JMIR Publications

ABSTRACT

Background:

Physical inactivity represents a major global health concern, for which artificial intelligence (AI)-driven interventions offer promising potential through their ability to deliver personalized and adaptive support.

Objective:

The purpose of our paper is to systematically review the current state of AI-driven physical activity (PA) intervention research and to quantitatively synthesize evidence on their effectiveness.

Methods:

A systematic review and meta-analysis following the PRISMA guidelines was carried out. Pubmed, Scopus, SPORTDiscus and Web of Science were searched first in November 2025 for journal articles and later in May 2026 for conference articles. Additionally, studies from relevant previously published reviews were assessed. All studies that utilized AI methods in their intervention and reported sufficient pre- and post-intervention measurements for objective PA levels were included. Risk of bias was assessed with the RoB 2 framework and studies judged to have high risk of bias were excluded from the meta-analyses but included in narrative synthesis. In addition, we review the consideration of AI model fairness, explainability and AI ethics, utilized data types and participant feelings towards the interventions.

Results:

A total of 15 studies were included with ten studies (12 comparisons) included in the meta-analyses. We found a statistically significant positive overall effect in favor of AI-driven approaches (Hedges' g = 0.53, p < .001), with all included studies reporting higher effects for the AI interventions. More advanced AI methodologies were associated with larger intervention effects than simple rule-based approaches. Notably, none of the studies reported considering model fairness, explainability or AI ethics in their interventions. All studies utilized multimodal data with signal, tabular and text being the most common. Assessment of participant feelings was limited with substantial measurement variability but in general, moderate to high satisfaction was reported and majority felt the intervention was helpful for increasing PA levels.

Conclusions:

AI-driven interventions show potential for PA promotion and especially the use of recent advanced AI technologies has undiscovered potential. The research is still limited, lacking standardized protocols and adequate reporting of methodologies. Moreover, the lack of considerations for model fairness, explainability or AI ethics is a serious concern that risks the reliability, validity and non-discriminativeness of the field of research.