LAUSR

The accurate identification of a patient’s movement intention using electromyography (EMG) signals plays a pivotal role in wrist rehabilitation training. However, existing EMG signal classification methods often fail to adequately capture essential features, including critical channel information, temporal variations, and multiscale feature representations. To overcome these limitations, this article proposes a novel multibranch residual network (ResNet) architecture that combines bidirectional long short-term memory (BiLSTM) with two attention mechanisms, specifically the convolutional block attention module (CBAM) and squeeze-and-excitation (SE), for enhanced EMG signal classification. Experiments on the public Ninapro DB4 dataset show that the proposed multibranch residual attention network with BiLSTM achieves classification metrics of 93.66% mean accuracy, 92.01% F1-score, and 90.32% recall across six fundamental wrist rehabilitation movements. Notably, the model exhibits significant performance improvements of 8.24% for wrist supination and 6.48% for wrist pronation movements, compared to conventional convolutional neural network (CNN) baselines, demonstrating enhanced capability in classifying complex muscle activation patterns. This work demonstrates that integrating multibranch attention and residual BiLSTM architectures is highly effective for decoding complex EMG patterns.