image: Large-scale cohort evidence may be derived from national surveys, longitudinal cohorts, biomarkers and omics, wearable and linked health data and multidimensional health outcomes. AI can support the organisation of this evidence through time-resolved behavioural representation, multimodal data integration, structured use of unstructured data and knowledge extraction for decision support. These processes may inform context-sensitive physical activity guidance by integrating age, sleep patterns, chronic disease status, functional capacity, behavioural phenotypes and subgroup-specific needs. Translation into practice requires data harmonisation, external validation, interpretability, causal reasoning, feasibility and fairness.
Credit: Yanwei You. Reproduced from You Y. Med Research. Published online November 10, 2026. doi:10.1002/mdr2.70099. © 2026 The Author(s). Creative Commons Attribution License (CC BY).
Artificial intelligence could help researchers turn evidence from large population studies into physical activity guidance that reflects a person’s daily routines, health and practical constraints. In a correspondence published online in Med Research on 10 November 2026, Yanwei You of the National University of Singapore and Tsinghua University proposes a framework connecting detailed behavioural data, evidence organisation and professional decision support. The article presents a conceptual framework and research agenda; it does not report a clinically validated prescription system.
Population studies have strengthened the evidence linking physical activity, sedentary behaviour and sleep with health. Their findings often describe average associations, however, leaving a practical question: which changes are most relevant for a particular person with a particular health profile and daily routine?
The correspondence uses a hypothetical example of two older adults who each accumulate 30 minutes of moderate-to-vigorous activity and about nine hours of sedentary time per day. One walks regularly in the morning and keeps stable sleep times; the other has fragmented late-evening activity, long uninterrupted sitting periods and irregular sleep. Daily totals may conceal these differences. The example illustrates information that detailed behavioural analysis could reveal, without establishing that one pattern causes better health.
AI could help represent movement and sleep across the full day, combine wearable signals with health information, and extract useful information from clinical text or other unstructured records. Generative AI could also help organise findings from research papers and guidelines and draft possible recommendations. These outputs would need clear links to their underlying evidence and review by appropriately trained professionals.
The proposed roadmap has four connected stages. Data alignment makes measurements and outcome definitions more comparable across studies. Model development uses richer behavioural information while testing performance in independent populations. Decision-support design brings together health status, functional capacity and everyday constraints to identify feasible options. Translation into practice then requires professional review, implementation and follow-up to assess whether guidance is useful and should be adjusted.
For an older adult with low daytime activity, prolonged sitting and irregular routines, possible options might include gradually increasing daytime walking or interrupting long sitting periods. In the proposed process, a clinician or other trained practitioner would assess safety and feasibility before implementation. Follow-up information on behaviour, symptoms, function and adherence would help determine whether to continue or modify the guidance.
The correspondence emphasises that better prediction alone cannot establish that a behaviour change will improve health. Candidate recommendations require causal evaluation and prospective or intervention testing. Systems also need to account for mobility limitations, work schedules and access to safe places for activity. The author identifies an expert-supervised prototype as a realistic goal over the next three to five years. Before clinical or public-health use, it would require external validation and calibration across independent populations and relevant subgroups, followed by prospective evaluation of safety, feasibility, acceptability, adherence and predefined behavioural or health outcomes compared with usual guidance.
Journal
Med Research
Method of Research
Commentary/editorial
Subject of Research
Not applicable
Article Title
Artificial Intelligence for Precision Physical Activity Guidance: Connecting Cohort Evidence With Practice
Article Publication Date
10-Nov-2026
COI Statement
The author declares no conflicts of interest.