Spectral entropy feature subset selection using SEPCOR to detect alcoholic impact on gamma sub band visual event related potentials of multichannel electroencephalograms (EEG)

Highlights

• The proposed feature selection method is based on maximization of class separability and minimum correlation between selected features.

• ICA is applied on Physionet alcoholic EEG database to separate pure EEG epochs from artifacts such as eye blink, body and cranial muscle movements.

• As per our knowledge, this is the first attempt to remove eye blink and muscle activity (> 30 Hz) from the EEG epochs of the alcoholic EEG database.

• Results show better classification accuracy compared to previous studieson the same Physionet alcoholic EEG database.

• The normalized spectral entropy feature is used for extensive analysis in this study for the first time.

Abstract

The problem of analyzing and identifying regions of high discrimination between alcoholics and controls in a multichannel electroencephalogram (EEG) signal is modeled as a feature subset selection technique that can improve the recognition rate between both groups. Several studies have reported efficient detection of alcoholics by feature extraction and selection in gamma band visual event related potentials (ERP) of a multichannel EEG signal. However, in these studies the correlation between features and their class information is not considered for feature selection. This may lead to redundancy in the feature set and result in over fitting. Therefore in this study, a statistical feature selection technique based on Separability & Correlation analysis (SEPCOR) is proposed to select an optimal feature subset automatically that possesses minimum correlation between selected channels and maximum class separation. The optimal feature selection consists of a ranking method that assigns ranks to channels based on a variability measure (V-measure). From the ranked feature set of highly discriminative features, different subsets are automatically selected by heuristically applying a correlation threshold in steps from 0.02 to 0.1. These subsets are applied as input features to multilayer perceptron (MLP) neural network and k-nearest neighbor (k-NN) classifiers to discriminate alcoholic and control visual ERP. Prior to feature selection, spectral entropy features are computed in gamma sub band (30–55 Hz) interval of a 61-channel multi-trial EEG signal with multiple object recognition tasks. Independent Component Analysis (ICA) is performed on raw EEG data to remove eye blink, motion and muscle artifacts. Results indicate that both classifiers exhibit excellent classification accuracy of 99.6%, for a feature subset of 22 optimal channels with correlation threshold of 0.1. In terms of computation time, k-NN classifier outperforms multilayer perceptron-back propagation (MLP-BP) network with 7.93 s whereas MLP network takes 55 s to perform the recognition task with the same accuracy. Compared to feature section methods used in previous studies on the same EEG alcoholic database, there is a significant improvement in classification accuracy based on the proposed SEPCOR method.

Introduction

It is well-established that there will be structural and functional changes in the neuronal activities of human brain upon consumption of alcohol and even after long term abstinence [1]. Event related potentials (ERP) exhibit cerebral activity that characterize spatio-temporal changes in the human brain over a period of time due to alcoholism [2], [3], [4]. The dynamic processes of brain such as memory, attention and cognitive processing [5], [6] are correlated with synchronizations of phase locked peaks generated by ERP which are characterized as delta, theta, alpha, beta and gamma waves. Changes in characteristics of these waves due to alcoholism are reported extensively in literature [7], [8], [9], [10], [11].

Numerous studies have reported the use of linear, non-linear and statistical measures (features) for the classification of alcoholic and control EEG epochs achieving very high classification accuracies. The extraction of power spectral density (PSD) of the γ band (> 30 Hz) has been extensively used for classification of alcoholics and controls in the literature using FFT and other parametric techniques [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23]. The γ sub band between 30–50 Hz is assumed to contain the visually evoked responses that differentiate an alcoholic from a control subject. These studies confirmed that the parametric methods were superior in their performance as compared to FFT due to their better frequency resolution, leading to achieve better SNR accuracy. Also among all the parametric methods used, Burg’s method was reported to extract the most distinguishable parameters for classifying alcoholics and controls

Studies have also shown that there is a reduction in evoked γ oscillations [3] in alcoholics during the processing of a visual object recognition task in the occipital and frontal regions. The deficits in γ oscillations manifest themselves as cognition deficits in selective attention and working memory tasks. In order to study cognitive impairments in alcoholics, it is necessary to identify specific regions of brain that are largely influenced by alcohol consumption during these event related oscillations. Therefore selection of channels with high discrimination between alcoholic and control subjects using prominent features extracted from visual ERP’s of a multichannel EEG recording in γ band needs to be investigated. Selection of optimal subset from 64 channel EEG recordings of alcoholics and controls has been explored by a few researchers [15], [19], [24], [25], [26], [27], [28], [29]. The spectral power ratios in multiple sub gamma bands and mean gamma band power in the visual evoked potentials of multichannel EEG signal are reported to be good discriminators for optimal feature selection and classification of alcoholics and controls [14], [25]. An interesting study [15], reported on the selection of seven out of 61 channels using genetic algorithm (GA) optimization to effectively discriminate alcoholics and controls. Within the entire spectrum of 2–50 Hz, only γ2 (37–43 Hz) and γ3 (44–50 Hz) bands of the seven optimized channels were reported to be effective features providing an average classification accuracy of 94.3% and 81.8% with MLP-BP network and fuzzy art map (FAM) classifier respectively. Studies by Ong et al. [24] have shown that the use of PCA (Principal Component Analysis) for selecting optimal channels to detect alcoholics resulted in classification accuracies of 95.83%, 94.06%, 86.01% and 75.13% for 61, 16, 8, and 4 channels respectively. The disadvantage here is that the use of PCA requires large memory while handling large datasets.

