Cross-Day Myoelectric Gesture Recognition with Hybrid Multistream CNN-Bidirectional LSTM

Myoelectric hand gesture recognition is an effective and promising strategy for controlling prosthetic hands. Currently, most commercial systems rely on manual feature extraction, which is based on prior understanding of surface electromyography (sEMG) signals. While effective, this traditional method is limited by its dependence on handcrafted features. With recent advancements in artificial intelligence, deep learning techniques have been explored for myoelectric gesture recognition, offering the potential for more abstract and automated feature extraction. In this paper, we propose a new model called the Hybrid Multi-stream Convolutional Neural Network-Bidirectional Long Short-Term Memory (Multi-stream CNN-Bidir.LSTM) for myoelectric hand gesture recognition. Our model achieved state-of-the-art performance, with 89.03% accuracy in within-session testing and 40.66% accuracy in cross-day uncalibrated testing. Additionally, we demonstrate the importance of incorporating linear decision boundaries, proper feature selection, and deep model architecture design to optimize recognition performance.