Classification of normal and epileptic seizure EEG signals using wavelet transform, phase-space reconstruction, and Euclidean distance.

COMM-3795; No. of Pages 16

ARTICLE IN PRESS c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 4 ) xxx–xxx

journal homepage: www.intl.elsevierhealth.com/journals/cmpb

Classification of normal and epileptic seizure EEG signals using wavelet transform, phase-space reconstruction, and Euclidean distance Sang-Hong Lee a , Joon S. Lim b , Jae-Kwon Kim c , Junggi Yang b , Youngho Lee b,∗ a b c

Department of Computer Science & Engineering, Anyang University, Anyang-si, Republic of Korea IT College, Gachon University, Seongnam-si, Republic of Korea Department of Computer Science & Engineering, Inha University, Inchon-si, Republic of Korea

a r t i c l e

i n f o

a b s t r a c t

Article history:

This paper proposes new combined methods to classify normal and epileptic seizure EEG

Received 21 August 2012

signals using wavelet transform (WT), phase-space reconstruction (PSR), and Euclidean dis-

Received in revised form

tance (ED) based on a neural network with weighted fuzzy membership functions (NEWFM).

18 April 2014

WT, PSR, ED, and statistical methods that include frequency distributions and variation, were

Accepted 21 April 2014

implemented to extract 24 initial features to use as inputs. Of the 24 initial features, 4 minimum features with the highest accuracy were selected using a non-overlap area distribution

Keywords:

measurement method supported by the NEWFM. These 4 minimum features were used as

Epileptic seizure

inputs for the NEWFM and this resulted in performance sensitivity, specificity, and accuracy

Feature selection

of 96.33%, 100%, and 98.17%, respectively. In addition, the area under Receiver Operating

Phase space reconstruction

Characteristic (ROC) curve was used to measure the performances of NEWFM both without

Euclidean distance

and with feature selections.

ROC curve

1.

Introduction

Human brain disorders are responsible for many physiopathological diseases, especially epilepsy. Epilepsy is a chronic neurological disorder that is generally characterized by sudden and recurrent seizures [10,13,20]. Until recently, the occurrence of an epileptic seizure was unpredictable and its mode of action was little understood [9]. The electroencephalogram (EEG) is a highly complex signal and it provides one of the most commonly used information

© 2014 Elsevier Ireland Ltd. All rights reserved.

sources for studying brain function and neurological disorders [2]. Study of the brain’s electrical activity, using electroencephalographic records, is one of the most important steps in the diagnosis of neurological diseases [1,14,31,32,41]. EEG signal processing uses nonlinear methods, such as the Lyapunov exponent, and the correlation dimension is used to measure the complexity or degree of the EEG signal disorder [8,13,23]. Apart from EEG, Empirical mode decomposition (EMD) was developed specifically for nonlinear and nonstationary signal analysis [36]. Time and frequency domains are commonly used in feature extraction methods for the EEG

∗

Corresponding author. Tel.: +82 328204100. E-mail addresses: [email protected], [email protected] (S.-H. Lee), [email protected] (J.S. Lim), [email protected] (J.-K. Kim), [email protected] (J. Yang), [email protected] (Y. Lee). http://dx.doi.org/10.1016/j.cmpb.2014.04.012 0169-2607/© 2014 Elsevier Ireland Ltd. All rights reserved.

Please cite this article in press as: S.-H. Lee, et al., Classification of normal and epileptic seizure EEG signals using wavelet transform, phase-space reconstruction, and Euclidean distance, Comput. Methods Programs Biomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.04.012


2


signals and cross-correlations of EEG time series have been used for high performance classification of the time domain features [7,12,15,45,51]. Past research shows that, to analyze the frequency domain features, discrete Fourier transform (DFT) was applied first followed by the calculation of power spectral density (PSD) of the EEG signals [15,34]. Grammatical evolution (GE) [35,44], eigenvector methods [12,49,50], and the Fourier transform [39] are also other reported approaches for extracting features when classifying epileptic signals from the EEG recordings. EEG signals have been decomposed into time–frequency representations using discrete wavelet transform (DWT) [5,11]. First, EEG signals are decomposed into several sub-signals by DWT, and then wavelet coefficients of the several sub-signals are produced. The wavelet coefficients were used as inputs of Adaptive neuro-fuzzy inference system (ANFIS) [11]. Genetic programming is applied to extract new features, which are nonlinearly transformation from the wavelet coefficients [9]. Wavelet entropy was obtained from the wavelet coefficients on the several sub-signals, and then the wavelet entropy was used as features [20]. Statistical methods were also used to extract features from the wavelet coefficients [46]. Certain features based on DWT were obtained and applied as different classifiers during epileptic EEG classification, including methods, such as adaptive neuro-fuzzy inference [11], relative wavelet energy (RWE) with artificial neural network (ANN) method [26], multistage approach for clustering [16], new phase space-based approach to automatically classify sleepwake cycle [4], mixture of expert model [46] and support vector machine (SVM) [7]. In the current study, a neural network with weighted fuzzy membership functions (NEWFM) [21,22,24] that is known to deliver excellent performance in prediction and classification was used to classify the normal and the epileptic seizure signals from the EEG signals and 3 preprocessing steps were implemented for this purpose. The 3 preprocessing steps, applied sequentially, comprised of wavelet transform (WT), phase-space reconstruction (PSR), and Euclidean distance (ED). NEWFM and non-overlap area distribution measurement method were used for classification and feature selection. In this regard, the area under the Receiver Operating Characteristic (ROC) curve (AUC) was used to compare the performances of NEWFM without feature selection and NEWFM with feature selection. The remainder of this paper is organized as follows. Section 2 describes the related works. Section 3 describes EEG datasets used in this paper. Section 4 describes feature extraction and selection method. Section 5 provides experimental result. Discussion and conclusion have been included in Section 6.

