Electromyographic (EMG) gesture recognition remains challenging because EMG signals are non-stationary, subject-dependent, and difficult to model over long temporal ranges while maintaining computational efficiency. This study proposes EMamba, a topology-aware EMG gesture recognition framework that combines a topology-preserving EMG heatmap representation with a hybrid ResNet-SSM architecture. The proposed EMG-to-image transformation preserves inter-channel spatial topology and generates stable feature representations from complementary time-domain descriptors. The ResNet-SSM backbone integrates convolutional feature extraction with linear-complexity state-space modeling to capture both local spatial patterns and long-range temporal dependencies. Extensive experiments on Ninapro DB1, DB3, DB4, DB5, and the MYO dataset demonstrate that EMamba achieves improved recognition accuracy, reduced inference latency, and strong cross-subject generalization for both intact and amputated subjects. These results indicate that topology-aware representation and state-space modeling provide an effective and efficient solution for EMG-based gesture recognition in wearable human-machine interaction systems. The overall diagram of EMG classification • A topology-preserving EMG heatmap representation is proposed. • A hybrid EMamba framework combining ResNet and state-space modeling. • State-space modeling captures long-range dependencies in EMG signals. • Experiments include amputee, intact, and wearable EMG datasets. • High accuracy with strong cross-subject generalization is achieved.
Zhang et al. (Sun,) studied this question.