Profound Ventures | Guidance Global Consulting (Brian Meshkin, Managing Partner; Anita Monteiro, CEO)

See attached file(s)

PROFOUND VENTURES • A Social Impact Venture Studio

September 14, 2026

Dockets Management Staff (HFA-305)
U.S. Food and Drug Administration
5630 Fishers Lane, Rm. 1061
Rockville, MD 20852

Re: Comment on “Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback” (August 2026)
Docket ID FDA-2026-N-7874

Dear FDA Leadership:
We are writing on behalf of Profound Ventures Corporation, a social impact-focused venture studio, and
Guidance Global Consulting (GGC), a strategic consulting and solutions firm in healthcare, AI, and digital
transformation, to provide comment on the Center for Devices and Radiological Health's (CDRH) discussion
paper on the regulation of generative AI (GenAI)-enabled medical devices. Profound Ventures builds and
incubates a portfolio of social-impact focused healthcare, diagnostics, and applied-AI ventures. GGC has
extensive experience advising federal health agencies and health systems on AI adoption, quality measurement,
and regulatory strategy. This comment draws on the collective operating experience of our portfolio, not the
perspective of any single company alone.

I, Brian Meshkin, offer this comment as a Tri-Sector Leader whose career spans the public, private, and
nonprofit sectors in ways that bear directly on the evidentiary and governance questions CDRH has posed. As
a teenager, I led a successful child bicycle helmet safety campaign that changed laws nationwide - an early
lesson in how evidence translates into public policy. I later served as an elected official on the Howard County,
Maryland Board of Education where I pioneered creation of the digital education policy, and I have served on
nonprofit boards in the healthcare and social-impact space. In the private sector, I spent years as an executive
at Johnson & Johnson and Eli Lilly, two of the most heavily regulated life-sciences companies in the world,
before becoming a founder and operator in my own right. At Eli Lilly, I helped start the e.Lilly Venture Fund
and was the Team Leader on e.Business Customer Connectivity where we had an incubator launching 8
companies, invested in many others, and conducted pilots in the United States and seven of the largest markets
worldwide to evaluate the impact of digital health technologies in healthcare and life sciences. At Johnson &
Johnson, I helped launch Cancer.com with leading patient advocacy groups. Eventually, I went back out on my
own and founded and scaled Proove Biosciences, which built what was, at the time in 2017, the largest clinicalgenetic databank in chronic pain - a real-world evidence infrastructure used to validate pharmacogenomic
associations against accepted clinical and guideline standards (CPIC, ACOEM/MTUS).

My own portfolio at Profound Ventures, includes but is not limited to:

• Labarium Diagnostic Systems, Inc., a CLIA-certified/COLA accredited clinical laboratory platform
building an AI-enabled diagnostics infrastructure across multiple acquired laboratories;

• Proove Genomics, a pharmacogenomics diagnostics business operating under close guidelineanchoring and fraud-and-abuse discipline;

• HumAInity, the first AI tool certified in the world as a healthcare worker extender, and a CMS
Ecosystem partner; and Arockia AI, a clinical decision support venture co-founded by both of us

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 1 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

focused on delivering world-class quality improvement and outcomes insights at the point-of care
decision-making.

I offer this comment alongside Anita Monteiro, CEO of Guidance Global Consulting and a former SES federal
executive who spent 12 years at CMS overseeing a large Medicare beneficiary advocacy and innovation
portfolio. She is Co-Founder of Arockia AI, Inc., alongside me; Co-Founder of Outcomeus; and serves on the
industry advisory boards of HumAInity, Proove Genomics, and Labarium Diagnostic Systems giving her direct
visibility and influence on the evidentiary, deployment, and quality questions each of those ventures raises for
this discussion paper. She is also academic expert and program leader for Executive Education at Harvard
Medical School, Emory University and Wharton School of Business at the University of Pennsylvania.

Together, our trajectories - activist, elected official, nonprofit board member, large pharmaceutical executive
and entrepreneur on one side; federal regulator, Medicare beneficiary advocate, and quality-improvement
innovator on the other - are reflected in the ventures and advisory relationships described above.