2.

Related works

Until now, several techniques have been proposed for the classification of normal and epileptic seizure EEG signals in the literature and EEG classification accuracy have been reported. A description of the previous research follows. Güler and Übeyli [11] proposed the ANFIS to classify EEG signals using wavelet coefficients. The presented ANFIS model combined the neural network adaptive capabilities and the

fuzzy logic qualitative approach. Guo et al. [9] applied genetic programming (GP) to perform automatic feature extraction since GP can improve the discrimination performance of K-nearest neighbor (KNN) classifier and decrease the input feature dimension. Kumar et al. [20] reported entropy based detection of neurological disorders, where three entropies, called as wavelets, were used as feature selection parameters and the neurological disorders were classified using two neural network models, viz. recurrent Elman network and radial basis network. Subasi [46] proposed an ANN based expert model, with a double-loop Expectation-Maximization (EM) algorithm, for the detection of epileptic seizures. This study employed DWT for feature extraction on the same two sets of the epileptic EEG signals. Ling et al. [26] proposed DWT, RWE along with ANN to classify the EEG signals. The EEG signals were first decomposed into different frequency bands with DWT. The RWE provided information about the relative energies associated with the different frequency bands of the EEG signals. A feed-forward ANN was finally used for classification. Abawajy et al. [16] proposed a novel multistage approach for classifying large and highly dimensional dataset. The multistage approach combined dimensionality reduction algorithms, multiple unsupervised clustering algorithms and several supervised classification algorithms. Arnaud et al. [4] proposed a new phase space-based approach to automatically classify sleep–wake cycles in humans, using only two EEG electrode positions. However, this study did not classify each sleep stage using phase space analysis (mainly on Poincare plot), which is generally applied in non-linear and relatively short time series analysis. Chandaka et al. [7] utilized a crosscorrelation aided SVM classifier for classifying EEG signals of healthy subjects with eyes open and epileptic patients during seizure with the epileptic EEG signals. This study involved least square support vector machine (LS-SVM) and feature extraction using a clustering technique. In the literature, numerous techniques have been used to obtain representations and extract features of interest for the classification purposes. Until now, there is no study that is related to the phase-space reconstruction and ED approach for feature extraction of the EEG signals. In this paper, we propose a new approach that combines the wavelet transform, phasespace reconstruction, and ED comprising of feature selection based neural network, with weighted fuzzy membership function (NEWFM), to classify the normal and the epileptic seizure EEG signals.

3.

Electroencephalogram (EEG) datasets

This study used EEG signal datasets (http://epileptologie-bonn .de/cms/front content.php?idcat=193&lang=3&changelang=3) to classify normal and epileptic seizure signals [3,46]. The experimental dataset was first divided into five sets denoted as A, B, C, D, and E, each containing 100 single-channel EEG signals and each signal is of 23.6 s in duration [3,46]. These segments were selected and cut out from continuous multichannel EEG recordings after visual inspection for artifacts, e.g., due to muscle activity or eye movements. Sets A and B consisted of segments taken from surface EEG recordings



