Discussion
This study engaged patients and caregivers, providers, decision-makers, digital health and research participants in a nominal group technique process to identify priority areas for AI and PC. Small group discussions identified barriers, implementation issues, and necessary resources for progress. The final list of nine priority areas included physician, patient and system-level supports; and foundational areas that are necessary for the success of AI in other priority areas.
The consultation session revealed foundations that need to be improved to support progress in AI development and application, such as communication between PC and AI stakeholders, intersectoral collaboration, data standards and interoperability, and legal issues. In addition to longer-term foundational work, participants encouraged the initiation of AI projects that align with priorities and offered suggestions about conducting these projects in settings where the data and culture are in place to support the continuum of AI development through to careful evaluation and implementation. Core values to maintain throughout these processes included collaboration between diverse participants to maintain suitability of a candidate tool for practice settings, and to attend to equity concerns and patient-centredness.
Most of the ranked priorities from the consultation session include areas wherein AI may support PC by performing relevant functions or tasks, such as using AI to predict patients at high risk of poor health where early intervention is useful. It is noteworthy that despite this consultation session happening during the COVID-19 pandemic, with instructions that the pandemic should be considered in responses, only two COVID-19 specific priorities were identified in the first small group discussion and none remained in the final ranked list. It is also interesting that the environmental scan found 110 AI-driven tools that may be relevant to PC, yet among our participants selected for their engagement in AI and PC, few examples of AI-driven tools implemented in Ontario PC settings were discussed. Appraising specific tools to see if they are available in Ontario and whether they are suitable to meet priority areas as delineated in this study could be an avenue for future research.
The two ‘foundation-related’ priorities (4 and 5) will support progress of AI for all areas of PC. Many AI applications for PC will rely on data generated by PC, which was the topic of priority 5 and in small group discussions about data access, quality and consent. Initiatives supporting use of Ontario PC data for research include Institute for Clinical Evaluative Sciences (ICES)21 and the emerging PC Ontario Practice-Based Learning and Research Network,22 as well as national databases such as those housed by the Canadian PC Sentinel Surveillance Network23 and the Canadian Institute for Health Information.24 Paprica et al explored views of Ontario general public regarding the use of linked administrative health data held by ICES, finding that people are generally in favour of using these data for public benefit, assuming privacy and security; however, positive attitudes towards data use are more mixed or negative when there is private sector involvement.25 These findings are congruent with our findings regarding the importance of data ownership and oversight for AI-driven tool development, which has substantial private sector involvement.
Data sharing and communication also was a priority in a study by Shaw et al that used the nominal group technique to elicit priorities for virtual care-related policy planning for Ontario PC.26 Similar to AI, at the time of their consultation (before the COVID-19 pandemic) virtual care was considered a novel technology with potential benefit. Their recommendations included the need for a patient-centred focus and system- or social-level changes.26 One suggestion for engaging patients was in outcome measure selection,26 which is relevant for AI applications as well which ties together the themes emerging from our study around patient centredness and the need for rigorous evaluation. Another relevant suggestion regarding virtual care implementation was the use of a sociotechnical model of care,26 which is also cited as important for AI to contribute to a learning health system framework whereby data are used in feedback loops to improve care.12 The idea of a sociotechnical model aligns with themes from our small group discussions about the importance of system culture and communication between stakeholders. Previous work towards improving multidisciplinary collaborations includes guidelines produced by Saleh et al for AI-clinical collaborations,27 codesign of a documentation assistant for PC consultations by Kocaballi et al in Australia,28 and a ‘code to bedside’ framework for quality improvement methods by Smith et al in the USA.29
Given the breadth and complexity of PC, there are many perceived opportunities for AI to be useful-focused efforts on tangible projects are needed for the field to mature. Our consultation session identified priority PC challenges, which AI is well suited to support given the current or near-term capabilities of AI and the Ontario PC context. Although the consultation sessions focused on Ontario, the environmental scans suggest there may be similarities in terms of AI-driven tools and PC needs in other jurisdictions. Other sectors may use our list of priorities as a starting point to refine based on their context. Together, the findings from our study can be used to guide future research and evaluation efforts, as well as to guide organisations and decision-makers in guiding the allocation of resources towards advancing AI for PC.
Strengths and limitations
Strengths of the study, which is the first to identify priorities for AI and PC in Ontario, included bringing together multiple types of stakeholders and designing small group sessions to facilitate equal participation. The study prework to both appraise the present context based on environmental scans and to provide primer documents provided foundational knowledge that supported strong engagement by all participants, regardless of prior AI knowledge. Limitations include representation mainly from academic communities as opposed to industry and private practice. The environmental scans were not limited in this way; therefore, help to balance these findings.