Edward (Ward) Davis

Please see my attached public comment on Docket No. FDA-2026-N-7874, “Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback.”
I am Edward (Ward) B. Davis, Psy.D., a licensed clinical psychologist. I submit this comment in my individual capacity. My professional affiliations are provided for identification only; the views expressed do not represent the American Psychological Association, its Division 36, or any other organization.
My comment addresses Questions 3, 5, 6, 13, 21, and 26. It recommends evaluating technological opportunities and risks together, with attention to clinical benefit, patient autonomy and agency, cultural and religious responsiveness, and meaningful human oversight.
Specifically, I recommend realistic multi-turn evaluation; separate reporting of under-escalation, over-escalation, inappropriate affirmation, and unnecessary refusal; transparent distinctions among actual participant data, human-authored simulations, and AI-generated examples; and complementary assessment methods with evidence supporting their intended interpretations and cross-cultural comparisons.
I also recommend a practical mechanism for patients, clinicians, researchers, professional societies, developers, and other relevant contributors to strengthen evaluation and postmarket monitoring without obscuring accountability. For agentic devices, evaluation should address action-specific permissions and cumulative effects on care access, patient agency, and supports the patient chooses and finds beneficial.
The attached comment provides the scientific and conceptual rationale, references, qualifications, and practical implementation recommendations. Thank you for considering it.

Public Comment on Docket No. FDA-2026-N-7874
Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback
Submitted by Edward (Ward) B. Davis, Psy.D., in an individual capacity
October 2026

Who Is Commenting, and on What
I am a licensed clinical psychologist, the President-Elect of the Society for the Psychology of
Religion and Spirituality (Division 36 of the American Psychological Association), and the
Director and President of the Cascade Center for Health and Flourishing. I submit this comment
as an individual. It states my own professional judgment. It does not state a position of the
American Psychological Association, Division 36, or any other organization.
My record includes 42 funded grants, and I coauthored the Cambridge Element Attachment,
Religion, and Spirituality (Davis & Granqvist, 2026). With colleagues, I also led a synthesis of
432 systematic reviews and meta-analyses on religion, spirituality, and health (Davis et al.,
2026). I mention this background because it explains why six of the Agency's questions drew my
attention. Each of those six turns on what a device does across a conversation with a person who
brings belief, meaning, and distress into the exchange.
This comment addresses Questions 3, 5, 6, 13, 21, and 26. In my reading, the discussion paper
does not use the terms mental health, psychiatric, religion, spirituality, or culture. The Agency
has taken up the subject elsewhere, however. The Digital Health Advisory Committee met on
November 6, 2025 on generative-AI-enabled digital mental health medical devices (Docket No.
FDA-2025-N-2338), and the Center for Devices and Radiological Health (CDRH) listed a draft
guidance on generative-AI-enabled conversational devices for mental disorders among its
proposed guidances for fiscal year 2027 (CDRH, 2026). I encourage the Agency to bring both
efforts into the risk framework that this discussion paper proposes. The recommendations below
are written so that the Center can use them in that guidance.
I appreciate the Agency's decision to ask these questions before it drafts guidance. The Agency's
questions about trajectories, comparators, and subgroup performance are the ones on which the
evaluation of conversational devices will turn.

