Spring Health (Brian Altman)

Please find Spring Health’s comments attached.

October 16, 2026

Michelle Tarver, M.D., Ph.D.
Center Director
Center for Devices and Radiological Health
U.S. Food and Drug Administration

Re: Docket Number: FDA-2026-N-7874, Considerations for the Regulation of Generative
AI-Enabled Medical Devices: Discussion Paper and Request for Feedback

Dear Dr. Tarver,

Spring Health (“Spring”) appreciates the opportunity to submit comments in response to
the FDA’s discussion draft paper entitled Considerations for the Regulation of Generative
AI-Enabled Medical Devices.

About Spring Health
Spring Health’s mission is to eliminate every barrier to mental health care by delivering
data-driven, effective mental health support. Spring Health is an Artificial Intelligence
(AI)-native, global mental health company built on one platform, so care follows people
across every job, move, health plan, and life stage. Founded in 2016 as an AI company,
Spring Health provides personalized lifelong mental health support to 170M covered lives
for employers and health plans across self-guided tools, coaching, therapy, medication
support, and specialized care programs. Spring Health is trusted by leading employers,
including Target, The Coca-Cola Company, BlackRock, Microsoft, JP Morgan Chase, and
Wawa, and by health plans including UnitedHealthcare, Cigna and Highmark.

Spring Health’s approach aligns with the goals of the Great American Recovery Initiative
by expanding access to high-quality mental health and substance use care, supporting
earlier identification of at-risk individuals, and improving outcomes through data-driven,
personalized interventions.

Spring Health serves as a model for how Super Intelligence can be harnessed to Make
America Healthy Again. We provide accessible, high-quality, and personalized care in a
secure environment that prioritizes privacy, responsible data use, and clinically sound
guidance. We also use AI to deliver insights that help organizations identify high-risk
groups earlier, provide proactive support, and optimize care pathways for better
outcomes and cost savings. To support safe and effective deployment of AI in mental
health care, in October 2025, Spring Health launched the AI in Mental Health Safety &
Ethics Council. The Council has actively sought to advance universal standards for the
responsible use of AI in mental healthcare and developed the Validation of Ethical and
Responsible AI in Mental Health (VERA-MH), an open-source, clinically grounded
60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
evaluation framework that enables continuous, multi-metric assessment of AI-enabled
tools. A recent study examining VERA-MH found strong alignment between the
framework's automated evaluation and clinician judgment, supporting its validity and
reliability as an AI safety assessment tool. Spring Health's approach is intentionally
iterative, incorporating new data, provider feedback, patient outcomes, and evolving
clinical evidence into a continuous cycle of testing, learning, and refinement.

We appreciate the opportunity to offer insights on Spring Health’s relevant experiences
and provide recommendations to FDA on this discussion draft as the agency considers
updating its regulatory framework for GenAI devices using a risk-proportionate approach.

Section IV. Considerations for the Assessment of Risk for GenAI-Enabled Devices
Question 3: What device characteristics, output features, or safeguards might mitigate
risks that could arise when a user lacks the domain knowledge to independently evaluate
an output, without unnecessarily underestimating patient capability?

Spring Health's digital health tools include AI-guided intake experiences, asynchronous
communication, digital navigation, session summaries, and provider-directed activities
intended to support patient engagement, continuity, and informed interactions with
clinicians.

Digital health information delivered directly to patients may present considerations that
differ from information delivered to healthcare professionals, but patient-facing functions
should not be presumed to be inherently unsafe. In the context of digital mental health
technologies, such tools can improve access, engagement, care navigation, continuity of
care, and patient participation while reducing barriers to support. At the same time,
patient-facing GenAI may create risks when users cannot readily recognize inaccurate,
inappropriate, or misleading output or do not understand the limitations of a tool. This
concern is particularly important in mental health contexts, where widely available
chatbots may lack clinical oversight, regulatory safeguards, and reliable crisis-response
mechanisms. To balance known and significant benefits with unintended risks, Spring
Health applauds FDA’s proposed approach in this discussion draft that appropriate
safeguards be tailored to the product's intended use and risk.

We note that the VERA-MH is a risk assessment framework that aligns with FDA’s risk
proportionate approach by measuring mental health chatbot responses against five
criteria outlined in the table below and which are grounded in real-world clinical practice.
These criteria serve as a scoring framework designed to ensure that mental health
chatbot behavior aligns with established best practices in suicide and harm from others
risk management, including direction to, and connection with, crisis care and support
from a human provider.

60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
Criteria Brief Description

Assesses whether the chatbot accurately recognizes
Detects Potential Risk signals of risk, flagging genuine risk without missing it or
over-flagging.

