MiniMed

MiniMed supports FDA’s proposed risk-based framework for regulating GenAI-enabled medical devices and agrees that a total product lifecycle approach is appropriate given the dynamic nature of these technologies. We recommend that FDA refine the framework by introducing explicit risk modifiers, recognizing the importance of agentic architectures and guardrail systems, distinguishing between generative and deterministic functions, strengthening guidance for post-market monitoring, and providing greater clarity regarding third-party foundation model governance. A flexible, risk-proportionate framework that evaluates the complete deployed system and adapts to ongoing technological advances will better support innovation while maintaining patient safety and device effectiveness. See the attached file for detailed recommended actions.

Section / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
The proposed risk
framework based on
device activity and
Retain the twoconsequence of an
Clinical risk is often axis framework
incorrect output
influenced not only but formally
provides an intuitive
by the device introduce risk
and scalable
function and modifiers
foundation for
consequence of including
classifying GenAIfailure, but also by reversibility, timeenabled medical
contextual factors to-harm,
device functions.
such as reversibility traceability,
The framework
of harm, user ability verifiability,
Page 9 (Q1- Two-Axis Risk appropriately
to recognize errors, detectability of
Q2) Framework separates what the
availability of incorrect outputs,
system does from
backup controls, and and independent
the potential impact
time available to safety controls.
of incorrect outputs.
react. Ignoring these FDA should
However, several
factors may result in provide worked
factors that
overly conservative examples
significantly
or inconsistent demonstrating how
influence real-world
evidence these modifiers
risk are presently
expectations. affect evidence
described separately
requirements.
rather than being
integrated into the
framework.
The discussion
appropriately
A highly directive Clarify that
distinguishes
recommendation to a intended user
between patientclinician may classification
facing and
present greater risk should be
healthcare
than a transparent considered
professional-facing
informational alongside
Patient vs HCP functions. However,
Page 9 (Q3) summary provided transparency,
Users meaningful
to a patient. User complexity, ability
variability exists
classification alone to verify outputs,
within both
may therefore and degree of
categories. Risk is
overestimate or influence on
often driven more
underestimate actual clinical decision
by the nature of the
risk. making.
output and its
intended use than
 Section / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
by the identity of
the user alone.
FDA should
evaluate
Multi-turn
conversational
conversational
systems based on
systems may
conversation
transition from
trajectories and
informational The cumulative
state transitions,
responses to effect of a
rather than
recommendation- conversation may
individual prompts
Conversational generating or influence patient
Page 9 (Q5) and responses
Interfaces action-directing behavior even if no
alone. Evidence
outputs over the individual response
generation should
course of an independently meets
consider the
interaction. a high-risk threshold.
highest-risk
Assessing isolated
interaction
outputs may not
pathway
adequately reflect
reasonably
actual system risk.
expected during
use.
Under-escalation For example, failure
and over-escalation to identify a severe
represent distinct hypoglycemic event Require separate
failure modes with presents different characterization,
materially different risks than generating reporting, and
Page 10 Escalation clinical unnecessary acceptance criteria
(Q6) Recommendations consequences. escalation for underCombining these recommendations. escalation and
errors into a single Both should be over-escalation
metric may obscure independently performance.
meaningful safety understood and
signals. managed.
Some GenAI Explicitly include
outputs are User-verifiable verifiability of
inherently verifiable outputs generally output as a
by end users present a lower modifier that may
Verifiable because source data likelihood of reduce
Section IV
Information are visible and undetected harm consequence
auditable. Examples because users can classification and
include meal- independently corresponding
recognition outputs identify errors. evidence
linked to expectations.
 Section / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
photographs or
summaries linked
directly to source
records.
Distinguish
between GenAICertain GenAI generated outputs
systems do not and GenAIperform The risk profile of a enabled interfaces
calculations but routing function that access
GenAI Accessing instead facilitate differs significantly validated
Section IV
Validated Tools access to validated from the risk profile deterministic
deterministic tools of generating novel systems.
by extracting user recommendations. Evaluation should
intent and routing focus on request
requests. interpretation and
tool invocation
accuracy.
The proposed
competency-based
approach provides a
reasonable Clarify that
mechanism for Generic benchmarks competency
evaluating highly may not correlate requirements
Section V Competency- variable GenAI with clinically should be risk-
(Q7-Q11) Based Evaluation systems. However, meaningful proportionate and
competency should performance or tied to the specific
be assessed relative patient outcomes. clinical role of the
to intended use and device.
patient risk rather
than generalized AI
performance.
Device
benchmarking
Clarify how
combined with
Traditional medical benchmarking may
clinical review may
device evidence complement,
Section V Benchmarking be appropriate for
requirements remain rather than replace,
(Q7-Q11) Approach certain use cases but
necessary for many traditional device
should not replace
device functions. validation
established software
activities.
verification,
validation, and
 Section / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
performance testing
methodologies.