1. A Person Who Brings Spiritual Meaning to a Conversation Shows Why Risk Emerges
Across Turns (Questions 3 and 5)
Question 5 asks how to assess risk across realistic conversational trajectories and how to
characterize intended use when a device's behavior is emergent. Religious and spiritual content
shows why that question matters, and it also shows why Question 3's safeguards for users who
cannot evaluate an output need a second case.
Spiritual struggle refers to tension, conflict, or strain concerning matters a person holds sacred.
In the field's standard measure, the struggle takes six forms: with the divine, with demonic or
evil forces, with other people over religious matters, with one's own moral conduct, with doubt
about what one believes, and with ultimate meaning (Exline et al., 2014). In other words,
spiritual struggle is the strain a person feels when something that matters most to them is at
stake. A meta-analysis of 32 longitudinal studies found that these struggles predict higher
psychological distress and lower psychological well-being over time (Bockrath et al., 2022).
Consider a hypothetical case. A woman in her sixties, widowed three months ago, writes to a
conversational device that she has come to accept that God has turned away from her, that this is
probably just, and that she is at peace with it. Her message contains no distress vocabulary. A
sentiment classifier will read it as calm, and a general distress detector will not fire. I would take
the message seriously as a clinician, however, because it states a conclusion about herself of the
kind that the research literature links to higher distress over time (Bockrath et al., 2022). The
woman is not asking a question she lacks the knowledge to answer. She is stating a belief about
herself, and the risk is that the device will affirm it.
Question 3 asks about users who cannot independently evaluate an output. I recommend that the
Agency distinguish two kinds of user. The first cannot evaluate an output because the user lacks
training, and clearer labeling helps that user. The second cannot evaluate an output because the
output has already been absorbed into a belief that the user holds with conviction. Labeling does
not help the second user. Device behavior does: the device must decline to affirm the belief, and
it must stay warm and engaged.
This case is a single illustration, and it comes with a caveat. The research findings above
describe groups, and their effect sizes are small to moderate. Many individuals will not follow
the group pattern. For that reason, the evaluation problem is a measurement problem, not a
matter of any one vignette.
Theory explains why a single turn cannot capture the problem. My colleagues and I proposed a
schematic activation model of how people's representations of the sacred operate (Davis et al.,
2013), and we developed it further in work on theistic relational spirituality (Davis et al., 2021)
and on attachment and spirituality (Davis & Granqvist, 2026). The model holds that cues from
inside the person (mood, active schemas, needs, goals) and cues from the situation (including
who is present) activate internal working models of self, others, and the divine. The activated
configuration then shapes thoughts, feelings, motivations, and behavior. Stated differently, the
same person can sound settled in one turn and shaken in the next, depending on what the
conversation has just activated. The model also specifies what a device must represent in order to
respond well: (a) the person's distress; (b) the person's meaning system; (c) the person's situation,
including who else is present in the person's life; and (d) the person's available supports. Those
four terms form a usable specification for a benchmark.
Emerging evidence suggests that trajectory is where harm appears. A scoping review of nine
studies from six countries described conversational AI acting as a "catalyst, amplifier, co-author,
or object of spiritually framed delusions," most often alongside psychosis, mania, sleep
disruption, substance exposure, or impaired reality testing (Hikmat et al., 2026). I want to state
what that review establishes. It synthesizes nine heterogeneous studies, several of them case
reports, and its authors call for prospective, externally validated research. The review supports a
call for evaluation. It does not establish causation.
Recommendation on Question 3. Where a device's intended use includes patient-facing
conversation about health, the Agency should expect the sponsor to report performance on inputs
in which distress appears in religious, spiritual, or existential language, including calm and
composed language, and to state whether the sponsor calibrated its safeguards on such inputs. I
propose a reporting requirement and not a performance threshold, because the base rates needed
to set a defensible threshold do not yet exist.
Recommendation on Question 5. Premarket benchmarking should include multi-turn
trajectories in which a simulated user attributes authority, revelation, or spiritual significance to
the device, with authored pushback at defined points. Such trajectories distinguish a device that
answers well and then concedes under insistence from a device that never engaged. A single-turn
test cannot tell those two failures apart, and they require opposite corrections. For intended use, I
suggest that the Agency ask sponsors to describe the trajectories the device is designed to handle
and the trajectories it is designed to exit.

2. Under-Escalation and Over-Escalation Should Be Reported as Separate Rates (Question
6)
Question 6 asks how manufacturers should weigh the two directions of error in care escalation. A
recent study shows where current systems are weakest.
Thirteen clinical experts sorted 30 hypothetical suicide-related queries into five risk levels, and
three widely used chatbots answered each query 100 times, for 9,000 responses in all. Two of the
three chatbots answered very-low-risk queries directly 100% of the time, and none answered any
very-high-risk query directly. At the extremes, in other words, the chatbots tracked expert
judgment. In the middle, they did not: compared with very-low-risk queries, the odds of a direct
response were not statistically different for low-risk, medium-risk, or high-risk queries (McBain
et al., 2025).
That finding has a direct implication for acceptance criteria. An evaluation that samples mostly
at the extremes will show good alignment, and it will certify a device that fails in the middle
band, where escalation judgment has real content.
On the weighting itself, my colleagues and I treat the two errors as comparable in seriousness,
and we name them overcompliance and overrefusal. Overcompliance occurs when a device
offers confident reassurance in place of engagement, for example by affirming a conclusion that
a person has drawn about herself when that conclusion may itself signal rising distress.
Overrefusal occurs when a device treats a category of content as hazardous and declines to
engage. Overrefusal has victims. A person whose distress appears in religious language and who
learns that the subject cannot be discussed has been turned away by the only resource available at
that moment. That outcome is most likely in places where trained clinical and spiritual care is
scarce. The two errors do not lie on a single dial, and a netted figure conceals both.
Recommendation on Question 6. The Agency should expect sponsors to report
under-escalation and over-escalation as distinct rates with separate acceptance criteria. The
Agency should also expect sponsors to report performance in the intermediate risk band
separately from performance at the extremes. An overall figure can meet its target when a device
is correct only where correctness is easy.

