An optimized feature selection and classification method for using electroencephalographic coherence in brain–computer interfaces

https://doi.org/10.1016/j.bspc.2014.11.001Get rights and content

Highlights

  • EEG coherence is used to blindly select sensors in a brain–computer interface.

  • In addition to reveal connectivity, coherences are used to discriminate mental tasks.

  • Our method selects optimal sensors without prior knowledge about the brain processes.

  • The proposed method achieves good accuracy rates with a reduced number of sensors.

  • In spite of previous reports, we show that EEG coherence is relevant for classifying mental tasks.

Abstract

We propose a method to use electroencephalographic (EEG) coherences as features in a brain–computer interface (BCI). The coherence provides a sense of the brain's connectivity, and it is relevant as different regions of the brain must communicate between each other for the integration of sensory information. In our case, the process of feature selection is optimized in the sense that only those statistically significant and potentially discriminative coherences at a specific frequency are used, which results in a feature vector of reduced-dimension. Next, those features are classified through an optimized linear discriminant, where the best discriminating hyperplanes are selected such that the area under the receiver operating characteristics (ROC) curve is maximized. Overall, the proposed EEG coherence selection and classification method can provide efficiency rates similar to those obtained with other methods in BCI, but with the advantage of blindly selecting and optimal combination of features out of all the possible pairwise coherences. We demonstrate the applicability of the proposed method through numerical examples using real data from motor and cognitive tasks.

Introduction

A brain–computer interface (BCI) is a communication system that allows a subject to act on his/her environment solely by means of his/her thoughts, i.e. without using the brain's normal output pathways of muscles or peripheral nerves [1]. Non-invasive BCIs rely on electroencephalographic (EEG) measurements of the brain's activity to read out the intentions of the subject and translate them into commands for a computerized system.

The translation from the brain activity to a command is usually achieved by means of a feature generator that extracts feature values from the EEG signals that correspond to the underlying neurological mechanism employed by the user for control. Next, a feature translator classifies the features into logical control signals, such as a two-state discrete output. Many methods have been proposed so far to carry out the extraction/classification processes in BCI, and a very comprehensive review about them can be found in [2]. In general, feature extraction methods are closely related to specific neuromechanisms, while feature classification algorithms are determined by the type of features that they discriminate.

Here, we examine the use of the EEG coherence as feature in a BCI. The coherence provides a sense of the brain's connectivity, and it is relevant to measure it as different regions widely distributed over the brain must communicate between each other in order to provide the basis for integration of sensory information, as well as for many functions that are critical for learning, memory, information processing, perception, and behavior. Transient periods of synchronization of oscillating neural discharges have been proposed to act as an integrative mechanism that may bring a widely distributed set of neurons together into a coherent ensemble that underlies a cognitive act [3], and many studies have used the EEG coherence to quantify such synchronization process (see [4] and references therein). In [5], the patterns in the coherence were studied during sequential and simultaneous tasks, while in [6], signals corresponding to spontaneous EEG, imaginery movement, and movement execution were classified based on the coherence using hidden Markov models and a multilayer perceptron. Nevertheless, the only attempt known to us of using the coherence in the context of BCI can be found in [7]. There, the use of the coherence as a feature was assessed for the case of measuring the mean coupling between signals recorded from an electrode and its neighbors, and a few individual electrode pairs reflecting connectivity between fronto-centro-parietal and temporal lobes. Given the limited number of subjects tested and the coherences that were assessed, their results do not allow for a statistical conclusion regarding general performance of the proposed measures. Nevertheless, the results in [7] suggest that coherence-based features might not perform as well as other features, but still could be relevant for classifying mental tasks.

Therefore, in this paper we propose an optimized method for feature selection and classification which is customized for the EEG coherence. The process of feature selection is optimized in the sense that only those statistically significant and potentially discriminative coherences at a specific frequency are used, which results in a feature vector of reduced-dimension. Next, those features are classified through an optimized linear discriminant, where the best discriminating hyperplanes are selected such that the area under the receiver operating characteristics (ROC) curve is maximized. Based on these ideas, the paper is organized as follows: the coherence is briefly reviewed in Section 2, then the proposed coherence-based feature selection and classification process is introduced; in Section 3, we show the applicability of our method through a series of numerical examples using real EEG data; in Section 4, we discuss the results and future work.

Section snippets

Methods

In this section we briefly review the concept of coherence, then we explain our proposed coherence-based feature selection and pose a classification procedure customized for those features.

Numerical examples

We performed a series of numerical experiments to show the applicability of the proposed method using two different data sets of real EEG measurements. The first one corresponds to motor imaginery, and the second to cognitive tasks. For both data sets, we evaluated the search for optimal feature vectors from combinations of L = 3 sensors as previous studies have revealed that at least three modular structures of effective brain networks are involved in cognitive control [15]. Therefore, we

Conclusions

We presented a method based on EEG coherence to select an optimal feature vector with reduced-dimension for BCI applications. The selection was performed without prior knowledge about brain activity related to the tasks involved. However, by assessing the significance of the EEG coherence, we assure that the signals from selected channels indeed provide information related to the brain connectivity associated to the task. Our experiments with real EEG data show that the selected sensors are

Acknowledgment

This work was supported by the Mexican Council of Science and Technology (CONACyT) under Grant 220145.

References (22)

Cited by (20)

  • Development of a brain computer interface based on steady-state visual evoked potential with multiple intermittent photo stimulation

    2020, Biomedical Signal Processing and Control
    Citation Excerpt :

    The explosive growth of this field is evident by the fact that the majority of all BCI research articles ever published have appeared in the past 10 years. The majority of such BCIs use the steady-state visual evoked potential (SSVEP) as the input signal, from which features are extracted by means of signal processing techniques, such as canonical correlation analysis [2], coherence function [3] and singular spectrum analysis [4], to provide the output commands. Depending on the kind of brain activity signal used as input, the BCI may be classified as either active or passive.

  • Evaluating the effect of the cutoff frequencies during the pre-processing stage of motor imagery EEG signals classification

    2019, Biomedical Signal Processing and Control
    Citation Excerpt :

    Therefore, it is necessary to select carefully the pre-processing techniques to be applied according to the features of the signal, as well as the parameters for each technique to obtain a filtered signal that guarantees the success of the further stages. There are proposals for selecting an appropriate band pass filter before feature extraction to achieve better results in the classification [19,20]. However, these works only explore a reduced range of cutoff frequencies of the band pass filter, losing relevant information from other frequencies that could increase the performance of the classification stage.

  • Evaluating spiking neural models in the classification of motor imagery EEG signals using short calibration sessions

    2018, Applied Soft Computing Journal
    Citation Excerpt :

    So, in this section both methods are briefly exposed. For more detailed information see [30,19,20]. Finally, the feature vector yOPT is built by the coherences values computed from the selected L electrodes.

  • A self-adaptive frequency selection common spatial pattern and least squares twin support vector machine for motor imagery electroencephalography recognition

    2018, Biomedical Signal Processing and Control
    Citation Excerpt :

    The core of an MI-BCI system is its discrimination algorithm. Many useful discriminant methods have been proposed to improve the robustness and accuracy of BCIs, such as the locality preserving projection method based on a self-regression model [19], the linear discriminant algorithm [20], the probabilistic methods [21,22], and various neural networks[23–25]. Support vector machines (SVMs) have also been used frequently for classification of EEG patterns, including an SVM combined with intelligent optimization algorithms that were certified to give high quality results in BCI [14,26,27].

View all citing articles on Scopus
View full text