Nolla Health (Sean Geiger)

Please see the attached comments from Nolla Health (Magic Health, Inc.) regarding FDA’s discussion paper, “Considerations for the Regulation of Generative AI-Enabled Medical Devices,” Docket No. FDA-2026-N-7874.

We appreciate the opportunity to provide feedback and welcome continued discussion with the Agency.

October 18, 2026
Submitted electronically via Regulations.gov

Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative
AI-Enabled Medical Devices: Discussion Paper and Request for Feedback

Dear FDA Review Team:
Generative AI has the potential to improve access to care and the quality of care patients
receive. FDA plays an essential role in helping patients benefit from these technologies
safely. Nolla Health appreciates the opportunity to comment on the Discussion Paper and
welcomes continued collaboration with the Agency.[1]
Nolla provides dermatology care in more than 40 states through a clinician-reviewed care
model. Licensed clinicians review treatment recommendations and make prescribing
decisions. Our experience using AI to support this care informs our interest in how
increasingly capable models may take on greater responsibility for defined clinical tasks,
supported by appropriate evidence, safeguards, and oversight.
Nolla has worked with the Utah Office of Artificial Intelligence Policy (OAIP) to develop an
oversight framework for a pilot of AI-initiated topical acne prescribing for eligible patients in
Utah age 18 or older. Under a Regulatory Mitigation Agreement, OAIP authorized a staged
pilot with a defined patient population, treatment options, eligibility checks, escalation rules,
and human-review requirements.[2] We offer this experience as one example of how a
regulator and developer have worked through these questions. Particular elements may be
useful to FDA's discussion.
Our comments address selected questions about clinical risk, evidence, supervision,
monitoring, and model updates.

I. Assess the deployed clinical function
Questions 1–3 and 26

FDA's two-axis framework provides a useful starting point. We recommend using it to inform
a function-specific analysis of the severity and probability of harm, the effectiveness of tested
controls, and the residual risk after mitigation. Autonomy affects how failures arise and
whether they can be intercepted; it does not, by itself, establish the severity of harm.
Reversibility, time to harm, detectability, and access to corrective care also matter. A
bounded autonomous function may present less risk than an advisory function whose
incorrect output could cause serious injury.
The assessment should address the finished clinical function, including its eligible
population, permitted recommendations and actions, exclusions, escalation rules, and the
role and timing of human review. Capabilities of the underlying foundation model should not
be attributed to a product that does not expose them. Conversely, references to guardrails or

Nolla Health · FDA-2026-N-7874
human oversight should reduce the assessed risk only when evidence demonstrates that
those controls work in the intended setting.
Output wording is only part of the assessment. A system may present a suggestion that
users routinely follow without meaningful review. Another may issue a directive but be
technically limited to a small set of permitted actions. Sponsors should explain how scope is
enforced, whether prohibited recommendations or tool calls can pass through, and how the
product defers safely when it cannot complete a task.
Patient-facing status warrants attention to comprehension, health literacy, accessibility,
overreliance, and access to human assistance. It should modify the assessment rather than
automatically determine the risk tier. Clinician-facing systems also require evidence about
actual use, including time pressure and the ability to evaluate an output independently.
Ordinary labeling and consent materials should explain the system's clinical role, limitations,
and human alternatives; AI alone should not trigger a separate consent regime. A
patient-facing system that initiates treatment should generally provide a practical route to
human clinical review, with response time proportionate to urgency and potential harm.
The same approach applies to agency. A system selecting one permitted treatment within a
fixed protocol differs from one that develops open-ended plans, orders tests, and changes
treatment over time. Broader agency requires evaluation of the entire action sequence,
including tool permissions, patient identity, memory, duplicate actions, interruption,
resistance to prompt injection and tool manipulation, and recovery from external-service
failures. Additional requirements should follow from these capabilities and their
consequences, rather than from an “agentic” label alone.