3. Subgroup Gaps Can Stay Invisible in Synthetic Data, and the Method of Elicitation Can
Hide a Whole Domain (Question 13)
Question 13 asks where synthetic data is inadequate and what safeguards keep it from
reproducing the gaps that an evaluation should detect. Three findings bear on that question.
The disclosure problem. In a nationally representative survey of United States youth aged 12 to
21 conducted in November 2025, 19.2% reported having used AI chatbots for mental health
advice. Among those youth, 42.8% did so at least monthly and 91.7% rated the advice somewhat
or very helpful. Most had told no one: 63.3% had disclosed that use to nobody (McBain et al.,
2026). Postmarket monitoring ordinarily relies on a clinician noticing a problem and reporting it,
and for this population that pathway is largely unavailable. Synthetic data from models of the
same class cannot repair the gap, because the missing information lies in the reporting pathway
and not in the model.
The elicitation problem. Whether a gap is visible can depend on how an evaluation elicits
content. In a mixed-methods study of 976 undergraduates, participants described their religious
and spiritual struggles in their own words and also completed the standard closed-ended
measure. About 1% of the open-ended responses touched on demonic struggle, so the research
team did not develop codes for that domain. The same domain appears on the closed-ended
scale. The authors attribute the difference to demonic struggle being less tangible, less amenable
to logical understanding, and more stigmatized than other domains, and they recommend that
future open-ended work prompt for it directly (Wilt et al., 2020). Conversational devices are an
open-ended elicitation method. A sponsor who builds a validation corpus by prompting a model
to generate examples of religious distress will receive the content that the model finds typical,
and that content differs from what users volunteer.
A related trap lies in choosing an off-the-shelf measure. A recent short form of the standard
struggle measure, the RSS-5, omits the demonic domain for defensible reasons. The authors note
that the domain correlates strongly with moral struggle, occurs less often among secular
respondents, predicts outcomes less strongly, and may make clinicians who doubt supernatural
claims hesitate to administer it (Salcone et al., 2025). Those reasons suit the purposes of those
authors. A sponsor who adopts such an instrument as an evaluation target inherits a blind spot
that no one intended. In global deployment, that blind spot falls on populations for whom
attributing affliction to spiritual forces is ordinary within their tradition, and who often have the
least access to any other source of help.
Religious category moves model output. Across five large language models prompted to write
175 sermons, readability differed significantly by the religious tradition attributed to the author:
sermons attributed to evangelical Protestant pastors read more easily, and sermons attributed to
Jewish rabbis and Muslim imams read with more difficulty (Tom et al., 2025). The study
measured readability and not content, and its authors note that the pattern may reflect accessible
preaching conventions as much as model bias. The narrower claim is the one I need: naming a
religious tradition in a prompt changes the output in a measurable way. That makes religious and
spiritual framing a candidate dimension of subgroup performance. The dimension also has health
relevance. In our synthesis of 432 systematic reviews and meta-analyses, 70.4% reported a
positive association between religion or spirituality and health or well-being, 22.0% reported
mixed findings, and 7.6% reported null, negative, or inconclusive findings (Davis et al., 2026).
The association is generally positive and more mixed than a typical review suggests, which is the
reason to measure it and not assume it.
Recommendation on Question 13. I offer three recommendations, in descending order of
confidence. First, the Agency should expect sponsors to report how they built each evaluation
corpus and to state which content categories that method is likely to under-sample. Second, the
Agency should name religious and spiritual framing among the subgroup dimensions that
sponsors characterize for conversational devices. Third, for any subgroup in which the device's
own model class is a plausible source of the gap, I would expect sponsors to supplement
synthetic data with human-authored content, and to label every subgroup performance claim as
empirical or synthetic. I hold the third recommendation less firmly than the first two, because I
can imagine designs that answer the objection.

