David Liebovitz

Please see the attached comment letter regarding the discussion paper on generative AI-enabled medical devices. — David Liebovitz, MD, Professor of Medicine, Northwestern University Feinberg School of Medicine (personal capacity).

Re: Docket FDA-2026-N-7874 — Considerations for the Regulation of Generative
AI-Enabled Medical Devices

I submit this comment in my personal capacity as an academic internal medicine
physician whose work is in clinical informatics and the evaluation of AI in clinical
care, including research on bias and stigmatizing language in AI- and
clinician-generated clinical text. I support the discussion paper's two central moves —
a function-specific, two-axis risk framework and a competency-based evaluation
model applied to the deployed device rather than the foundation model alone — and
offer six specific suggestions.

1. Add detectability and reversibility as explicit modifiers to the two-axis
framework.
The two axes capture how independent the function is and how bad an error would
be, but not whether anyone can catch the error or undo it. A measurement output a
clinician cannot independently verify, and an action whose effect cannot be reversed,
are risks of a different kind, not just degree. Making detectability and reversibility
named modifiers — as several early commenters have also urged — would keep the
framework two-dimensional while preventing its most dangerous quadrant from being
understated. This is also familiar ground: detectability and reversibility sit naturally
within failure mode and effects analysis (FMEA), where risk priority numbers (RPNs)
are computed from severity, occurrence, and detectability. Grounding the modifiers in
FMEA gives sponsors and reviewers a shared, quantitative vocabulary for ranking
failure modes rather than treating them qualitatively and inconsistently.

2. Benchmarking should include a language-and-equity competency, reported
by subgroup.
A GenAI device can be clinically proficient on average while producing stigmatizing,
dismissive, or lower-quality language for identifiable patient subgroups, and while
performing unevenly across the ways different patient populations describe
symptoms. "Generalizability" should expressly include (a) performance stratified by
relevant patient subgroups and input styles, and (b) evaluation of generated language
for stigmatizing or bias-carrying content, using published lexicons and clinician
adjudication. Average-case benchmarking will not surface these failure modes; they
are precisely the ones sample-based postmarket review is worst at catching after the
fact. In our own internal analyses of clinical documentation, we have found
bias-carrying and stigmatizing language in both human-written and AI-generated
notes. The problem is not unique to AI — but AI can reproduce it at scale and with an
unwarranted appearance of neutrality, which is why generated-language evaluation
belongs in premarket benchmarking rather than being left to postmarket discovery.

3. Premarket evaluation should include adversarial stress-testing in
simulated complex clinical contexts.
Benchmarking on representative cases measures average performance; deployment
failures tend to come from the tails — ambiguous presentations, conflicting or
erroneous data, deteriorating patients, time pressure, and users who over-trust the
output. Before deployment, a GenAI device should be stressed with the toughest
simulated clinical scenarios its designers can construct, including adversarial cases
built to elicit failure. This is now practical in a way it was not a few years ago:
realistic simulated clinical datasets and encounters can be generated at scale, so
systems can be pushed to their limits without exposing patients to risk. Such testing
will never be complete — no finite set of scenarios exhausts clinical reality — but it is
hard to justify real-world use of a system that has not first been stress-tested in
simulation to identify its gaps. FDA should expect sponsors to report the classes of
scenarios tested, the failure modes found, and how the device behaves when it
reaches its limits.

4. Assess risk at the level of the conversation, not the turn.
FDA rightly flags multi-turn conversations that migrate from informational to
action-directing. The regulatory unit should match: a function that is low-risk at turn
one can become action-taking by turn ten as context accumulates and the user defers.
Benchmarking should include scripted multi-turn trajectories designed to test
whether the device recognizes that migration and escalates its own caution — refusal,
uncertainty communication, or handoff — as the conversation crosses into higher-risk
territory.

5. The unit of analysis for clinical confirmation should presumptively be the
human–AI team, compared against care without the device.
The clinically meaningful question is rarely whether the model outperforms a clinician
in isolation; it is whether care delivered with the device is better than the care that
patient would otherwise have received — including delayed specialist input or no
intervention. Clinical confirmation designs should state the comparator explicitly, and
where the intended user is a generalist, confirmation should be conducted with
generalist users, since the same output carries different risk in generalist and
specialist hands.

6. If postmarket monitoring offsets premarket uncertainty, require a
detectability statement — and govern the adjudicators.
Trading premarket evidence for postmarket monitoring is defensible only for failure
modes the monitoring can actually detect. Sponsors should be required to state,
before authorization, which failure modes each monitoring mechanism can and
cannot detect, and the trade-off should not be available where plausible harm is
severe and irreversible. For sample-based clinician review, FDA should specify the
adjudication infrastructure the paper currently leaves open: how adjudicators are
credentialed, what conflicts disqualify them, and whether review operates under FDA
or institutional oversight. On supervisory agents: a machine supervisor should be held
to the same competency domains as the device it monitors, evaluated on sequestered
datasets it cannot have been trained against — an unevaluated supervisor converts
monitoring into theater. For my part, I would be glad to assist with AI output review,
including serving as a clinician adjudicator, as FDA develops this infrastructure.

I appreciate CDRH undertaking this while the framework can still be shaped, and I
am glad to provide further detail on any of the above.

David Liebovitz, MD
Professor of Medicine, Northwestern University Feinberg School of Medicine
*(Submitted in a personal capacity; institutional affiliation listed for identification
only.)*