Better Place Wellness Company (David Green)

Better Place Wellness Company, operating as Allday Health, is a Nashville-based public benefit corporation developing clinician-integrated, AI-enabled technology for behavioral health care, with an initial focus on perinatal mental health and episodic depression. Our full comments are attached.

In summary, we ask CDRH to:
(1) address behavioral health explicitly, since none of the paper’s examples comes from mental health.
(2) recognize clinician-governed architectures as a meaningful risk control within device regulation.
(3) encourage fit-for-purpose validation of measurement functions against prospectively defined reference measures, including ecological momentary assessment and validated instruments, with evidence proportionate to intended use and risk.
(4) weigh over- and under-escalation in clinical context, with clinician confirmation recognized where consistent with required response times;
(5) accept privacy-preserving postmarket monitoring where raw inputs are processed on the device and discarded.
(6) expect foundation models to change and govern those changes at the level of validated device performance, with a PCCP able to cover model substitution against prespecified acceptance criteria.
(7) expect subgroup reporting on language, dialect and literacy, with synthetic data supplementing rather than replacing real-patient evidence.
(8) apply only the relevant benchmarking elements to bounded classification and measurement functions.
(9) make independent evaluation accessible to small companies.

October 19, 2026
Division of Dockets Management (HFA-305)
Food and Drug Administration
5630 Fishers Lane, Room 1061
Rockville, MD 20852
Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative
AI-Enabled Medical Devices: Discussion Paper and Request for Feedback
Better Place Wellness Company appreciates the opportunity to comment on the Center for
Devices and Radiological Health’s discussion paper on generative AI-enabled medical
devices.

About the commenter
Better Place Wellness Company, operating as Allday Health, is a Nashville-based public
benefit corporation developing clinician-integrated, AI-enabled technology for behavioral
health care.
With an initial focus on perinatal mental health and episodic depression, Allday Health is
developing a privacy-preserving platform that derives longitudinal behavioral health
measurements from inputs processed on the patient’s device and presents those
measurements within clinician-guided care workflows.
As a developer working toward FDA-authorized digital therapeutics, the company brings a
practical perspective to measurement validation, clinical oversight, postmarket monitoring
and the governance of third-party AI models. These considerations inform the
recommendations that follow.

Summary of comments
1. The framework should address behavioral health explicitly.
2. Clinician-governed architectures should be recognized as a meaningful risk control.
3. Measurement functions need fit-for-purpose reference measures and proportionate
evidence.
4. Escalation errors in behavioral health should be weighed in clinical context.
5. Postmarket monitoring expectations should be compatible with privacy-preserving
design.
6. Foundation model changes should be expected and governed at the level of validated
device performance.
7. Subgroup performance should cover language, dialect and literacy, and synthetic data
should supplement rather than replace real-patient evidence.
8. Bounded classification and measurement functions should be evaluated on the
elements that apply to them.
9. Independent evaluation should be accessible to small companies.
1. Behavioral health should be addressed explicitly
Discussion paper questions 1 and 9
None of the discussion paper’s worked examples comes from mental health, although FDA’s
Digital Health Advisory Committee devoted its November 6, 2025 meeting to generative AIenabled digital mental health devices, and behavioral health is among the areas where
generative AI is reaching patients fastest. Behavioral health raises considerations the
current examples do not: outputs are often estimates of internal states rather than
verifiable physical findings, the reference standards often rely on patient-reported
instruments and other imperfect clinical references, and the most safety-critical events are
rare, heterogeneous and time-sensitive.
Recommendation: Include behavioral health examples in any future guidance, covering at
least a clinician-facing measurement function, a patient-facing informational function, and
a care-escalation function, and draw on the November 2025 advisory committee record in
doing so.

2. Clinician-governed architectures as a risk control
Discussion paper questions 1, 3 and 14
The paper recognizes that functions operating under continuous healthcare professional
supervision differ meaningfully from fully autonomous ones, and it asks whether
downstream safeguards should be an additional risk dimension. We support adding that
dimension. In our field, the strongest safeguard is structural: an accountable, treating
clinician receives the output inside an existing care relationship, and that clinician governs
what reaches the patient, whether by reviewing outputs directly or by authorizing in
advance the protocols under which the device acts. The clinician safeguard should scale
evidence expectations within device regulation, not serve as a route out of it.
Recommendation: Add “availability of downstream safeguards” to the risk framework,
and recognize clinician-governed architectures as a meaningful risk mitigation, with
safeguards proportionate to the function’s clinical significance. These may include
identified clinical accountability, prospectively authorized intervention protocols,
documented review of clinically consequential outputs, and auditable escalation pathways.
Evaluation should consider the effectiveness of the combined human-AI workflow,
including the degree of clinician involvement appropriate to the device’s intended use and
risk, shown through human factors and workflow evidence, not only through model
performance.