Each portfolio company in our partnership and individually has required us to build, defend, or operate under
evidentiary standards for clinical validity and clinical utility - the same distinction we believe this discussion
paper needs to make more explicit.

Given the breadth of the discussion paper's 26 questions, this comment focuses on the ten questions where
Profound Ventures' and GGC’s portfolio experience is most useful to CDRH's thinking, organized by the paper's
own section structure. This paper advances CDRH's 'Innovation and Global Leadership' public health pillar and
builds on the Digital Health Center of Excellence's 2021 AI/ML-Based SaMD Action Plan and its December
2024 Predetermined Change Control Plan (PCCP) guidance.

Our recommendations are offered in that continuity as an extension of tools CDRH has already built, not a
request to build new ones from scratch. Throughout, and consistent with the statutory least-burdensome
principle CDRH itself invokes in the paper's introduction, our recommendations favor risk-proportionate, post
market-weighted oversight that preserves a competitive, innovation-friendly path to market for early-stage
companies - the companies Profound Ventures exists to build.

A Cross-Cutting Recommendation: Separate Clinical Validity from Clinical Utility as Distinct
Evidentiary Bars

Before addressing specific questions, we urge CDRH to import a distinction long used in diagnostic test
evaluation - most familiar from the ACCE framework (Analytic validity, Clinical validity, Clinical utility)
applied to genetic and genomic tests - into its competency-based approach for GenAI-enabled devices. In my
experience validating pharmacogenomic associations at Proove Biosciences and relaunching Proove Genomics,
these are not interchangeable concepts, and conflating them is a recurring source of both regulatory risk and
marketplace harm:

• Analytic/technical validity - whether the device does what it claims to do at a mechanical level
(accuracy, reproducibility). This maps closely to the paper's Generalizability elements (R.1, R.2).
• Clinical validity - whether a device's output correlates with the true clinical state or accepted body of
evidence (e.g., whether an escalation recommendation reflects the actual guideline-supported standard
of care). This maps to Clinical Proficiency (E.1–E.4), but the paper does not yet require sponsors to
disclose the strength of the underlying evidence base per claim.

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 2 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

• Clinical utility - whether use of the device's output, in the hands of its intended user, actually improves a
health outcome or decision relative to not using it. This is closer to what Section V.C's clinical
confirmation approaches are reaching for, but the paper does not clearly distinguish it from clinical
validity as a separate evidentiary requirement.
In pharmacogenomic panel commercialization, we learned this distinction the hard way: guideline support is
not uniform across clinical claims even within a single panel - some associations carry strong guideline-level
clinical validity anchoring, while others lack an equivalent evidentiary anchor for the genetic component and
require materially different, more conservative claims language. The same discipline should apply to GenAIenabled devices at the individual-function level, not just the device level. A device may be clinically valid
for one function (e.g., E.1 knowledge recall) while lacking clinical utility evidence for another (e.g., whether
its escalation output changes downstream care in a way that improves outcomes). We recommend CDRH
require sponsors to benchmark and label each device function against both bars separately, rather than allowing
a single aggregate competency assessment to imply utility that has not been demonstrated.

A Second Cross-Cutting Recommendation: Least-Burdensome, Risk-Proportionate Oversight
to Preserve Competitive Entry for Early-Stage Companies

As a venture studio that builds and funds early-stage healthcare and diagnostics companies, Profound Ventures
has a direct, practical stake in a question the discussion paper touches on but does not fully resolve: evidentiary
rigor and regulatory burden are not the same variable, and a framework that conflates them will entrench
incumbents rather than protect patients.

We support a competency-based approach that is genuinely risk-proportionate - but we urge CDRH to design
it so that the fixed cost of compliance does not itself become the dominant barrier to market entry, which would
work against the innovation and patient-access goals the paper states as its own north star.

