Christian Péan

Submitted by Christian A. Péan, MD, MS
Orthopaedic surgeon; Co-Founder and CEO, RevelAi Health
To the Food and Drug Administration:
I appreciate FDA’s invitation to comment on the regulation of generative AI-enabled medical devices. I support the discussion paper’s two-axis framework, which considers device activity and the consequences of an incorrect output, and its exploration of competency-based evaluation.[1] My central recommendation is to establish a practical pathway through which clinical AI systems can earn graduated autonomy, with explicit performance limits, meaningful patient disclosure, and verifiable clinical accountability.
I submit these comments in my personal capacity as an orthopaedic surgeon and healthcare technology founder. My leadership and financial interest in RevelAi Health are relevant to this topic. These views do not represent Duke University, Duke Health, or another institution.

A pathway to graduated clinical autonomy
Comment on FDA’s generative AI medical-device discussion paper
Docket FDA-2026-N-7874 | September 18, 2026

Submitted by Christian A. Péan, MD, MS
Orthopaedic surgeon; Co-Founder and CEO, RevelAi Health

To the Food and Drug Administration:

I appreciate FDA’s invitation to comment on the regulation of generative AI-enabled medical devices. I
support the discussion paper’s two-axis framework, which considers device activity and the consequences of
an incorrect output, and its exploration of competency-based evaluation.[1] My central recommendation is
to establish a practical pathway through which clinical AI systems can earn graduated autonomy, with explicit
performance limits, meaningful patient disclosure, and verifiable clinical accountability.

I submit these comments in my personal capacity as an orthopaedic surgeon and healthcare technology
founder. My leadership and financial interest in RevelAi Health are relevant to this topic. These views do not
represent Duke University, Duke Health, or another institution.

1. Recognize the opportunity for substantive clinical care
In my opinion, large language models have reached a level of capability that makes near-autonomous
delivery of selected elements of clinical care a credible, immediate development objective. In
musculoskeletal care, this includes managing routine perioperative concerns, adjusting therapeutic exercise
within an established plan, following chronic low back and large-joint pain, and coaching patients on
mobility, fall prevention, and bone health. This is my assessment of the technology’s potential, not a claim
that any particular product has established safety or effectiveness for these uses.

A system with a clearly defined clinical role, persistent and verified patient context, reliable follow-up, and
access to appropriate clinical tools could function as a clinical AI associate. Its contribution could extend
beyond answering questions to carrying out portions of a longitudinal care plan. A convincing clinical persona
is insufficient: competence must be demonstrated in the complete system, including its memory, scheduling,
tool use, and escalation workflow.

FDA should provide a route for such systems to progress from clinician-reviewed recommendations to
independently completing validated tasks. The permissible scope should be specific to the function,
population, and deployment setting. Progression should depend on evidence; a general label such as
“assistant” should neither establish safety nor prevent an appropriately evaluated system from providing
meaningful care.
2. Define risk-specific error limits and hallucination thresholds
FDA should convene clinicians, patients, statisticians, and developers to establish clinically justified ranges of
acceptable residual error for defined categories of use. These should be measured acceptance limits, not
assumed tolerances. A single acceptable hallucination percentage across all clinical applications would be
misleading. An inaccurate educational detail, an invented medication history, and false reassurance about a
possible surgical complication have different consequences.

Sponsors should prespecify an error taxonomy distinguishing unsupported or fabricated claims; incorrect
recommendations; omissions; failures to escalate; unnecessary escalation; and failures to execute an
intended action. Hallucination should be defined operationally as an assertion presented as factual that is
unsupported by, or contradicts, the relevant verified record or authoritative evidence. Clinically reasonable
disagreement should be adjudicated separately.

For each consequential function, I recommend reporting: the observed error rate and confidence interval;
the severity and reversibility of potential harm; the frequency with which an error reaches a patient or
triggers an action; and the proportion detected and corrected before harm. Reports should distinguish errors
generated before safeguards from residual errors after safeguards, without double-counting overlapping
categories.

Denominators matter. Assertion-level accuracy can conceal a hazardous patient journey containing many
individually accurate statements. Evaluation should therefore include rates per clinically consequential
recommendation, per conversation, and per patient episode over a specified period. Repeated observations
within patients should be accounted for statistically. Rare but serious failures require targeted testing and
sufficient exposure; zero observed events does not establish zero risk.

Acceptance criteria should include justified upper confidence bounds for serious errors, separate limits for
missed escalation, and minimum performance within relevant patient subgroups. Common low-consequence
errors and rare catastrophic failures should not be averaged into a reassuring overall score. Sponsors should
define sentinel failures that trigger immediate investigation and suspension of the affected function,
alongside rules for restoring it.

I would not propose unsupported numerical cutoffs in this comment. FDA should publish worked examples
explaining how to derive thresholds from clinical consequences, exposure, comparator performance, and
uncertainty. Such examples would give developers actionable targets while allowing evidence-based
differences between applications.

