Background: Stroke is the leading cause of long-term disability in the world and there is a compelling need for scalable and effective rehabilitation strategies. Combined with the help of artificial intelligence (AI) based pose estimation technologies, telerehabilitation is a promising paradigm for taking evidence-based motor rehabilitation beyond the hospital. But, clinical validity, accuracy benchmarks and patient compliance of such systems are poorly defined.

Objective: This study aimed to systematically review the accuracy of AI-based pose estimation systems, their clinical effectiveness on validated motor and functional outcomes in stroke survivors in the context of telerehabilitation, and the rate of patient adherence.

Methods: A systematic review with PRISMA 2020 guidelines. The six electronic databases (PubMed, Scopus, Web of Science, IEEE Xplore, Embase, and the Cochrane Library) were searched from January 2015 to April 2026, with the specific Boolean search strings defined for each of the six databases. Studies included in the eligible studies recruited adult stroke survivors and used AI pose estimation (OpenPose, MediaPipe, Kinect/Azure Kinect, or similar system) and had at least one quantitative outcome reported. The Cochrane RoB 2 tool, ROBINS-I and Newcastle–Ottawa Scale were used as appropriate for appraising risk of bias.

Results: Twenty-seven studies (1,482 participants) included. The accuracy of the pose estimation was between 78% and 97%, with Azure Kinect and the deep learning pipelines (OpenPose, MediaPipe) achieving a lower joint angle root mean square error (RMSE 2.1°-5.8°) compared to the first generation of Kinect v2 (6.3°-12.4°). The mean change in Fugl-Meyer Assessment (FMA) score across 21 of the 25 reporting studies was 8.3 points (95% CI 6.1–10.5) to improve significantly. Improvements in Berg Balance Scale and Barthel Index scores also were similar. Compliance averaged 78.4% (range 52%–96%), much better than standard therapy compliance rates for conventional outpatient therapy. For the most part, the Technology Usability scores (System Usability Scale) were above 70.

Conclusion: AI-based pose estimation systems show clinical meaningful accuracy and are able to facilitate measurable improvement in motor function, balance and activities of daily living in stroke survivors in telerehabilitation settings. Patients' adherence is comparable to face-to-face therapy. These findings need to be consolidated into clinical practice guidelines through standardisation of accuracy metrics, longitudinal designs and prospective randomised trials.