Lune Health Inc. (Tina Latif)

See attached file(s)

16 October, 2026

Submitted electronically via Regulations.gov

Re: Docket FDA-2026-N-7874; Considerations for the Regulation of Generative AIEnabled Medical Devices

Dear Dr. Tarver and Dr. Abramson,

Lune Health is a women’s health startup. We build AI tools to help women understand
their cycles, fertility and treatment, organize their records and prepare for clinical
appointments. Our work includes egg freezing and IVF, where explaining a clinic
instruction must remain distinguishable from making a new medical recommendation.

We welcome the FDA’s recognition that generative AI can expand access and strengthen
patient empowerment. We particularly welcome its acknowledgement that regulation
should not underestimate patients’ ability to engage with their own health information.

Our submission makes one central point: women’s health should be an explicit
consideration throughout the FDA’s framework for generative AI, from how risk is judged
to how devices are tested, monitored and updated.

1. Context: why women’s health needs explicit attention

Women are already using AI for their health, whether or not it was built for them.

OpenAI's January 2026 report estimates that more than 40 million people worldwide use
ChatGPT for health questions every day, and that in the United States about seven in ten
of those conversations happen outside standard clinic hours.1

For a woman in fertility treatment, that is usually when the questions come: What did my
clinic mean? How do I take this injection? Is this symptom normal? She will use whatever
tools are available to answer them. The regulatory goal should be to make it possible to
build better ones: specialized, properly tested, clear about their limits, and held to
standards proportionate to the consequences of being wrong.

Emerging evidence reinforces this need. MIT researchers found that models made about
7% higher error rates for female patients and were more likely to tell them to manage their
OpenAI. AI as a Healthcare Ally. January 2026. Available at: https://cdn.openai.com/pdf/2cb29276-68cd-4ec6-a5f4-c01c5e7a36e9/OpenAI-AI-as-a-Healthcare-AllyJan-2026.pdf. (The daily-use figure is global; the out-of-hours figure is drawn from US users.)
symptoms at home.2 A randomized study in Nature Medicine found that models which
score well on medical benchmarks do far worse when ordinary people use them in live
conversation.3 And independent testing of consumer health AI found that it under-triaged
about half of true emergencies, and that its advice shifted towards less urgent care when
a patient reported that family or friends had played down their symptoms.4

These findings matter more for women. Women in emergency departments are less likely
than men to be given pain relief for the same reported pain,5 and more than half of
younger women later hospitalized with a heart attack had been told by a provider that
their symptoms were not heart-related.6 An AI that becomes less urgent when a
symptom is played down will therefore make the same mistakes, and under-triage
women more often than men.

AI is being built on evidence and data that underrepresents women.

In 1977, FDA guidance effectively excluded many women of childbearing potential from
early clinical research. The FDA reversed that in 1993, the year Congress required women
to be included in NIH-funded research. Significant progress has followed, but important
evidence gaps remain. This matters because AI systems inherit the limitations of the
evidence and data available to them.

Women are making bigger reproductive decisions on too little information.

Many women lack basic information about fertility decline, miscarriage risk and
reproductive aging. In a recent survey of US “Gen Z” women (aged 18-27), 59% felt
uninformed about their reproductive health, even though 74% said they wanted children.7

