Meadows Mental Health Policy Institute

Please see the attached public comment on behalf of Meadows Mental Health Policy Institute.

Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Meadows Mental Health Policy Institute

Considerations for the Regulation of Generative AI-Enabled Medical Devices

October 19, 2026

The Honorable Kyle Diamantas
Acting Commissioner
U.S. Food and Drug Administration
10903 New Hampshire Avenue
Silver Spring, MD 20993

RE: FDA-2026-N-7874-0001

Dear Acting Commissioner Diamantas:

Thank you for the opportunity to share feedback on the discussion paper, “Considerations for
the Regulation of Generative AI-Enabled Medical Devices”. We appreciate this opportunity to
provide input on how to assess risk, evaluate safety and effectiveness before marketing, and
monitor performance after deployment for generative AI-enabled medical devices.

Independent and nonpartisan, the Meadows Mental Health Policy Institute (Meadows Institute)
works at the intersection of policy and programs to create equitable systemic changes so all
people in Texas, the nation, and the world can obtain the healthcare they need. The Meadows
Institute provides state and national leaders with data-driven, trusted policy and program
guidance, and works to shift the focus of new investments toward early intervention for
children and families, address the mental health crisis in our jails and emergency rooms, and
help all people with mental health needs recover and be well.

The Meadows Institute, in collaboration with leading mental health technology companies ,
offers a behavioral health perspective on these questions, drawing attention to the potential
benefits of these technologies and the safeguards needed to support their responsible use. We
begin with the broader landscape of AI in behavioral healthcare and then discuss important
considerations across the product lifecycle, including 1) model development and training, 2)
pre-market safety benchmarking, and 3) post-market surveillance.

Background

As detailed in the Meadows Institute’s “The Positive Impact of AI in Behavioral Health” use
cases, AI encompasses a range of technologies, and AI tools are designed and configured to
perform different functions. In behavioral healthcare, these functions may include organizing
 Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

intake information, supporting clinical documentation, helping clinicians interpret symptom
measures, or providing structured support between appointments. Tools developed and
validated for a defined clinical purpose, with safeguards and oversight appropriate for that
purpose, have the potential to reduce workload, improve care coordination, and support more
timely responses to patient needs.

Risks arise when a tool that has not been designed and tested for clinical treatment is used for
that purpose inappropriately. For example, a general-purpose conversational system may be
used for mental health advice or as a substitute for therapy even though it has not been
developed and validated to provide that care. Its responses may then offer clinically
inappropriate guidance, reinforce harmful beliefs or behaviors, or fail to respond appropriately
to safety concerns. Tools developed specifically for behavioral health also require rigorous
evaluation; a stated clinical purpose alone does not establish safety or effectiveness. These
distinctions underscore the importance of assessing each tool according to its intended
function, actual use, and supporting evidence.

FDA’s Technology-Enabled Meaningful Patient Outcomes for Digital Health Devices (TEMPO)
pilot offers one illustration of how a generative AI tool can be built for a defined clinical
purpose and deployed within a structured framework of oversight and evidence generation.
Limbic, a developer of AI-led behavioral health tools, has been selected to deliver cognitive
behavioral therapy through an AI voice agent under licensed clinician supervision, with
outcome and safety data informing FDA’s assessment. Evidence from the pilot is still emerging,
but the example suggests that purpose-built tools can operate with meaningful safeguards. It
also points to the right response to the risks above: rather than lowering expectations for
evidence and oversight, FDA can help developers of tools intended for clinical use understand
what is required to demonstrate safety and effectiveness.

The FDA’s emphasis on a risk-based approach that considers intended use and technological
characteristics provides an important foundation for addressing these opportunities and
concerns. The benefits, risks, and evidence needed to evaluate a tool depend on what it does,
who uses it, and how it is incorporated into care. Treating AI applications as a single category
risks obscuring consequential differences and discouraging beneficial uses alongside those that
warrant greater scrutiny. Our comments therefore emphasize distinctions that can inform the
FDA’s assessment of risk and its consideration of evidence across the total product life cycle of
generative AI-enabled devices.

Distinguishing Between General Purpose and Purpose-Built Tools

Generative AI chatbots typically combine an underlying foundation model with application
components that shape how it interacts with users. Foundation models are trained on large,
 Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

diverse datasets to learn patterns in language and generate responses. Developers can then
adapt the model through additional training and configure the application by specifying its role,
providing access to reference materials, managing conversational memory, and establishing
rules for responding or taking actions. Together, these choices shape the chatbot’s functionality
and behavior.

