Recurrent Neural Network for electromyographic gesture recognition in transhumeral amputees

Abstract Gesture recognition is a key aspect of myoelectric control of upper-limb prostheses and is rather complex to achieve for transhumeral amputees. The prosthesis control of upper arm movements must rely only on the arm muscles, which were not involved in these gestures before the amputation. For decades, machine learning has been used in research for upper-limb gesture recognition. However, reported classification accuracies for transhumeral amputees have not improved significantly since the 1990s. Latest developments in deep learning suggest it can outperform classical machine learning both in accuracy and processing time. This study aims to determine if a deep learning approach, specifically a Recurrent Neural Network (RNN), could better recognize the movement intents in transhumeral amputees. To do so, the classification accuracy and the processing time of the RNN were measured and compared to two state-of-the-art approaches that use a linear discriminant analysis (LDA) and a multilayer perceptron (MLP) respectively. All three approaches were used to classify the signals of five transhumeral amputees between 6 upper-limb gestures. For subjects 1, 3 and 5, the classification accuracy was significantly higher (p = 0.0002) for the RNN (79.7%) compared to the LDA (67,1%) and the MLP (74,1%). Additionally, the RNN had a much smaller processing time, under 7 ms, compared to 385 ms and 377 ms for the LDA and the MLP respectively. Consequently, the RNN is better suited for a real-time prosthesis control that occurs between 100–250 ms. Results suggest deep learning as a viable solution for gesture recognition in transhumeral amputees.