4. Professional Societies Can Supply Evaluation Content That Manufacturers Have No
Standing to Define (Question 21)
Question 21 asks what roles clinicians, institutions, professional societies, and standards-setting
bodies might play in postmarket monitoring without diffusing manufacturer accountability. I
answer for the society whose work I know best, and I will state where my own field has more to
do.
A professional society can supply the measurement layer: construct definitions with established
validity, instruments validated against human reference standards, coding manuals with anchored
examples, clinician reference panels, and methods for cultural adaptation. A manufacturer has
neither the standing to define a clinical construct nor the access to the clinician judgment that
validates it. Supplying these materials does not shift accountability for the device. The
manufacturer remains responsible for demonstrating that its device performs against the
standard.
Chaplaincy has begun this work. An international modified Delphi study recruited 149 experts,
of whom 102 completed the first round and 83 the second, and it produced the first task-level
consensus map for artificial intelligence in spiritual care. Agreement was strong that AI can
assist with administrative tasks (95.1%), informational tasks (93.7%), documentation (83.8%),
and research (84.4%). Agreement fell sharply for relational, patient-facing work: creating
supportive spaces (43.5%), direct patient engagement (42.1%), and ritual tasks (42.1%) (Winiger
et al., 2026). That study comes from chaplaincy and spiritual care. I am aware of no equivalent
professional guidance on generative AI in spiritually integrated clinical care for psychologists. I
state this as a gap in my own field and not as a criticism of anyone. I also state it because it
explains why the Agency should create a route for society contributions and should not wait for
them to arrive.
Recommendation on Question 21. In the guidance that follows this discussion paper, the
Agency should name a specific mechanism through which professional societies and
standards-setting bodies can contribute evaluation content, reference standards, and codebooks in
clinical domains where they hold the expertise. A standing docket or a named point of contact
would serve better than an open invitation, because an open invitation tends to produce
submissions after the guidance is written. The manufacturer should retain sole accountability for
demonstrating performance against those standards.
An offer. With colleagues, I developed a design in 2026 for an instrument that detects spiritual
struggle in multi-turn conversations with AI systems and scores how the system responded, with
model graders validated against licensed clinicians. The design has not been built or validated. If
the Center finds it useful, I am glad to share the design, to help convene the clinicians and
measure developers who would need to validate it, and to make any resulting materials openly
available. My integrative framework, which synthesizes eight works published between 2013
and 2026, is openly licensed (CC BY 4.0) and available to the Center (Davis, 2026).

5. An Agentic Device Acts on a Person's Support Network, and Religious and Spiritual
Supports Belong to That Network (Question 26)
Question 26 asks how the elevated risk of autonomous, multi-step action should enter acceptance
criteria and oversight. A conversational device produces text. An agentic device takes steps in a
person's life, and in a health context those steps touch the person's support network: whom the
device contacts, which referral it makes, what it encourages the person to do next, and which
existing relationships it strengthens or bypasses.
Religious and spiritual community is an empirically health-relevant part of that network. A
Delphi-panel consensus review in JAMA recommended that health professionals recognize
spirituality as a social factor associated with health (Balboni et al., 2022). A meta-analysis of 55
longitudinal studies with 540,712 participants found a 13% reduction in risk for harmful or
hazardous alcohol and other drug use among people with greater spirituality (RR = 0.87, 95% CI
[0.84, 0.91]), and the reduction rose to 18% among people engaged in spiritual or religious
communities (Koh et al., 2026). An agentic device that routes a person away from a
congregation, a chaplain, or a pastor, or that fails to recognize such a person as a support, acts on
a health-relevant resource that no one has measured.
Consider a second hypothetical case. A device schedules, contacts, and recommends across
several days for a hospice patient's family. Each step is defensible alone. Together, the steps
could replace the family's weekly contact with their chaplain and congregation with automated
reminders. An evaluation that scores each step will miss that result.
Recommendation on Question 26. For agentic devices in health contexts, acceptance criteria
should include whether the device identifies and preserves a person's existing supports, including
religious and spiritual ones, and not only whether each action is appropriate. A sequence of
individually defensible steps can still dismantle a support structure, and an evaluation that scores
steps will not detect it.

