Introduction

Epilepsy disease is related to neurological disorder manifested with sudden electrical disturbance of the brain, which may be transient or long haul (Şengür et al., 2016). WHO reported that approximately 50 million people of the world are affected with epilepsy (Epilepsy, 2019). Manual detection of epileptic seizure from EEG signals is not easy task. There is always a deficiency of high skills of the neurophysiologists in detecting epileptic seizure from EEG of patients admitted in neurology centre. An automatic system for identification of epilepsy may solve this issue. Electroencephalogram (EEG) is the most common instrument for recording of brain’s electrical activity and constructs a visual representation that containing information about the brain activities (Khandpur, 2019). EEG based brain computer interface (BCI) using machine learning approach is the most common technique for automatic diagnosing and detection of brain related disease.

Related works

Dennis et al. (2014) applied spectrogram image based method on classification of sound event. Three different types of temporal approaches namely frame based, global and local feature were applied for feature extraction from spectrogram images. In this work, global feature namely Sub-band Power Distribution Image Feature (SPD-IF) based classifier were outperform in classification of spectrogram images related to different sounds. An optimum allocation based multiclass EEG signals classification was reported by Siuly & Li (2014). In this work, EEG signals of all classes were segmented into epochs of fixed time period. Further, optimum allocation (OA) algorithm was applied on epochs to select the optimum samples from epochs. These optimum samples of different classes of EEG signals were utilized as input features for multiclass classification. Multiclass least square support vector machine (MLS-SVM) classifier with radial basis kernel function (RBF) was used for proposed OA based EEG signal classification study.

A survey on entropy based features used for epileptic seizure EEG signals classification was reported in Acharya et al. (2015). This survey, found that the Renyi’s Entropy (REN), Sample Entropy (SampEn), Spectral Entropy (SEN) and Permutation Entropy (PE) were the most promising features for automated epileptic seizure detection system. Das & Bhuiyan (2016) proposed epileptic seizure EEG signal detection technique using combination of empirical mode decomposition (EMD) and discrete wavelet transform (DWT).  In this work, EEG signals were decomposed using EMD and intrinsic mode function (IMF) of level 1 and 2 are extracted. Further, level 4 DWT decomposition was carried out on IMF 1 and IMF 2 with Daubechies 4 (db4) mother wavelet. Spectral entropy based features, like shanon entropy, log-energy entropy and Renyi entropy were calculated from EMD-DWT decomposed EEG signals. A classification accuracy of 89.40% was obtained by k-nearest neighbour (kNN) classifier using EMD-DWT with log-energy entropy feature. Şengür et al. (2016) proposed epileptic seizure identification using time-frequency (t-f) image based texture features. In this work, EEG signals were converted into t-f images using Short Time Fourier transform (STFT). Grey Level Co-occurrence Matrix (GLCM), Texture Feature Coding Method (TFCM) and Local Binary Pattern (LBP) were applied on the t-f images for texture features extraction. Classification accuracy (CA) of 92.50% was achieved with support vector machine (SVM) using GLCM texture features as learning inputs.

Chandani & Kumar (2017) reported discrete wavelet transform and neural network based epilepsy detection model. In this work, Daubechies db4 mother wavelet was utilised for decomposition of EEG signals up to level 5. Features namely mean, skewness, kurtosis, entropy, standard deviation and median were extracted from decomposed EEG signals. These statistical features were used as training and testing attributes for multi layer perceptron neural network (MLPNN) with hold out data division protocol. Mohammadpoory et al. (2017) reported epileptic seizure identification method based on weighted visibility graph entropy (WVGE). In this work, EEG signals were mapped into WVG (weighted visibility graph) and from the structure of WVGs, WVG entropy was computed. Some statistical measures namely mean, maximum, standard deviation and minimum of WVGEs were subsisted as features. Four classifiers namely SVM, kNN, decision tree (DT) and Naïve Bayes (NB) were used for WVGE based classification. Classification accuracy of 97.00% was shown by DT classifier.

