JMIR Publications

ABSTRACT

Background:

In primary care, artificial intelligence (AI) is increasingly being developed to support personalised physical activity recommendations for people living with long-term conditions by integrating clinical, behavioural and contextual information. As integration occurs, patients, carers and healthcare professionals must decide whether AI-generated recommendations should influence clinical care. Existing qualitative studies describe stakeholder perspectives but provide limited explanation of the reasoning processes underpinning these decisions.

Objective:

This study aimed to develop a stakeholder-informed conceptual framework describing how people living with long-term conditions, carers and healthcare professionals evaluate AI-generated physical activity recommendations in primary care.

Methods:

Framework analysis was used to explore how stakeholders evaluated AI-generated physical activity recommendations. We conducted a secondary analysis of transcripts from four focus groups, which were inductively coded to identify participants’ perceptions of AI-generated recommendations. Through iterative researcher–Large Language Model dialogue, codes were synthesised into higher-order concepts across stakeholder groups. These concepts were subsequently developed into an explanatory framework describing the reasoning processes underpinning stakeholders’ evaluation of AI-generated physical activity recommendations in primary care. Finally, concepts from the COM-B model were mapped onto the framework to examine how capability, opportunity and motivation may influence responses to, and engagement with, AI-generated physical activity recommendations.

Results:

Stakeholders evaluated AI-generated physical activity recommendations through four interacting constructs. Decisions depended on confidence in the recommendation (construct one = credibility), its alignment with individual health and circumstances (construct two = personal fit), appropriate future clinician involvement in sustaining recommended physical activity (construct three = supported use), with any behaviour change underpinned by the individual’s capacity, motivation and opportunity to act on recommendations (construct four = behavioural appraisal). Rather than relying on a single consideration, participants integrated these constructs before deciding whether AI-generated recommendations should influence care.

Conclusions:

The stakeholder-informed conceptual framework developed as part of this study demonstrates that successful implementation of AI-enabled physical activity recommendations depends on how recommendations are evaluated within clinical decision-making, rather than on algorithmic performance alone. By identifying the reasoning processes used by patients, carers and healthcare professionals, and considering personal behaviour change barriers and facilitators, the framework provides practical guidance for designing AI systems that support shared decision-making, facilitate clinical implementation and improve the acceptability of AI-generated recommendations in primary care. The study also demonstrates the potential of researcher-directed LLM-assisted secondary qualitative analysis to support transparent conceptual framework development while maintaining researcher oversight.