Machine Learning in Rehabilitation Training: Traditional and Deep Learning Approaches

This study aims to review and analyze the application of machine learning techniques in rehabilitation training, particularly in motor function recovery after stroke. The methodological framework is classified into two primary categories: traditional machine learning techniques and deep learning approaches. Traditional machine learning methods include: using multiple linear regression to predict gait parameters, applying Support Vector Machines (SVM) combined with surface electromyography (sEMG) signals to recognize upper limb movement intentions, and employing decision trees to assist in developing rehabilitation training plans. Deep learning methods involve using Deep Neural Networks (DNNs) to process electroencephalogram (EEG) signals, Time-series sEMG signals can be effectively captured using Long Short-Term Memory (LSTM) models, and CNN-LSTM hybrid models for skeletal motion sequence recognition. Some methods, such as SVM optimized with genetic algorithms and CNN-LSTM models incorporating attention mechanisms, have achieved recognition accuracy exceeding 90% without manual feature extraction. Although current models perform well in recognizing standard rehabilitation movements, challenges remain in terms of poor interpretability, weak generalization to rare movements, and limited deployability in low-resource environments. Future efforts should focus on interpretable network design, transfer learning strategies, and the application of model distillation algorithms to facilitate the real-world deployment of rehabilitation training systems.