations, authorization
boundaries, monitoring, and incident response procedures for model degradation or
compromise.
For agentic functions that initiate communications, coordinate care steps, or trigger
downstream operational actions, sponsors should define accountable roles, escalation
ownership, exception handling, and criteria for human takeover. Manufacturer accountability
for the finished device should remain clear and non-delegable, while clinical workflow
ownership and response responsibilities should be defined for healthcare organizations,
users, or care teams where applicable.

VI. Conclusion
Kestra supports FDA’s development of a risk-based, total product life cycle approach for
GenAI-enabled medical devices. The Discussion Paper provides a strong foundation by
recognizing the need to evaluate the deployed device, apply risk-proportionate evidence
expectations, incorporate clinical confirmation, monitor real-world performance, and address
third-party foundation models and agentic functions.
To make the framework feasible for manufacturers to implement, FDA should crystallize four
priority expectations: auditable reconstruction of AI-driven recommendations,
communications, and actions; explicit controls for patient- and provider-facing outreach and
Kestra Medical Technologies, Inc. | 12
 KESTRA MEDICAL TECHNOLOGIES, INC. | FORMAL COMMENT

care-coordination agents; practical governance of third-party foundation model changes;
and AI-specific cybersecurity and model supply-chain controls. These expectations should
be applied through a risk-proportionate total product life cycle framework that preserves
least burdensome principles, supports effective quality system controls, and does not
depend on information or vendor commitments manufacturers cannot reasonably obtain.
Clear, objective, and consistently applied expectations will help manufacturers generate
appropriate evidence, manage third-party dependencies, maintain postmarket
accountability, and support innovation without compromising patient safety.
Kestra appreciates FDA’s consideration of these comments.
Respectfully submitted,

Beverly Magrane
Vice President, Quality & Regulatory Affairs
Kestra Medical Technologies, Inc.

Kestra Medical Technologies, Inc. | 13