ARTICLE IN PRESS 3

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e x x x ( 2 0 1 4 ) xxx–xxx

Table 1 – Numbers of training and test sets. Class

Training set

Epileptic Normal Total

500 500 1000

Test set

Total set

300 300 600

800 800 1600

that were carried out on five healthy volunteers using a standardized electrode placement scheme (according to the international 10–20 system) [39]. Sets A and B have been acquired using surface EEG recordings of five healthy volunteers with eyes open and closed, respectively. Sets C, D, and E originated from the EEG archive of presurgical diagnosis. The EEGs from five patients were selected, all of whom had achieved complete seizure control after resection of one of the hippocampal formations, which was therefore, correctly diagnosed to be the epileptogenic zone. Segments in set D were recorded from the epileptogenic zone, and those in set C from the hippocampal formation of the opposite hemisphere of the brain. While sets C and D contained activities measured during seizure free intervals, set E contained only seizure activities. All EEG signals were recorded with the same 128-channel amplifier system, using an average common reference. The data were digitized at 173.61 samples per second using 12 bit resolution. Band-pass filter settings were 0.53–40 Hz (12 dB/oct). Furthermore, in this study, sets A and E have been used for comparison with Chandaka’s performance [7], Ling’s performance [26] and Subasi’s performance [46]. In a previously report by Kumar, performances of sets A and E were compared with those of sets B and E, and finally it was concluded that A and E are better in performance than B and E [20]. Sets A and E, depicted in Fig. 1, were used to classify the normal and epileptic seizure signals. These sets contained 100 files and each file contained 4097 successive EEG signals. The 4097 successive EEG signals were further divided into eight sets, each with 512 signals (4096 total) and the remaining one EEG signal was deleted. Thus, 800 sets of 512 signals were created from the 100 files. The 800 sets were then mixed at random. As shown in Table 1, the first 500 sets were used as training sets, while the remaining 300 sets were used as test sets, as done previously by Chandaka, Ling, and Subasi.

4.

Methods

4.1. Overview of the classification of normal and epileptic seizure EEG signals Fig. 2 shows the process for classifying normal and epileptic seizure signals. The proposed model consists of two parts. The first part is preprocessing part for feature extraction and second part deals with classification and feature selection using NWFM. The methods involved in the first part are as follows: as a first step, wavelet coefficients are created from the EEG signals using WT (Fig. 2). In the second step, wavelet coefficients are plotted in a two-dimensional phase space diagram using PSR. In the third step, the ED is calculated between the wavelet coefficients plotted in a two-dimensional phase space diagram and the origin (0, 0), and then initial features were extracted from the EDs using statistical methods. In the second part, the

Table 2 – Frequencies corresponding to different levels of wavelet decomposition with a 173.6 Hz sampling rate. Decomposed signal D1 D2 D3 D4 D5 A5

Frequency range (Hz) 43.4–86.8 21.7–43.4 10.85–21.7 5.425–10.85 2.7125–5.425 0–2.7125

minimum features with the highest accuracy were selected using a non-overlap area distribution measurement method implemented with the NEWFM. The minimum features were used as weighted fuzzy membership functions that preserved the disjunctive fuzzy information and characteristics [21,24].

4.2.

Preprocessing for feature extraction

4.2.1.

Wavelet transform

The Fourier transform [28] is frequently used for signal processing to analyze signals within certain frequency domains. However, it is not suitable for processing any time-frequency information since they cannot be analyzed simultaneously. In order to overcome this disadvantage, a discrete wavelet transform (DWT) was recently utilized to facilitate efficient time-frequency analysis. The DWT decomposes non-station signals at different frequency intervals with various resolutions [5]. The original signal is decomposed into a set of coefficients describing the frequency content at given times. The DWT can be defined as follows [30]: S2i x(n) =

hk S2i−1 x(n − 2i−1 k)

(1)

k∈Z

W2i x(n) =

gk S2i−1 x(n − 2i−1 k)

(2)

k∈Z

where S2i is a smoothing operator, W2i is the digital signal x(n), i ∈ Z is the integral set, and hk and gk are coefficients for the corresponding low-pass and high-pass filters. A filtered signal at level i is down-sampled, thereby reducing the length of the signal at level i − 1 by a factor of two and generating the detail (di ) and approximation coefficients (ai ) at level i. The primary advantage of DWT is that, a short or a long time-interval either can be used to observe the characteristics of the signals. For a signal with rapid variation, a small window can be applied, whereas, for a signal with slow variation, a larger window can be used [38]. Additionally, both the approximation coefficients and the detail coefficients can be calculated at the same time [29]. Selection of suitable wavelet and Decision of the number of decomposition levels are very important in analysis of signals using the DWT. The number of decomposition levels is chosen based on the dominant frequency components of the signal [46]. In this study, we produced wavelet coefficients, including detail and approximation coefficients at levels 1–5, using Daubechies 4 (DB4), which is the most frequently used method for EEG signal analysis with WTs [1,18,40,52]. Table 2 shows the frequencies corresponding to different levels of wavelet



4


Fig. 1 – Examples of sets A and E of EEG signals.

Fig. 2 – Diagram of the model proposed for classifying normal and epileptic seizure EEG signals.