3. Measurement functions need fit-for-purpose reference measures
Discussion paper questions 1, 2 and 14
We agree with the paper that measurement and signal-processing functions can be higher
risk even when their outputs are non-directive, because the user usually cannot evaluate
the basis for the output. A clinician reviewing an ambient affect estimate cannot inspect the
raw input, particularly when raw input is processed on the device and discarded after
inference, as privacy-protective designs increasingly do. That places the burden on
validation, and validation needs agreed reference measures.
Recommendation: For longitudinal affect and symptom measurement, encourage a fit-forpurpose validation framework using prospectively defined reference measures. These may
include ecological momentary assessment (EMA) for moment-level patient-reported states,
validated instruments such as EPDS, PHQ-9 and GAD-7 for established clinical constructs,
and independent clinical assessment where appropriate. Sponsors should demonstrate the
validity, reliability, calibration and clinical relevance of derived measurements, with
attention to temporal alignment, uncertainty and subgroup performance, and should show
clinicians calibrated uncertainty alongside each estimate. Evidence requirements should be
proportionate to intended use and clinical risk, distinguishing analytical and clinical
validation of measurement functions from demonstration of therapeutic effectiveness. For
clinical confirmation of measurement functions (questions 11 and 12), study size should
generally follow the precision needed to show agreement with the reference measures
rather than power for an effectiveness endpoint, unless the measurement’s role in clinical
decisions calls for prospective evidence of clinical utility.

4. Escalation errors in behavioral health, weighed in context
Discussion paper question 6
The paper is right that under-escalation and over-escalation are both harms and are not
directly comparable. In behavioral health, under-escalation can mean a delayed response to
a patient in crisis. Over-escalation has costs specific to this field: alarm fatigue among
clinicians, which erodes the attention the signal depends on; unnecessary contact with
emergency or crisis services; stigma; and patients disclosing less once they learn that
disclosure triggers escalation. Loss of disclosure undermines the measurement itself.
Recommendation: Ask sponsors to characterize both error directions separately, by
subgroup, at the operating point used; to state the clinical rationale for that operating
point; and to describe how thresholds are governed (for example, set or confirmed by the
clinical site under its existing protocol). Recognize clinician confirmation as a risk
mitigation where clinically appropriate and consistent with required response times. For
time-sensitive safety events, sponsors should also demonstrate predefined escalation
pathways, contingency procedures, and safeguards against delayed response. Escalationrelated safety features within a broader device should be evaluated according to their
actual intended function within that device, without requiring that they be represented as
stand-alone risk-assessment or crisis-prediction claims.

5. Postmarket monitoring compatible with privacy-preserving design
Discussion paper questions 18 and 19
Periodic re-benchmarking, sample-based clinician review and drift monitoring are sound
approaches. Sample-based review of real-world inputs and outputs, however, assumes the
manufacturer keeps those inputs. For devices that process sensitive speech or text on the
patient’s device and discard the raw input, retaining it for monitoring would work against
the privacy protection that makes passive sensing acceptable to patients in the first place.
Allday Health’s architecture is designed around this principle, so for us the question is
practical, not theoretical.
Recommendation: Recognize privacy-preserving monitoring methods as acceptable
where raw inputs are not retained, including audits of derived outputs against later
instrument and EMA scores, review of samples patients have separately consented to
share, replay of fixed benchmark and synthetic sets on every deployed model version, and
on-device or federated evaluation. These alternative methods should be supported by
prespecified performance metrics, documented validation procedures, and demonstrated
ability to detect clinically meaningful degradation without routine retention of sensitive
source data. Monitoring expectations should not become a reason to collect or keep more
sensitive data than the intended use requires.

