Advanced TSGL-EEGNet for Motor Imagery EEG-Based Brain-Computer Interfaces

Abstract

Deep learning technology is rapidly spreading in recent years and has been extensive attempts in the field of Brain-Computer Interface (BCI). Though the accuracy of Motor Imagery (MI) BCI systems based on the deep learning have been greatly improved compared with some traditional algorithms, it is still a big problem to clearly interpret the deep learning models. To address the issues, this work first introduces a popular deep learning model EEGNet and compares it with the traditional algorithm Filter-Bank Common Spatial Pattern (FBCSP). After that, this work considers that the 1-D convolution of EEGNet can be explained by a special Discrete Wavelet Transform (DWT), and the depthwise convolution of EEGNet is similar to the Common Spatial Pattern (CSP) algorithm. Therefore, this work improves the EEGNet by using the algorithm Temporary Constrained Sparse Group Lasso (TCSGL) to enhance its performance. The proposed model TSGL-EEGNet is tested on the BCI Competition IV 2a and BCI Competition III IIIa datasets that both are 4-classes classification MI tasks. The testing results show that the proposed model has achieved 78.96% (0.7194) average classification accuracy (kappa) on the dataset BCI Competition IV 2a, which are greater than EEGNet, C2CM, MB3DCNN, SS-MEMDBF and FBCSP, especially on insensitive subjects. The proposed model has also achieved 85.30% (0.8040) average classification accuracy (kappa) on the dataset BCI Competition III IIIa, which are greater than the EEGNet, MFTFS et al. At last, this work uses average-validation and stacking to further enhance the effect of the model. The 4-classes classification average accuracy rates reach 81.34% and 88.89%, and the kappas reach 0.7511 and 0.8519 on dataset BCI Competition IV 2a and BCI Competition III IIIa, respectively. Additionally, this work also uses the Grad-CAM to visualize the frequency and spatial features that are learned by the neural network.

Brain-Computer Interface using temporal- spectral features and neural network classifier

“…It has not escaped our notice that as SVMs were previously shown to be superior with respect to feature classification (whereas deep learning networks were shown to be superior in BCI feature selection [15,23,24,27]) a combination of both methods might improve our algorithm further and allow it to generalize to tasks outside of motor imagination or control (i.e., non-verbal communication, language decoding, or parsing of thoughts).…”

“…Our method is not the first to consider the temporal structure of EEG signals during MI task classification. For example, previous work [15] has combined temporally constrained group LASSO with Convolutional Neural Network aiming at interpretating the mechanism of the EEGNet [28] model. Similarly, a framework for time frequency CSP smoothing was recently implemented to improve EEG decoding performance through ensemble learning [29].…”

“…The TSGSP algorithm uses Support Vector Machines (SVM) for the classification of new trials to their corresponding action class. The inclusion of temporal data was shown to improve the performance of CSP-based BCIs [10,11,15].…”

A multi-scale fusion CNN model based on adaptive transfer learning for multi-class MI-classification in BCI system

“…Since our input EEG data representation is 1D time-series signal, 1D-CNN has been employed which is relatively easier to train and offers minimal computational complexity compare to its 2D counterparts whilst achieving state-of-the-art performance [57]. The convolution layer consists of 1D-convolution filters of a specified kernel stride which perform convolution operations sliding along the time axis of EEG signal to obtain feature maps and time-frequency information of the time series data [58, 59]. In 1D-CNN, forward propagation can be expressed as follows: where is the input; represents the bias of k th feature information in l th layer; is defined as the connecting weight between i th feature of the l − 1 th layer and k th feature of the l th layer; represents output of the i th feature of l − 1 th layer; conv 1 D denotes convolution operation.…”

Investigating Feature Ranking Methods for Sub-Band and Relative Power Features in Motor Imagery Task Classification

“…Discriminative feature learning, sliding window common spatial patterns are some recent approaches used for MI task classification [ 21 , 22 ]. EEG-Net with Temporary Constrained Sparse Group Lasso also proves its efficiency in MI task classification [ 23 ]. Akbulut et al [ 24 ] proposed alpha and beta frequency power for MI task classification as the frequency represented as most responsible frequency of motor tasks.…”