ARTICLE IN PRESS



5

Fig. 3 – Comparisons of the features extracted from decomposed signals.

decomposition for the EEG signals with a 173.6 Hz sampling rate. The extracted features, shown in Table 3 were used to represent the frequency distribution of the wavelet coefficients of the EEG signals. These features were calculated for the decomposed signals, shown in Table 2. Fig. 3 shows a comparison of the features of Table 3 on the sets A and E. It was observed that, these features extracted from the decomposed EEG signals for the sets A and E are different from each other. Therefore, we decided that wavelet coefficients can be served as useful parameters in classifying normal and epileptic seizure EEG signals.

4.2.2.

Phase space reconstruction

It is sometimes necessary to search for patterns in a time series and in a higher dimensional transformation of the time series. Phase space reconstruction (PSR) is a method used to reconstruct the so-called phase space. The concept of phase space is a useful tool for characterizing any low-dimensional

Table 3 – Description of the extracted features from decomposed signals. No 1 2 3 4

Extracted features Maximum of the wavelet coefficients in each sub-band. Minimum of the wavelet coefficients in each sub-band. Mean of the wavelet coefficients in each sub-band. Standard deviation of the wavelet coefficients in each sub-band.

or high-dimensional dynamic system. A dynamic system can be described using a phase space diagram, which essentially provides a coordinate system where the coordinates are all the variables comprising mathematical formulation of the system. A point in the phase space represents the state of the system at any given time [42,43]. For a time series Xi , where i = 1, 2, . . ., N, the phase space can be reconstructed according to: Yj = (Xj , Xj+ , Xj+2 , . . ., Xj+(m−1) )

(3)

where j = 1, 2, . . ., N − (m − 1), m is the dimension of the phase space, and is the delay time [37,48]. In this study, we used a wavelet detail coefficient Xi at levels 1–5 and a wavelet approximation coefficient Xi at level 5, as shown in Table 2 where the x-axis plots Xi and the y-axis plots Xi+1 . This plot is known as a two-dimensional phase space diagram. Fig. 4 shows the wavelet coefficients plotted in a twodimensional phase space diagram. PSRs for the wavelet coefficients of normal and epileptic seizure EEG signals are shown in Fig. 4. These figures show that the patterns related to the higher dimensional transformations can be more discriminate than those in the time series itself. Fig. 4 also shows that PSR’s points of normal EEG signals accumulate more around the origin than those of the epileptic seizure EEG signals, and therefore, the normal signals are in the phase portraits and are more regular than the epileptic seizure EEG ones.



6


Fig. 4 – Wavelet coefficients plotted by PSR.



ARTICLE IN PRESS


7

Fig. 4 – (Continued ).



8


Fig. 5 – Comparisons of the features extracted from the Euclidean distance and a statistical method.

4.2.3. Feature extraction using Euclidean distance and a statistical method Euclidean distance measures can be used in decision-making. The Euclidean distance has been studied and applied in many fields, such as clustering algorithms and induced aggregation operators [19,33,47]. The Euclidean distance is the distance between two points in an Euclidean space. The Euclidean distance between the origin (0, 0) and PSR (Xi , Xi+1 ) is calculated according to:

4 represents the amount of variation in the frequency distributions [17,46]. Fig. 5 compares the features in Table 4 on the sets A and E. From Fig. 5, the Euclidean distances calculated from the PSR for the wavelet coefficients in the sets A and E are different from each other. This implies that the Euclidean distances calculated from the PSR can also serve as useful parameters in classifying the normal and the epileptic seizure EEG signals, similarly as Fig. 3.

E(i) =

2

X(i) + X(i + 1)

2

(4)

Based on the Euclidean distance values E(i) in Eq. (4), 24 features were used as initial inputs for NEWFM and these were extracted using the statistical methods described in Table 4. Features 1, 2, and 3 in Table 4 represent the frequency distributions of Euclidean distance E(i), while feature

Table 4 – Description of the features extracted from the Euclidean distances. No 1 2 3 4

Feature extracted Mean of the Euclidean distances in each sub-band. Median of the Euclidean distances in each sub-band. Average power of the Euclidean distances in each sub-band. Standard deviation of the Euclidean distances in each sub-band.

4.3. Neural network with weighted fuzzy membership function (NEWFM) for classification and feature selection 4.3.1. Neural network with weighted fuzzy membership function (NEWFM) This study used a NEWFM for classifying normal and epileptic seizure EEG signals. The NEWFM is a supervised classification neuro-fuzzy system that uses the bounded sum of weighted fuzzy membership functions (BSWFM) [21,22,24]. Fig. 6 shows that the NEWFM contained three layers, i.e., the input, hyperbox, and class layers. The 24 features extracted using the feature extraction method explained in Table 4 were used as inputs for the NEWFM, as shown in Fig. 6. As shown in Fig. 7, the weights and the center of each membership functions are adjusted by the Adjust(Bl ) operation in Algorithm 1, e.g., W1 , W2 , and W3 are moved down, v1 and v2 are moved toward ai , and v3 remains in the same location



ARTICLE IN PRESS 9


Fig. 6 – Structure of the NEWFM.

