Perception, assessment, and coaching: a systematic review and taxonomy of computer vision-based physical rehabilitation techniques.

The review was not registered in advance in PROSPERO and no inter-rater agreement statistics were recorded, so the protocol cannot be independently checked. Studies with successful results are more likely to be published than those with failed validation, poor usability or low adherence, so the picture presented is probably more optimistic than the field actually is. The included work is very mixed (clinical studies, engineering benchmarks, prototypes, conference papers, reviews and preprints) and could not be pooled, so no single estimate of accuracy or benefit can be drawn from it. Most of the data behind these systems comes from healthy people or from people mimicking an impairment; datasets containing real patient movement are rare and usually small, typically fewer than 50 people. For most included records no clinical comparator was coded at all, so accuracy claims often rest on public engineering datasets rather than on clinical measurement. Patient-reported outcomes such as pain, confidence, satisfaction and quality of life were rarely reported, so it is unclear whether technically accurate systems produce benefits patients notice. Only English-language literature in major international databases was searched, and paediatric and cognitively impaired populations were barely represented.

The study was funded by the Interdisciplinary Research Program in Medicine and Engineering at The First Affiliated Hospital of University of South China, the Natural Science Foundation of Hunan Province, and the Scientific Research Project of Hunan Provincial Department of Education. The authors declared that the funding bodies had no role in the design of the study, the collection, analysis and interpretation of data, or the writing of the manuscript. No competing interests are stated in the text supplied.

Do not read the accuracy figures as proof that camera-based rehabilitation is ready for your clinic. The review's own appraisal graded no included publication as high-level evidence, and only 2 of the perception-domain records were validated against laboratory motion capture, so most systems have been checked against engineering benchmarks rather than clinical standards. The coaching and generative AI sections describe prototypes and proposals, not tested treatments.