sEMG-based deep learning framework for the automatic detection of knee abnormality

Knee abnormality is a vital issue that can be diagnosed utilizing a sEMG signal to detect muscle abnormalities. Manually analyzing EMG data is time-consuming and requires skilled doctors. Hence, this paper aims to provide an automated system for the diagnosis of knee abnormality. Here, sEMG signal acquired from four different lower limbs muscles of 22 volunteers with three activities: walking, sitting, and standing, of which 11 seem healthy, and the rest were diagnosed clinically with knee abnormality. Noises are present during the sEMG signal recording, so a multi-step classification approach is proposed here. At first, wavelet denoising was implemented to denoise the sEMG signals. Further, the overlapping windowing method with a window size of 256 ms along with an overlapping of 25% was utilized to minimize the computational complexity. Afterward, a hybrid convolutional neural network with long short-term memory (Conv-LSTM) model is used for screening abnormal subjects. In this hybrid approach, a convolutional neural network (CNN) is used for temporal learning, while long short-term memory (LSTM) is for sequence learning. The results exhibit that the proposed wavelet-based denoising followed by Conv-LSTM model is the most precise and convenient model used for the detection of knee abnormality using sEMG signals so far.