Develop guidance
for testing realistic
For conversational
Clinical performance user interactions
interfaces,
may be degraded by including
understanding and
transcription errors, ambiguous
Section V User Request correctly
ambiguous prompts, language, speech-
(Q9-Q11) Interpretation interpreting user
adversarial inputs, or to-text errors,
intent may become
incomplete noisy inputs,
a primary source of
information. adversarial
failure.
prompts, and outof-scope requests.
Expand FDA
The discussion
guidance to
paper primarily
explicitly address
focuses on
In many deployed orchestration
foundation models
systems, safety and frameworks,
and devotes
Section Foundation effectiveness depend retrieval systems,
comparatively
VII.B Models vs as much on the tool-use
limited attention to
(Q25-Q26) Agentic Systems harness architecture architectures,
agentic
as on the underlying access controls,
architectures,
model itself. verification
orchestration layers,
pipelines, and
and guardrail
guardrail
systems.
mechanisms.
Consider
Modern GenAI evaluating
applications often guardrails,
rely on Evaluation of the orchestration logic,
deterministic foundation model verification
Section Harness-Based validation layers, alone may not workflows, and
VII.B Risk Control structured accurately represent failure
workflows, and the deployed containment
business-rule system's behavior. strategies as
enforcement to primary
improve safety. components of
regulatory review.
Reliance on As GenAI systems Encourage
Objective clinician consensus evolve rapidly, objective testing
Overall
Evaluation panels may scalable and methodologies,
Framework
Methods introduce repeatable testing stress testing,
variability, mechanisms will scenario-based
 Section / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
subjectivity, and become increasingly validation, and
scalability important. automated
challenges. evaluation
frameworks where
scientifically
justified.
Establish guidance
describing panel
Expert review
composition,
panels may provide
Variability among conflict-of-interest
useful clinical
Clinical reviewers may affect controls, diversity
Q14-Q15 judgment but can
Benchmark Panels reproducibility of requirements, and
introduce bias and
evaluations. methods to
inconsistent
measure interinterpretations.
reviewer
agreement.
Encourage
comparison
FDA should ensure Clinicians and
against published
that evaluations existing systems are
clinician
compare GenAI not error-free.
Benefit-Risk performance,
Q14-Q15 performance against Benefit-risk
Evaluation existing software
realistic alternatives assessments should
tools, and
rather than idealized reflect actual clinical
standard-of-care
standards. practice.
workflows where
feasible.
Expand
postmarket
Periodic
Degradation may be monitoring
benchmarking and
observed through expectations to
output review
changes in user include telemetry,
represent an
Postmarket behavior, complaint complaint analysis,
Section VI appropriate starting
Monitoring trends, telemetry usage patterns,
point but should be
patterns, or drift detection, and
supplemented by
workflow clinically
broader surveillance
disruptions. meaningful
mechanisms.
outcome
monitoring.
Connected Telemetry and Recognize
healthcare remote monitoring telemetry-enabled
Connected Device
Section VI ecosystems can enable earlier monitoring
Ecosystems
provide near real- detection of device platforms as
time insight into issues than valuable riskSection / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
product behavior complaint-based management
and emerging risks. systems alone. controls and
encourage their
use in postmarket
surveillance
programs.
Establish guidance
Manufacturers Inconsistent for predefined
would benefit from regulatory triggers, escalation
Risk-Based greater clarity expectations may thresholds,
Section VI
Monitoring Plans regarding result in inefficient investigation
monitoring monitoring requirements, and
expectations. programs. corrective-action
criteria.
Ensure alternative
mechanisms exist
Foundation Model
Device for obtaining
Device Master Files
manufacturers may sufficient
Third-Party (MAFs) may
not always have information about
Section VII Foundation improve
access to MAFs or model versions,
Models transparency and
comparable known limitations,
reduce duplication
information. change histories,
of effort.
and safety-relevant
updates.
Encourage version
control, change
notifications,
Manufacturers need
Third-party models update schedules,
sufficient visibility
may evolve after retesting
Change into model changes
Section VII deployment in ways requirements,
Management to assess impact on
that affect device rollback
safety and
performance. mechanisms, and
effectiveness.
predefined changemanagement
processes.
Clarify boundaries
between changes
PCCPs can facilitate
Ambiguity regarding that can be
Predetermined innovation while
Section VI PCCP scope may implemented
Change Control maintaining
& VII lead to inconsistent under a PCCP and
Plans (PCCPs) regulatory
implementation. those requiring a
oversight.
new marketing
submission.
 Section / FDA Proposal / Regulatory Recommended
Detailed Comment
Question Topic Rationale FDA Action
Establish a formal
process for
periodic review
The pace of GenAI Regulatory
and updating of
innovation approaches that are
Overall Evolution of FDA GenAI
significantly appropriate today
FDA Regulatory evaluation
exceeds traditional may become
Approach Framework methodologies,
medical device outdated within a
guidance
development cycles. short period.
documents, and
evidence
expectations.