NUST Institutions Library Catalogue

This research investigates the use of electromyography (EMG) signals for real-time control in rehabilitation applications. Utilizing the Myo armband, we captured EMG signals corresponding to 12 distinct hand and finger movements. We compared the performance of two machine learning classifiers, Long Short-Term Memory (LSTM) networks and Vanilla Neural Networks (VNN), in accurately classifying these movements. LSTM networks demonstrated superior performance, achieving higher accuracy and robustness in signal classification compared to VNN. To address adaptability and reduce training time for new users, we employed transfer learning techniques. Our research also incorporated transfer learning techniques to enhance model performance, leveraging both a broad dataset collected from multiple subjects and a focused dataset from a single individual over an extended period. Our results show that transfer learning significantly improves the adaptability of the system, allowing for quicker and more efficient integration of new subjects into the model. The study further includes statistical analysis to validate the performance improvements, with paired t-tests and ANOVA confirming the significance of our findings. This work highlights the potential of LSTM networks and transfer learning in enhancing the usability and effectiveness of EMG-based control systems for rehabilitation, paving the way for more responsive and adaptable prosthetic devices. The integration of advanced machine learning techniques into EMG signal processing presents a promising avenue for future research and clinical applications.