Clinician-curated ChatGPT-4o– and DeepSeek-R1–assisted exercise programmes added to standard physiotherapy for subacromial pain syndrome: a triple-masked randomized controlled trial

Large language models may assist clinicians in developing exercise programmes and could offer an alternative approach to conventional physiotherapist-designed rehabilitation in musculoskeletal disorders such as subacromial pain syndrome. The aim of this triple-masked randomized controlled trial was to compare clinician-curated ChatGPT-4o– and DeepSeek-R1–assisted exercise programmes with a conventional physiotherapist-designed exercise programme in individuals with subacromial pain syndrome. This 3-arm randomized controlled trial was conducted in the outpatient physical medicine and rehabilitation clinic of a university-affiliated hospital. Adults aged 18 to 65 years with subacromial pain syndrome were randomized in a 1:1:1 ratio to receive conventional, clinician-curated ChatGPT-4o–assisted, or clinician-curated DeepSeek-R1–assisted exercise programming. All groups received identical physical therapy modalities and supervised exercise sessions 3 times per week for 6 weeks. The primary outcome was the Shoulder Pain and Disability Index. Secondary outcomes included pain during activity, Quick Disabilities of the Arm, Shoulder and Hand, and the 12-Item Short Form Health Survey. Between-group comparisons were based on change scores from baseline to week 6. Seventy-five participants were enrolled, and 69 completed the 6-week follow-up. Shoulder Pain and Disability Index scores improved in all groups, with no significant between-group difference over 6 weeks ( P = 0.428). No significant between-group differences were found for key secondary outcomes after Holm adjustment, including pain during activity ( P = 0.744), Quick Disabilities of the Arm, Shoulder and Hand ( P = 0.744), and physical component summary scores ( P = 0.374). Resting pain differed significantly between groups ( P = 0.005), favouring DeepSeek-R1 over conventional programming (mean difference, 2.74 [95% CI, 1.26 to 4.22]). No adverse events were reported. Clinician-curated ChatGPT-4o– and DeepSeek-R1–assisted exercise programmes did not produce superior 6-week improvements in shoulder pain-related disability or key functional outcomes compared with conventional programming in subacromial pain syndrome; these findings should not be interpreted as establishing equivalence. The observed resting-pain difference in favour of DeepSeek-R1 was exploratory and requires confirmation. Clinical trial registration : ClinicalTrials.gov (NCT07148687) First Posted 3 August 2025.