Local pattern transformation based feature extraction techniques for classification of epileptic EEG signals

Highlights

• Two feature extraction techniques are introduced for epileptic EEG signal classification.

• Both the techniques (LNDP and 1D-LGP) are based on local pattern transformation.

• The histogram based features lead to fast execution speed and high classification accuracy.

• Both techniques are better than 1D-LBP for preserving the structural property of patterns.

• LNDP and 1D-LGP techniques can be adopted for one dimensional signal processing.

Abstract

Background and objective

According to the World Health Organization (WHO) epilepsy affects approximately 45–50 million people. Electroencephalogram (EEG) records the neurological activity in the brain and it is used to identify epilepsy. Visual inspection of EEG signals is a time-consuming process and it may lead to human error. Feature extraction and classification are two main steps that are required to build an automated epilepsy detection framework. Feature extraction reduces the dimensions of the input signal by retaining informative features and the classifier assigns a proper class label to the extracted feature vector. Our aim is to present effective feature extraction techniques for automated epileptic EEG signal classification.

Methods

In this study, two effective feature extraction techniques (Local Neighbor Descriptive Pattern [LNDP] and One-dimensional Local Gradient Pattern [1D-LGP]) have been introduced to classify epileptic EEG signals. The classification between epileptic seizure and non-seizure signals is performed using different machine learning classifiers. The benchmark epilepsy EEG dataset provided by the University of Bonn is used in this research. The classification performance is evaluated using 10-fold cross validation. The classifiers used are the Nearest Neighbor (NN), Support Vector Machine (SVM), Decision Tree (DT) and Artificial Neural Network (ANN). The experiments have been repeated for 50 times.

Results

LNDP and 1D-LGP feature extraction techniques with ANN classifier achieved the average classification accuracy of 99.82% and 99.80%, respectively, for the classification between normal and epileptic EEG signals. Eight different experimental cases were tested. The classification results were better than those of some existing methods.

Conclusions

This study suggests that LNDP and 1D-LGP could be effective feature extraction techniques for the classification of epileptic EEG signals.

Introduction

Epilepsy is a central nervous system disorder and it has been reported that approximately 45–50 million people suffer from this disorder [1]. EEG captures the neurological activity inside the brain by placing electrodes on the scalp and helps in detection of epileptic seizure [2], [3]. Epileptic seizure detection can be considered as a classification problem where the task is to classify an input signal either as an epileptic seizure signal or as a non-seizure signal. EEG signals are usually recorded for long durations to carry out an analysis. Detection of epileptic seizure in these long duration EEG signals requires expertise. In the absence of expert, particularly in emergencies, seizure detection becomes a challenging task. Therefore, providing a framework for automated epileptic seizure detection is of great significance. A number of methods have been proposed in literature for the classification of epileptic EEG signals. The methods can be categorized into following domains of signal analysis:

Time domain analysis: Some techniques in the field of epileptic EEG signal classification belong to this category. Techniques like linear prediction [4], Fractional linear prediction [5], Principal component analysis based radial basis function neural network [6], etc. have been proposed for epileptic seizure detection.

Frequency domain analysis: With the assumption that the EEG signals are stationary signals, Polat and Gunes [7] introduced a hybrid framework based on frequency domain analysis with Fourier transform and decision tree. In the hybrid model, Fourier transform was used for feature extraction and decision tree was used for the classification. Srinivasan et al. [8] used features extracted from time domain and frequency domain for seizure detection in EEG signals. The extracted features were fed to the Elman neural network to detect epileptic seizure.

Time–frequency domain analysis: It is well known that EEG signals are non-stationary signals. Considering the non-stationary property of EEG signals, feature extraction based on time–frequency image analysis with Hilbert-Huang transformation (HHT) was introduced for seizure classification [9]. Tzallas et al. [10] introduced a combined approach with time–frequency domain analysis for seizure detection. A number of methods have been proposed based on wavelet transform [11], [12], [13], [14], [15], multi-wavelet transform [16] for the detection of epileptic seizure. With the consideration of both the non-stationary and non-linearity properties of EEG signals, empirical mode decomposition (EMD) based method was introduced for the classification between normal and epileptic EEG signals [17].

Even though the above techniques have been used for epileptic EEG signal classification and other applications, one of the major issues associated with these techniques is the computational cost involved in the feature extraction step. Feature extraction techniques based on the local pattern transformation are computationally simple and widely used in different pattern recognition applications. One of such technique is the Local Binary Pattern (LBP). LBP has gained popularity in the field of face recognition [18], signal processing [19], speech processing [20], and epileptic EEG signal classification [21], [22]. However, one limitation of LBP is its sensitiveness to local variation.

In this study, two effective feature extraction techniques called LNDP and 1D-LGP have been proposed for the classification of epileptic EEG signals. Both the techniques are computationally simple and insensitive to local and global variations. The insensitiveness of LNDP and 1D-LGP overcomes the limitation of 1D-LBP. These proposed techniques work in two phases. In the first phase, the local patterns are transformed and histogram formation is done. The histogram represents the feature vector of the corresponding EEG signal. Histogram classification is completed in the second phase. The histogram contains the structural information of the EEG signal. The classification has been carried out with four different machine learning classifiers. The classification performance is evaluated with 10-fold cross validation considering the sensitivity, specificity and accuracy.

The remaining content of this paper is organized as follows: methodology and materials used are discussed in Section 2. Experimental results are shown in Section 3. Finally, Section 4 concludes the article with future direction.