The rest of the paper is structured as follows: In materials and methods section, we explain the dataset used in present work, about spectrogram and briefly describe the gray scale feature extraction methods used in present work. Further, we also explain about the proposed method in details. Performance evaluation subsection briefly describes the parameters used for assessment of proposed method. Next consecutive section is results and discussions. Finally, we conclude the present work in conclusions and future aspects section.    

Materials and methods

In this section, various steps for epilepsy detection method of present work are discussed in detail. Epileptic seizure detection is done using spectrogram based images of EEG signals. All the steps are explained in following subsections.  

Dataset

In the present work, an open source EEG dataset are used for proposed an epileptic seizure detection model. This standard dataset is obtained from the University of Bonn (EEG time series download page, 2019). This dataset contains EEG data of five set A, B, C, D and E. Here all the five sets are having 100 signals of EEG and each EEG signal has 4097 samples. Time duration of each EEG signal of all five sets is 23.6 second. Table 1 shows brief information about epileptic seizure EEG signal dataset cited from Andrzejak et al. (2001). From Table 1, it is clear that the first two set A and B are having normal EEG signals. C and D are having seizure free EEG signals and set E is containing the epileptic seizure EEG signals.

Table 1. Description table of dataset

Spectrogram

Spectrograms are time-frequency (t-f) images of time domain signals. An illustration of spectrogram image of normal and epileptic EEG signals is given in Figure 1. Short Time Fourier transform (STFT) is utilized to formation of spectrogram (Şengür et al., 2016). STFT is one of the mostly used techniques for analysis of non-stationary signals like EEG signals. STFT of EEG signals provides information about frequency components available in particular time interval simultaneously (Polikar, 2019). To obtain time and frequency information simultaneously, whole signal is divided into small parts using window function. Further, Fourier transforms are taken of each small parts of signal. Mathematical formula of STFT and window function used is shown below in consecutive equations.    

                                               {1}

                                ,                       {2}

Where  of equation 1 shows the small part of the signal at time . Equation 2 shows the formula of hamming window function and is number of discrete frequencies.

          (a)                                                                                           (b)

Figure 1. Spectrogram images of EEG signals (a) normal EEG and (b) epileptic seizure EEG

Feature Extraction

In pattern recognition, feature extraction is one of the vital tasks. In the present work, STFT t-f image based feature such as colour, texture, shapes and statistical features are extracted. Features of STFT t-f image contain time and frequency information of EEG signal that helps to fulfil the pattern classification task. Huge documentations are available on image features related algorithms. Table 2 summarizes the number of feature, name and category of feature extracted from STFT t-f images. In the present work, 472 gray scale features of STFT t-f images are extracted.          

Table 2. Summary of features extracted from STFT t-f images

Proposed methodology

The proposed model is designed for spectrogram image based epileptic seizure classification using STFT technique as depicted in Figure 2. This model is divided into two major parts using dashed vertical line. The left part of the model represents the training part called as offline system while right part of the model portrays the testing part called as online system. Training portion of model contains various blocks like preprocessing of EEG signals, feature extraction and training of classifier. In the present work, first block represents signal preprocessing. Elimination of noises from EEG signals is the initial step of the preprocessing. All noise free EEG signals are transformed into spectrogram images using STFT technique. In the next consecutive block, 472 gray scale features are extracted from these spectrogram images. Brief information of these 472 features is tabulated in Table 2. These features are taken as input for supervise training of classifiers in next consecutive block of the training part. After iterative training of classifier, optimized learning parameters are generated. The trained classifier, with optimized learning parameters is then used in online system of the model. Testing part of the model has same consecution signal preprocessing and feature extraction blocks as described for training part. Now the next step in testing part is taking decision on EEG signals using extracted feature as input by trained classifier.                    

Figure 2. Proposed methodology of epileptic seizure detection

Performance evaluation

In the present work, performances of classification models are evaluated using two class confusion matrix as shown in Table 3. Here, symbol  subscript with TP, FP, TN and FN indicates true positive, false positive, true negative and false negative respectively. Some performance measures like accuracy, sensitivity, specificity, Area under receiver operating characteristic curve (AUC) and Matthew’s correlation coefficient (MCC) are used in the present work. These performance measures can be calculated from equation 3 to equation 7 respectively (Singh et al., 2017).