Four specific recommendations follow from this principle:

• Default to enforcement discretion and the existing clinical decision support exclusion for low-activity,
non-directive, HCP-facing informational functions. CDRH should resist the temptation to use the
novelty of GenAI as a reason to narrow existing statutory carve-outs (e.g., section 520(o)(1)(E)) that
already keep large categories of lower-risk clinical software outside device regulation. Scope creep into
previously excluded functionality raises the cost of entry for every early-stage company building in this
space, regardless of actual patient risk. The laboratory-developed test (LDT) experience is instructive
here and is discussed further in response to Question 16 below: courts and Congress have recently and
affirmatively reinforced CLIA's quality-system-based, non-FDA-premarket-gated model as sufficient
oversight for complex, high-stakes diagnostic testing, and we believe that same model has direct
application to GenAI-enabled functions embedded in laboratory workflows.
• Make the most burdensome elements of the competency-based approach - prospective clinical studies,
mandatory independent third-party adjudication, and extensive subgroup benchmarking - the exception
reserved for the highest-risk tier (fully autonomous, severe-consequence functions), not the default
expectation for the broad middle of the risk matrix. A well-capitalized incumbent can absorb a
prospective trial as a cost of doing business; an early-stage company frequently cannot, and a
framework that requires trial-level evidence for moderate-risk functions will simply select for capital
rather than for safety or quality.

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 3 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

• Keep any third-party accreditation, adjudication, or Foundation Model MAF program strictly voluntary
and non-gating, with shared or subsidized infrastructure (sequestered benchmark datasets, pooled
independent adjudicator panels) available on reasonable terms to small entities. A mandatory third-party
gate, even a well-intentioned one, functions economically like a licensing fee and will be felt most
acutely by companies without an existing compliance department - precisely the companies most likely
to bring genuinely novel approaches to market.
• Rely on post-market monitoring, not expanded premarket evidentiary burden, as the primary mechanism
for managing the residual uncertainty inherent to GenAI-enabled devices. This is not merely a
competitiveness argument: because GenAI-enabled devices can evolve after deployment in ways
bounded software historically has not, continuous post-market signal is often a more accurate safety
mechanism than a point-in-time premarket study, and it achieves that safety outcome at materially lower
fixed cost to the innovator.
We elaborate on several of these points below in response to specific questions, and we address the post-marketmonitoring point directly in response to Question 18 in Section VI.

A Third Cross-Cutting Recommendation: Extend Risk-Proportionate Burden-Sharing to the
Deployment Side, Not Only the Sponsor Side

The FDA discussion paper's competency-based framework is built almost entirely around sponsor obligations.
In our combined operating experience spanning device commercialization and, from the CMS side, oversight
of how Medicare-participating providers actually adopt new clinical technology, deployment-side capacity is
frequently the binding constraint on safe use, not sponsor-side evidence. A rural hospital, critical access
facility, or federally qualified health center can lawfully deploy a well-benchmarked device without the
informatics staffing to meaningfully run the re-benchmarking, degradation monitoring, or human-oversight
review this framework presumes.

We recommend CDRH's post market monitoring guidance explicitly address minimum deploying-site
capacity as a condition of higher-tier device functions and look to CMS Conditions of Participation and
existing quality-reporting infrastructure - which already reaches every Medicare-participating hospital - as a
coordination point, rather than building monitoring expectations that assume sponsor-side infrastructure will
substitute for site-level capacity.

Just as we urge CDRH to keep sponsor-side compliance from becoming a de facto barrier to entry for earlystage companies, we urge CDRH to avoid a framework that only a well-resourced health system can
operationally sustain.

Section IV - Assessment of Risk (Questions 1 and 3)

Question 1 - Does the two-axis risk framework capture the dimensions most relevant to risk? What
additional dimensions should be represented?

We support the two-axis framework as an organizing heuristic, but recommend CDRH add evidentiary
traceability as an explicit third dimension or consequence-axis modifier: whether a given output is grounded
in a validated biomarker, guideline, or clinical reference standard, versus general model reasoning without such
Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 4 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

an anchor. At Proove Genomics, associations anchored to CPIC and ACOEM/MTUS guidelines supported
materially different claims and risk posture than associations without an equivalent guideline-level
endorsement. A GenAI-enabled device output that traces to a validated reference standard is meaningfully
lower-risk, for a given activity and consequence level, than one that does not. I recommend the risk framework
explicitly reward - through reduced evidentiary burden - device functions built on validated, traceable clinical
anchors, and correspondingly require heightened scrutiny for functions relying on emergent or unanchored
model reasoning.