[25]. After learning, each of n fuzzy sets in hyperbox node Bl contains three weighted fuzzy membership functions (WFMs, gray membership functions in Fig. 8). The bounded sum (one of operations on fuzzy set) of WFMs (BSWFM) in the ith fuzzy set of Bil (x), denoted as ib (x) (bold line in Fig. 8), is defined by:

ib (x) =

3

Bil (j (x)).

(5)

j=1

Fig. 7 – Example of before and after Adjust(Bl ) operation.

Fig. 8 – Example of bounded sum of the 3 weighted fuzzy membership functions (BSWFM, bold line).



10


Algorithm 1. Adjust(Bl ) operation algorithm.


ARTICLE IN PRESS



4.3.2.

Feature selection

It has been found in recent years that a large number of features available in many small devices (say, smart phones) are the primary cause of slow operation speed, function failure and inefficiency of the device memory, which makes the device unusable. Irrelevant data can cause confusion in algorithms and lead to false conclusions, thereby producing poor results. The primary aim of feature selection is to eliminate redundant or irrelevant features and enhance the performance of a device with high accuracy and lower the operating costs, using minimum features. To select the best features in this study, the 4 minimum features with the highest accuracy were selected from 24 initial features using the non-overlap area distribution measurement method implemented with the NEWFM. This method measures the degree of salience of the ith feature based on the non-overlapped areas in the area distribution using the following equation [21,24]: f (i) =

(AreaiE + AreaiN ) i

2 i

1/(1 + e−|AreaE −AreaN | )

(6)

11

where AreaE and AreaN are the epileptic seizure superior area and the normal superior area, respectively. Higher f(i) values indicate better feature characteristics. Fig. 9 shows examples of both good and bad candidate features selected from the 24 initial features. Specifically, the figure shows the features of two BSWFMs, obtained during the training process with the NEWFM program. The two BSWFMs graphically demonstrate the difference between a normal and an epileptic seizure signal for each input feature. The features shown in Fig. 9(b) are in a state where the two BSWFMs almost overlap with each other and differ from the features shown in Fig. 9(a). In this regard, Fig. 9(b) features appears ambiguous because they cannot be easily divided into normal and epileptic seizure signals. The f(i) values are obtained using Eq. (6) and these values are greater for the features in Fig. 9(a), √ compared to those in Fig. 9(b). In Table 5, “ ” indicates features Algorithm 2 finally selected from the 24 initial features.

Fig. 9 – Examples of good and bad candidate features during feature selection. Please cite this article in press as: S.-H. Lee, et al., Classification of normal and epileptic seizure EEG signals using wavelet transform, phase-space reconstruction, and Euclidean distance, Comput. Methods Programs Biomed. (2014), http://dx.doi.org/10.1016/j.cmpb.2014.04.012


12


Algorithm 2. Non-overlap area distribution measurement algorithm.


ARTICLE IN PRESS


13


Table 5 – Description of the features selected. Euclidean distance

No in Table 4

Euclidean distance at D1

1 2 3 4


1 2 3 4


1 2 3 4

Selected

√

Euclidean distance

1 2 3 4


1 2 3 4

Euclidean distance at A5

1 2 3 4

√ √

No in Table 4


Selected

√

5.

Experimental results

Table 6 – Confusion matrix of classification results for NEWFM without feature selection.

5.1.

Statistical tests

Epileptic seizure EEG signals Normal EEG signals

In the current study, TP (True Positive) refers to cases where epileptic seizure signals were classified as epileptic seizure signals and FN (False Negative) refers to cases where epileptic seizure signals were classified as normal signals. FP (False Positive) refers to cases where normal signals were classified as epileptic seizure signals and TN (True Negative) refers to cases where normal signals were classified as normal signals. The sensitivity, specificity, and accuracy are defined in the following equations (7):

Specificity = Accuracy =

TN × 100 TN + FP

(7)

TP + TN × 100 TP + FN + TN + FP

We employed the ROC curve to compare the classification performances of NEWFM both without and with feature selections. The ROC curve is a graphical plot of true positive fraction with sensitivity as the y-coordinate and false positive fraction with specificity as the x-coordinate. The ROC can also be represented equivalently by plotting the fraction of true positives out of the positives (True Positive Rate) and the fraction of false positives out of the negatives (False Positive Rate) [6,27]. The area under the ROC (AUC) can be used to measure the performances of NEWFM both without and with feature selections, based on the principle – greater the area, higher is the accuracy.

