Enhanced Bilateral EMG Classification Through Integration of Graph Convolution, Transformer, and BiGRU Networks

Bilateral electromyography (EMG) signal classification presents unique challenges due to inter-hand correlations and signal interference. This paper introduces Transformer GCN-BiGRU, a novel hybrid deep learning architecture designed to address these complexities using the B-EMGDT Dataset, which contains synchronized EMG recordings from both hands performing six fundamental movements. Our model integrates three key components: Graph Convolutional Networks (GCN) for spatial feature extraction, Transformers for capturing long-range dependencies, and Bidirectional Gated Recurrent Units (BiGRU) for temporal sequence modeling. Unlike traditional approaches, Transformer GCN-BiGRU treats EMG channels as a graph structure, allowing GCN to extract spatial dependencies between electrodes, while Transformer-BiGRU models temporal dynamics and inter-hand interactions. Additionally, an adaptive channel attention mechanism mitigates signal interference by dynamically weighing channel contributions. Experimental results demonstrate that Transformer GCN-BiGRU significantly outperforms existing methods, surpassing the previous 85% accuracy threshold on the B-EMGDT dataset. This work establishes a new benchmark for bilateral EMG signal classification, highlighting the advantages of hybrid architectures in biomedical signal analysis and their robustness against signal variability and noise.