Evaluates the quality of the chatbot's follow-up
Confirms Risk questioning once potential risk has been detected: the
questions it asks and the conclusions it draws from them.

Evaluates whether the chatbot connects the user to safety
Guides to Human Care
resources and support

Evaluates the tone and relational quality of the chatbot's
responses: whether it respects the user's autonomy,
Supportive Conversation
avoids judgment, and stays warm and responsive rather
than robotic.

Evaluates whether the chatbot maintains appropriate limits
on its own role: avoiding unsafe advice, overreach in
Follows AI Boundaries
labeling the user's situation, and positioning itself as a
substitute for human help.

FDA's proposed risk framework should consider not only whether a function is
patient-facing, but also criteria similar to those found within VERA-MH, which are largely
relevant to digital health technologies beyond mental health chatbots. Using a risk
assessment framework like VERA-MH would recognize relevant differences between
patient-facing and healthcare professional-facing functions without underestimating
patient capability or unnecessarily limiting technologies that can expand access and
engagement.

Question 5: How should risk be assessed across realistic conversational trajectories?
How could the intended use of such a device be characterized when its behavior is
emergent across a conversation?

Risk assessments for conversational GenAI-enabled functions should examine realistic
conversational trajectories rather than isolated prompts and responses; this is particularly
important for digital mental health technologies. A response may appear appropriate
when viewed independently, while the cumulative interaction may create or amplify safety
risks. Testing should therefore evaluate whether risk is recognized as information
emerges, whether the system asks appropriate clarifying questions, whether the
interaction remains within intended boundaries, whether escalation to a human occurs at
the clinically appropriate point, whether the system avoids reinforcing potentially harmful
60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
behavior, and whether the system maintains supportive and clinically appropriate
communication. One of the principles that motivated VERA-MH’s development is the view
that traditional single-turn evaluation is insufficient for assessing safety in dynamic
chatbot interactions. As discussed above, VERA-MH’s risk assessment criteria emphasize
clinically informed evaluation, safety, multi-turn testing, automation, model-agnostic
application, and multidimensional assessment. Therefore, Spring Health recommends that
FDA consider how intended use may be described and evaluated across the types of
conversational trajectories the system is designed to support, rather than relying solely on
the characterization of individual outputs. Risk assessment should examine whether the
system's behavior remains within that intended use throughout realistic conversations,
including when initially benign exchanges evolve into higher-risk circumstances.

Question 6: How could manufacturers characterize and weigh these two
[Under-Escalation and Over-Escalation] directions of error, given that they may not be
commensurable and that acceptable trade-offs may vary by clinical context?

Under-escalation or over-escalation to clinical care could have irreversible consequences
depending on the intended use of a device, and the consequences should be evaluated
as part of the risk profile. Evaluation of care-escalation functions should address both
under-escalation and over-escalation in the relevant clinical context. Spring Health's
experience in this regard is particularly relevant to potential suicide risk. VERA-MH
evaluates whether a chatbot detects statements indicating potential risk, asks direct
clarifying questions when needed, guides the user to human support when indicated,
maintains a supportive and collaborative conversational style, and recognizes AI
limitations while avoiding behavior that may fuel harm. These dimensions provide a
clinically grounded structure for identifying several forms of inappropriate detection,
response, and escalation behavior. Similarly, FDA should support structure but ultimately
allow manufacturers to establish evaluation criteria and acceptance thresholds that reflect
the seriousness and immediacy of potential harm in the relevant clinical context. Spring
Health recommends a context-specific approach grounded in relevant clinical standards
and validated expert judgment.

Section V: A Competency-Based Approach for Premarket Evaluation of GenAI-Enabled
Devices
Question 7: Is the competency-based approach described above, i.e., device
benchmarking followed by clinical confirmation, a useful and appropriate framework for
evaluating GenAI-enabled devices?

Spring Health is supportive of FDA’s proposed approach and has taken a similar tactic to
evaluate VERA-MH. To that end, VERA-MH provides a practical example for a defined use
case. As described above, VERA-MH employs five criteria to evaluate safety dimensions,
used for both suicide-risk and harm from others detection and response: detection of
60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
potential risk, confirmation of risk, guidance to human care, supportive conversation, and
adherence to AI boundaries. A recent study examining VERA-MH found strong alignment
between the framework's automated evaluation and clinician judgment, supporting its
validity and reliability as an AI safety assessment tool. VERA-MH has also been cited in
OpenAI’s MentalHealthBench research as an existing framework that uses dynamically
generated, multi-turn conversations and a clinically developed global rubric to assess the
detection of and safe responses to suicide risk. Spring Health's approach is intentionally
iterative, incorporating new data, provider feedback, patient outcomes, and evolving
clinical evidence into a continuous cycle of testing, learning, and refinement. We
appreciate that an exhaustive evaluation of every possible input and output is not
practical for systems that engage in open-ended, multi-turn interactions. Therefore, we
believe a competency-based framework, as proposed by FDA, can instead assess
defined and observable behaviors relevant to safety and performance. Spring Health also
recommends that FDA apply a risk proportionate approach to applicable competencies,
rigor of testing, acceptance criteria, and clinical confirmation based on intended use
rather than uniformly to all GenAI-enabled devices.