We also recommend CDRH account for deployment-context modifiers on the consequences axis. Through
Arockia AI's work in clinical decision support for small and rural hospitals, I have seen firsthand that the same
informational output carries different real-world consequence depending on whether the deploying facility has
on-call specialty coverage or relies on the device to partially close an access gap. A risk framework calibrated
only to output content, without reference to deployment environment, will systematically underweight risk in
exactly the settings - under-resourced facilities - where GenAI-enabled devices are likely to see the most
consequential use.

This concern is not ours alone. HHS's December 2025 AI Strategy names democratizing AI technologies and
access as one of its four core goals, and OMB Memorandum M-25-21 requires HHS divisions to classify and
manage 'high-impact AI' with defined risk controls. We recommend CDRH harmonize its two-axis risk
framework with HHS's high-impact AI classification rather than developing an independent taxonomy, so that
sponsors and deploying health systems are not made to reconcile two non-aligned federal risk frameworks for
the same device and so that CDRH's rural and under-resourced-facility risk calibration serves HHS's own access
goal directly.

Question 3 - When could patient-facing delivery of clinical information present different or higher risk,
while preserving the benefits of patient empowerment?

HumAInity, which I co-founded, is the first AI tool certified in the world as a healthcare worker extender, and
is a CMS Ecosystem partner. That certification model - positioning the device explicitly as an extension of a
credentialed human worker's scope, rather than as an autonomous, freestanding source of clinical information
- is, in my experience, one of the more effective mitigations CDRH could consider for the patient-facing risk
question. A “worker extender” framing does two things simultaneously: it preserves the democratization and
access benefits the paper rightly wants to protect, while anchoring accountability and scope to an identifiable,
licensed human role rather than diffusing it. We recommend CDRH explore “worker extender” or comparable
scope-anchored certification models, and the CMS ecosystem-partner precedent for how a federal payer
program has already evaluated and integrated this type of certified AI extension into a program of record, as a
template for how FDA might structure patient-facing risk mitigation without resorting to blanket restriction of
patient-facing information.

From the CMS side, this scope-anchored model also aligns with how Medicare has historically drawn
accountability lines around licensed practitioners rather than the tools they use; a precedent CDRH may find
useful in framing this question.

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 5 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

Section V - Competency-Based Premarket Evaluation (Questions 7, 9, 11, 14, and 16)

Question 7 - Is device benchmarking followed by clinical confirmation a useful and appropriate
framework?

Yes, and we would go further: this structure mirrors, almost exactly, the pathway pharmacogenomic diagnostics
already travel - analytic/technical validation, followed by clinical validity established against guideline or
reference standards, followed by real-world clinical utility evidence. CDRH is not inventing a new evaluation
philosophy; it is correctly recognizing that GenAI-enabled devices should be held to the same layered
evidentiary discipline that diagnostic testing has used for years, adapted for open-ended inputs and outputs. We
recommend CDRH say so explicitly in any resulting guidance, both because it will aid sponsor understanding
and because it will allow CDRH to draw on decades of diagnostic evidentiary practice rather than building the
competency-based approach from first principles.

We would add one condition: the framework's value depends entirely on both stages scaling down, not just up,
with risk. For the lower and middle bands of the two-axis matrix, benchmarking alone - without a mandatory
clinical confirmation layer - should presumptively satisfy the competency-based approach. Reserving the full
weight of both benchmarking and clinical confirmation for the higher-risk bands keeps the framework
genuinely least-burdensome rather than uniformly burdensome and avoids imposing trial-level evidentiary cost
on functions that do not carry trial-level risk.

Question 9 - Would the benchmarking structure provide adequate evidence of clinical knowledge, safety
behavior, and generalizability? Are there missing or redundant elements?