6. Foundation model changes should be expected and governed at the level of
validated device performance
Discussion paper questions 22, 23, 24 and 25
The paper asks how manufacturers can detect and respond to changes initiated by a thirdparty model developer. Any regulated device built on a foundation model should expect
that model to change, and often: foundation models are revised and superseded on a
timescale of weeks to months, while device development and premarket review take years.
A framework that ties authorization to a single fixed model version would leave such
devices running on outdated or unsupported models by the time they reach patients. The
question is especially sharp for models provided by a device’s operating system, which are
updated on the vendor’s schedule and are sometimes retired and replaced outright.
Manufacturers can design for this. Where a foundation model performs a bounded task
with constrained, near-deterministic outputs, a device can be built so that its validated
function does not depend on any one model: the manufacturer defines the task, the input
handling and the permitted outputs, and qualifies each candidate model against the same
acceptance criteria before it is allowed to produce clinical output. In that design the
regulated object is the validated function and its acceptance criteria, and the model is a
qualified, replaceable component. The manufacturer remains accountable for device
performance whichever model is in use.
Recommendation: Recognize model-agnostic designs of this kind, and confirm that a PCCP
can cover substitution or update of the underlying foundation model, including vendorinitiated changes, where the PCCP prespecifies the task and output constraints, the
acceptance criteria every model version must meet, run-time identification of the model
version in use, and a fallback that prevents an unqualified version from producing clinical
output (for example, withholding that output and notifying the clinician). Model changes
that meet the prespecified criteria should not require a new marketing submission. We also
encourage FDA to invite operating-system and platform providers to take part in the
proposed Foundation Model Device Master File program, and to provide advance notice of
model changes, model version identifiers accessible to applications, and a period during
which prior versions remain available.

7. Subgroup performance across language, dialect and literacy, and the role of
synthetic data
Discussion paper questions 12 and 13
We support the paper’s inclusion of non-standard dialects, accents, colloquialisms and
varying literacy in subgroup performance. For language-based behavioral health functions,
these factors change how distress is expressed, not only how words are recognized, and the
populations most affected (including Medicaid-enrolled perinatal patients and patients
speaking languages other than English) are often the least represented in training data.
Synthetic data can help fill these gaps in benchmarking, but the paper correctly notes that
synthetic data produced by the same class of model as the device can reproduce the
device’s own blind spots.
Recommendation: Expect subgroup reporting on language, dialect and literacy for
language-based devices. Where synthetic data is used, ask sponsors to show that it is
sufficiently independent of the model under test, through data provenance, generation by a
different model family where practical, bias assessment, and independent or blinded
clinician adjudication of its realism, and to use it to supplement, not replace, real-patient
evidence for the subgroup in question.

8. Bounded classification and measurement functions evaluated on the
elements that apply to them
Discussion paper questions 9 and 17
Many clinical uses of foundation models do not generate open-ended text for users. A
function may use a language or multimodal model to assign an input to one of a fixed set of
labels, with deterministic decoding and no conversational interface. Such functions are still
GenAI-enabled in the sense that their behavior depends on a foundation model, but several
benchmarking elements designed for open-ended output, such as multi-turn conversation
and communication quality, may not apply, while others (reproducibility, calibration,
subgroup performance, robustness to input variation, and recognition of inputs outside the
validated operating domain) apply fully. Using a foundation model to produce a categorical
label also does not by itself make the clinical function generative, and that boundary would
benefit from clarification.
Recommendation: Clarify how competency-based evaluation applies to bounded
classification and measurement functions that incorporate generative foundation models.
Allow sponsors to justify the applicability of individual benchmarking elements based on
intended use, output characteristics and clinical risk, while retaining appropriate
requirements for robustness, uncertainty communication, scope adherence and subgroup
performance.
9. Independent evaluation accessible to small companies
Discussion paper questions 10 and 16
Independent adjudicators and sequestered evaluation datasets would strengthen
confidence in benchmark results, especially where public benchmarks are saturated or
contaminated. If independent evaluation becomes an expectation, its cost and availability
will shape who can build these devices. Behavioral health has few public, clinically
adjudicated datasets, and small companies and public benefit corporations are often the
ones building for underserved populations.
Recommendation: Build on the ASCA and Medical Device Development Tool programs so
that qualified independent evaluators and shared benchmark assets are available on
reasonable, published terms; support academic and public-interest stewardship of
sequestered behavioral health benchmark sets; and keep sponsor-developed benchmarks
acceptable when their construction, prespecification and adjudication are documented.

Closing
We appreciate CDRH’s early and open engagement on these questions and would welcome
the opportunity to take part in further discussion, workshops or pilot programs on
generative AI-enabled devices in behavioral health.
Respectfully submitted,
David Green
Founder and Chief Executive Officer
Better Place Wellness Company
Nashville, Tennessee