Analysis and Comparison of 3D Human Pose Kinematics in Stroke Rehabilitation Using Deep Learning

Stroke often leads to reduced upper limb mobility, causing patients to adapt their movements to compensate for muscle weakness or joint stiffness. However, if these compensatory movements go unmonitored, they can negatively impact the rehabilitation process. This study aims to tackle that issue by analyzing 3D human pose data collected during stroke rehabilitation exercises using deep learning techniques. The research compares the movement patterns of stroke patients and healthy individuals as they perform seated upper-limb rehabilitation tasks with the aid of a robotic system. Key movement patterns such as forward leaning, trunk rotation, and shoulder elevation are identified and assessed using Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) models. Data were collected using a Microsoft Kinect sensor from 10 healthy participants and 9 stroke patients. Healthy participants also mimicked common compensatory behaviors typically observed in stroke patients. Advanced processing methods were applied to detect these movements. The performance of RNN and LSTM models was evaluated using metrics such as accuracy, precision, recall, and F1-score, with the RNN model achieving superior results 95% test accuracy along with strong precision and recall. These findings suggest that RNN-based systems hold promise for real-time feedback applications in stroke rehabilitation, potentially supporting better recovery outcomes.