FN 15 TN 295

Table 7 – Confusion matrix of classification results for NEWFM with feature selection. Epileptic seizure EEG signals Normal EEG signals

5.2. TP Sensitivity = × 100 TP + FN

TP 285 FP 5

TP 289 FP 0

FN 11 TN 300

Results

In this study, normal and epileptic seizure signals were classified from the EEG signals following Chandaka et al. [7], Ling et al. [26], and Subasi [46]. Fig. 9(a) and (b) shows the BSWFMs for two feature inputs out of the 24 feature inputs. Differences between the normal and epileptic seizure signals for the 24 feature inputs can be visualized and related characteristics can be analyzed. Classification performances are shown in Tables 6 and 7. The NEWFM performance was compared with the support vector machine, RWE with ANN, and the mixture of expert model, which were used by Chandaka, Ling and Subasi as classification models, respectively. Table 8 shows that NEWFM with feature selection was compared with the NEWFM without feature selection, support vector machine, RWE with ANN, and the mixture of expert model. NEWFM with feature selection method outperforms other methods. But, Ling’s Sensitivity [26] is higher than the current proposed method. To this end, further study is needed to increase the sensitivity. As the AUC

Table 8 – Comparisons of performance results (%) for support vector machine, RWE with ANN, mixture of expert, and NEWFM for classification of normal and epileptic seizure EEG signals. Accuracy Support vector machine [7] RWE with ANN [26] Mixture of expert [46] NEWFM without feature selection NEWFM with feature selection

95.5 95.2 94.5 96.66 98.17

Specificity 98.6 92.12 94 98.33 100

Sensitivity 92.4 98.17 95 95 96.33



14


Fig. 10 – ROC curves for comparisons of performance results with feature selection and without feature selection.

of the NEWFM with feature selection is wider than that of the NEWFM without feature selection in Fig. 10, NEWFM with feature selection performed better than NEWFM without any feature selection, in classifying normal and epileptic seizure EEG signals.

6.

Discussion and conclusion

This study used the WT, the PSR, and the Euclidean distance in the preprocessing steps to extract useful features from EEG signals. First of all, we produced wavelet coefficients from EEG signals using the WT, and then the PSR was used to plot wavelet coefficients in a two-dimensional phase space diagram. Next, the Euclidean distances were calculated between wavelet coefficients plotted by the PSR and the origin (0, 0). Finally, 24 features were extracted from the Euclidean distances using statistical methods. 4 minimum features with the highest accuracy were selected from the 24 initial features by the NEWFM. The NEWFM obtained BSWFMs for the 4 minimum features to identify the fuzzy membership functions for the 4 features. These fuzzy membership functions were used to classify the normal and the epileptic seizure signals from the EEG signals. Using the 4 minimum features as inputs for the NEWFM, 96.33% sensitivity, 100% specificity, and 98.17% accuracy were obtained. These values represented a much better performance when compared to those of Chandaka, Ling or Subasi. Furthermore, when the ROC curve was tested, it was observed that the AUC of the NEWFM with feature selection is wider than the NEWFM without feature selection, which confirms that the former is better and outperforms the latter. To conclude it can be stated that NEWFM is a useful technique to represent the characteristic of EEG signals. Improvement in feature selection method in view to enhancing the sensitivity is planned as the future work.

Acknowledgements This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (NRF-2013M3C8A2A02078403). This work was supported by the Industrial Strategic Technology Development Program, 10037283, funded by the Ministry of Trade, Industry & Energy (MI, Korea).

references

[1] H. Adeli, Z. Zhou, N. Dadmehr, Analysis of EEG records in an epileptic patient using wavelet transform, J. Neurosci. Methods 123 (2003) 69–87. [2] R. Agarwal, J. Gotman, D. Flanagan, B. Rosenblatt, Automatic EEG analysis during long-term monitoring in the ICU, Electroencephalogr. Clin. Neurophysiol. 107 (1998) 44–58. [3] R.G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, C.E. Elger, Indications of nonlinear deterministic and finite dimensional structures in time series of brain electrical activity: dependence on recording region and brain state, Phys. Rev. E 64 (2001) 061907. [4] B. Arnaud, A. Tarik, D. Xavier, Phase space and power spectral approaches for EEG-based automatic sleep–wake classification in humans: a comparative study using short and standard epoch lengths, Comput. Methods Programs Biomed. 109 (2013) 227–238. [5] E. Avci, D. Hanbay, A. Varol, An expert discrete wavelet adaptive network based fuzzy inference system for digital modulation recognition, Expert Syst. Appl. 33 (2007) 582–589. [6] R.M. Centor, Signal detectability: the use of ROC curves and their analysis, Med. Decis. Making 11 (1991) 102–106. [7] S. Chandaka, A. Chatterjee, S. Munshi, Cross-correlation aided support vector machine classifier for classification of EEG signals, Expert Syst. Appl. 36 (2009) 1329–1336. [8] F. Shayegha, S. Sadria, R. Amirfattahia, K. Ansari-Aslb., A model-based method for computation ofcorrelation dimension, Lyapunov exponents andsynchronization from



