A Multiparty Collaboration to Engage Diverse Populations in Community-Centered Artificial Intelligence Research

Abstract

Artificial intelligence (AI)-enabled technology has the potential to expand access to high-quality health information and health care services. Learning how diverse users interact with technology enables improvements to the AI model and the user interface, maximizing its potential benefit for a greater number of people. This narrative describes how technology developers, academic researchers, and representatives from a community-based organization collaborated to conduct a community-centered project on emerging health technologies. Our project team comprised representatives from Stanford Medicine, Google, and Santa Clara Family Health Plan’s Blanca Alvarado Community Resource Center. We aimed to understand the usability and acceptability of an AI-driven dermatology tool among East San Jose, California, community members. Specifically, our objectives were as follows: to test a model for cross-sector research of AI-based health technology; to determine the utility of the tool in an ethnically and age-diverse population; to obtain in-depth user experience feedback from participants recruited during community events; to offer free skin health consultations; and to provide resources for receiving follow-up care. We describe a collaborative approach in which each party contributed expertise: knowledge of the community from the community health partner, clinical expertise from the academic research institution, and software and AI expertise from the technology company. Through an iterative process, we identified important community needs, including technological, language, and privacy support. Our approach allowed us to recruit and engage a diverse cohort of participants, over 70% of whom preferred a language other than English. We distill learnings from planning and executing this case study that may help other collaborators bridge the gap between academia, industry, and community in AI health care innovation.

Artificial Intelligence (AI)-enabled technology has the potential to expand health care information and access.¹ The impact of AI tools on individuals’ health decision-making is a key area of research requiring further investigation.2, 3, 4, 5, 6 Moreover, evaluating AI performance in diverse populations is essential to obtaining a robust, representative understanding of people’s perceptions.⁷ Learning how real-world users, including those with limited digital and health literacy, interact with technology enables improvements to the AI model and user interface, helping to maximize its potential benefit for a broader group of people.^2,8 Our team, comprising representatives from Stanford Medicine, Google, and Santa Clara Family Health Plan (SCFHP), aimed to understand the usability and acceptability of a research-only, AI-driven dermatology tool (the application) among East San Jose, California, community members. This tool allows users to upload 3 smartphone images of a skin concern at different angles, along with optional metadata and identifies matching conditions based on a deep learning model. Our objectives were as follows: to determine the utility of the tool in an ethnically and age-diverse population; to obtain in-depth user experience feedback from event attendees; to offer free skin health consultations; to provide resources for follow-up; and to test a model for cross-sector, community-engaged collaboration. This narrative describes how technology developers, academic researchers, and representatives from a community-based organization conducted a community-centered project on emerging health technologies.

Setting

The project was conducted over 4 community events hosted by SCFHP’s Blanca Alvarado Community Resource Center (CRC) in East Valley, San Jose, California, between October 2022 and May 2023. SCFHP serves over 320,000 members throughout Santa Clara County through Medi-Cal and Medicare health plans. The East San Jose CRC offers local, community-responsive resources to SCFHP members, including case management and multilingual health education. East Valley or East San Jose is one of the most diverse neighborhoods in Santa Clara County. Approximately 81% of households speak a language other than English.⁹ Residents in East Valley face substantial wealth disparities compared with other neighborhoods in Santa Clara County. Moreover, 35% of East Valley families and nearly half (46%) of children fall below 185% of the Federal Poverty Level. Comparatively, only 2% of families in the county’s highest-income neighborhood exceed this threshold.⁹ Inadequate health insurance and limited access to health care services, including dermatologists, contribute to health inequities throughout the county.

Application

The application tested is a research-only informational application that helps users understand their skin concerns (Figure 1). When uploading a case, users are asked to provide 3 images of their concern with their smartphone, at different angles. They are also prompted to provide optional metadata, including their demographic characteristics (age, sex at birth, skin reactivity to sun exposure), location and appearance of the condition, duration, and any symptoms associated with the concern. Images and metadata are analyzed by a deep learning model similar to that described previously.⁴ The model returns 3 to 11 matching conditions, in order of prediction confidence. The application displays these conditions as a horizontal carousel of tappable cards, with textbook images of each condition and descriptions. Selecting a card reveals detailed condition information, vetted by dermatologists.

Figure 1.

Application design and workflow.

Approach

During an introductory meeting, CRC leaders agreed that the proposed project, including free skin education and counseling, would be useful for health plan members and community residents. CRC staff invited the project team to set up a booth at a series of upcoming community events. Before the first event, representatives from Stanford and Google conducted 2 site visits and met weekly with CRC staff to discuss event logistics (eg, the physical space, anticipated attendees, and the theme or purpose of each event).