At the same time, women are having children later and increasingly making consequential
reproductive decisions before they have a fertility specialist to ask. Egg-freezing cycles
have risen sharply, and hundreds of thousands of IVF cycles are performed in the United
States each year.8
Gourabathina A, Pan E, Gerych W, Ghassemi M. The Medium is the Message: How Non-Clinical Information Shapes Clinical Decisions in LLMs. In: Proceedings of the 2025
ACM Conference on Fairness, Accountability, and Transparency. 2025. Available at: https://doi.org/10.1145/3715275.3732121.
Bean AM, Payne RE, Parsons G, et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine. Published online
February 9, 2026. Available at: https://doi.org/10.1038/s41591-025-04074-y.
Ramaswamy A, Tyagi A, Hugo H, et al. ChatGPT Health performance in a structured test of triage recommendations. Nature Medicine. 2026;32:1671-1675. Available at:
https://doi.org/10.1038/s41591-026-04297-7.
Guzikevits M, Gordon-Hecker T, Rekhtman D, et al. Sex bias in pain management decisions. Proceedings of the National Academy of Sciences. 2024;121(33):e2401331121.
Available at: https://doi.org/10.1073/pnas.2401331121.
Lichtman JH, Leifheit EC, Safdar B, et al. Sex differences in the presentation and perception of symptoms among young patients with myocardial infarction: evidence from
the VIRGO study. Circulation. 2018;137(8):781-790. Available at: https://doi.org/10.1161/CIRCULATIONAHA.117.031650.
Clements ML, Sawicki S, Segree N. What are my options? Fertility awareness, infertility treatment awareness, and education needs among Generation Z women. F&S
Reports. 2026. Available at: https://www.fertstertreports.org/article/S2666-3341(25)00172-2/fulltext.
Society for Assisted Reproductive Technology. Preliminary National Summary Report for 2022 (29,803 egg-freezing cycles, up from 24,560 in 2021; 389,993 IVF cycles).
Reported in: SART data indicates rise in IVF use in 2022. Contemporary OB/GYN. April 24, 2024. Available at: https://www.contemporaryobgyn.net/view/sart-dataindicates-rise-in-ivf-use-in-2022.
Women are therefore being asked to make increasingly complex reproductive decisions
while often lacking clear, individualized information about their own bodies. Properly
designed AI can help close this gap. Poorly designed AI can compound it.

2. Recommendations

2.1 Recognize traceability to a verified source as a meaningful risk mitigation
Questions 1, 2 and 9; Appendix A E.3

Question 1 asks whether traceability to source material belongs in the risk framework. It
does. Showing a woman the fertility-medication dose her clinic has already prescribed is
different from deciding what her dose should be. The same goes for personalization:
telling a woman, from her own records, that her cycle has varied over six months restates
her data. It does not direct her care.

Where a system draws on a medical record, clinic instruction or other verified source, the
manufacturer should demonstrate that the information is retrieved accurately, attributed
clearly and kept visibly separate from any AI-generated explanation. Where the product's
job is to carry a clinic's instruction to the patient, getting the dose and its units right
should be tested directly.

This can be shown through benchmarking: a prespecified test set in which the source is
known, scored for retrieval accuracy and for whether the output visibly distinguishes the
source from the AI's explanation, with results included in the premarket submission.
Wellness products sit outside this framework, but the same test is good practice for any
product that relays clinical instructions, and we would welcome a recognized standard
they could test against.

Done properly, this lets AI help patients understand and follow the care they already have,
without every personalized answer being treated as if the model were practicing medicine
on its own.

2.2 Do not assume that patient-facing information is inherently higher risk;
assess it against the real-world alternative
Questions 3, 14 and 15

Patient-facing AI may carry different risks, but not automatically higher ones. A woman
undergoing IVF with a question about tonight's injection usually cannot reach a specialist.
Her alternatives are an internet search, an internet forum, a general-purpose chatbot,
delayed care or nothing.
The FDA should weigh benefit and risk against that real alternative, not against an ideal of
instant clinical access. Otherwise, the framework risks holding back tools that are better
than what she would use instead.

2.3 Test across reproductive states, not women as a single subgroup
Questions 9-13; Appendix A E.1, E.4 and R.2

We support the approach the FDA is considering: assess a generative AI device the way a
doctor is assessed, first on standardized tests, then in realistic clinical conditions, with
more evidence required as the risk rises.

However, women's health shows why testing by demographic group is not enough. A
woman may move through menstruation, trying to conceive, fertility treatment,
pregnancy, pregnancy loss, postpartum, perimenopause and menopause, and the same
symptom, measurement or question can mean something different in each. A device
could score well on women as a broad category and still fail in specific phases of a
woman's life.