ARTICLE IN PRESS


[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

depth-EEG signals, Comput. Methods Programs Biomed. 113 (2014) 323–337. L. Guo, D. Rivero, J. Dorado, C.R. Munteanu, A. Pazos, Automatic feature extraction using genetic programming: an application to epileptic EEG classification, Expert Syst. Appl. 38 (2011) 10425–10436. T. Gülay, Y. Özbay, A new approach for epileptic seizure detection using adaptive neural network, Expert Syst. Appl. 36 (2009) 172–180. I˙. Güler, E.D. Übeyli, Adaptive neuro-fuzzy inference system for classification of EEG signals using wavelet coefficients, J. Neurosci. Methods 148 (2005) 113–121. I˙. Güler, E.D. Übeyl˙i, Expert systems for time-varying biomedical signals using eigenvector methods, Expert Syst. Appl. 38 (2011) 9991–9999. N.H. Güler, E.K. Übeyl˙i, I˙. Güler, Recurrent neural networks employing Lyapunov exponents for EEG signal classification, Expert Syst. Appl. 25 (2005) 506–514. N. Hazarika, J.Z. Chen, A.C. Tsoi, A. Sergejew, Classification of EEG signals using the wavelet transform, Signal Process. 59 (1997) 61–72. Z. Iscan, Z. Dokur, T. Demiralp, Classification of electroencephalogram signals with combined time and frequency features, Expert Syst. Appl. 38 (2011) 10499–10505. J.H. Abawajy, A.V. Kelarev, M. Chowdhury, Multistage approach for clustering and classification of ECG data, Comput. Methods Programs Biomed. 112 (2013) 720– 730. A. Kandaswamy, K.C. Sathish, R.P. Ramanathan, S. Jayaraman, N. Malmurugan, Neural classification of lung sounds using wavelet coefficients, Comput. Biol. Med. 34 (2004) 523–537. S. Kar, M. Bhagat, A. Routray, EEG signal analysis for the assessment and quantification of driver’s fatigue, Transport. Res. F 13 (2010) 297–306. N. Karayiannis, Soft learning vector quantization and clustering algorithms based on ordered weighted aggregation operators, IEEE Trans. Neural Networks 11 (2000) 1093–1105. S.P. Kumar, N. Sriraam, P.G. Benakop, B.C. Jinaga, Entropies based detection of epileptic seizures with artificial neural network classifiers, Expert Syst. Appl. 37 (2010) 3284–3291. S.H. Lee, J.S. Lim, Forecasting KOSPI based on a neural network with weighted fuzzy membership functions, Expert Syst. Appl. 38 (2011) 4259–4263. S.H. Lee, J.S. Lim, Parkinson’s disease classification using gait characteristics and wavelet-based feature extraction, Expert Syst. Appl. 39 (2012) 7338–7344. K. Lehnertz, C.E. Elger, Spatio-temporal dynamics of the primary epileptogenic area in temporal lobe epilepsy characterized by neuronal complexity loss, Electroencephalogr. Clin. Neurophysiol. 95 (1995) 108–117. J.S. Lim, Finding Features for Real-Time Premature Ventricular Contraction Detection Using a Fuzzy Neural Network System, IEEE Trans. Neural Networks 20 (2009) 522–527. J.S. Lim, T.W. Ryu, H.J. Kim, S. Gupta, Feature Selection for Specific Antibody Deficiency Syndrome by Neural Network with Weighted Fuzzy Membership Functions, FSKD 2005 (LNCS 3614), 2005, pp. 811–820. G. Ling, R. Daniel, A.S. Jose, P. Alejandro, Classification of EEG Signals Using Relative Wavelet Energy and Artificial Neural Networks, GEC, 2009, June, pp. 177–183. C.J. Lloyd, The use of smoothed ROC curves to summarize and compare diagnostic systems, J. Am. Stat. Assoc. 93 (1998) 1356–1364. R. Lyons, Understanding Digital Signal Processing, Prentice Hall, second ed., 2004, pp. 49–156.

