Comparison of feature selection and classification methods for a brain–computer interface driven by non-motor imagery

Abstract

The aim of this study was to compare methods for feature extraction and classification of EEG signals for a brain-computer interface (BCI) driven by auditory and spatial navigation imagery. Features were extracted using autoregressive modeling and optimized discrete wavelet transform. The features were selected with exhaustive search, from the combination of features of two and three channels, and with a discriminative measure (r (2)). Moreover, Bayesian classifier and support vector machine (SVM) with Gaussian kernel were compared. The results showed that the two classifiers provided similar classification accuracy. Conversely, the exhaustive search of the optimal combination of features from two and three channels significantly improved performance with respect to using r(2) for channel selection. With features optimally extracted from three channels with optimized scaling filter in the discrete wavelet transform, the classification accuracy was on average 72.2%. Thus, the choice of features had greater impact on performance than the choice of the classifier for discrimination between the two non-motor imagery tasks investigated. The results are relevant for the choice of the translation algorithm for an on-line BCI system based on non-motor imagery.

Classification of Mental States Based on Spatiospectral Patterns of Brain Electrical Activity

“…For the feature selection or reduction, we used the sequential floating forward selection (SFFS) that overcomes a well-known drawback of sequential forward selection (SFS), referred to as a nesting effect, to select the best feature subset as well as to reduce the dimensionality of the feature vectors [41]. Since it has been frequently reported that classification accuracy does not significantly depend on the types of classification algorithms [31], the Bayesian classifier that is one of the most frequently used classification algorithms in EEG-based mind reading studies [25,28,30,31,37] was used for classification.…”

“…Traditionally, spectral powers and their asymmetry between the two hemispheres have been frequently adopted as feature vectors for discriminating different mental tasks or cognitive states, such as mental calculation, internal speech, motor imagery, spatial navigation imagery, and so on [23][24][25][26][27][28][29]. In the majority of previous EEG-based mind reading studies, brain electrical activities recorded from a few electrodes, which produce the most distinct brain activity patterns, were chosen for analysis [23][24][25]27,[29][30][31]. However, similar to the individual voxel-based analysis in fMRI studies, this approach might also lead to loss of potentially valuable information associated with rather complex cognitive states because a variety of brain areas can be involved even when a participant is conducting a simple mental task [32].…”

Development of a hybrid mental spelling system combining SSVEP-based brain–computer interface and webcam-based eye tracking

“…Brain-computer interfaces (BCIs) are non-muscular communication and interaction technologies that allow these disabled individuals to communicate with the outside world using their brain signals [1]. To date, a variety of neural signals have been used with the aim of implementing practical BCI applications, such as electroencephalography (EEG) [1][2][3][4][5][6][7][8][9][10][11][12][13], magnetoencephalography (MEG) [14,15], electrocorticography [16,17], near-infrared spectroscopy [18][19][20], functional magnetic resonance imaging [21,22], and transcranial Doppler ultrasound [23,24]. In particular, the number of EEG-based BCI applications has increased markedly during the past five years [5].…”

A comparison of univariate, vector, bilinear autoregressive, and band power features for brain–computer interfaces

“…Selecting suitable features is crucial to obtain good overall BCI performance [7, 8]. In this study, we focus on BCIs based on event-related desynchronization [28] and explore extensions of the simple AR model and compare the resulting features with logBP features.…”