General-purpose chatbots are conversational applications designed to assist users across a
broad range of topics and tasks. Their training, instructions, and safeguards support broad
utility to consumers rather than establish clinical competence in behavioral health. Purposebuilt chatbots, on the other hand, are developed or adapted for a defined function, intended
population, and care setting. Developers may use additional training, clinical instructions,
curated reference materials, and safeguards tailored to that function. These safeguards may
include privacy protections, limits on the guidance provided, protocols for recognizing and
responding to clinical risk, and pathways for human review and support. General-purpose and
purpose-built chatbots may share a conversational interface and even the same underlying
foundation model. The distinction lies in the functions they are designed to perform, how they
are configured, and the evidence supporting their use.

Distinguishing Between Patient- and Provider-Facing Tools

Among purpose-built behavioral health AI tools, another important distinction is whether the
tool delivers information directly to patients or supports healthcare providers in delivering care.
Patient-facing tools may provide education, guide structured therapeutic exercises, or offer
recommendations that users act on without prior clinician review. Their evaluation should
consider whether users can understand the information, recognize the tool’s limitations, and
access appropriate human support when needed. Patients’ knowledge of their own experiences
and treatment goals is an important resource, but they should not be expected to
independently verify the clinical accuracy of AI-generated guidance. Safeguards should reflect
the tool’s function and the consequences of relying on its outputs.

Provider-facing tools may organize clinical information or generate assessments and
recommendations for review by a healthcare professional. This review can offer an additional
safeguard when the provider has the relevant expertise, sufficient information, and time to
assess the output before it affects care. However, labeling a tool as provider-facing does not
establish that meaningful oversight occurs. Evaluation should examine whether providers can
verify or challenge its outputs, whether recommendations fall within their clinical expertise,
and whether workflows support correction of errors before they reach patients.

These categories should inform risk assessment alongside the tool’s specific function and
clinical context. A patient-facing educational tool and a patient-facing treatment tool may
 Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

warrant different evidence and safeguards, as may a provider-facing documentation tool and a
system that recommends treatment changes. For tools used by both patients and providers,
evaluation should address each interaction and clarify who is responsible for reviewing outputs
and responding to identified risks.

Model Development and Training

Representativeness

Purpose built AI tools for behavioral health should be developed with their intended users and
clinical purpose in mind. The data used to train, adapt, and test a tool should adequately reflect
the populations, purposes, and settings in which it will be used, including relevant variation in
mental health conditions, symptom severity, age, language, cultural context, and
socioeconomic circumstances. A large dataset alone does not establish that a model is suited to
its intended users; the relevance, quality, and coverage of the data are essential. Developers
should assess gaps in that representation and evaluate whether the tool performs safely and
effectively across these groups.

Privacy

While general-purpose tools are not positioned to protect user privacy, purpose-built
behavioral health AI tools should protect sensitive patient information and comply with
applicable federal and state privacy requirements, including HIPAA and, where applicable, 42
CFR Part 2 protections for substance use disorder records. Tools can be deployed within secure
environments that control access to patient information and limit its use and disclosure. When
a tool relies on a third-party foundation model, these protections must extend to that
relationship, with technical safeguards and contractual agreements that prevent unauthorized
disclosure or reuse of patient information, including its use to train general-purpose models.

Pre-Market Safety Benchmarking

Once a tool’s intended function and users are defined, safety benchmarking provides a
structured way to evaluate whether it behaves as intended. Benchmarking is a process by
which a system is tested against a defined set of tasks or scenarios using explicit scoring criteria.
Approaches may include testing individual responses, simulating conversations over multiple
exchanges, and deliberately challenging safeguards with ambiguous or adversarial inputs.
Evaluation may use clinician review, automated scoring, or a combination of both.

Benchmarking should reflect the tool’s function and whether its outputs reach patients directly
or undergo provider review. Patient-facing conversational tools should be tested for their
ability to recognize indications of clinical risk and respond appropriately as a conversation
 Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

develops. Provider-facing tools require assessments tailored to their role, including the
accuracy of information presented to clinicians and, where applicable, the reliability of risk
alerts. In either setting, testing should examine both missed risks and unnecessary escalation,
including differences in performance across populations and ways of expressing distress.
Research demonstrates promise for AI-assisted risk identification in defined settings,1 but
detection performance must be evaluated for the specific task, population, and deployment
setting.

Open-source frameworks such as VERA-MH offer developers a practical way to examine
chatbot behavior before exposing patients to a tool. VERA-MH uses simulated, multi-turn
conversations and a clinical scoring rubric to assess how systems respond to safety concerns,
such as suicidality or harm from others.2 This approach can reveal weaknesses that a single
question-and-answer test may miss and gives developers an opportunity to revise and retest
their products before deployment.

