Real-time Compensatory Movement Detection in Upper Limb Rehabilitation Using Deep Learning Methods

This paper presents a real-time approach for detecting compensatory movements in upper limb rehabilitation for stroke patients using deep learning algorithms. The study applied Recurrent Neural Networks (RNN), Gated Recurrent Unit (GRU), Long-Short-Term-Memory (LSTM), and Transformer to analyze Microsoft Kinect data from the Toronto Rehab Stroke Pose dataset. The models were trained with focal loss to address imbalanced data distribution. The simulation results showed that the proposed deep learning algorithms are effective in detecting compensatory movements. The GRU-based models provide the fastest results and the transformer models exhibit the best accuracy and fastest inference time on the employed CPU .