The AI4DPP project is being implemented in Peel region, Ontario [22], which has a high prevalence of T2D, particularly among South Asian, Asian, Arab, African, and Hispanic populations [23]. In 2015, the incidence of diabetes in Peel was 1192/100,000, reflecting an increase of 182% within 2 decades [24].

Our team works with health system practitioners, decision-makers, community organizations, and people with lived and living experience of T2D in Peel region. This partnered approach aims to ensure that the 2 machine learning models can be implemented in a way that is responsive to local needs [25] when applied at a population level. The AI4DPP Solution Network hosts a biannual full-day meeting for patient and public stakeholders in Peel region to share project updates. The May 2024 in-person edition of the network meeting served as the setting for the participatory workshop we conducted for this study.

This study was reviewed by Trillium Health Partners Research Ethics Board and was determined to not qualify as human participant research (reference #247). The AI4DPP study also underwent ethics review and approval by the Health Sciences Research Ethics Board at the University of Toronto, Dalla Lana School of Public Health (protocol number: 46174). When invitations to the network meeting were sent out, we notified participants that anonymized handwritten notes for research purposes would be taken during the workshop and that no participant would be identified or quoted by name in the analyses or outputs. Invitees were requested to contact the research team if they had questions or would like to opt out ahead of the meeting. Verbal group consent and confidentiality of discussions were established at the beginning of the meeting. Community representatives and people with lived and living experience were each given a CAD $150 (US $107.60) gift card as honorarium after the meeting. We also reimbursed their parking and transportation costs, where applicable.

We conducted a literature scan [26] to identify strategies for engaging community organizations and people with lived and living experience on AI for health care. This was followed by a 90-minute participatory workshop at the AI4DPP Solution Network meeting.

First, we present findings from the preworkshop activities—literature scan and the conceptual framework for community engagement on AI developed by the research team. Next, workshop findings are presented under 3 subheadings: participants’ perspectives on community engagement on AI, key principles for community engagement on AI, and practical considerations for operationalizing the key principles. A total of 30 people attended the workshop: 6 people with lived and living experience (patients and caregivers), 2 community organization representatives, 3 senior leaders from Peel Public Health, 4 health care providers, 12 AI4DPP team members (including 4 trainees), and 3 representatives from the project funder, the Canadian Institute for Advanced Research.

Synthesis across the 6 papers resulted in a compilation of 10 community engagement principles for AI: (1) trust, (2) power-sharing, (3) empowerment, (4) value alignment, (5) equity, (6) codesign, (7) transparency, (8) education, (9) early engagement, and (10) accountability. Practices identified in the literature for engaging patients and communities in AI and technology interventions included a need for early and meaningful engagement of patients and communities, ensuring diversity of participants [34-36] and adopting transparent and bidirectional communication where the community feels empowered and able to participate in decision-making [36]. Patient education was emphasized as being crucial to meaningful engagement to inform patient-centered design and build trust [34,36]. The papers drew from diverse approaches, particularly community-based participatory research and citizen science.

The International Association for Public Participation spectrum of engagement—a framework that defines levels of public engagement starting from being informed to being consulted, being involved, collaborating, and finally, empowering—the highest level of engagement [36,37], was explicitly adopted or referenced in most of the 6 papers. At the first level of the International Association for Public Participation spectrum—inform—patients and community organizations are informed about an AI-related health project including the benefits, risks, and potential impact. At the second level, consult, patients’ and community organizations’ perspectives on the feasibility and utility of the project are solicited. At involve, the third stage, project teams should be working with patients and community organizations to consider and address any concerns related to the AI intervention, and at the collaborate stage, patients and community organizations should be active decision-makers regarding the project. When at the empower level, patients and community organizations should ideally take the lead in setting priorities and making decisions on the project. Collectively, these insights from the literature captured in Table 1 informed the discussion prompts for the workshop.

In this study, we engaged community members, people with lived and living experience, and other health sector stakeholders to identify 6 priority principles and accompanying practices for engaging communities in population-wide deployment of AI tools, using the AI4DPP project as an example. The approach of workshop activities aligned primarily with the Involve and Collaborate levels of the IAP2 framework. Participants’ insights were reflected directly in the prioritization and operationalization of principles (Involve), and they actively influenced decision-making on which principles would be prioritized and how they would be operationalized (Collaborate). Since population level use of AI/machine learning models in health systems is geared primarily toward health care decision-makers and health care providers, our study’s findings are relevant for incorporating the missing perspective in the AI development cycle. That is, the perspectives of patients, caregivers, and other end beneficiaries. Our findings align with and extend prior research on community engagement in AI and digital health, particularly around themes of trust, diversity, transparency, and codesign. In this section, we compare our results with existing literature and reflect on their implications for health care AI implementation.