A chatbot’s ability to detect risk is important, but it is only one part of a safe response.
Developers should specify what human involvement follows that detection: who receives that
alert, what information they receive, how quickly they must respond, and how the patient is
connected to appropriate care. Evaluation should test the full workflow, including whether
alerts reach the responsible person and result in timely follow-up.

We urge the FDA to consider incentivizing rigorous, transparent benchmarking and workflow
evaluation.

Post-Market Surveillance

Premarket benchmarking is important, but it is not sufficient to establish continued safety and
effectiveness in clinical use.3 Differences in patient histories, symptom presentations, language,
and care environments cannot fully be reproduced in simulated scenarios. A tool may perform
well on a standardized test and still respond unsafely or ineffectively in real-world encounters.

Postmarket evaluation, or the ongoing assessment of a product’s safety and effectiveness after
it becomes available for use, should address two distinct issues: whether a tool remains safe in
actual use and whether it achieves its intended clinical outcomes. This approach is analogous to
postmarketing surveillance for medicines, including Phase IV clinical trials, which can identify
risks and assess outcomes after a product enters broader use.

For a chatbot used to deliver psychotherapy, for example, safety monitoring should examine
whether it avoids reinforcing maladaptive behaviors, stays within its intended role, and
responds appropriately to clinical risk. Effectiveness evaluation should examine whether it
 Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

delivers the intended evidence-based psychotherapy and whether patients experience
meaningful improvements in symptoms, functioning, or other relevant outcomes. Appropriate
delivery of a therapeutic approach and improvement in patient outcomes both require
evaluation; neither should be assumed from the other.

Clinician involvement is important to both post-market safety and effectiveness evaluations.
Qualified mental health professionals can review encounters to assess whether responses
follow the intended therapeutic approach, account for clinical context, and identify departures
that could undermine treatment. These reviews should be paired with systematic measurement
of patient outcomes to determine whether the tool provides meaningful clinical benefit. Safety
and effectiveness findings should be assessed across patient populations and following changes
to the model or application.

Postmarket monitoring allows manufacturers to track how their products perform in everyday
care and take corrective action when problems emerge. Depending on the severity and scope
of a problem, these actions may include revising safeguards, restricting a function, reverting an
update, or suspending use. Manufacturers should establish who is responsible for these
decisions and how affected providers and patients will be informed.

Professional societies representing mental health professionals can contribute to both safety
and effectiveness evaluation by helping define relevant harms and clinical outcomes,
developing structured methods for reviewing encounters, and supporting clinician education
and reporting of concerns. Their involvement can strengthen the clinical relevance and
consistency of evaluation while manufacturers remain accountable for their products.
Consistent with the FDA’s total product life cycle approach, postmarket evidence should
complement adequate premarket evaluation and inform ongoing assessment of a device’s
benefits and risks.

Conclusion

Realizing the potential of generative AI-enabled technology depends on distinguishing tools
built and tested for clinical use from those that are not, and on evidence that continues from
development through postmarket use. The Meadows Institute appreciates your attention to
these questions and your commitment to a risk-based, total product life cycle approach.

We look forward to working with the FDA as it refines its approach. The Meadows Institute
would welcome the opportunity to convene FDA staff, behavioral health clinicians, researchers,
technology developers, professional societies, and other collaborators and experts to discuss
the considerations raised in this letter and how they can be put into practice.
 Public Comment: Considerations for the Regulation of Generative AI-Enabled Medical Devices

Please contact John Snook at jsnook@mmhpi.org or (571) 331-5725 with any questions.

Sincerely,

Chief Policy Officer, Meadows Mental Health Policy Institute

References

1. Holley, D., Daly, B., Beverly, B., Wamsley, B., Brooks, A., & Zaubler, T. (2025). Evaluating
generative pretrained transformer (GPT) models for suicide risk assessment in synthetic patient
journal entries. BMC Psychiatry, 25, Article 753. https://doi.org/10.1186/s12888-025-07088-5
2. Belli, L., Bentley, K., Alexander, W., Ward, E., Hawrilenko, M., Johnston, K., Brown, M., &
Chekroud, A. (2025). VERA-MH concept paper: Validation of ethical and responsible AI in mental
health. Spring Health. https://8056012.fs1.hubspotusercontent-na1.net/hubfs/8056012/VERAMH%20Concept%20Paper.pdf
3. Nelson, B. W., Kalinich, M., & Torous, J. (2026). Specialized or general-purpose—The wrong
question for mental health AI safety. JAMA, 336(4), 281-282.
https://doi.org/10.1001/jama.2026.11448
Positive Use Cases of AI in Mental Health

OVERVIEW OF POTENTIAL APPLICATIONS
OF AI IN MENTAL HEALTH CARE
Artificial intelligence (AI) is reshaping health care, offering an unprecedented opportunity to positively
transform mental health services when used to support clinicians and other mental health providers
under the direction of human clinicians. Recognizing the rapidly evolving landscape of AI in mental
health care and that many applications still require more rigorous testing and validation in clinical
settings, this document is intended to provide a broad overview of current and potential opportunities
for AI to support the continuum of mental health care.