Subsequently, we mapped the participant workflow, defined roles and assignments, and trained personnel. All Stanford and Google staff and volunteers (approximately 50 across all events) were asked to complete cultural humility training. A total of 8 pre-event training sessions were conducted for research staff and volunteers. Stanford clinicians also had the opportunity to interact with the application, guided by Google researchers.

All community event attendees aged 18 years or older who spoke English, Spanish, Vietnamese, Mandarin, or Tagalog and had a skin concern were eligible to participate in the project. For all events, participants had the opportunity to use the application on their skin concerns and then see a Stanford-affiliated clinician for an educational screening and consultation. Complete findings are detailed in separately¹⁰; Figure 1 outlines the overall process. After each event, we gathered feedback from staff and volunteers via surveys, iterated on the workflow, edited the participant surveys and interviews, and conducted in-depth debrief sessions with Google researchers and Stanford clinicians.

Team

Our multidisciplinary team comprised SCFHP directors and staff, including the Director of Community Engagement; Google researchers, program managers, product managers, and volunteers; Stanford-affiliated board-certified clinicians, operational staff, and medical student volunteers. Supplemental Table (available online at https://www.mcpdigitalhealth.org/) details the specific roles and responsibilities of different team members. The protocol was reviewed and approved by Stanford (IRB-67027) and Advarra (e-Protocol-67027) institutional review boards.

Hurdles

We encountered several hurdles in collecting detailed feedback from the community. We characterize these as either hurdles related to accommodating the participant pool or hurdles related to the project environment.

First, we identified multilanguage support as a priority, which required translating medical information in the application, including information on more than 200 conditions, surveys, and informed consent materials into 4 languages. Additionally, our team recruited Google and Stanford volunteers who spoke English and either Spanish, Vietnamese, Mandarin, or Tagalog to guide participants through the event. Because we anticipated a high volume of Vietnamese speakers at our past event, we recruited 2 additional Vietnamese interpreters from one of the CRC’s resident advisory groups. Even so, across all events, peak hours resulted in an influx of interested participants where demand for interpreters temporarily exceeded availability. This increased our attrition rate among those who approached the consent table but could not wait for language support to complete the project.

Key aspects of the project centered on ensuring participant safety and privacy. The event location and demographic makeup of the attendees necessitated a heightened sensitivity to the health and social needs of the participants. To ensure participant safety, the process defaulted to all participants consulting with a clinician after using the application. Free clinician consultations were available to community members who elected not to participate in the project. To safeguard participants’ privacy, all consultations took place in a separate, adjoining area of the booth or indoor mock examination rooms. Outdoors, we used dividers and tent walls to provide a designated private space for the clinician-participant encounter. In addition, the project team supplied informational brochures about health care coverage and eligibility, as well as the phone numbers and locations of free or sliding-scale clinics so that participants could seek care irrespective of financial, privacy, or other concerns.

We encountered additional hurdles related to the physical environment. Our team was stationed in tents outside of the CRC, which posed challenges related to photograph quality, internet connectivity, weather, and ambient noise. During the first event, participants used an indoor room to consult with a clinician or take photographs of a sensitive area. Moving between outdoor and indoor spaces caused connectivity interruptions requiring the use of additional hotspots. Because we were collocated with other booths and event activities including cultural and musical performances, ambient noise proved challenging for interviews and compounded existing language barriers.

The outdoor setting also decreased photograph quality among users with limited smartphone familiarity. Incorporating tent roofs and sidewalls helped reduce glare and increase screen visibility. Allowing participants’ family members or volunteers to assist with photograph-taking increased the image quality. This process mimicked real user workflows where people may ask a family or friend to take a photograph of areas that they cannot readily reach, such as the back of the torso or limbs.

Findings

A primary goal was the quantitative assessment of user satisfaction and attitudes about their condition (eg, level of concern or intended next step) before and after using the application, alongside qualitative feedback about their concerns and experiences. The detailed results are discussed separately; but in brief, we observed several potential benefits of application use, including helping participants better align their level of concern and planned next steps with dermatologist assessments; and survey feedback from participating clinicians that the application results were mostly consistent with their assessments.¹⁰ The robustness of these results depended on having a sufficiently large and diverse participant pool across the event days. As such, we tracked and aimed for approximately 100 participants and qualitatively monitored insights to evaluate when we were seeing convergence in learnings and diminishing returns for additional insights (ie, saturation).^11,12

Across the events, 133 participants were consented. Of these, 19 chose to consult with a clinician but declined to use the application, for reasons such as the skin concern not being currently visible. The remaining 114 used the application and consulted with a clinician. Owing to the time commitment, 2 participants withdrew partway through the process. Both participants were in the first event. After streamlining the flow, we had no subsequent withdrawals. The remaining 112 participants completed all steps in the process and provided a wealth of qualitative and quantitative feedback. Although most feedback about the application is summarized in a separate report,¹⁰ we report in this study the distributions of participant characteristics (Table), which reflect a range of spoken languages, self-reported ethnicities, and health and technical literacy.

