BlogGuides
AI for physiotherapists: accurate, but not yet reliable

One click, and Google will show my articles more often in your results.
Artificial intelligence in physiotherapy is concentrated in two areas: camera-based movement assessment and home exercise apps with real-time feedback. But the broadest review of the first, presented in 2025, found that none of the ten studies it examined showed adequate reliability on the COSMIN scale, despite accuracy as high as 97%.
Movement assessment: high accuracy, unproven reliability
A research team, Tamer Burjawi, Oren Tirosh, Doa El-Ansary, Joshua Farragher, and Adrian Pranata, presented a systematic review at the 2025 World Physiotherapy Congress in Tokyo on machine learning for assessing functional movement in people with low back pain, ten studies included, protocol registered on PROSPERO. Seven of those ten studies report accuracy between 75% and 97%, a figure that sounds encouraging on its own. But the review also measures reliability against the COSMIN scale, the standard for judging whether a measurement tool is consistent over time and across different raters, and there the picture changes: none of the ten studies demonstrated adequate or good reliability. Only five of the ten had a low risk of bias on the Newcastle-Ottawa Scale. It's worth noting this is, for now, a congress ePoster, not yet a peer-reviewed journal article.
Knee valgus: AI agrees with a manual app, not yet with the gold standard
A narrower case comes from Germany. Lukas Janisch and Lilo-Marie Maack, shared first authors alongside Carlos J. Marques, published a 2026 study in BMC Musculoskeletal Disorders comparing a manual 2D app (PhysioMaster) against two AI pose-estimation models, OpenPose and BlazePose, for measuring dynamic knee valgus angle during a single-leg functional test. OpenPose showed excellent agreement with the manual measurements, better than BlazePose. The authors conclude that both the manual method and the AI-based ones give reliable, largely consistent estimates, but that the concurrent validity of all three methods against 3D motion capture, the gold standard for this kind of measurement, still needs to be verified: the study didn't collect motion-capture data, only monocular video.
Home exercise for knee osteoarthritis: function yes, pain no
In Shanghai, a team led by Dian Zhu published a 2025 study in JMIR mHealth and uHealth, a randomized trial of a computer-vision exercise app for people with knee osteoarthritis: 60 participants, 32 in the app group and 28 in the control group, six weeks with three sessions a week at community centers. The app group showed a significant improvement in physical function and self-efficacy compared with the control group, but no significant difference in pain or stiffness. The app was built by the same university research team, not an outside company, and no conflict of interest is declared. One limitation is worth flagging: the authors describe an "intention-to-treat" analysis, but the reported results cover only those who completed the study, 41 of 60 people, with higher dropout in the control group.
The shoulder app: better results, built by the people who studied it
A Spanish team led by Olga Villar-Alises published a 2026 study in Musculoskeletal Science and Practice, an assessor-blinded RCT on 46 adults with rotator cuff-related shoulder pain, comparing conventional physiotherapy against the same physiotherapy plus an AI-based home rehabilitation app that uses computer vision to give real-time feedback during exercises. At four weeks, the app group showed a pain reduction above the clinically relevant threshold and improved arm function, and after the intervention fewer participants in the app group needed further healthcare, 3 versus 10. That result should be read alongside a disclosure from the authors themselves: two of the six, including the first author, are employees or founders of Healthinn, the spinoff that built the app (reportedly called RehBody, according to trade press), while the other four declare no ties.
What it means for a physiotherapy practice or rehabilitation center
Four independent sources, all published or presented between 2025 and 2026, describe the same picture: AI movement assessment is accurate on narrow tasks but hasn't yet shown reliability at scale, and home exercise apps show promising results on still-small samples, in one case with a conflict of interest disclosed by the authors themselves. It's the same caution that applies to any regulated healthcare profession facing a new tool: I wrote about this for veterinarians, where the broadest independent validation didn't find a single one of six X-ray reading platforms ready for clinical use, another case where lab-measured accuracy isn't yet enough to guarantee everyday use.
If you run a physiotherapy practice or a rehabilitation center and want to know how to be found by people searching online for a physiotherapist in your area before evaluating a specialized AI tool, get in touch: let's talk it through.