The correlation between selected optimal subset of channels for alcoholics was not explored until a specific study [25] that was carried out to select a set of active channels based on mean γ band power for each channel per subject. These active channels were later used to determine correlation coefficient among them and only those which scored a correlation coefficient greater than 0.6 with other selected channels were retained as candidate channels to train least square SVM classifier for the discrimination of alcoholics and controls. An average classification accuracy of 80% was reported among different pairs of 45 optimal channels. The problem in this method is that the highly correlated channels become redundant as far as classification accuracy is considered which may also lead to over fitting.

A study [19] using non-linear parameters such as ApEn (Approximate Entropy), Sample Entropy, Lyapunov Exponent and Higher order spectra for feature extraction and detection of alcoholics reported an accuracy of 91.7% using SVM classifier. The selection of channels was done by performing a statistical t-test on all the features and seven of them are used for classification The study used the same alcohol EEG dataset [30] as in the proposed study with 30 alcoholic and 30 control subjects visual ERP’s.

In all these studies, the class information between groups is not considered while choosing features. Even though the selection of optimal channel subset is based on certain criterion such as the gamma band power or the statistical t-test or correlation among them, the merit of each individual channel within the subset is not weighed in terms of its ability to separate alcoholics from controls. Hence the main objective of this proposed research is to choose only those optimal spectral entropy features which maximize the separation between classes with low correlation among them by using a statistical pattern recognition method called Separability and Correlation (SEPCOR) method [31], [32], [33]. It computes a variability measure (V-measure) for each channel which is defined as a ratio of the variance of class means to the mean of variance within the class. A high V-measure for a channel implies that the variance between class outputs is large while the mean of variance within the class output is small. Thus the high V-measure of a channel is indicative of maximum class separability and minimum correlation value and hence more discriminative of groups. The channels are arranged in a descending order with respect to their V- measure. A feature with highest V-measure is the first candidate feature chosen and the second feature is selected if its correlation coefficient with the first chosen channel is less than a preset (user defined) correlation threshold called, MAXCOR else it will be rejected. Proceeding further, the next sorted channel is tested for correlation threshold with the previously selected channel with highest V-measure. This method is repeated until all the features are exhausted. The channels selected thus represent a minimum set of (subset) highly uncorrelated channels with large V-measures The validity of the proposed method for feature selection is evaluated by applying SEPCOR reduced feature set as inputs to MLP and k-NN classifiers. The classifier performances are cross validated using 50% hold out cross validation.

Recently, entropy estimation of EEG signals have been used extensively to explain how EEG signals change with time either in frequency or in phase domain [34], [35], [36]. It implies many possible different rates of change of summed pyramidal cell membrane potentials. Commercial entropy monitors measure the depth of anesthesia as EEG activity is more regular (orderly) under its influence than in wakefulness [37], [38]. Higher values of entropy are indicative of more complexity involved in the dynamic processes of different cortical regions. It is reported that the spectral peak of increasing amplitude in α-frequency band causes spectral entropy to increase when low frequencies are present in the signal [39]. Thus the entropy estimation provides a measure of disorderliness and hence some important information regarding the complexity of processes involved in a system. These studies lead us to explore the potential use of entropy as an effective discriminating feature for classifying alcoholic/control visual ERP in sub gamma band range. In the present study, Spectral Entropies (SE) of artifact free EEG recordings are estimated in the gamma sub band range (30–55 Hz) for both alcoholic and control subjects’ 61-channel EEG recordings. The estimated SE features from 61 channels of both alcoholic and control gamma sub band visual ERP form the entire feature set. Traditionally, EEG data epochs containing any or/and all of these artifacts are rejected and this may lead to loss of important information from the dataset. To overcome this, artifacts are separated from actual EEG signals by performing linear transformations using methods such as Principal Component Analysis (PCA), Independent Component Analysis (ICA) etc. In the current study, ICA is used to separate EEG signal from eye blink, motion and muscle artifacts.

The rest of the paper is organized as follows: Section 2 describes the proposed methodology. Results and discussion are reported in Section 3 and 4 respectively. Conclusion is reported in Section 5.