We therefore ask that benchmarking element R.2 name sex and reproductive state as
subgroups, and that E.1's population-specific considerations extend beyond pregnancy to
the states above. Manufacturers should also test across care pathways and health
literacy and language, and state any gaps in labeling and the premarket summary rather
than hide them in a good average. Testing should also run across repeated conversations,
since that is how a woman actually uses these tools, and check that the device neither
talks her out of escalating when she plays a symptom down nor amplifies alarm when she
fears the worst.

The same goes for synthetic data. Invented cases can help test rare situations, but
manufacturers should document in their benchmarking methods that generating them
does not repeat the historical underrepresentation of women or flatten female physiology
toward a male default.

2.4 Treat privacy-preserving postmarket monitoring as a safety requirement,
not a trade-off against safety
Questions 18-21

Postmarket monitoring matters, because premarket testing cannot anticipate every real
conversation. But routine human review of real patient conversations creates a risk of its
own: privacy.
That risk is sharpest for consumer products holding reproductive health information,
which often sit outside HIPAA. It is also a safety issue. A woman who fears that someone
may later read her conversation may leave out a pregnancy, a pregnancy loss, a
medication, her sexual-health history or a symptom, and get a less safe answer as a
result.

The FDA should make clear that effective monitoring does not require keeping or
reviewing every conversation. It should expressly allow privacy-preserving methods: ondevice processing where appropriate, consent-based sampling, de-identification,
restricted-access clinical review, structured safety-event capture and aggregate
monitoring.

Developers should not have to build an archive of intimate health conversations to prove
they are monitoring. The same expectations should extend to the foundation-model
providers through which these conversations pass.

2.5 Do not let the median clinician set the pass mark for women
Question 14

The paper asks whether device performance should be compared to a panel of clinicians
or to the median clinician in practice. For women’s health this needs care. Where clinicians
already underestimate women’s symptoms, a device that matches the median clinician
inherits that gap and passes. Acceptance criteria should be anchored to the standard of
care, and sponsors should show that the comparator itself performs consistently for
women before it is used as the benchmark.

2.6 Require greater stability and transparency from third-party foundationmodel providers
Questions 22-25

Most specialized health products (particularly those built by smaller companies) are built
on third-party foundation models. The downstream developer is fully accountable for
safety but has little visibility into changes to the model underneath.

The FDA’s proposed Foundation Model Device Master Files could help, but as described
they are voluntary and confidential: the FDA reviewer sees the file, the accountable
developer does not (Section VII.A). Developers need, at a minimum, an identifiable model
version, notice of material changes, known healthcare limitations and failure modes, and
relevant performance data, broken down by sex and reproductive state where relevant.
Foundation-model providers should also give reasonable notice of material changes. A
developer cannot run a credible change-control process if a safety-critical dependency
changes without warning. Revalidation should then be proportionate to the change. A
modification that affects clinical reasoning, escalation or safety behavior may require rebenchmarking; a change limited to style or other low-risk behavior should not
automatically trigger wholesale revalidation.

3. Closing

Generative AI creates a rare opportunity to address some of the information gaps that
have characterized women’s health for decades. Used well, it can help women
understand their own records and bodies, make complex reproductive decisions, prepare
for clinical care, and obtain reliable information when a specialist is not available.

None of this will happen by default. It takes a framework that is explicit about women at
every step: one that separates relaying a clinic's instruction from giving new advice; that
judges a tool against what a woman would otherwise use; that tests across the
reproductive states and real conversations she will actually have; that monitors without
making her afraid to speak; and that keeps devices safe as the models beneath them
change.

Women are already using AI for health. The FDA’s framework will help decide whether it is
built for them.

Thank you for the opportunity to comment.

Tina Latif
CEO and Founder
Lune Health