Background
A brain-computer interface (BCI) is a system that makes communication between the brain and an external device. The electroencephalogram (EEG) is the most favourable tool for extracting neural signals from the brain. Motor Imagery (MI) based BCI with EEG signals is an active BCI paradigm. The performance of MI-based BCI is easily affected by noise and redundant information. To decrease noisy and redundant information and increase the spatial resolution of the EEG signals, a multichannel EEG-based BCI system is used. However, high-dimensional data from multichannel BCI systems has serious impact on the classification performance. Therefore, for better classification performance of EEG-based BCI systems, channel selection methods are used. Generally, many traditional signal processing techniques such as correlation and power spectrum have been used for feature-based channel selection. However, the estimation of the power spectrum discards the phase relationship among frequency components.
Methods
To solve this problem, a bispectrum (BSPKTM) based channel selection technique is used to overcome the drawback of the power spectrum. It effectively provides the frequency domain information of MI related brain activities. Therefore, in this study, a bispectrum-based channel selection algorithm is proposed for the MI-based BCI system. The most relevant channels from bispectrum analysis are selected from bispectrum analysis using a set-based integer-coded fuzzy granular evolutionary algorithm (SIFE). The features are extracted from the selected channels using wavelet scattering transform (WST).
Results
Finally the experiments are tested on multiple classifiers and best performance is obtained using the SVM classifier. The best results are obtained as accuracy 96.78%, sensitivity 93.58%, specificity 94.64%, F1-score 0.9403, and kappa value 0.8821. The other classifiers also attained significant results using minimum number of EEG channels.
Conclusions
The proposed work explores the utility of channel reduction using BSPKTM-SIFE and WST based features extraction for the classification of left hand and right hand MI tasks EEG signals.