Blaine Warkentine, MD

Thank you for issuing this discussion paper and inviting early input. I am a physician who has spent 20+ years at the boundary between clinicians and health technology, and I now build and operate software through which licensed physicians review and attest to AI-generated clinical outputs. The through-line of these comments is one distinction the paper is already reaching for: the question is not only how accurate a generative model is, but who is accountable for the consequences of acting on its output. Accuracy is a property of a model; accountability is a property of a person.

1. Endorse the two-axis risk framework - and make a named human-in-the-loop and reversibility first-class, risk-LOWERING variables, not only a way to sort risk upward. Inserting a qualified human who reviews the basis for an output and signs or declines it before it becomes consequential moves a device DOWN the effective-risk surface, the mirror image of autonomy moving it up. Two devices with identical capability should not carry identical risk if one routes every consequential output through an accountable licensed reviewer and the other acts alone.

2. Make the clinical-decision-support boundary explicit. Connect the lower risk of HCP-facing use to the independent-review criterion at section 520(o)(1)(E): when a qualified professional can independently review the BASIS for a recommendation before acting, the device occupies a distinct, lower-risk class. Naming that "name-the-human" pattern gives developers a clear, safety-promoting target.

3. In competency-based evaluation, score the human-AI TEAM as deployed - and be precise that the human’s role is accountability, not higher accuracy. Adding a human to an already-strong model does not reliably raise task accuracy and can lower it on some decision tasks (Vaccaro et al., Nature Human Behaviour, 2024). The inference is not "remove the human" but that the human’s function in a regulated device is ownership of and answerability for the consequence, which no benchmark confers. Benchmark the model honestly on its own; let clinical confirmation measure the team as deployed. The paper’s escalation and clinical-deferral competencies are exactly right.

4. Postmarket "shared-ecosystem responsibility" works only with a portable, tamper-evident record of who reviewed what and what they decided. Periodic re-benchmarking, sample-based clinician review, and drift monitoring function only if a durable, privacy-preserving record of the human decision travels with the determination. One workable pattern: each reviewed output becomes a signed, timestamped, PHI-free attestation whose fingerprint is anchored to an external tamper-evident ledger, so the determination and its sign-off can be verified after the fact without exposing patient data.

5. For agentic systems, anchor oversight to whether a named, licensed human signs the consequential action - not to autonomy in the abstract. At each point where the system takes a consequential or hard-to-reverse action, is there a qualified human who reviews the basis and is accountable for it? That checkpoint scales across architectures and is more robust to prompt injection than autonomy labels. The paper’s principle - that a foundation-model master file is not a device approval, and the sponsor deploying the device remains responsible - is exactly right and should stay central.

A great deal of what a generative model does (retrieval, drafting, summarization, triage) is well suited to automation, and the framework should keep that path light. The narrow, load-bearing exception is the consequential clinical or coverage decision, where the durable requirement is not a better score but an accountable, licensed human who reviews the basis, signs, and answers for the outcome. I would be glad to provide further detail, including the attestation-and-ledger pattern in point 4.

Respectfully submitted,
Blaine Warkentine

Comment on Docket FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices — Discussion
Paper and Request for Feedback

To: Digital Health Center of Excellence, Center for Devices and Radiological Health, U.S. Food
and Drug Administration
Submitted via: Regulations.gov, Docket No. FDA-2026-N-7874
From: Blaine Warkentine, MD, MBA
Date: August 27, 2026

Thank you for issuing this discussion paper and for inviting early input. I am a physician
(MD, Medical College of Wisconsin; MBA, University of Utah) who has spent more than
twenty years at the boundary between clinicians and health technology — including
building the orthopedic vertical of an image-guided surgical-navigation company
(BrainLAB) to roughly $250M, and named inventor on patents in image-guided
navigation. I now build and operate production software through which licensed
physicians review and attest to AI-generated clinical outputs. I write in support of the
paper's central instincts, and to offer five refinements grounded in systems I actually
run.

The through-line of these comments is a single distinction the paper is already reaching
for: the question is not only how accurate a generative model is, but who is
accountable for the consequence of acting on its output. Accuracy is a property
of a model; accountability is a property of a person. The framework will be strongest
where it keeps those two separate and builds around the second.