II. Require evidence that reflects clinical use
Questions 5–10 and 12–16

Nolla supports combining competency-based benchmarking with clinical confirmation.
Benchmarks should establish the capabilities required for the intended use; clinical
confirmation should establish whether those capabilities translate into safe and effective use
in the intended workflow. A product should satisfy its benchmark gates before progressing to
clinical confirmation, and strong benchmark results should not offset poor clinical
performance. Evaluation should cover the deployed configuration, including post-training,
prompts, retrieval, tools, interface, and restrictions. This approach is consistent with the
emphasis on clinically relevant testing and human–AI interaction in IMDRF's Good Machine
Learning Practice principles.[3]
We support a common set of competencies while allowing benchmarks suited to the
intended use. Sponsors may draw on qualified external resources, independent evaluations,
and intended-use-specific tests. FDA should assess whether the evidence measures the
claimed competency and represents the relevant population and workflow. A single
mandatory benchmark would become a training target and could lose value through
contamination, saturation, or changes in clinical practice. Independent evaluation should
provide credible assurance without unnecessary cost or delay. FDA should permit multiple

Nolla Health · FDA-2026-N-7874
qualified evaluation routes and avoid making a small group of designated providers the
exclusive option.
Safety-critical competencies need separate, prespecified acceptance criteria. Strong
average performance should not compensate for unacceptable failures involving
contraindications, emergencies, prohibited actions, or scope restrictions. Criteria should
reflect the clinical consequences of failure and statistical uncertainty. For judgments that
admit reasonable alternatives, disagreement with a reference clinician should not
automatically count as error. Controls intended to prohibit an action should be tested against
that prohibition, rather than evaluated only through an overall accuracy score.
Conversation testing should follow the clinical purpose and boundaries of the product. A
triage conversation may appropriately become more directive as it gathers information; an
out-of-scope recommendation remains a failure. Complete encounters should be evaluated
for missed new information, persistent incorrect assumptions, contradictions, unsafe
reassurance, and instructions issued before a later correction. A safe final response should
not erase a harmful intermediate output. Where memory or repeated use is part of the
intended workflow, testing should cover those interactions. FDA's human factors guidance
similarly emphasizes evaluating sequential tasks within realistic workflows.[4]
Escalation should be assessed by trigger and error direction. Missing an emergency,
unnecessarily referring a routine case, and failing to respond to a patient request for human
review have different consequences. Sponsors should define clinical reference standards
and acceptable error rates for each material trigger, with uncertainty estimates and
appropriate prevalence-sensitive measures. Rare-event stress tests should be distinguished
from estimates of routine performance. Clinical confirmation should also test handoffs,
latency, incomplete inputs, and operational failures, because detecting a need for escalation
does not by itself establish that the care pathway works.
Credible benchmarking requires prespecified endpoints, separation of development and
evaluation data, assessment of contamination, and qualified clinical adjudication. Statistical
plans should account for related outputs within an encounter, repeated patients, missing or
indeterminate results, and adequate subgroup representation. Pivotal results should be
auditable. If an LLM grades clinical outputs, its judgments should be validated against
independent qualified human review; agreement between models or an absence of
self-preference does not establish clinical validity.
Synthetic cases can help test rare contraindications, emergencies, and boundary conditions.
Sponsors should explain how cases were generated, how realism was assessed, and
whether related models created both the test material and evaluated outputs. Results from
real, synthetic, and semi-synthetic cases should be reported separately. Any pooled estimate
should disclose its weighting and rationale, and material discrepancies should be
investigated. Synthetic evidence should not ordinarily be the sole clinical confirmation for a
patient-facing system that takes clinical action. Generating more examples of an
underrepresented population does not establish that those examples reflect that population's
clinically relevant characteristics.

Nolla Health · FDA-2026-N-7874
The comparator should match the claimed clinical role and use comparable information. A
generalist task calls for appropriately qualified generalists; a specialist claim requires the
relevant specialist comparison. When the intended use includes clinician review, evaluation
should consider the clinician–AI workflow. Acceptance criteria should distinguish unsafe
advice from clinically acceptable alternatives and consider errors that clinical reviewers
themselves may make. Sponsors should report the full distribution of clinician performance,
not only a mean or median. Two acne regimens may both be appropriate for the same
patient; disagreement about the selected treatment need not imply error. Clinical
equivalence should reflect the patient's circumstances and prespecified criteria, not simply
membership in the same drug class.
Access can be part of the benefit assessment. Credible comparisons may include delayed or
forgone care when those are realistic alternatives for the intended population. Such
comparisons should use defensible evidence and should not lower the safety floor or require
individual patients to prove that conventional care is unavailable. The objective is to measure
the care the product can improve while preserving appropriate standards for the clinical
function it performs.