3. Compare the complete care service with realistic alternatives
Evaluation should consider both clinical quality and whether care is actually delivered. For longitudinal MSK
care, useful endpoints include pain and function, completion of appropriate care, timely reassessment,
patient understanding, clinician workload, and avoidable utilization. Comparisons should reflect the intended
setting and an appropriate clinical standard. Limited access to existing services is relevant to benefit-risk
assessment, but should not justify weaker safety standards for underserved patients.
4. Make verification an operational requirement
Verification should occur before consequential advice or action, after execution, and during ongoing
surveillance. The necessary verification should increase with the potential harm and the independence of the
system. A second model agreeing with the first is not, by itself, independent confirmation; both may share
the same error.

Before action, the system should verify patient identity, the current care plan, relevant restrictions, and the
freshness and provenance of decision-critical information. Patient reports, imported records, and model
inferences should remain distinguishable. Conflicting or missing information should trigger clarification or
clinical review when it could change the decision. Retrieval of a document is not enough: the cited evidence
must actually support the recommendation.

For exercise adjustment, verification could include checking the proposed change against the prescribed
program, functional status, symptom trajectory, and applicable precautions. Changing repetitions within a
validated progression protocol differs from changing postoperative weight-bearing restrictions. Bone-health
medication initiation, interruption, substitution, or dosing decisions should have a distinct authorization and
verification pathway; success at exercise coaching should confer no medication-management authority.

After an action, the system should confirm the intended result. A reminder does not establish medication
administration; a referral request does not establish an appointment; an escalation message does not
establish clinician receipt or response. Each consequential task should have an accountable owner, a
completion state, a response deadline appropriate to urgency, and a fallback when the expected response
does not occur.

The audit record should preserve the relevant input sources, system version, applicable protocol,
recommendation or action, verification result, and human intervention. Reviewers need an inspectable
clinical basis and execution history, rather than reliance on a model’s account of its own reasoning.

5. Evaluate continuity, memory, and changing autonomy
A longitudinal system should be tested across weeks and months of care, including missed contacts,
conflicting clinician instructions, new diagnoses, treatment changes, and transitions between settings. Tests
should examine whether an earlier error persists in memory and influences later decisions, whether
corrected information replaces outdated assumptions, and whether scheduled follow-up reliably occurs.

Autonomy should decrease when verified context becomes inadequate, symptoms change materially, or the
system encounters an unfamiliar situation. Conversely, evidence may support independent execution of
specific tasks for an eligible patient. Repeated human approval of every low-risk interaction should not
become a default substitute for validation. Human oversight must be feasible in practice, with measured
response capacity and explicit responsibility for unresolved concerns.
6. Require patient disclosure that supports informed use
Patients should receive a clear explanation that they are interacting with AI, what it is authorized to do, when
a clinician reviews its work, what information it uses, and how to reach a person. Disclosures should explain
that errors and incomplete context are possible, identify service availability and response expectations, and
state the appropriate route for urgent concerns. These disclosures should recur when the system’s role
changes materially.

For older adults, disclosure should work in a telephone conversation as well as on a screen. Testing should
include comprehension across levels of health and technology literacy, language, hearing, vision, and
cognition. Brief teach-back and consented caregiver participation can make disclosure meaningful. Patients
should be able to correct their history and understand how their information is retained and used.

A disclaimer cannot compensate for unsafe functionality or shift responsibility for detecting errors to the
patient. A system that directs treatment should accurately describe that role. Boilerplate language suggesting
that it provides no clinical advice would be confusing if its actual function is to guide a patient’s care.

7. Link continued autonomy to monitored performance
The permitted level of autonomy should remain contingent on performance after deployment. Sponsors
should define monitoring intervals, independent clinical review, thresholds for restricting a function, and
procedures for notifying affected patients and clinicians when an error changes care. Monitoring should
include missed care and failed follow-up, alongside incorrect content and adverse actions.

Changes to the underlying model, prompts, retrieval sources, memory logic, exercise protocols, and
connected tools should undergo testing proportionate to their effect on clinical behavior. A modification that
expands medication authority or changes escalation behavior warrants more scrutiny than a cosmetic
interface change. Material performance deterioration should trigger reassessment, with a documented route
to rollback or reduced autonomy.

I encourage FDA to develop practical examples for chronic disease management and perioperative follow-up,
including exercise progression, medication continuity, and escalation of changing symptoms. These would
help translate the framework into an achievable development and evaluation pathway.

The regulatory objective should be to enable clinical AI to assume meaningful responsibility when the
evidence supports it, while making the limits of that responsibility explicit and enforceable. Patients stand to
benefit from care that is more continuous, responsive, and accessible. Clear error limits, credible verification,
and graduated autonomy would help make that promise testable and accountable.

Respectfully,
Christian A. Péan, MD, MS

Reference
[1] U.S. Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback. Sections IV–VI. Discussion paper; not draft or final
guidance. https://www.fda.gov/media/194242/download