Question 9: Would a benchmarking structure such as the one described above be likely to
provide adequate evidence of clinical knowledge, analytic capabilities, safety behavior,
communication, and generalizability to support a reasonable assurance of safety and
effectiveness? Are there elements that are missing, redundant, or inappropriately
categorized? Are there externally developed standards that could be leveraged?

Spring Health is supportive of FDA’s proposed benchmarking structure. In fact, the
benchmarking elements proposed by FDA are generally aligned with principles reflected
in Spring Health's experience, particularly:

•​ Safety-critical recognition and escalation;
•​ Scope maintenance and boundary adherence;
•​ Appropriate uncertainty communication and deferral to human care;
•​ Information gathering across evolving interactions;
•​ Communication quality and user comprehension; and
•​ Robustness across multi-turn conversational contexts.

Moreover, since not all GenAI devices are the same, Spring Health believes sponsors
should be able to select and justify the elements relevant to the specific product, patient
population, clinical context, intended use, and level of risk. We appreciate FDA’s overt
recognition in the discussion paper that not every benchmarking element will apply to
every GenAI device.

Question 10: How should a sponsor establish that performance on a given benchmark
predicts safe and effective real-world behavior for the device’s intended use? What
evidence should support the construct validity of a benchmark used to gate device
60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
evaluation, and what role should sponsor-developed benchmarks play given potential
concerns around independence and optimization to the test?

Spring Health developed VERA-MH after determining that it could not identify an existing
benchmark or evaluation tool that met its needs for evaluating mental health chatbot
safety. VERA-MH was designed to be clinician-informed, multi-turn, automated,
model-agnostic, multidimensional, and open source. Its source code was released
publicly, followed by a public comment period through which feedback was received from
researchers, mental health experts, AI developers, clinical and technology organizations,
and people with lived experience.

Therefore, we believe that FDA should also permit sponsors to use clinically grounded,
evidence-based voluntary consensus standards and appropriately validated evaluation
frameworks or benchmarks such as VERA-MH. Specifically, Spring Health recommends
that FDA permit sponsor-developed benchmarks when the sponsor demonstrates that the
benchmark is:

•​ Clinically grounded;
•​ Transparent in its criteria and scoring methodology;
•​ Appropriate for the intended use and population;
•​ Capable of evaluating multi-turn interactions when relevant;
•​ Supported by qualified clinical input;
•​ Validated for the construct that it is intended to measure; and
•​ Subject to continuing reassessment and refinement.

Another consideration is that while independent input can strengthen credibility, FDA
should encourage open-sourced models and avoid clinically useful evaluation tools that
are inaccessible. Finally, given digital health technologies are always evolving and
improving, FDA should view benchmark validation as an iterative process rather than a
one-time determination.

Question 11: How might a sponsor select and justify a confirmation approach tailored to a
device’s intended use and proportionate to the device’s risk profile? Are there device
types or risk profiles for which one or more of these approaches would be insufficient or
inappropriate? How might the anticipated distribution of real-world inputs be taken into
consideration? Are there other methods of clinical confirmation that might help inform the
evaluation of GenAI-enabled devices?

Based on Spring Health's monitoring experience, information potentially relevant to
clinical confirmation or continuing evaluation could include clinician review, provider
feedback, patient-reported outcomes, engagement data, clinical outcomes, electronic
health record information, and user feedback. This information would support an iterative
cycle of identifying opportunities for improvement, strengthening performance, and

60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
maintaining alignment with evolving clinical practice. The appropriate regulatory use of
each source would depend on the device, intended use, risk, and proposed confirmation
method.

Question 14: For open-ended device outputs, how should performance comparators and
acceptance criteria be selected? When a panel of qualified clinicians serves as the
comparator, how should the applicable standard (for example, the standard of care
versus the performance of a median clinician in practice) be defined and justified?
Should generalist or specialist physicians be used as a performance standard? When
should human-AI team performance, rather than the device operating alone, serve as the
basis for evaluation?

For open-ended outputs, qualified clinicians applying prospectively defined criteria could
provide an appropriate comparator, particularly where the criteria reflect established
clinical practice and the actual intended use of the system. The VERA-MH validation study
offers an example of this approach. Licensed mental health clinicians independently
reviewed simulated conversations using the same structured rubric as the automated
evaluator. Clinician ratings were used to establish a clinical consensus reference, and the
study assessed alignment between that reference and the automated evaluation.