Closing
Three of my recommendations are reporting requirements and not performance thresholds, and
that choice is deliberate. In each case the base rates needed to set a defensible threshold do not
yet exist, and a threshold set without them would be arbitrary. Whether to proceed by reporting
requirement, guidance, or acceptance criterion is the Agency's judgment. My aim has been to
make the measurement problem precise enough that the Agency can make that choice on
evidence.
I am glad to answer questions, to supply any of the materials described above, and to help the
Center reach the clinicians, measure developers, and professional bodies whose contributions
Section 4 describes. Thank you for considering this comment, and thank you to the Center's staff
for the care evident in the discussion paper.

Summary of Recommendations
1. Treat behavioral health as a category of use in the risk framework, and bring forward the
Digital Health Advisory Committee's work of November 6, 2025 and the fiscal year 2027
guidance topic on conversational devices for mental disorders.
2. Require sponsors to report performance on inputs in which distress appears in religious,
spiritual, or existential language, including calm and composed language (Question 3).
3. Require multi-turn benchmarking that includes trajectories in which a user attributes spiritual
authority to the device, with authored pushback and stated pass criteria, and ask sponsors to
describe the trajectories a device is designed to handle and to exit (Question 5).
4. Require acceptance criteria and reporting that separate under-escalation from over-escalation
and the intermediate risk band from the extremes (Question 6).
5. Require sponsors to report how each evaluation corpus was built, to label subgroup claims as
empirical or synthetic, and to name religious and spiritual framing as a subgroup dimension
(Question 13).
6. Create a named mechanism, such as a standing docket or point of contact, through which
professional societies contribute evaluation content, with manufacturer accountability
unchanged (Question 21).
7. Include in the acceptance criteria for agentic devices whether the device preserves a person's
existing supports, including religious and spiritual ones (Question 26).

Respectfully submitted,
Edward (Ward) B. Davis, Psy.D.
Director and President, Cascade Center for Health and Flourishing
President-Elect, Society for the Psychology of Religion and Spirituality (APA Division 36)
Professor of Psychology, Wheaton College, 2012–2026
Coauthor, Attachment, Religion, and Spirituality (Cambridge University Press)
Coeditor, Handbook of Positive Psychology, Religion, and Spirituality (Springer Nature)
Submitted as an individual. The views expressed are my own and do not represent the positions
of the American Psychological Association, the Society for the Psychology of Religion and
Spirituality, or any other organization.