Clinical Application Description

AI tools can transcribe and summarize session notes and automate intake data, including recording

Workflow
Clinical Documentation symptoms, documenting medical history, and documenting screening results, greatly relieving
provider workload.

Automation AI tools can streamline administrative tasks, facilitating care coordination through record sharing,
Administrative Tasks
referrals, registering patients, billing, and scheduling.

Provider Education &
AI tools can provide more consistent and reliable high-quality training for clinicians, peers, and
Provider Training other mental health providers through content personalization, specific feedback and coaching,
and realistic, adaptive, patient-simulated training to increase learning and retention.

Quality Improvement
Provider Quality AI tools can assess providers’ fidelity to evidence-based practices and support continuous learning
Monitoring and feedback to improve outcomes.

AI tools can screen and stratify patients by risk level to prioritize urgency, assess need to
Intake Assessment
determine level of care needed, and connect patients to the best provider based on specialty,
Clinical Decision
& Care Placement
availability, insurance, and more.

Support & AI tools can analyze facial, voice/speech patterns, movement, or other data (e.g., screening results,

Predictive Tools
EHR data) to inform clinicians and help them make more precise diagnostic decisions regarding
Diagnostic Support
mental health conditions. AI can also be used to keep prescribers and patients mindful of potential
medication side effects.

Learn more at mmhpi.org
 AI tools can support more reliable and seamless data collection for clinicians and help them better
use data to inform clinical decision-making by aggregating and analyzing data, and by providing

Clinical Decision Support
Measurement & visualization of symptom measures used in measurement-informed care to support more
Outcomes Monitoring personalized treatment. Tools can also help clinicians better identify and predict emerging risks,
monitor patient engagement, encourage medication adherence, improve outcomes, and ensure
patients are receiving appropriate care.

& Predictive Tools
AI tools can help translate evidence-based treatment guidelines to support consistent delivery of
Treatment-Protocol best practices into everyday practice. AI tools can make recommendations to clinicians on next
Support steps in care pathways, suggest treatment adjustments based on symptom changes, identify
potential side effects from medications, and help clinicians utilize standardized protocols.

When a patient opts in, AI can analyze patient data from smartphones or wearables to screen
or monitor for mental health issues, response to treatment, and/or relapse. This may include
Patient Monitoring
tracking a variety of physiological, social, and behavioral data, such as heart rate, sleep, movement
patterns, and facial and vocal characteristics.

Patient-Facing Tools
Personalized Digital AI tools can help personalize digital mental health apps to tailor intervention activities, such
Mental Health as mindfulness, cognitive behavioral interventions, or psychoeducation, to act as an adjunct or
Support Tools enhancement to clinician care.

Chatbots specifically designed and validated for mental health applications can be employed and
directed by clinicians across a range of applications, from guiding patients through skill-building
Purpose-built Mental between therapy sessions to conducting therapeutic conversations under clinician direction and
Health Chatbots1 with control to reduce mental health symptoms. While these tools hold promise in extending clinical
Clinician Oversight care, additional research on effectiveness and safety is necessary. Purpose-built chatbots differ
from other chatbot models by design, expressly keeping the clinician in the control seat for close
monitoring and quality control.

Population
AI tools can help health systems manage population health goals by detecting population-level
Population & Systems trends to inform decision-making, such as how to target programs or interventions. AI can also

Health
Level Health Analytics better enable and reduce clinician burden on quality measurement and reporting for value-based
payment models.

Research
AI can enhance computing power to analyze large volumes of complex data, such as brain images,
AI & Neuroscience to better understand potential causes and impacts of mental health conditions and predict
responses to treatment.
Chatbots refer to computer programs designed to simulate conversation with human users.

©2026 Meadows Mental Health Policy Institute

Learn more at mmhpi.org
Positive Use Cases of AI in Mental Health Care

THE DATA POWERING AI
SOLUTIONS IN MENTAL HEALTH

AI technologies hold the potential to transform many aspects of mental health care. Across use cases – from screening
and diagnosis to monitoring and treatment – the transformative power of AI derives from its ability to leverage massive
amounts of data from multiple sources and make it more actionable. In doing so, AI can identify meaningful patterns and
generate insights to support better decision-making for clinicians and patients.1, 2 This means earlier detection of mental
illness, enhanced quality of care, better medication management, improvements in treatment adherence, improved care
coordination, and ultimately, improved outcomes, reduced costs, and less suffering.

