Machine learning-based adaptive personalization in virtual reality stroke rehabilitation: a systematic review

Stroke is the second leading cause of disability worldwide. Recently, VR has been seen as a new therapeutic tool. Many existing VR systems rely on rule-based difficulty adjustment that may not adequately capture the non-linear dynamics of stroke recovery. Machine learning algorithms can provide adaptive personalization in virtual reality stroke rehabilitation.

The aim is to systematically review available evidence on ML-based adaptive personalization mechanisms in VR stroke rehabilitation, focusing on algorithms and adaptation strategies, clinical outcomes and implementation considerations.

In accordance with the PRISMA 2020 statement six databases were searched (January 2015-December 2025). Included studies conducted with the use of ML algorithms for VR. The ML algorithms, adaptation mechanisms, therapeutic parameters, clinical outcomes, and implementation factors are covered in data extraction. Risk of bias was assessed with validated tools; meta-analysis was performed when appropriate.

In total, twenty-five studies were included for analysis. Reinforcement learning (n = 8) and deep learning (n = 7) were prominent algorithms. Around 80% of studies implemented real-time closed-loop adaptation. The results of the meta-analysis show that the upper limb motor function (FMA-UE) improved by 7.47 points compared with the control group (MD: 7.47 points; 95% CI: 5.38-9.57; p < 0.001). This improvement exceeds the minimal clinically important difference. The usability scores were positive (SUS=72.5), with few adverse events reported.

ML-based adaptive VR rehabilitation demonstrates clinical efficacy and safety for stroke recovery, though implementation barriers in resource-limited settings require attention.

https://www.crd.york.ac.uk/PROSPERO/view/CRD420261298450, PROSPERO CRD420261298450.