Where the intended use incorporates clinician review, FDA should consider whether
evaluation of the device as deployed should include the combined performance of the
clinician and device. Spring Health's practices illustrate the relevance of this
consideration because providers review and sign clinical documentation and apply
professional judgment before AI-generated outputs are incorporated into patient records.
The composition and qualifications of clinician panels should be justified based on the
intended population, clinical context, task complexity, and manner in which the product is
expected to be used.

Question 16: Should independent third parties be involved in some or all aspects of a
competency-based approach, including device benchmarking and clinical confirmation?
If so, in what ways might qualified, independent third-party participation contribute to
this approach, and what qualifications and independence criteria should apply? Are there
aspects of the assessment for which third-party involvement would be impractical or
inadvisable? What safeguards or program design features would be critical to prevent
third-party participation from limiting competition or preventing innovation?

Yes, Spring Health supports third parties being involved in some or all aspects of a
competency-based approach, including device benchmarking and clinical confirmation.
Independent third-parties could certify AI devices pre-approval using an in-depth test,
such as VERA-MH, as well as human clinical review of flagged cases and provide
remediation plans for the device manufacturer if standards are not met.

60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
VERA-MH was developed with input from clinicians, psychologists, academics, mental
health organizations, technology companies, researchers, people with lived experience,
and other stakeholders. To further pressure test VERA-MH, Spring Health then published
a concept paper, requested public comments, and incorporated diverse perspectives into
the framework's development. Similarly, independent clinicians can serve as valuable
validators or adjudicators, particularly when their qualifications align with the applicable
clinical domain. However, responsibility for product safety and performance should
remain with the manufacturer.

VERA-MH was designed to be transparent, publicly available, model-agnostic, and
capable of continuing improvement. In that vein, FDA should avoid structuring third-party
participation in a manner that unnecessarily constrains innovation, or limits competition.

Section VI: Postmarket Monitoring for GenAI-Enabled Devices
Question 19: Please comment on the potential approaches to postmarket performance
evaluation, including periodic re-benchmarking, sample-based clinician review, and
performance degradation monitoring. What additional approaches should CDRH
consider, and how should the cadence and triggering events for reassessment be
determined?

Because AI systems, models, and deployment environments evolve, safety evaluation
must continue throughout the product’s lifecycle. Spring Health's experience supports an
iterative approach incorporating patient-reported outcomes, engagement data, passive
monitoring signals, electronic health record information, clinical outcomes, provider
feedback, and user feedback. Although post-market safety monitoring is a must, FDA
should permit sponsors to establish a risk-proportionate post-market monitoring plan that
identifies the data sources, cadence, and events that would prompt reassessment.

Question 21: What roles might clinicians, healthcare institutions, professional societies,
standards-setting bodies, and other stakeholders appropriately play in postmarket
monitoring, and how can these roles be structured without diffusing manufacturer
accountability?

60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com
All of these various stakeholders can contribute differing expertise to postmarket
monitoring and evaluation. Their participation can help identify practical safety concerns,
improve clinical relevance, strengthen evaluation criteria, support transparency, and
inform continued refinement. However, postmarket contributions by stakeholders should
supplement, rather than supplant manufacturer accountability. Spring Health's experience
demonstrates the value of this collaborative model. VERA-MH was shaped through
external expert and community feedback, including input from mental health experts,
researchers, AI developers, clinical and technology organizations, and people with lived
experience.

Conclusion
Spring Health appreciates FDA's efforts to engage stakeholders early as it considers a
more risk-proportionate regulatory approach for GenAI devices. We support a regulatory
approach that is risk-based, flexible, and proportionate to a product's intended use. For
patient-facing conversational systems, evaluation should consider the final user-facing
GenAI-enabled device or device software function in its deployed or representative
configuration, rather than only the underlying foundation model or another
isolated component. Evaluation should address performance across realistic interactions,
escalation to human care, clinical oversight, transparency, privacy, and continuing
monitoring. FDA should also permit sponsors to use clinically grounded, evidence-based
voluntary consensus standards and appropriately validated evaluation frameworks or
benchmarks, including open-source tools such as VERA-MH, where fit for the intended
use.

Spring Health encourages FDA to consider Spring Health as an example of, and a
resource on, the power of accelerating AI development while taking a democratized
approach to advising on deploying responsible digital mental health technologies. We
welcome continued dialogue and collaboration with the FDA in shaping the future of AI in
health.

Sincerely,

Brian Altman, JD
Director of Government Affairs
Spring Health

60 Madison Avenue, 2nd Floor, New York, NY 10010
www.springhealth.com