The Data Underpinning AI Tools in Mental Health Care
One of the most promising aspects of leveraging AI is the potential to create a more complete treatment picture by pulling in data
from a variety of sources. AI technologies can analyze data from sources inside and outside the health care setting to generate
meaningful insights, many of which are not currently being used, including:3

Source Data Extracted

Electronic Data on prescriptions, refills, procedures, appointments, admissions, observations,
health
patient reported symptom assessments, and laboratory results4
records

Wearables and Multimodal sensing data (such as sleep patterns, physical activity, heart rate,
smartphones and smartphone usage) to support real-world monitoring5

Voice data to capture acoustic and speech patterns or facial expression data to
Audio, text, analyze visible emotional cues and psychomotor indicators that may indicate
and video
psychiatric symptoms and symptom severity6,7

Clinical notes and Unstructured data (such as therapy notes, clinical observations, secure messaging,
patient-generated or patient journals) to detect changes in cognition, affect, or functioning and
text identify emerging risk signals8

ECG or heart Heart rhythm and electrical activity to assess physiological responses associated
rate monitor with stress, anxiety, and depression9

Neurophysiological data to identify biomarkers and neurocognitive profiles of
Brain images
mental conditions10

Physiological data from laboratory tests to analyze biological markers
Laboratory tests (such as genetic expression and metabolic indicators) associated with mental
health conditions11

Learn more at mmhpi.org
AI uses multiple technologies to extract, analyze, and synthesize data from these sources, transforming raw data
into meaningful insights and organizing them so clinicians and patients can easily put them into action.

Data Extraction Data Analysis Data Synthesis

Extracts information from Recognizes patterns, detects Organizes, visualizes, and
a variety of sources using changes, and predicts trajectories, synthesizes data to make the
technologies such as computer often using machine learning and analyzed data readily usable by
vision, speech recognition, or deep learning providers and patients
natural language processing

While technologies and their evidence-bases are constantly evolving, the insights from these data can be employed
across a range of mental health care settings to help identify, diagnose, monitor, and treat mental health conditions.
By unlocking these new sources of data, mental health care has the potential to reach the level of precision already
common in other areas of medicine, delivering more personalized care and better outcomes.12

Responsible Use Considerations for Data in AI Mental Health Technologies
The Meadows Institute supports the responsible use of AI in mental health care. While many safeguards and
governance practices are necessary to ensure AI is used ethically and effectively, some considerations that are
particularly important when evaluating the data that underpin these technologies include:

Datasets as the building blocks of AI: Ensuring that AI applications are developed and tested using robust,
clinically grounded, high-quality datasets is foundational to developing effective and ethical tools. Datasets must
be representative across care settings, mental health conditions, demographics, languages, and uses.13

Real-world application research: Although current evidence for AI’s performance in research settings is
promising, additional longitudinal, real-world trials are needed to assess utility, sustainability, and fairness when
new data streams (such as voice analysis and multimodal sensing) are used in routine clinical practice.14

Privacy and security: As with all health care technology, strong privacy and cybersecurity protections must be in place
to safeguard sensitive information. Additional precautions are necessary when handling sensitive biometric data.15

Koutsouleris et al. (2022). https://doi.org/10.1016/S2589-7500(22)00153-4, 2Stade et al. (2024). https://doi.org/10.1038/s44184-024-00056-z, 3Baydili et al (2025). https://doi.org/10.3390/diagnostics15040434,
4
Koutsouleris et al. (2022). https://doi.org/10.1016/S2589-7500(22)00153-4, 5Hau et al. (2025). https://doi.org/10.1016/j.schres.2025.05.019, 6Muddaloor et al (2025). https://doi.org/10.3390/s25113424,
Ramanarayanan (2024). https://doi.org/10.1044/2024_JSLHR-24-00142, 8Kim et al. (2025). https://doi.org/10.1007/s40501-025-00359-8, 9Linardon & Torous. (2025). https://doi.org/10.1002/eat.24468,
Lee et al. (2021). https://doi.org/10.1016/j.bpsc.2021.02.001, 11Baydili et al (2025). https://doi.org/10.3390/diagnostics15040434, 12Le Tourneau, and Bièche. (2018). https://doi.org/10.2217/pme-2018-0036,
Mandal et al. (2025). https://doi.org/10.48550/arXiv.2508.09809, 14Nilsen etal. (2022). https://doi.org/10.1177/26334895221112033, 15Magee, M. & Farahany (2024). https://doi.org/10.1016/j.neuron.2024.09.004

©2026 Meadows Mental Health Policy Institute

Learn more at mmhpi.org
Positive Use Cases of AI in Mental Health Care

ACCELERATING OPTIMAL
CARE PLACEMENT AT INTAKE
Mental health care is most effective when people receive support early and at the right level of care. Yet, long
wait times and mismatches between individual needs and available services continue to delay treatment, worsen
symptoms, and add strain to our already overburdened health systems.

