The role of Artificial Intelligence in interventions to reduce physical inactivity: A systematic review

Technical Summary: The Role of Artificial Intelligence in Interventions to Reduce Physical Inactivity

Problem Statement
Physical inactivity is a growing public health challenge, with age-standardised prevalence rising to 32.7% in high-income countries as of 2022. While digital behaviour change interventions offer a scalable solution, the specific efficacy of Artificial Intelligence (AI) in enhancing these interventions remains unclear. AI, defined broadly as algorithms that learn from data to perform automated tasks without explicit human programming, holds theoretical promise for personalising interventions and reducing engagement barriers. However, there is a lack of synthesized evidence regarding whether current AI-assisted public health interventions effectively reduce physical inactivity or increase physical activity levels.

Methodology
This systematic review adhered to the Cochrane Handbook and PRISMA 2020 guidelines, registered under PROSPERO (CRD42025642339). The authors conducted a comprehensive search across Medline, Embase, Web of Science, and Scopus for Randomised Controlled Trials (RCTs) published between January 2010 and January 2025. The search strategy included direct RCT searches, searches of recent systematic reviews (2023–2025) to capture missed trials, and forward/backward citation searching.

• Eligibility: Included studies were RCTs set in high-income countries (World Bank definition) within a public health context, targeting physical inactivity. Interventions had to utilize AI (excluding simple expert systems or computer-tailoring as the sole mechanism). Studies were restricted to English language publications.
• Data Synthesis: Due to significant heterogeneity in interventions, comparators, outcomes, and timepoints, meta-analysis was not feasible. The authors employed a narrative synthesis approach.
• Quality Assessment: Two reviewers independently assessed the risk of bias using the Cochrane Risk of Bias 2 (RoB2) tool, focusing on the most objective outcome at the longest timepoint.
• Stakeholder Engagement: The review incorporated feedback from two UK-based communities of practice (practitioners and strategic directors) regarding protocol refinement and synthesis findings.

Key Contributions and Results
The review identified and included 17 RCTs (comprising 48 reports and 3,282 randomised participants). The studies were categorized by AI technique:

1. Chatbots (5 trials): Evaluated apps using chatbots (rules-based or generative) to promote activity.
2. Adaptive Systems (12 trials): Evaluated recommender systems, reinforcement learning, machine learning, and case-based reasoning for selecting motivational messages or tailoring interventions.

Key Findings:

• Overall Effectiveness: There is no strong evidence that AI-assisted interventions generally reduce physical inactivity or increase overall physical activity (subjective or objective) in high-income countries.
• Chatbots: Five trials assessing chatbots showed little evidence of effectiveness. Results were imprecise due to small sample sizes (36–954 participants) and high risks of bias in four of the five trials.
• Adaptive Systems: Twelve trials using recommender systems, reinforcement learning, or machine learning showed little evidence of increasing general physical activity. However, there was some evidence suggesting these technologies may increase step counts specifically. Seven trials reported beneficial effects on step counts (increases ranging from ~650 to 2,220 daily steps), though results were often imprecise. One trial noted an increase in step count without a commensurate increase in time spent walking.
• Bias and Precision: The majority of included trials (12 of 17) were judged to have a high risk of bias, primarily due to issues with randomisation, missing outcome data (attrition), and outcome measurement. Consequently, most results were statistically imprecise.
• Equity: Only one trial reported subgroup analyses by PROGRESS-Plus characteristics (gender, age, education), finding no differential effectiveness. No trials reported on equity considerations for other technologies.
• Technology Gap: The review found no trials utilizing Large Language Models (LLMs), despite their recent proliferation. The only trial involving a generative model component (Oh, 2025) was small (n=36), short-term (1 week), and lacked a non-chatbot control, preventing an estimation of the chatbot's additive value.

Significance and Claims
The authors conclude that while AI-assisted interventions may increase step counts, there is currently no strong evidence of a beneficial effect on reducing physical inactivity generally. The paper emphasizes that the current evidence base is limited by:

1. Heterogeneity: Diverse interventions and outcome measures prevent quantitative synthesis.
2. Methodological Quality: High risks of bias and small sample sizes render many findings inconclusive.
3. Reporting Gaps: AI components are often poorly described in trial reports, hindering replication and understanding.

The review argues that future research must move beyond assuming AI is inherently beneficial. It calls for:

• Better description of AI interventions in public health, potentially adopting guidelines like CONSORT-AI or SPIRIT-AI.
• The embedding of equity considerations in design and analysis to ensure AI does not exacerbate health inequalities.
• The use of standardized, objective outcome measures beyond self-reporting.
• Further investigation into the potential for AI to cause unintended harms or maintain existing disparities, a risk identified in broader public health AI literature but not yet fully theorised in this specific domain.

The authors maintain a modest stance, stating that the lack of strong evidence does not prove AI is ineffective, but rather that it should not be assumed to be beneficial or risk-free without further rigorous, well-designed trials.