Only recently has the AI community started shifting from model-centric AI (where the accuracy of the models is the most important parameter) to data-centric AI (where the understanding of how data are collected, collated, questioned, analyzed, and interpreted is paramount) [42]. The social power that patients and their caregivers indirectly exert in shaping the sociotechnical milieu of AI adoption in health care has not been sufficiently prioritized in addressing population-level adoption of AI tools [43]. Yet, evidence shows that social acceptability of AI in health care can improve patient engagement in decision-making and enhance patients’ compliance with AI-informed care plans [44,45]. Trust was identified as the most important principle of community engagement. This is perhaps not surprising since AI has often been referred to as a “Blackbox,” increasing skepticism about its negative and unintended impact. In turn, lack of patient and public trust hinders AI implementation in health [46]. Trust is also central to how societies perceive health research and interventions, and trustworthiness is a key tenet in the ethical and responsible use of AI. There is a role, therefore, for AI researchers to open this “Blackbox” by using methods that enhance the explainability of machine learning models [47]. However, explainability from the perspective of patients and communities may differ significantly from those of technical audiences. While a full discussion of community-centered explainability is beyond the scope of this paper, we raise this point to emphasize the need for ongoing dialog on how transparency and trust in AI can be codeveloped with affected communities.

Open dialog has been reported as critical to meaningful community engagement in AI [48], and regular communication and feedback throughout a project can facilitate understanding. Patients and community partners expect honest communication about the limitations and results—positive and negative—of AI in accessible ways. They want AI researchers to commit to regularly monitor and share the results of AI implementation in health care, and to act on feedback received from community members [8]. Transparent governance structures, for example, through community advisory boards, have also been recommended to ensure accountability and build trust when involving communities on AI for health [8,49]. Because building trust takes time, necessitating ongoing engagement [48,50], it is recommended that patients and community members are engaged early, ideally at the problem identification stage and throughout the AI life cycle [51,52]. Furthermore, early engagement ensures that implementers are asking the right questions and that critical contextual factors are considered sooner than later [53]. Other researchers have also demonstrated that early and continuous involvement of community members helps to avoid tokenism [38]. Ultimately, proper alignment of researcher and community values is necessary for successful collaboration between AI researchers, developers, and communities [54].

Addressing diversity has been reported as a neglected area in patient and public involvement in health care [55,56]. Therefore, it is not surprising that inclusivity and diversity of participants in community engagement was a recurrent point raised in our study, similar to the work of others [49,51]. This is strongly related to equity, the second-ranked principle. Workshop participants stressed the need to include diverse participants across demographics, especially minorities, not only for community engagement activities but also in governance. Similarly, participants in our sessions noted that inclusivity will require culturally responsive approaches to engagement and an understanding of diverse backgrounds which may be rooted in historical and structural harm. Bringing all voices to the table can help ensure fairness and could potentially limit the perpetuation of systemic biases inherent in data and AI algorithms [47,57]. Although education was not prioritized as a standalone principle, workshop participants noted that meaningful community engagement would require a digitally informed and literate citizenry, making it nonetheless foundational for community-engaged AI interventions. The importance of AI patient education has been highlighted as crucial to meaningful community engagement and codesign [36,50].

Although the starting point of our study was a rapid literature scan, which could have resulted in missing some relevant papers, the use of a participatory approach involving diverse actors including people with lived and living experience helped strengthen our findings. An additional strength of our approach is that stakeholders have now been involved early—during the pre-implementation phase of the project, which sets the stage for their contributions in other phases. Third, using a modified NGT encouraged participation and openness, elevating the perspectives of community members. Additionally, although power dynamics are often a concern in mixed-stakeholder engagement, we did not observe this. This may be due to the experience of participants with lived and living experience in patient engagement and our teams’ efforts to foster an inclusive setting before and during the workshop.

Our process has some limitations worth mentioning. While previous studies have established the significant burden of T2D in Canada [11] and Peel [24], patients and community partners were not involved in problem identification (ie, define the stage of the AI life cycle) and machine learning model development (ie, design and test stage of the AI life cycle) of the AI4DPP project. As such, community engagement was not fully embedded as conceptualized in Figure 1. Also, workshop participants represent only a small subset of the broader community—only 6 of the 30 workshop attendees were people with lived and living experience. While this limits the representativeness of perspectives reflected in this formative study, the workshop was intended as an initial step in a continual process of community engagement. Finally, using a predefined conceptual framework (Figure 1) to guide workshop discussions may have constrained the emergence of fresh ideas. Despite these limitations, our community engagement approach is a valuable contribution to the field since most existing community engagement frameworks in AI for health focus solely on the design phase of the AI life cycle [9].

Building on this work, our next step is to establish a community advisory group composed of people with lived and living experience and community organization representatives. This group will support the implementation and evaluation phases of the AI4DPP project, providing feedback on how well the principles and practices identified in this study are being enacted. This will inform the development of community-informed guidelines to support AI deployment in public health. The identified principles and practices are early inputs for a community engagement framework to deploy AI prediction tools for population health, which will be refined and tested in subsequent phases of the AI4DPP project. Future research could further refine participatory strategies and embed longitudinal engagement mechanisms throughout the AI life cycle.

This study described a formative qualitative process used to identify principles and practices for meaningful community engagement in the deployment of AI tools for diabetes prevention and management. Using the context of Peel region, these insights reflect the priorities and expectations of community members, patients, and public health stakeholders. While the outputs are specific to our study context, they may be useful to others designing or adapting engagement strategies for AI implementation in public health settings. Our findings offer practical considerations for researchers, AI developers, clinicians, and community partners seeking to embed equity-oriented engagement practices into the design and deployment of AI-driven interventions.