How AI Streamlines Patient Intake and Care Placement
When implemented responsibly, AI can help strengthen intake processes under the supervision of clinicians by analyzing
large volumes of available data to better assess urgency and match people with the most appropriate level of care:

Referrals
Patient narratives Patient reported
and messages medical history

Past medical
Validated measures records (diagnoses,
and symptom reports AI Risk Stratification medications, etc.)
Platform
STEP 1. A patient completes an electronic STEP 2. AI analyzes the data alongside
intake form with symptoms, needs, and history. relevant electronic health records (EHRs).

Recommended Care Plan Options
Digital cognitive behavioral
LOW
therapy, peer coaching

MEDIUM Individual therapy

HIGH Psychiatry, care coordination

STEP 4. A clinician reviews the recommendations and STEP 3. AI predicts likely level of severity, recommends
meets with the patient to finalize the care plan. appropriate care, and flags patients with urgent needs.

Potential Benefits of AI-Supported Intake1,2,3

COMMON CHALLENGES IMPROVEMENTS WITH AI

Time-intensive manual review of patient intake information More efficient processing of high volumes of information using minimal staff time

Urgent cases flagged for immediate clinician review; less urgent cases are
First-come, first-served waitlists
matched with other clinical resources on a routine pace

Data-driven matching of patient needs and preferences to the right
Multiple appointments to reach the appropriate provider
clinical expertise

A limited number of specialty providers cannot meet Helps clinicians direct high-need patients to specialty care and steers
patient needs lower-severity cases to lower-intensity services

These tools are designed to support intake, risk identification, and care placement by humans. Clinical diagnosis and treatment decisions remain the responsibility of licensed care providers.

Learn more at mmhpi.org
Early Evidence Points to AI’s Opportunity in Mental Health Triage
Risk prediction is widely used in many areas of health care to determine the appropriate level of care, with more
than 70% of hospitals using predictive AI.4 Research shows that AI-supported risk prediction can help clinicians
identify co-occurring conditions to better support people with complex chronic needs.5 Early evidence from around
the world suggests similar promise in mental health care.
2X RATE OF
RECOVERY

9 13 50
58%
HOURS MINUTES

A U.S. telehealth network cut crisis
71%
An outpatient mental health clinic
27%

A UK-based telehealth platform
specialist review times from ~9 in Canada saw a 71% drop in wait doubled recovery rates (27% to 58%)
hours to 13 minutes using AI risk times with AI-supported intake.7 with AI-enabled intake versus
assessment.6 standard referral.8

Core Principles for Responsible AI Use Responsible Use Spotlight

Responsible use of AI should: Especially when used for risk prediction, AI
• Demonstrate safety and effectiveness across systems must use outcome metrics that directly
populations
reflect the problem being addressed, rather than
• Protect patient privacy and confidentiality
indirect proxy measures. For example, prior health
• Easily integrate into routine care settings
care spending may correlate with medical need
• Have been trained and tested on high-quality,
clinically validated data but should not substitute for assessing actual

• Augment– never replace– clinical judgment and clinical need, even if it is easier to measure.9
decision making

Koutsouleris et al (2022) doi.org/10.1016/S2589-7500(22)00153-4; 2Lee et al (2024) doi.org/10.2196/58129; 3Lee et al (2025) doi.org/10.1016/j.jval.2025.06.002; 4Chang et al (2025) www.ncbi.nlm.nih.gov/books/NBK618497/;
5
Soufa et al doi.org/10.3390/jcm14103434; 6Swaminathan et al (2023) doi.org/10.1038/s41746-023-00951-3; 7Stephenson et al (2026). https://doi.org/10.1177/07067437251355641 8Rollwage (2023) doi.org/10.2196/44358;
Obemeyer et al (2019) doi.org/10.1126/science.aax2342;