III. Permit evidence-supported changes in supervision
Questions 11, 18, and 21

For bounded functions, we encourage FDA to consider clinical confirmation through
supervised stages. Review could begin before a clinical action, move to complete
retrospective review, and later use risk-based sampling. Each transition should depend on
prespecified evidence, with the ability to return to more intensive supervision or pause the
function. The required evidence should reflect the possible harm, time available to intervene,
reversibility, population, and effectiveness of the safeguards.
The Utah protocol provides a concrete example. Its first stage requires clinician review
before prescription submission. Later stages may change the timing and frequency of review
after defined criteria are met and OAIP provides the required prior written approval.[2] The
patient population, first-line topical treatment options, contraindications, and escalation rules
remain fixed. Progression changes the supervision of a bounded clinical function; it does not
permit unrestricted prescribing or expansion into other conditions.
We recommend that an authorized staged plan specify the evidence required for
progression, the deployed configuration being evaluated, the permitted changes during
evaluation, and the thresholds for increasing supervision or pausing. The progression
package should explain performance against the agreed criteria, relevant subgroup findings,
updates made during the supervised period, and unresolved safety signals. Review can then
focus on whether those conditions have been met. Monitoring, escalation, adverse-event
reporting, and change control should continue after a transition.
A planned reduction in review within an authorized pathway should follow its evidence and
approval conditions. It should be distinguished from an increase in autonomy outside that
pathway, even when the clinical indication stays the same.

Nolla Health · FDA-2026-N-7874
Greater reliance on postmarket evidence still requires an adequate premarket foundation.
The sponsor should demonstrate the required competencies and the operation of its
safeguards before relying on monitoring to address remaining uncertainty. Monitoring cannot
compensate for inadequate premarket evidence when a plausible failure is catastrophic,
irreversible, or difficult to detect. Randomized trials may be necessary when the clinical
claim, potential harm, or unresolved uncertainty requires them. They should not be the
default solely because a bounded, lower-risk function uses generative AI or initiates an
action.
State-supervised pilots may also produce evidence relevant to federal review. We encourage
prospective coordination: a sponsor and state regulator define the pilot, the sponsor seeks
early FDA feedback through an appropriate mechanism, and FDA identifies the
evidence-quality and patient-protection requirements for a later submission. FDA's real-world
evidence guidance provides a basis for considering fit-for-purpose evidence.[5] Early
coordination could prevent a pilot from collecting endpoints or using methods that cannot
answer the federal review questions.
State authorization would not equal FDA authorization. The opportunity is to make evidence
generated under a qualifying protocol useful beyond the initial pilot, while retaining FDA's
independent assessment of the finished product.

IV. Connect monitoring to action and responsibility
Questions 19–21

A monitoring plan should identify both the information collected and the response when a
safety signal appears. Measures should reflect the function's material failure modes,
including adverse events, missed and unnecessary escalation, scope failures, clinician
overrides, subgroup performance, and tool or service failures. Reports need denominators
and information about the exposed population. A count of events without the number and
type of patients exposed cannot establish an event rate.
Monitoring should combine periodic benchmarking, review of live encounters, and checks for
performance deterioration. Sampling should include routine cases as well as cases selected
for elevated risk, with reporting that distinguishes targeted findings from overall error rates.
Stable reference tests allow comparisons over time; contemporary cases help establish
whether those tests still reflect clinical use. Clinical outcomes and follow-up should inform
the assessment alongside output quality. Cadence should reflect patient volume, severity,
and the time available to prevent harm.
Prespecified thresholds should trigger investigation, additional review, return to an earlier
stage, restriction, rollback, or suspension, as appropriate. A worsening trend may warrant
action before a minimum threshold is crossed. The plan should identify who is responsible
for each response and when the regulator must be notified. Monitoring is useful only if the
sponsor can investigate and act on what it finds.
Supervisory AI may help prioritize review, but its performance must be established for the
monitoring task. Sponsors should measure missed failures and false alarms, examine
whether errors are correlated with those of the clinical model, and compare findings with

Nolla Health · FDA-2026-N-7874
independent clinical adjudication. The supervisor also needs monitoring and revalidation
after relevant changes. Model agreement should not be presumed equivalent to clinical
review or used to remove required human oversight without supporting evidence.
Manufacturer accountability should remain clear. Clinicians retain responsibility for their own
decisions; participation by clinicians, institutions, independent reviewers, or
foundation-model providers should not transfer the manufacturer's obligations for the device.
A clinician who did not review an AI-initiated decision should not become responsible for it
merely because the clinician participates in the broader service. Oversight arrangements
should define each party's duties and preserve the sponsor's responsibility for the finished
product.
The Utah agreement illustrates one negotiated structure, with reporting to OAIP and
provision for external validation.[2] Any federal reporting requirements should be
proportionate and sufficiently informative for the responsible regulator. Mandatory public
reporting should be standardized and reasonably consistent with comparable devices and
healthcare institutions; AI alone does not justify an ad hoc disclosure regime.

