An efficient feature extraction scheme for classification of mental tasks based on inter-channel correlation in wavelet domain utilizing EEG signal

Highlights

• Alternative approach of obtaining band limited signals using filter.

• Effect of using various classical BCI classifiers.

• Generalization of the proposed method for BCI.

• Detail explanation of how the number of levels in wavelet decomposition is chosen and the equivalent range of frequencies obtained by each wavelets function.

Abstract

In this paper, an efficient scheme of extracting features from EEG signal is proposed for mental task classification based on inter-channel relationship in wavelet domain. It is shown that use of wavelet domain inter-channel relationship can drastically improve the classification performance obtained by conventional wavelet statistics. Both multi-level wavelet decomposition and node reconstruction are utilized for proposed inter-channel correlation feature extraction. It is expected that the correlation obtained from different combination of channels will be different for various mental tasks depending on the nature of the stimulus generated in the brain and thus can provide distinctive features. Support vector machine (SVM) classifier is used to carry out classification of five different mental tasks obtained from an openly accessible EEG dataset. It is found that the proposed scheme can classify mental tasks with a very high level of accuracy compared to that obtained by some existing methods.

Introduction

Classification of various mental tasks, such as geometrical figure rotation, mathematical multiplication, mental letter composing, visual counting, base-line resting plays an essential role to handle various applications of brain–computer interface. For the purpose of classification of these mental tasks, electroencephalogram (EEG) is one of the most promising tools, which is widely used now-a-days. Different methods are available in literature where distinctive features are extracted from EEG signals and then used for mental task classification. For example, in [1], [2], spectral power and asymmetry ratio based feature extraction scheme is proposed. These methods offer comparatively satisfactory classification performance but lacks consistency for all cases. In [3], a dictionary consisting of power spectral density and common spatial pattern algorithm is introduced to classify various mental tasks. Mental task classification approach based on autoregressive (AR) modelling is reported in [4], [5], [6], [7]. Among these methods, in [7], generalized Higuchi fractal dimension spectrum apart from AR parameters is used as feature. In [8], instead of AR parameters, reflection coefficients are used for feature extraction. In [9], discrete wavelet transform (DWT) and empirical mode decomposition (EMD) based feature extraction approach is proposed where feature selection method is utilized for better classification performance. Besides DWT and EMD, wavelet packet decomposition is investigated for the decomposition of EEG signal for motor-imagery (MI) based brain–computer interface (BCI) system in [10]. In [11], interaction dynamics of neuronal oscillations are analyzed using wavelet transforms and shown that wavelet based coherence features provide better performance in classifying neuronal oscillations. In [12], both conventional Fourier and wavelet coherence is applied to sensory-evoked electrical brain activity and relatively better performance is obtained when coherence was calculated using wavelet approach rather than classical Fourier approach. A parametric feature extraction and classification method introduced in [13] use fast Fourier transform to obtain features for BCI system. In [14], stockwell transform based feature extrraction method utilizing statistical parameter is introduced. Cross-correlation based approach of classifying MI tasks is presented in [15] where a subset of possible channel pairs are considered to obtain cross correlation sequence. In these channel pairs, one signal is chosen as fixed reference signal obtained from a particular channel of a particular MI task, as a result, the effect of obtaining cross correlation sequence from all possible combination of channel pairs is not considered. Moreover, reference signal selection is not universal for all sorts of mental and MI tasks and requires empirical selection of a channel signal from different types of tasks. Most of the reported feature extraction except [15] are performed on the data taken from various channels while the inter-channel relationship has not been utilized. It is considered that stimulation of a particular location of brain depends on the types of mental tasks performed. With a view to capturing the difference between both stimulated and non-stimulated EEG signals of brains obtained from various channels placed in the skull, measuring inter-channel relationship in some efficient spectro-temporal domains may play a significant role. Thus, development of a proficient method capable of detecting and classifying different types of mental tasks utilizing the inter-channel relationship is still undiscovered.

One of the objectives of this paper is to exploit the distinctive nature of inter-channel relationship to obtain effective features. Both multi-level wavelet decomposition and node reconstruction of EEG data are utilized to obtain the proposed inter-channel correlation features. Moreover, in the proposed method, unlike [15], reference signal selection based on empirical reasoning is not required and only the information obtained from the channels of the test frame is utilized to extract correlation features. Effect of using statistical features obtained from various wavelet approaches along with proposed inter-channel correlation features on classification performance is also investigated. SVM classifier is used to carry out classification process utilizing a publicly available EEG dataset on various mental tasks.