Automatic Classification of Movement Imagination using Electroencephalography (EEG) Signals with Application in Brain-Computer Interface

Background and aims: Following spinal cord damage from trauma or disease, there is a greatneed for a method that can substitute for voluntary control in order to regain self-mobility, environmentalcontrol, and computer access. One of the potential solutions is brain-computer interface(BCI) that records brain activity in the form of electroencephalography (EEG) signals andextracts the movement intention or imagination by artificial intelligence methods and finally providesthe possibility of communicating with the surrounding. High density EEG acquisition systemsinclude many recording electrodes. However, selecting the optimum EEG channels and theappropriate classifier is crucial for accurate detection of the intended movement in BCI systems.Method: In the current paper, we used the motor imagery EEG dataset ۱ provided by BCI competitionIV. It consisted of EEG signals recorded from seven subjects. The subjects had to imaginethe movements of the left (class ۱) and right (class ۲) hands during a cue-based BCI experiment.This dataset includes ۲۰۰ trials for each subject.In the preprocessing step, first, EEG signals are band-pass filtered in the frequency range of ۸-۳۰Hz using a ۳rd order Butterworth filter. Then, a low Laplacian filter is applied for source localization.To select the optimal channels containing the most useful information of the imaginedmovement, the wrapper-based method called sequential forward feature selection (SFFS) is used.After applying the regularized common spatio-spectral pattern (RCSSP) filter for better separabilityof the features, the variance of the channels and their logarithm are extracted as time domainfeatures. Finally, the support vector machine (SVM) and weighted extreme learning machine(WELM) were used for the classification of the performed motor imagery tasks.Results: The K-fold cross validation (K = ۱۵) was used to evaluate the performance of the proposedmethod. The quantitative criteria include accuracy, precision, and recall.The average (±standard deviation) classification accuracy, precision, and recall obtained for allsubjects by SVM classifier were approximately ۸۴.۲۰±۹.۲۸%, ۸۷.۱۷.۷۶±۹.۳۷%, and ۸۱.۵۰±۹%,respectively. On the other hand, using the WELM classifier, the average accuracy, precision,and recall were equal to ۸۳.۸۰±۸.۴۰%, ۸۴.۸۱±۸.۵۲%, and ۸۳.۱۷±۷.۲۷%, respectively. Therefore,based on the precision criterion, the SVM classifier applied to the EEG channels selected by theSFFS method has approximately three percent higher precision compared to WELM.Compared to one of the recent papers in the field of motor imagery classification (MohammadNorizadeh Cherloo, ۲۰۲۱), our purposed method has approximately two percent more accuracyvalue.Conclusion: In this paper, we classified two motor imagery tasks by two classifiers using EEGsignals. Although different studies have been conducted to classify motor imagery tasks, our resultsshow that the proposed method outperforms the previous studies.

Keywords:

Brain-Computer Interfaces , motor imagery , Common spatial patterns , Channel selection , support vector machines