V. Support timely adoption of validated model improvements
Questions 22–25

Advances in foundation models may improve interpretation of patient information,
recognition of risk, and communication. When a replacement model demonstrates
improvement in the product's intended clinical use, unnecessary delays may postpone
benefits for patients. We encourage FDA to support timely updates through defined testing
and change-control plans, with continued monitoring and the ability to reverse a change if
performance deteriorates.
A sponsor may be unable to name every future model architecture or set of weights at the
time of submission. It can define categories of permitted modifications, their limits, possible
hazards, and the evidence each release must satisfy. Building on FDA's predetermined
change control plan framework, we recommend that authorized plans be able to address
bounded categories of sponsor-trained checkpoints, fine-tuned models, replacement models
from another developer, and specified changes to prompts, clinical rules, retrieval, or
architecture.[6]
Each category should have explicit limits and tests of the complete deployed system. The
plan should define the clinical scope that remains fixed, safety-critical and subgroup
acceptance criteria, regression testing, triggers for additional clinical confirmation, release
controls, monitoring, and rollback. A new model's overall score should not conceal
deterioration in an important task or population. Similarly, a rule change that corrects one
error may create unnecessary exclusions or alter escalation elsewhere in the workflow.
Changes that satisfy an authorized plan should be implementable without another marketing
submission. Changes outside it, including unplanned expansion of the patient population,
permitted actions, effective autonomy, or clinical scope, should remain subject to the
applicable review requirements. As discussed above, this should preserve transitions in

Nolla Health · FDA-2026-N-7874
supervision that the authorization already covers. Benchmark representativeness also needs
review as clinical practice and the deployed population change.
The practical controls differ across model dependencies. A sponsor may version-pin open
weights, control a derivative model, or depend on a hosted service that can change
upstream. FDA should examine whether the sponsor can identify the deployed configuration,
prevent unreviewed changes, test before release, and restore a safe version or fallback.
Hosted services may change through filters, routing, or tool interfaces even when the
selected model identifier remains the same. Supplier notice arrangements and end-to-end
checks should address those dependencies, with a contingency plan if the service can no
longer perform the clinical task safely.
We support FDA's proposed voluntary foundation-model master files as a way to share
confidential information across downstream submissions. Their role should remain
informational: they should not certify a model, create a safe harbor, or substitute for
evaluation of the finished clinical product. The same model can be combined with very
different permissions, safeguards, and workflows.
A voluntary file should also not become a practical prerequisite for authorization. An
upstream developer might decline to create one, restrict access, or update it too slowly for
downstream needs. Sponsors should have a workable route to provide equivalent evidence
when a file is unavailable. A replacement model that falls within an authorized change plan
and satisfies its requirements should not be delayed solely because an upstream file has not
yet been created or amended. The downstream sponsor remains responsible for providing
sufficient evidence for the finished product.
Nolla appreciates FDA's consideration and welcomes further discussion of how these
approaches can expand access to safe, effective care.
Respectfully submitted,
Sean Geiger
sean@nollahealth.com
On behalf of:​
Nolla Health (Magic Health, Inc.)​

Nolla Health · FDA-2026-N-7874
References
[1] FDA. Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback. August 18, 2026.

[2] Nolla Health–Utah Regulatory Mitigation Agreement and incorporated pilot protocol.
September 22, 2026.

[3] IMDRF. Good machine learning practice for medical device development: Guiding principles.
N88 FINAL:2025. January 27, 2025.

[4] FDA. Applying Human Factors and Usability Engineering to Medical Devices. Final guidance;
updated August 3, 2026, originally issued February 3, 2016.

[5] FDA. Use of Real-World Evidence to Support Regulatory Decision-Making for Medical
Devices. Final guidance, December 2025.

[6] FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for
Artificial Intelligence-Enabled Device Software Functions. Final guidance, August 2025.

Nolla Health · FDA-2026-N-7874