The Safety, Clinical Proficiency, Generalizability, and Agentic AI Capability elements are well-organized, but
we recommend CDRH make explicit within the Clinical Proficiency category which elements are being used
to establish clinical validity (E.1, E.2, E.3 - is the underlying knowledge and reasoning accurate against
accepted clinical evidence) versus which are being reserved for clinical utility evaluation in Section V.C (does
use of the output change a decision or outcome). As currently structured, a sponsor could benchmark strongly
across E.1–E.4 and reasonably (but incorrectly) treat that as evidence of utility. We also recommend CDRH
add a benchmarking element addressing claims-evidence alignment - whether the strength of a device's outputfacing language (e.g., “recommend” vs. “consider” vs. general information) is calibrated to the actual strength
of the clinical evidence behind that specific function. In my experience, mismatches between claim strength
and evidentiary support are one of the more common and consequential failure modes in diagnostic and now
AI-enabled commercialization, and they are not obviously captured by S.1–S.3, E.1–E.4, or R.1–R.2 as
currently defined.

We also recommend CDRH add a benchmarking element for clinician usability and cognitive burden - human
factors testing conducted with front-line users, not accuracy benchmarking alone. Alert fatigue and override
behavior are well-documented failure modes for clinical decision support generally, and a device that
benchmarks well on Clinical Proficiency can still fail in deployment because it does not fit clinical workflow.

Question 11 - How might a sponsor select and justify a clinical confirmation approach proportionate to
risk? Are there device types for which the listed approaches would be insufficient?

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 6 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

Drawing on the Proove Biosciences experience building a large clinical-genetic databank in chronic pain, we
would encourage CDRH to recognize large, longitudinally maintained clinical-genetic and clinicaloutcomes databanks as a distinct, and in some cases superior, clinical confirmation approach - not only as a
supplement to premarket testing, but as an ongoing confirmation and subgroup-validation resource that
outperforms synthetic data for hard-to-sample populations, provided the databank's provenance, consent
structure, and adjudication independence are documented. Real, longitudinally collected clinical-genetic data
avoids the circularity risk the paper itself raises with synthetic data generated by models of the same class under
evaluation (see Question 13 of the discussion paper). We recommend CDRH explicitly list “large longitudinal
real-world clinical-genetic/clinical-outcomes databanks, independently maintained and governed” as a clinical
confirmation approach, positioned between retrospective evaluation and prospective clinical study in the paper's
rigor spectrum.

More broadly, we recommend CDRH establish a rebuttable presumption against requiring a prospective
clinical study for any device outside the highest-risk band of the two-axis framework, with the sponsor free to
offer one voluntarily but not required to default to it. Prospective studies are the single largest fixed cost in the
clinical confirmation spectrum and requiring them as a default - rather than an exception justified by genuinely
severe-consequence, high-autonomy use - would functionally price early-stage companies out of the moderaterisk middle of the market, leaving that space to incumbents best able to amortize trial costs across a larger
existing revenue base.

Question 14 - How should performance comparators and acceptance criteria be selected for open-ended
outputs? When should human-AI team performance, rather than the device alone, be the basis for
evaluation?

HumAInity's certification and deployment model - as a worker extender operating within a credentialed human's
scope of practice, rather than as a standalone diagnostic or decision-making agent - has given us direct operating
experience with the human-AI team question. In deployments where the device is functioning as an extender
of an identified worker's scope, we recommend CDRH treat human-AI team performance as the default
comparator and evaluation unit, not device-alone performance, because device-alone benchmarking in that
deployment model measures a configuration that will not exist in the field. Device-alone performance should
be reserved as the primary comparator for devices intended for fully autonomous, unsupervised use (the upperright of the two-axis framework).

We would also recommend the comparator selection be tied transparently to whichever configuration the
sponsor's labeling and intended use actually describe, so that a sponsor cannot benchmark in the easier
configuration (human-AI team) while marketing for the harder one (autonomous). Human-AI team
benchmarking should also measure override behavior and clinician trust calibration over time, not only at a
single point of certification but a device also that is initially well-calibrated can still drift toward over-reliance
or unwarranted disuse as clinicians gain experience with it.

Question 16 - What role should independent third parties play in benchmarking and clinical confirmation?

