Channel Selection for Optimizing Feature Extraction in an Electrocorticogram-Based Brain-Computer Interface

Abstract

Feature extractor and classifier are two major components in a brain-computer interface system, in which the feature extractor plays a critical role. To increase the discriminability of features or feature vectors used for classification, it is necessary to select a suitable number of task-related data recording channels. In this article, a machine-learning algorithm is proposed for optimizing feature extraction in an electrocorticogram-based brain-computer interface. Common spatial pattern was used for feature extraction, and channel selection was performed by genetic algorithm for optimizing the feature extraction. Fisher discriminant analysis was used as classifier, and the channel subset chosen at each generation was evaluated by classification accuracy. The algorithm was applied to three electrocorticogram datasets that were recorded during two kinds of motor imagery tasks. The results suggest that the channel number used for building a brain-computer interface system could be significantly decreased without losing classification accuracy, and the accuracy rate could be noticeably improved by using the optimal channel subsets chosen by genetic algorithm.

Unsupervised machine-learning classification of electrophysiologically active electrodes during human cognitive task performance

“…It remains unclear how the selection of active electrodes would affect performance in verbal memory task-related classification of, for example, trials with successful and unsuccessful recall of remembered and forgotten words based on iEEG signals. Various methods for active electrode selection have been proposed to improve task-related classification performance in other tasks, including recursive channel elimination 17,22 and genetic algorithms 23,24 . For recursive channel elimination 17 , the authors select active electrodes by recursively training a support vector machine classifier on the features of a subset of channels and eliminate the least-contributing channel until a pre-specified number of channels remain.…”

Decoding repetitive finger movements with brain activity acquired via non-invasive electroencephalography

“…GAs have been used in EEG studies to improve the detection of neural pathologies such as Alzheimer's Disease and Epilepsy (Kim et al, 2005; Ocak, 2008). GAs have also been used to improve classifiers that detect motor intent for BCI application by improving the quality of trials used to train the classifier (Wang et al, 2012), or by determining which channels are used with the classifier (Wei et al, 2010). In the context of this study, the GA was utilized to find the optimal set of channels that yielded the highest decoding accuracies.…”

User Adapted Motor-Imaginary Brain-Computer Interface by means of EEG Channel Selection Based on Estimation of Distributed Algorithms

“…More closely related to our approach, previous researches have addressed the channel selection problem using evolutionary algorithms [4,9,[14][15][16][17]. However, the application of EDAs has been constrained to the analysis of Magnetoencephalography (MEG) data in the context of multiobjective optimization [18], an approach with important differences with the one introduced in this paper.…”

“…Wei et al [17,19] apply genetic algorithms (GAs) for the analysis of multichannel electrocorticogram (ECoG) recordings in a ECoG-based BCI. They combine the application of the CSP method for feature extraction, Fisher discriminant analysis as classification method, with the use of a GA for feature selection.…”