Table.

Participant Characteristics

Characteristic	Breakdown	Participants	%
Language	Vietnamese	34	30.9
	English	33	30.0
	Spanish	30	27.3
	Mandarin	9	8.2
	NA	4	3.6
	Total	110	100.0
Race/ethnicity (optional)	Asian	53	48.2
	Black or African American	1	0.9
	Hispanic, Latino, or Spanish origin	17	15.5
	Native Hawaiian or other Pacific Islander	2	1.8
	White	1	0.9
	NA	39	35.4
	Total	110	100.0
Comfort filling medical forms (optional)	Not at all confident	2	1.8
	Slightly confident	10	9.1
	Moderately confident	14	12.7
	Very confident	31	28.2
	Extremely confident	13	11.8
	NA	40	36.4
	Total	110	100.0
Comfort using mobile applications (optional)	Not at all confident	9	8.2
	Slightly confident	14	12.7
	Moderately confident	13	11.8
	Very confident	25	22.7
	Extremely confident	8	7.3
	NA	41	37.3
	Total	110	100.0

NA indicates "not applicable", indicating the participant skipped the question (for example when they want to accelerate study progress due to time commitments).

The diversity of participant engagement allowed us to gather qualitative insights regarding application usability that may not have been observed in a less diverse group. For instance, 1 clinician in a debrief interview noted: “The app may just need a little bit more tailoring to have more representative [in-app, textbook] photos depending on like the skin phototype. A lot of the photos so far were in patients with lighter-colored skin. So, I think to be broadly used across different populations, it would be helpful to show a spectrum, especially for rashes, which can look a little bit more on the purple side rather than pink if somebody were to have a darker colored skin.” The desire for textbook photographs in the application to better match their skin characteristics was echoed by many participants: a common theme of feedback highlighted mismatches between participants’ images and application example images in terms of skin tone, body part, age, and condition extent or severity. Similarly, other feedback such as differing usages of medical or anatomical terms based on language or cultural background, and familiarity with technology was only obtained by recruiting a sufficiently broad participant pool.

Lessons Learned

The first step in a cross-sector, community-centered collaboration is thoughtful partnership selection, which includes priority alignment and a shared understanding of common goals. Stakeholder analysis should be used early in the partnership development phase to analyze each entity’s unique contributions, perspectives, and limitations.¹³

For example, Figure 2 visualizes our 3-way partnership, where the roles and responsibilities of each primary stakeholder reflect their relative capabilities and interests. For example, the academic institution aimed to develop a model for community-engaged research of AI-based health care technology; the technology company sought to improve the design of health AI tools by soliciting and incorporating diverse input from the community; the community partner wanted to provide useful services to their residents and ensure their voices were heard in developing technology that they may eventually use. Executing this project and ensuring each partner’s needs were met required building trust. We achieved this via regular meetings and inclusive and collaborative approaches to all project activities—event planning, day-of data collection, postevent feedback, and process iteration. After the last event, relevant findings in the form of posters and translated brochures were shared with the community and distributed at the CRC.

Figure 2.

Venn diagram of stakeholder roles, responsibilities and interests.

Our piloting approach allowed for the gradual scaling of enrollment. In our project, we expected that users less familiar with technology would need more time with the app, bottlenecking throughput. We adopted various strategies to help reduce this possibility. We began with internal dry runs and a pilot where we presented mocks of the application to get feedback and facilitate researchers’ and volunteers’ familiarization with the workflow. At the first event, we were intentionally overresourced. For example, the clinician-to-participant ratio and volunteer-to-participant ratio were higher than the subsequent events. We also used longer surveys to capture many potential user sentiments. As the process evolved over the next 3 events, the staffing ratio was fine-tuned, and the survey was trimmed to improve efficiency. With these changes, the time-normalized numbers of participants per researcher improved approximately 3-fold from the first to the last event.

Conclusion

This narrative highlights how cross-sector, multistakeholder partnerships can facilitate community engagement by accommodating local needs and meeting community members where they regularly receive health information. We offer a model for community-based collaboration that can be replicated by others seeking to bridge the gap between academia, industry, and community in AI health care innovation.

Potential Competing Interests

Author Devon-Sand reports payments made to the institution by Google LLC. Dr Sayres owns Alphabet stock and is a Google employee. Dr Liu owns Alphabet stock and is a Google employee. Author Strachan owns Alphabet stock and is a Google employee. Author Smith reports payments made to the institution by Google LLC. Dr Ko reports payments made to the institution by Google LLC. Dr Lin reports payments made to the institution by Google LLC.