Deep learning model cascade in electromyography decoding for gesture classification

Surface Electromyography (sEMG)-based gesture recognition plays a crucial role in advanced human-computer interaction (HCI) scenarios, including intelligent prosthetics, rehabilitation robotics, industrial product sorting, and immersive virtual/augmented reality control. The recognition accuracies of a multitude of gross and fine gestures from sEMG signals tend to be affected by complex temporal dynamics and inter-channel variability. This paper proposes a hierarchical hybrid deep learning model that cascades Transformer, channel attention (CA), and bidirectional long short-term memory (BiLSTM) networks for effective gesture classification. In the proposed architecture, the preprocessed sEMG signals are first projected into a latent feature space and fed into a Transformer encoder to capture long-range temporal dependencies and global contextual information. A channel attention mechanism is then employed to adaptively recalibrate inter-channel feature representations, enhancing discriminative muscle activation patterns while suppressing redundant information. Subsequently, a BiLSTM module models bidirectional temporal dynamics to capture the initiation, evolution, and termination characteristics of gestures. An attention-based temporal pooling strategy further emphasizes informative time steps before the classification head. Our proposed model was evaluated on the NinaPro DB5 and BandMyo dataset comprising various gestures, including fine finger and wrist movements. Results show that our method performs better on NinaPro DB5 dataset. An overall accuracy of 81.96% and a weight F1-score of 0.82 are achieved, outperforming individual Transformer, CA, and BiLSTM. The highest accuracy of certain gesture could reach 96.55%. The comparison demonstrates that the synergistic integration of global dependency modeling, channel-wise feature refinement, and bidirectional temporal learning improves overall recognition performance, particularly for certain fine gestures. Moreover, the parameter number of our proposed model is low (0.20 million) in comparison with existing approaches. Our work may provide a potential solution for effective sEMG-based gesture recognition for real-world human-computer interaction applications.