Gait Phase Classification from sEMG in Multiple Locomotion Mode Using Deep Learning

Gait phase classification holds significant importance in various domains, including exoskeleton and prosthetics control, assistive technology, disease monitoring, etc. This classification is achieved by utilizing wearable sensors such as inertial measurement units (IMUs), electromyography (EMG), or computer vision techniques. Nevertheless, EMG possesses greater use due to its ability to relay more crucial information related to neuromuscular activity and excitation of the muscles. Furthermore, the inclusion of gait phase classification for various locomotion modes, in addition to level-ground walking, can enhance the algorithm’s practicality in real-world scenarios. This work introduces a novel deep learning model called Bi-GRU-Transformer-Net, incorporating Bi-GRU and transformer-based encoders, together with a gating mechanism. The utilization of a gating module to integrate two encoders has been found to improve the accuracy of classification when compared to a single encoder across various locomotion modes. Our proposed model demonstrates a notable improvement in accuracy for the downstairs gait phase, with an improvement of up to 5.10%. Additionally, for upstairs trials, our model achieves an increase in accuracy of up to 11.59% when compared to existing deep learning models. While the Bi-LSTM-Net has superior performance in walking trials compared to our proposed model, our method exhibits better mean accuracy across three distinct locomotor modes leading to a performance improvement of up to 11.68% overall.