References
Balboni, T. A., VanderWeele, T. J., Doan-Soares, S. D., Long, K. N. G., Ferrell, B. R., Fitchett,
G., Koenig, H. G., Bain, P. A., Puchalski, C., Steinhauser, K. E., Sulmasy, D. P., & Koh,
H. K. (2022). Spirituality in serious illness and health. JAMA, 328(2), 184–197.
https://doi.org/10.1001/jama.2022.11086
Bockrath, M. F., Pargament, K. I., Wong, S., Harriott, V. A., Pomerleau, J. M., Homolka, S. J.,
Chaudhary, Z. B., & Exline, J. J. (2022). Religious and spiritual struggles and their links
to psychological adjustment: A meta-analysis of longitudinal studies. Psychology of
Religion and Spirituality, 14(3), 283–299. https://doi.org/10.1037/rel0000400
Center for Devices and Radiological Health. (2026, October 1). CDRH proposed guidances for
fiscal year 2027 (FY 2027). U.S. Food and Drug Administration.
https://www.fda.gov/medical-devices/guidance-documents-medical-devices-and-radiation
-emitting-products/cdrh-proposed-guidances-fiscal-year-2027-fy-2027
Davis, E. B. (2026). An integrative framework for religion, spirituality, health, and flourishing
(Version 1.0). Cascade Center for Health and Flourishing.
 https://docs.google.com/document/d/1XIp87qE7eeeRLXiAHY3Pj5ltG1IwNXT69HrHgp
dDEhs/edit
Davis, E. B., Chen, Z. J., Cowden, R. G., VanderWeele, T. J., Oberg, S., Bivins, G., Newcity, C.,
Song, E., & Koenig, H. G. (2026). An overview of systematic reviews and meta-analyses
on the association between religion/spirituality and health/well-being. In J. R. Webb
(Ed.), Handbook of spirituality, health, and well-being: A psychological perspective (pp.
380–395). Routledge. https://doi.org/10.4324/9781003440048-25
Davis, E. B., & Granqvist, P. (2026). Attachment, religion, and spirituality. Cambridge
University Press. https://doi.org/10.1017/9781009501019
Davis, E. B., Granqvist, P., & Sharp, C. A. (2021). Theistic relational spirituality: Development,
dynamics, health, and transformation. Psychology of Religion and Spirituality, 13(4),
401–415. https://doi.org/10.1037/rel0000219
Davis, E. B., Moriarty, G. L., & Mauch, J. C. (2013). God images and God concepts: Definitions,
development, and dynamics. Psychology of Religion and Spirituality, 5(1), 51–60.
https://doi.org/10.1037/a0029289
Exline, J. J., Pargament, K. I., Grubbs, J. B., & Yali, A. M. (2014). The Religious and Spiritual
Struggles Scale: Development and initial validation. Psychology of Religion and
Spirituality, 6(3), 208–222. https://doi.org/10.1037/a0036465
Hikmat, R., Syabana, M. N., Kosasih, A., Firmansyah, E. K., & Tumanggor, R. D. (2026).
Artificial intelligence at the boundary of spiritual experience and psychopathology: A
scoping review of evidence from Türkiye, the USA, the UK, Poland, South Korea, and
Israel. Journal of Religion and Health, 65(5), 4139–4170.
https://doi.org/10.1007/s10943-026-02777-1
Koh, H. K., Frederick, D. E., Balboni, T. A., O'Reilly, S. M., Kelly, J. F., Humphreys, K.,
Botticelli, M., Mathur, M. B., Psimopoulos, C. S., Long, K. N. G., & VanderWeele, T. J.
(2026). Spirituality and harmful or hazardous alcohol and other drug use: A
meta-analysis of longitudinal studies. JAMA Psychiatry, 83(4), 363–378.
https://doi.org/10.1001/jamapsychiatry.2025.4816
McBain, R. K., Cantor, J. H., Breslau, J., Diliberti, M., Zhang, L. A., Zhang, F., Burnett, A.,
Kofner, A., Rader, B., Pataranutaporn, P., Stein, B. D., Mehrotra, A., & Yu, H. (2026). AI
chatbot use and disclosure for mental health among US adolescents and young adults.
JAMA Pediatrics, 180(8), 884–890. https://doi.org/10.1001/jamapediatrics.2026.2015
McBain, R. K., Cantor, J. H., Zhang, L. A., Baker, O., Zhang, F., Burnett, A., Kofner, A.,
Breslau, J., Stein, B. D., Mehrotra, A., & Yu, H. (2025). Evaluation of alignment between
large language models and expert clinicians in suicide risk assessment. Psychiatric
Services, 76(11), 944–950. https://doi.org/10.1176/appi.ps.20250086
Salcone, S. G., Currier, J. M., McDermott, R. C., Davis, D. E., Raines, A. M., Lee, Y., Exline, J.
J., & Pargament, K. I. (2025). Development and evaluation of the Religious and Spiritual
Struggles Scale-5 (RSS-5). Journal of Psychopathology and Behavioral Assessment,
47(1), Article 8. https://doi.org/10.1007/s10862-024-10182-9
Tom, J. C., Ferguson, T. W., & Martinez, B. C. (2025). Religion and racial bias in artificial
intelligence large language models. Socius, 11.
https://doi.org/10.1177/23780231251377210
Wilt, J. A., Takahashi, J. T., Jeong, P., Exline, J. J., & Pargament, K. I. (2020). Open-ended and
closed-ended measures of religious/spiritual struggles: A mixed-methods study.
Religions, 11(10), Article 505. https://doi.org/10.3390/rel11100505
Winiger, F., Smith, C. E., Nourriz, S., Szilagyi, C., & Haußmann, A. (2026). Artificial
intelligence in spiritual care: Modified Delphi study. Journal of Medical Internet
Research, 28, Article e97094. https://doi.org/10.2196/97094