©2026 Meadows Mental Health Policy Institute

Learn more at mmhpi.org
Positive Use Cases of AI in Mental Health Care

AI-ENABLED DIAGNOSTIC
SUPPORT FOR PRECISION
MENTAL HEALTH CARE
An accurate diagnosis is the first step toward effective care. One of the complexities in diagnosing mental health conditions
is that they often share overlapping symptoms. This is further complicated by the fact that diagnoses often rely heavily on
patient-reported experiences and clinical observation. Consequently, mental health care is too often characterized by delays
in reaching diagnostic clarity and challenges in developing personalized treatment plans.1

How AI Supports Clinicians in Diagnosis and Assessment
AI-enabled technologies can augment traditional diagnostic methods by incorporating insights from multiple data sources – such
as voice and facial analysis, neuroimaging, and physiological and behavioral data from smartphones and wearables – into a more
integrated assessment process that can help clinicians make faster and more precise diagnoses.2

Facial analysis Voice Analysis
(Flat affect, gaze patterns) (Rate, tone, pitch)

AI extracts data from multiple sources, using techniques
Clinician interview Standardized questionnaires, such as facial and voice analysis.
& observation patient history scoring​

AI models analyze these behavioral
Clinician conducts traditional assessment with patient
signals and provide additional insights on
(e.g. clinical observation, standardized questionnaires).
behavioral patterns and symptom severity.

Clinicians use insights from all of these
sources to make an integrated assessment
and support diagnosis.

Potential Benefits of AI-Enhanced Multimodal Diagnostic Assessment3
• Increased insight into emotional states and behaviors that do not rely solely on patient self-report
• Differentiated diagnoses with similar clinical presentations, but different treatment needs
• Identification of previously unrecognized subgroups of patients within a broader diagnosis, such as bipolar
disorder or psychotic disorders
• Analyses of patterns over time to refine the accuracy of diagnoses as symptoms evolve

Learn more at mmhpi.org
Promising Research Points to the Transformative Potential of AI-Supported
Diagnostics in Mental Health
Precision medicine has reshaped cancer treatment by targeting the genetic and molecular drivers of disease rather
than tumor location, enabling more individualized therapies and drastically improving health outcomes.4 Although
gaps remain between research findings and routine clinical use, AI tools show promise in supporting similar
precision in the diagnosis of mental health and other brain disorders.

14 studies • multiple countries

75%
accuracy detecting depression
89%
accuracy distinguishing PTSD
85%
diagnostic accuracy across
via facial expressions in veterans by voice patterns psychiatric disorders

An Al-enabled app using a phone’s An Al tool was able to distinguish A systematic review of 14 studies from
front camera to capture facial between veterans’ voices with PTSD different countries found that Al tools
expressions correctly determined and those without PTSD with 89% for psychiatric disorders achieved 85%
diagnoses of depression with 75% accuracy.6 diagnostic accuracy.7
accuracy.5

Core Principles for Responsible AI Use Responsible Use Spotlight

Responsible use of AI should: While reviews from around the world have found
• Demonstrate safety and effectiveness across high levels of accuracy in AI-supported diagnostic
populations assessments, many studies are based on small sample
• Protect patient privacy and confidentiality sizes and limited population diversity. A broader
• Easily integrate into routine care settings research base is vital for the advancement of the
field and to ensure maximum safety for impacted
• Have been trained and tested on high-quality,
clinically validated data populations.8 Strong privacy and cybersecurity
protections must also be in place to safeguard
• Augment– never replace– clinical judgment and
decision making sensitive information, especially biometric data.9

Galatzer-Levy, Aranovich, Insel. (2023). https://doi.org/10.1162/daed_a_02040, 2Rony et al. (2025). https://doi.org/10.1177/20552076251330528, 3Lee et al. (2021). https://doi.org/10.1016/j.bpsc.2021.02.001,
4
Le Tourneau, and Bièche. (2018). https://doi.org/10.2217/pme-2018-0036, 5Nepal et al. (2024). https://doi.org/10.1145/3613904.3642680, 6Marmar. (2019). https://doi.org/10.1002/da.22890,
Ramanarayanan (2024). https://doi.org/10.1044/2024_JSLHR-24-00142, 8Mandal et al. (2025). https://doi.org/10.48550/arXiv.2508.09809, 9Magee, M. & Farahany (2024). https://doi.org/10.1016/j.neuron.2024.09.004

©2026 Meadows Mental Health Policy Institute

Learn more at mmhpi.org
Positive Use Cases of AI in Mental Health Care

DATA-DRIVEN MEASUREMENT
AND MONITORING TO
ENHANCE CARE QUALITY
Research shows that the consistent tracking of behavioral health symptoms, progress, and outcomes using brief, standardized
questionnaires improves outcomes, supports earlier interventions, and reduces costs. Despite its proven benefits, this
approach is not yet deployed throughout mental health care. AI can reduce the burdens associated with measurement and
enable greater use of measurement-informed care as a key component of everyday practice.1

How AI Improves Measurement and Monitoring to Enhance Care Quality
and Patient Outcomes
Systematic measurement of clinical markers – such as blood pressure, laboratory tests, and A1c levels – has long underpinned
treatment and ongoing care in other physical health conditions.2 AI can more seamlessly integrate measurement into clinician
workflows, supporting the meaningful use of mental health data in routine care. It can also incorporate additional data sources to
complement traditional symptom reporting.3

EHR data (diagnoses, medications,
Clinical notes & observations missed appointments, etc.)