We would point CDRH to a directly relevant and very recent precedent: the regulation of laboratory-developed
tests (LDTs). For decades, clinical laboratories have developed, validated, and continuously monitored
complex, high-complexity diagnostic tests - including genomic and interpretive tests functionally comparable
in sophistication to many GenAI-enabled outputs - under CLIA's quality-system framework: laboratory
Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 7 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

certification, defined analytical and clinical validation protocols, routine proficiency testing, and independent
accreditation audit (CAP and equivalent bodies), all without categorical FDA premarket device clearance. In
American Clinical Laboratory Association v. FDA (E.D. Tex., March 31, 2025), the court vacated FDA's final
rule that would have regulated LDTs as medical devices, holding that “the text, structure, and history of the
FDCA and CLIA make clear that FDA lacks the authority to regulate laboratory developed test services,” and
that Congress vested that authority in CMS through CLIA, not in FDA through the device provisions of the
FDCA. FDA did not appeal that decision. The court further noted that the vacated rule had threatened to shutter
laboratories and would have cost patients access to countless cutting-edge tests - a direct illustration of the
innovation and access costs that follow when a premarket device-style gate is layered onto a field already
governed by a rigorous, quality-system-based oversight model. Congress is now moving to further strengthen,
not weaken, that CLIA-based approach: the Enhancing Clinical Laboratory Innovation and Access Act
(“Enhancing CLIA Act”), introduced May 19, 2026, proposes to modernize CLIA oversight of LDTs
specifically to strengthen innovation, increase transparency, and improve access.

We believe this is a genuinely instructive analog, not merely an adjacent precedent. CLIA has demonstrated,
over decades and now reaffirmed in court, that rigorous scientific and quality standards, enforced through
accreditation and continuous proficiency testing rather than case-by-case federal premarket clearance, can
sustain both patient safety and a competitive, innovating field of laboratories. For GenAI-enabled device
functions that are developed, validated, and deployed as part of a CLIA-accredited laboratory's own testing and
interpretive services - for example, an AI-enabled interpretive layer integral to a laboratory-developed test - we
recommend CDRH explicitly recognize existing CLIA/CAP accreditation, validation, and proficiency-testing
obligations as sufficient to satisfy some or all of the benchmarking and third-party evidentiary elements
described in Section V, rather than layering an additional, FDA-specific device review requirement on top of a
quality-system framework Congress and the courts have already affirmed as fit for this purpose. This would
extend, to laboratory-embedded GenAI functions, the same innovation-preserving logic that CLIA has provided
the broader diagnostics industry.

For GenAI-enabled functions outside that laboratory-embedded context, we would still encourage CDRH to
look to CLIA/CAP-accredited infrastructure - the kind Labarium operates - as a template for hosting sequestered
benchmarking datasets, conducting periodic re-benchmarking, and supplying independently adjudicated realworld data, rather than constructing a wholly new third-party accreditation apparatus for GenAI-enabled devices
from scratch. Separately, CMS's evaluation and integration of HumAInity as an Ecosystem partner is itself an
example of a federal program already performing independent, cross-agency assessment of a certified AI
healthcare worker extender; we encourage CDRH to coordinate with CMS on lessons learned from that review
rather than developing an FDA-only evaluation model in isolation. In all cases, we would caution CDRH against
making any third-party pathway a mandatory gate to premarket clearance. A voluntary, optional third-party
pathway - similar in spirit to the existing ASCA program, and consistent with the CLIA precedent above - lets
sponsors with the resources to use it benefit from streamlined review, without imposing the same fixed cost on
a sponsor that instead submits its own well-documented in-house benchmarking directly to CDRH. If CDRH
does encourage third-party involvement, we recommend it accompany that encouragement with a shared or
subsidized infrastructure model - pooled adjudicator panels and sequestered benchmark datasets accessible
on standardized, published terms - so that third-party review is a genuinely lower-cost option for small entities
rather than a de facto toll that only well-capitalized sponsors can afford.

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 8 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

Section VI – Post market Monitoring (Questions 18 and 19)

Question 18 - Under what conditions should CDRH accept greater premarket uncertainty in exchange for
greater reliance on post-market monitoring?