1. Endorse the two-axis risk framework — and make a named human-inthe-loop and reversibility first-class, risk-lowering variables, not only
a way to sort risk upward. The Activity axis rightly treats fully-autonomous
action as higher-risk than an informational, non-directive output. We encourage
the framework to state explicitly the mirror image: inserting a qualified human
who reviews the basis for an output and signs — or declines — it before it
becomes consequential is a mitigation that moves a device down the effective-risk
surface, just as autonomy moves it up. Two devices with identical raw capability
should not carry identical regulatory risk if one routes every consequential output
 through an accountable licensed reviewer and the other acts on its own. The
reversibility of the action deserves the same first-class treatment. Making these
explicit gives developers a concrete, safety-promoting design target rather than a
penalty to be discovered late.
2. Make the clinical-decision-support boundary explicit in the
framework. The paper notes that HCP-facing use lowers risk because a clinician
can catch an incorrect output. We encourage CDRH to connect this directly to the
statutory independent-review criterion at section 520(o)(1)(E): when a device is
designed so that a qualified professional can independently review the basis for
its recommendation, and does so before acting, it occupies a distinct, lower-risk
class. Naming that "name-the-human" design pattern in the framework would let
developers build toward a clear line, rather than discovering it case by case, and
would reward exactly the architecture the paper's safety logic favors.
3. In competency-based evaluation, score the human–AI team as it will
be deployed — and be precise that the human's role is accountability,
not a guarantee of higher accuracy. The analogy to how clinicians are
trained and credentialed is apt. We add one caution from the evidence: adding a
human reviewer to an already-strong model does not reliably raise task accuracy,
and on some decision tasks can lower it (Vaccaro et al., Nature Human
Behaviour, 2024). The right inference is not "remove the human" — it is that the
human's function in a regulated device is ownership of and answerability for the
consequence, which no benchmark score confers. So benchmark the model
honestly on its own, but let clinical confirmation measure the team as actually
deployed. The paper's safety competencies for recognizing safety-critical
situations and escalating, and for calibration, uncertainty, and clinical deferral,
are exactly right: a device that reliably knows when to hand off to a human is
demonstrating the safety-critical behavior, not failing to be autonomous.
4. Postmarket "shared-ecosystem responsibility" works only with a
portable, tamper-evident record of who reviewed what and what they
decided. The paper's call for periodic re-benchmarking, sample-based clinician
review, and drift / performance-degradation monitoring is well-placed, and
distributing responsibility across manufacturers, clinicians, and institutions is
realistic. In practice it functions only if a durable, privacy-preserving record of the
human decision travels with the determination. We build and operate one such
pattern: each reviewed output becomes a signed, timestamped, PHI-free
attestation whose fingerprint is anchored to an external, tamper-evident ledger,
 so a determination and its human sign-off can be independently verified after the
fact without exposing patient data. We offer the pattern, not a product, as one
concrete way to make shared postmarket responsibility auditable, and would
welcome the opportunity to share the design with the Center in detail.
5. For agentic systems, anchor oversight to whether a named, licensed
human signs the consequential action — not to autonomy in the
abstract. The paper is right to give agentic AI elevated scrutiny for its "reduced
opportunity for human review." We suggest the operative test be functional and
local: at each point where the system takes an action that is consequential or
difficult to reverse, is there a qualified human who reviews the basis and is
accountable for it? Oversight tied to that checkpoint scales across architectures
and is more robust to prompt injection than autonomy-level labels alone. The
paper's own principle — that a foundation-model master file is not a device
approval, and that the sponsor deploying the device remains responsible for it —
is exactly right, and we encourage CDRH to keep it central: the party who deploys
the AI, and the licensed human who signs the consequential output, are where
accountability must sit.

These five points are one argument seen from five sides. A great deal of what a
generative model does — retrieval, drafting, summarization, triage — is genuinely well
suited to automation, and the framework should keep that path light. The narrow, loadbearing exception is the consequential clinical or coverage decision, where the durable
requirement is not a better score but an accountable, licensed human who reviews the
basis, signs, and answers for the outcome. A framework that names that human
explicitly — as a risk-lowering design element, as a distinct regulatory class, as the unit of
clinical confirmation, as the postmarket record, and as the agentic checkpoint — will age
well as the models keep improving.

I appreciate the Center's leadership in seeking input this early, and I would be glad to
provide further detail on any of the above, including the attestation-and-ledger pattern
described in point four.

Respectfully submitted,

Blaine Warkentine, MD, MBA
Reference cited: Vaccaro, Almaatouq & Malone, "When combinations of humans and AI are useful: A
systematic review and meta-analysis," Nature Human Behaviour, 2024. These comments reflect the
author's own views and describe design patterns in general terms; they are not legal advice and do not
disclose patient information.