Multiclass Gait Phase Classification From the Temporal Convolutional Network of Wireless Surface Electromyography Measurements

Recent advancements and developments in the field of rehabi- litation lead to the invention of myoelectric control interfaces for patients with disabilities. However, decoding the motion intent from the surface electromyography (sEMG) signals of hamstrings and quadriceps is challenging due to its complex mechanics associated with weight bearing joints and stochastic, nonstationary, and multicomponent behavior of signals. In this letter, a novel approach is proposed for multiclass gait phase classification during level walking using temporal convolutional network (TCN) of sEMG signals. For this purpose, sEMG and inertial measurement unit (IMU) data were recorded concurrently from 20 healthy participants during level walking on treadmill at a speed of 2.5 km/h. sEMG were collected from the muscles, namely, rectus femoris (RF), vastus lateralis (VL), biceps femoris (BF), and semitendinosus (SEM). The IMU measurements of knee flexion/extension data are utilized for labeling the four phases of gait cycle. The root mean square of sEMG epochs is used to design the TCN framework. The results show that the proposed framework has the ability to differentiate the four classes of gait with a maximum accuracy of 86.00% using the myoelectric activity from all the four muscles. The information from the muscle pairs SEM and VL, and RF and BF, yielded the correct detection rate of 83.00% and 84.00%, respectively. In addition, the accuracy is also improved by 6% with TCN when we compare accuracy obtained through convolutional neural network architecture. The findings suggest that the proposed approach is effective in decoding the motion intent of lower limb muscles, which may lead to the development of precise movement control of lower limb prosthesis.