We recommend CDRH accept this trade-off as the default posture for any GenAI-enabled device outside the
highest-risk band of the two-axis framework, not merely as an occasional accommodation. Three considerations
support this:

• It is more consistent with the least-burdensome statutory principle CDRH itself cites, which asks CDRH
to select the minimum evidence necessary for a reasonable assurance of safety and effectiveness - not
the maximum evidence theoretically obtainable.
• It is arguably more protective in practice for GenAI-enabled devices specifically, because these devices
can change after deployment in ways traditional locked software does not; a robust, continuous postmarket monitoring program (periodic re-benchmarking, degradation monitoring, sample-based clinician
review) can detect and respond to a real-world problem faster than a premarket study that only
characterizes the device as it existed at a single point in time, months or years before deployment.
• It materially lowers the fixed cost and time-to-market for early-stage companies, which is where a large
share of genuine GenAI clinical innovation is likely to originate, without abandoning safety oversight -
it relocates that oversight to where it is both more current and less front-loaded.
CDRH does not need to design this mechanism from first principles. The Predetermined Change Control Plan
framework already gives sponsors a pre-authorized pathway for planned model changes, subject to a defined
monitoring and verification protocol. We recommend CDRH extend the PCCP framework, rather than a new
parallel mechanism, to serve as the vehicle for the prespecified post-market monitoring plan this discussion
paper proposes for GenAI-enabled devices. The characteristics that should justify reduced premarket evidence
are, in our view, largely the ones already implicit in the paper's own post-market monitoring section: a
prespecified, adequately resourced monitoring plan with defined cadence and triggering events, a credible
mechanism for rapid corrective action (including market withdrawal or use restriction) if degradation is
detected, and transparency to CDRH sufficient for the agency to audit the monitoring program itself. Devices
in the highest-risk band - fully autonomous, severe-consequence functions - are the appropriate exception where
CDRH's traditional premarket-weighted posture should continue to apply.

Question 19 - Comment on periodic re-benchmarking, sample-based clinician review, and performance
degradation monitoring. What additional approaches, cadence, and triggering events should CDRH
consider?

Labarium's multi-site laboratory infrastructure already performs the operational analog of what CDRH is
proposing here - continuous proficiency testing, QC monitoring, and multi-site performance reconciliation
across an expanding footprint of acquired laboratories, under existing CLIA obligations. We recommend CDRH
treat accredited laboratory quality-system cadences (routine proficiency testing intervals, triggered reverification after any material change to method, reagent, or platform) as a proven, auditable cadence model for
periodic device re-benchmarking of GenAI-enabled devices, particularly for devices that are themselves
deployed within, or alongside, a CLIA-accredited laboratory workflow.

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 9 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

We further recommend CDRH require sample-based chart review conducted by practicing clinicians
independent of the sponsor, not technical re-benchmarking alone, as a standing element of post market
monitoring. Existing quality-improvement infrastructure including AHRQ-listed Patient Safety Organizations
and the peer-review and quality-measurement frameworks that are already operated inside health systems offers
CDRH a ready adjudication model rather than one newly built for this purpose.

We also recommend CDRH treat commercial and compensation structures tied to a GenAI-enabled device's
outputs as a post-market monitoring risk category in their own right, distinct from technical performance
degradation. In pharmacogenomic commercialization, percentage-of-volume or percentage-of-sales
compensation arrangements tied to clinician-facing test orders present structural Anti-Kickback Statute
exposure regardless of how compensation is routed - including through a marketing intermediary - and,
separately from the legal exposure, they create incentives that can distort the very real-world performance data
a post-market monitoring program relies on. A GenAI-enabled device whose adoption or usage volume is tied
to compensated referral relationships may show performance or utilization signals that reflect commercial
incentive rather than clinical merit. We recommend CDRH require sponsors to disclose, as part of post-market
monitoring plans, whether compensation structures exist that could influence device utilization, and to account
for that possibility when interpreting post-market performance signals.