Table 3. Arrangement of confusion matrix for two class problem

                                                 {3}

                                                                         {4}

                                                                         {5}

                                                {6}

      {7}

Results and Discussion

This section contains the results of experiments which are described in materials and methods section. In the present work, six SVM classifiers with different kernel functions are studied for classification of normal and epileptic EEG signals using spectrogram image features based technique. Three different data division protocols namely 5-fold, 10-fold and hold out cross validation (CV) are used. Table 4-6 shows performances of SVM classifiers using 5-fold, 10-fold and Hold out CV techniques respectively. Table 4 shows the performance of SVM classifiers with 5-fold cross validation on the basis of five statistical measures. From the Table 4, it is found that high classification accuracy (CA) of 90.00% shown by quadratic SVM classifier using 5 fold CV technique. Table 4 also shows the other performance measures for quadratic SVM classifier in terms of sensitivity, specificity, AUC and MCC are 91.00%, 89.00%, 90.00%, and 80.02% respectively.  Linear SVM with CA of 89.50% and cubic SVM with CA of 87.00% are ranked second and third position respectively.

Table 5 shows the performance of SVM classifiers along with 10-fold CV technique. From the Table 5, it can be observed that quadratic SVM shows high CA of 90.00%. Table 5 also shows the other performance measures for quadratic SVM classifier in terms of sensitivity, specificity, AUC and MCC are 93.00%, 87.00%, 90.00%, and 80.14% respectively.  Linear SVM shows CA of 88.50% and ranked second position in 10-fold CV. Third rank is achieved by cubic SVM with CA of 87.50% (see Table 5). Similarly, Table 6 illustrates the performance of SVM classifiers using hold out CV technique. From the Table 6, it is found that quadratic SVM shows high CA of 86.36%. Table 6 also shows the other performance measures for quadratic SVM classifier in terms of sensitivity, specificity, AUC and MCC are 87.88%, 84.85%, 86.36%, and 72.76% respectively.  Linear SVM is ranked second with CA of 83.33% and cubic SVM ranked third with CA of 81.82%. From Table 4-6, it is also observed that linear SVM, quadratic SVM and cubic SVM are shown consistency in performance.               

Table 4. Performance parameters of classifier models using 5-fold cross validation

Table 5. Performance parameters of classifier models using 10-fold cross validation

Table 6. Performance parameters of classifier models using hold out cross validation

Conclusions and Future aspects

In the present work, spectrogram image representation of the EEG signals with gray scale features has been used for identification of the epileptic seizure. In this work, six SVM classifiers with different kernel functions and three data division protocols are utilised. Highest classification accuracy of 90.00% is achieved using quadratic SVM with 5 fold and 10-fold CV techniques (see Table 4 and Table 5). Colour features of spectrogram images can improve the performance of present model. Near future, feature selection techniques can be applied on combination of colour and grey scale features for selecting the most appropriate features for classification of normal and epileptic EEG signals. New classification techniques like deep learning, extreme machine learning and hybrid machine learning can be used for classification purpose.

References

Acharya, U.R., Fujita, H., Sudarshan, V.K., Bhat, S. and Koh, J.E. (2015). Application of entropies for automated diagnosis of epilepsy using EEG signals: A review. Knowledge-Based Systems, 88: 85-96.

Amadasun, M. and King, R. (1989). Textural features corresponding to textural properties. IEEE Transactions on systems, man, and Cybernetics, 19(5): 1264-1274.

Andrzejak, R.G., Lehnertz, K., Mormann, F., Rieke, C., David, P. and Elger, C.E. (2001). Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Physical Review E, 64(6): 061907.

Chandani, M. and Kumar, A.  (2017). Classification of EEG Physiological Signal for the Detection of Epileptic Seizure by Using DWT Feature Extraction and Neural Network. International Journal of Neurologic Physical Therapy, 3(5): 38-43.