Patient engagement information Validated symptom scores
(homework, journals, symptom reports) (PHQ-9, GAD-7, etc.)
AI Model
STEP 1. Multiple measures and other data are gathered in the clinic and through patient monitoring tools.

STEP 3. The clinician reviews the AI-generated data and STEP 2. The AI platform continuously tracks measures from
insights and meets with the patient to discuss potential multiple sources, detects patterns, and, when appropriate,
adjustments to the care plan. generates recommendations for treatment adjustments.

Benefits of AI-Supported Measurement and Monitoring4,5,6

COMMON CHALLENGES IMPROVEMENTS WITH AI

Manual entry and scoring of symptom measures are Automated collection, synthesis, and visualization of data help clinicians
often cumbersome and difficult for clinicians to find track trends, progress, plateaus, or deterioration and consider data-informed
and use in practice adjustments to care

Utilization of unstructured data (e.g., clinical notes) and between-visit
Measurement based solely on symptom
monitoring through patient-facing tools, such as remote symptom
assessments taken in the clinic
assessments and digitally completed therapy exercises

Limited identification of patients at risk of crisis,
leading to poorer outcomes and strain on health Identification of patterns and early warning signals (e.g., emerging safety
systems concerns or potential treatment non-response) to prompt proactive follow-up

Learn more at mmhpi.org
Research Suggests Positive, Measurable Benefits
With the right tools, AI-enabled, measurement-informed care can help providers better track progress, individualize
treatment, and improve outcomes.
50 66% 58% 5%
5%
0 Reduced Reduction
inpatient in suicide
Questionnaire More patients with indicators admissions attempt
completion rates of suicidal ideation & ER Use probability

An algorithm-enhanced, measurement- A digital mental health company The VA’s REACH VET machineinformed care platform increased Patient identified 58% more patients with learning program was linked to fewer
Health Questionnaire (PHQ-9) completion indicators of suicidal ideation than inpatient mental health admissions,
rates from 5% to 66%, with 78% of would have been detected through less emergency room use, and a 5%
clinicians reporting use of the platform traditional self-reported depression reduction in the probability of a
to guide care.7 measures (PHQ-9).8 suicide attempt.9

Core Principles for Responsible AI Use Responsible Use Spotlight

Responsible use of AI should:
To allow for meaningful human oversight,
• Demonstrate safety and effectiveness across
AI outputs must be explainable. Providers must
populations
• Protect patient privacy and confidentiality be able to understand the basis for AI-generated

• Easily integrate into routine care settings risk scores and recommendations to provide
• Be trained and tested on high-quality, clinically meaningful oversight, and the use of AI should
validated data
be transparent to patients.10
• Augment– never replace– clinical judgment and
decision making

* The Patient Health Questionnaire-9 (PHQ-9) is a standardized, 9-question tool used to screen, diagnose, and monitor the severity of depression.

Koutsouleris. (2022). https://doi.org/10.1016/S2589-7500(22)00153-4, 2Bright Spots in Measurement-Informed Mental Health Care. (2025). Meadows Mental Health Policy Institute; West Health. https://mmhpi.org/project/
improving-outcomes-through-measurement-informed-care/, 3Galatzer-Levy, Aranovich, Insel. (2023). https://doi.org/10.1162/daed_a_02040, 4Steidtmann et al. (2025). https://doi.org/10.1176/appi.ps.20240135,
Hau et al. (2025). https://doi.org/10.1016/j.schres.2025.05.019, 6Xu et al (2023). https://doi.org/10.1176/appi.prcp.20220015, 7Steidtmann et al. (2025). https://doi.org/10.1176/appi.ps.20240135,
Hartz et al. (2024). https://doi.org/10.36401/IDDB-23-10, 9McCarthy et al (2021). https://doi.org/10.1001/jamanetworkopen.2021.29900, 10Joyce et al (2023) https://doi.org/10.1038/s41746-023-00751-9

©2026 Meadows Mental Health Policy Institute

Learn more at mmhpi.org