Section VII - Foundation Model MAFs and Agentic AI (Question 26)

Question 26 - What additional considerations apply to agentic GenAI-enabled devices, beyond non-agentic
devices, and how should elevated autonomy risk be reflected in acceptance criteria and oversight?

Through Arockia AI's clinical decision support work, we would encourage CDRH to anchor its agentic AI
oversight approach explicitly to the existing clinical decision support exclusion criteria under section
520(o)(1)(E) of the FD&C Act - in particular the independent-review criterion the paper itself references in
footnote 13. As agentic systems increasingly plan and execute multi-step action sequences, the practical
question for a deploying clinician or facility becomes whether a human retains a genuine, informed opportunity
to independently review and reject a planned action sequence before it executes, not merely whether a human
is nominally “in the loop.”

We recommend CDRH develop a bright-line operational test for what constitutes meaningful independent
review of an agentic action sequence - including minimum information the system must surface to the reviewing
human before an irreversible or high-consequence action - rather than treating “human-oversight checkpoint”
as self-defining. Small and rural facilities, which are Arockia AI's primary go-to-market channel, are the
deployment settings most likely to have thinner staffing available to perform that review function meaningfully,
and we recommend CDRH's acceptance criteria account for that reality.

We would also urge CDRH to hold the line on the device definition itself as agentic systems proliferate. Many
agentic healthcare AI functions - care coordination, documentation, scheduling, and administrative
orchestration - are, by the paper's own description, not necessarily functions that are the focus of FDA's device
regulatory oversight.

As these systems increasingly bundle device and non-device functions together, we recommend CDRH
reaffirm, rather than erode, the existing multiple-function device product framework: the presence of agentic
orchestration around a non-device function should not, on its own, pull that function into device regulation.

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 10 of 11
 PROFOUND VENTURES • A Social Impact Venture Studio

This distinction matters disproportionately to early-stage companies, which are more likely than incumbents to
be building narrow, single-purpose agentic tools that never touch a device function at all, and which cannot
absorb the cost of a premarket submission triggered by regulatory ambiguity rather than by genuine patient risk.

Conclusion and Offer of Collaboration

CDRH's discussion paper correctly recognizes that GenAI-enabled devices require an evidentiary approach
inspired by, but adapted from, how the healthcare system already evaluates clinical evidence and credentials
human clinicians. We would encourage CDRH to draw that inspiration not only from clinician credentialing,
but from the layered analytic validity / clinical validity / clinical utility discipline that diagnostic and genomic
testing has already developed and been regulated under for years.

At the same time, we would ask CDRH to hold itself to the same discipline it is asking of sponsors: match the
level of evidence and oversight to the level of risk, no more and no less. A framework that is rigorous where
risk is genuinely high and light-touch where it is not will do more to protect patients, and more to sustain a
competitive field of innovators capable of bringing safe GenAI-enabled devices to market, than one that applies
uniform, maximal evidentiary burden regardless of actual risk.

We would welcome the opportunity to make Labarium's multi-site CLIA/CAP laboratory infrastructure, Proove
Biosciences' and Proove Genomics' historical and ongoing clinical-genetic evidence generation, HumAInity's
certified worker-extender deployment and CMS Ecosystem experience, as well as Arockia’s ability to interject
and measure quality outcomes at the point-of-care available to CDRH as a case study, pilot partner, or public
workshop participant as this framework develops further.

Thank you for the opportunity to comment on this important discussion paper. We are available to provide
additional detail on any of the above at CDRH's convenience.

Respectfully submitted,

Brian Meshkin
Managing Partner, Profound Ventures
Chairman, President & CEO, Labarium Diagnostic Systems, Inc. | Co-Founder, HumAInity | Co-Founder,
Arockia AI, Inc.
Destin, Florida

Anita Monteiro, RN, MA, MSHCA, MBA, ACC
CEO & Co-Founder, Guidance Global Consulting
Co-Founder, Arockia AI, Inc.
Advisory Board Member, HumAInity | Proove Genomics | Labarium Diagnostic Systems, Inc.
Fulton, Maryland

Profound Ventures | Comment on Docket: FDA-2026-Considerations for Regulation of GenAI-Enabled Medical Devices | Page 11 of 11