Chary, R.V.R., Lakshmi, D.D.R. and Sunitha, D.K.N. (2011). Image Retrieval and Similarity Measurement based on Image Feature. International Journal of Computer Science & Technology, 2(4): 385-389.

Christodoulou, C.I., Pattichis, C.S., Pantziaris, M., Tegos, T. and Nicolaides, A. (1998). Texture Analysis for the Classification of Carotid Plaques.

Chu, A., Sehgal, C.M. and Greenleaf, J.F. (1990). Use of gray value distribution of run lengths for texture analysis. Pattern Recognition Letters, 11(6): 415-419.

Costa, A.F., Humpire-Mamani, G. and Traina, A.J.M. (2012, August). An efficient algorithm for fractal analysis of textures. In 2012 25th SIBGRAPI Conference on Graphics, Patterns and Images (pp. 39-46). IEEE.

Das, A.B. and Bhuiyan, M.I.H. (2016). Discrimination and classification of focal and non-focal EEG signals using entropy-based features in the EMD-DWT domain. biomedical signal processing and control, 29: 11-21.

Dennis, J., Tran, H.D. and Chng, E.S. (2014). Analysis of spectrogram image methods for sound event classification. In Fifteenth Annual Conference of the International Speech Communication Association.

Dyer, C.R. and Rosenfeld A. (1976). Fourier texture features: Suppression of aperture effects. IEEE Transactions on Systems, Man, and Cybernetics, 6(10): 703-705.

EEG time series download page. Retrieved June 12, 2019, from http://epileptologie-bonn.de/cms/front_content.php?idcat=193&lang=3&changelang=3

Epilepsy. Retrieved July 29, 2019, from https://www.who.int/news-room/fact-sheets/detail/epilepsy

Galloway, M.M. (1975). Texture classification using gray level run length. Computer graphics and image processing, 4(2): 172-179.

Gonzalez, R.C. and Woods, R.E. (2010). Digital image processing using MATLAB. 2nd ed., Pearson Prentice Hall.

Haralick, R.M. and Shanmugam, K. (1973). Textural features for image classification. IEEE Transactions on systems, man, and cybernetics, SMC-3(6): 610-621.

Hu, M.K. (1962). Visual pattern recognition by moment invariants. IRE transactions on information theory, 8(2): 179-187.

Khandpur, R. S. (2019). Handbook of Biomedical Instrumentation. Third ed., Chennai, Tamil Nadu: McGraw Hill Education (India) Private Limited.

Laws, K.I. (1980, December). Rapid texture identification. In Image processing for missile guidance. International Society for Optics and Photonics, 238: 376-381.

Mohammadpoory, Z., Nasrolahzadeh, M. and Haddadnia, J. (2017). Epileptic seizure detection in EEGs signals based on the weighted visibility graph entropy. Seizure, 50: 202-208.

Polikar, R. The wavelet tutorial. Retrieved August 05, 2019, from http://users.rowan.edu/~polikar/WTtutorial.html

Şengür, A., Guo, Y. and Akbulut, Y. (2016). Time–frequency texture descriptors of EEG signals for efficient detection of epileptic seizure. Brain informatics, 3(2): 101-108.

Singh, B.K., Verma, K., Panigrahi, L. and Thoke, A.S. (2017). Integrating radiologist feedback with computer aided diagnostic systems for breast cancer risk prediction in ultrasonic images: An experimental investigation in machine learning paradigm. Expert Systems with Applications, 90: 209-223.

Siuly, S. and Li, Y. (2014). A novel statistical algorithm for multiclass EEG signal classification. Engineering Applications of Artificial Intelligence, 34: 154-167.

Weszka, J.S., Dyer, C.R. and Rosenfeld, A. (1976). A comparative study of texture measures for terrain classification. IEEE transactions on Systems, Man, and Cybernetics, SMC-6(4): 269-285.

Wu, C.M. and Chen, Y.C. (1992). Statistical feature matrix for texture analysis. CVGIP: Graphical Models and Image Processing, 54(5): 407-419.