Systematic review of deep learning applications in physical rehabilitation for human movement analysis in compliance with PRISMA standards

In recent years, deep learning methods have had a profound impact on the analysis of human movement, particularly in the field of physical rehabilitation. These methods offer advanced capabilities for recognizing, classifying, and evaluating therapeutic exercises, paving the way for new approaches in remote rehabilitation and telerehabilitation. This work is part of a larger systematic review focusing on the analysis of human movement in sport and rehabilitation, considered as complementary fields sharing common methodological, biomechanical and technological issues. In accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocol, 1564 articles were initially identified from six major scientific databases. After applying the inclusion criteria, 35 studies were selected for detailed analysis in rehabilitation. An additional 9 studies addressing combined sport and rehabilitation contexts were also considered to provide complementary insights. The results reveal a preference for hybrid neural network architectures combining convolutional neural networks (CNN)(40%) and recurrent neural networks (Long Short-Term Memory and Bidirectional Long Short-Term Memory)(31%), there jointly the spatial and temporal dimensions of human movement, but the most studies focus on lower-limb rehabilitation(37%), particularly related to gait recovery and post-stroke motor disorders. Frequently, researchers used data-acquisition systems rely on computer-vision technologies (standard color cameras and color-depth cameras)(45%), followed by Inertial Measurement Units sensors(IMU) and biological sensing devices (electromyography and electroencephalography). This points to growing interest in multimodal sensor fusion. The integration of work from sport makes it possible to broaden the analysis to complex, dynamic and highly variable movements, thus strengthening the relevance and scope of the approaches developed for clinical rehabilitation. Although the analyzed studies demonstrate the strong potential of deep learning for automatic movement assessment and therapeutic exercise classification, several methodological limitations persist. A comprehensive review of rehabilitation therefore requires consideration of work from sport, whose methodological and technological contributions help to better address the overall challenges of human movement analysis. These limitations include small sample sizes, the use of proprietary datasets, and experiments conducted mainly in controlled laboratory settings rather than real-world rehabilitation environments. Such conditions restrict the generalizability of the findings and hinder their direct deployment in practical clinical scenarios. Consequently, while deep learning approaches show promising opportunities for more intelligent, personalized, and connected rehabilitation, their translation from technical performance to real-life therapeutic impact remains limited.