15

[29] S. Mallat, A theory for multiresolution signal decomposition: the wavelet representation, IEEE Trans. Pattern Anal. 11 (1989) 674–693. [30] S. Mallat, Zero crossings of a wavelet transform, IEEE Trans. Inf. Theory 37 (4) (1991) 1019–1033. [31] P. Matthew, T. Manolis, L. Norbert, Vision-based motion detection, analysis and recognition of epileptic seizures—a systematic review, Comput. Methods Programs Biomed. 108 (2012) 1133–1148. [32] I. Matthias, F. Hinnerk, A.T. César, W. Adrien, S. Björn, T. Jens, S. Andreas, EPILEPSIAE – a European epilepsy database, Comput. Methods Programs Biomed. 106 (2012) 127–138. [33] J.M. Merigó, M. Casanovas, Induced aggregation operators in the Euclidean distance and its application in financial decision making, Expert Syst. Appl. 38 (2011) 7603–7608. [34] T. Nguyen-Ky, P. Wen, Y. Li, Theoretical basis for identification of different anesthetic states based on routinely recorded EEG during operation, Comput. Biol. Med. 39 (2009) 40–45. [35] M. O’Neill, C. Ryan, Grammatical evolution. Evolutionary computation, IEEE Trans. Bio-med. Eng. 5 (2001) 349–358. [36] R.B. Pachori, Discrimination between ictal and seizure-free EEG signals using empirical mode decomposition, Res. Lett. Signal Process 5 (2008), Article ID 293056. [37] N.H. Packard, J.P. Crutchfield, J.D. Farmer, R.S. Shaw, Geometry from a time series, Phys. Rev. Lett. 45 (1980) 712–716. [38] S.Y. Pana, B.Z. Hsieha, M.T. Lub, Z.S. Lina, Identification of stratigraphic formation interfaces using wavelet and Fourier transforms, Comput. Geosci.-UK 34 (2008) 77–92. [39] K. Polat, S. Günes¸, Classification of epileptiform EEG using a hybrid system based on decision tree classifier and fast Fourier transform, Appl. Math. Comput. 187 (2007) 1017–1026. [40] J. Rafiee, M.A. Rafiee, N. Prause, M.P. Schoen, Wavelet basis functions in biomedical signal processing, Expert Syst. Appl. 38 (2011) 6190–6201. [41] O.A. Rosso, A. Figliola, J. Creso, E. Serrano, Analysis of wavelet-filtered tonic-clonic electroencephalogram recordings, Med. Biol. Eng. Comput. 42 (2004) 516–523. [42] B. Sivakumar, A phase-space reconstruction approach to prediction of suspended sediment concentration in rivers, J. Hydrol. 258 (2002) 149–162. [43] B. Sivakumar, A.W. Jayawardena, T.M.K.G. Fernando, River flow forecasting: use of phase-space reconstruction and artificial neural networks approaches, J. Hydrol. 265 (2002) 225–245. [44] O. Smart, I.G. Tsoulos, D. Gavrilis, G. Georgoulas, Grammatical evolution for features of epileptic oscillations in clinical intracranial electroencephalograms, Expert Syst. Appl. 38 (2011) 9991–9999. [45] F. Stephen, M. William, Dynamic, location-based channel selection for power consumption reduction in EEG analysis, Comput. Methods Programs Biomed. 108 (2012) 1206–1215. [46] A. Subasi, EEG signal classification using wavelet feature extraction and a mixture of expert model, Expert Syst. Appl. 32 (2007) 1084–1093. [47] E. Szmidt, J. Kacprzyk, Distances between intuitionistic fuzzy sets, Fuzzy Set Syst. 114 (2000) 505–518. [48] F. Takens, Detecting strange attractors in turbulence, in: Dynamical systems and turbulence, Lect. Notes Math. 898 (1981) 366–381.



16


[49] E.D. Übeyli, Analysis of EEG signals by combining eigenvector methods and multiclass support vector machines, Comput. Biol. Med. 38 (2008) 14–22. [50] E.D. Übeyli, Decision support systems for time-varying biomedical signals: EEG signals classification, Expert Syst. Appl. 36 (2009) 2275–2284.

[51] B. Varun, B.P. Ram, Automatic classification of sleep stages based on the time-frequency image of EEG signals, Comput. Methods Programs Biomed. 112 (2013) 320–328. [52] Q. Xu, H. Zhou, Y. Wang, J. Huang, Fuzzy support vector machine for classification of EEG signals using wavelet-based features, Med. Eng. Phys. 31 (2009) 858–865.


Epileptic seizure classification of EEG time-series using rational discrete short-time fourier transform.

Epileptic seizure detection from EEG signals using logistic model trees.

Detection of epileptic seizure event and onset using EEG.

A high-performance seizure detection algorithm based on Discrete Wavelet Transform (DWT) and EEG.

Classification of epileptic EEG signals based on simple random sampling and sequential feature selection.

Time-frequency texture descriptors of EEG signals for efficient detection of epileptic seizure.

Frequency Band Analysis of Electrocardiogram (ECG) Signals for Human Emotional State Classification Using Discrete Wavelet Transform (DWT).

Detection of motor imagery of swallow EEG signals based on the dual-tree complex wavelet transform and adaptive model selection.

Decoding objects of basic categories from electroencephalographic signals using wavelet transform and support vector machines.

Exploration of EEG features of Alzheimer's disease using continuous wavelet transform.

Fast and Accurate Calculation of Protein Depth by Euclidean Distance Transform.

Mobile healthcare for automatic driving sleep-onset detection using wavelet-based EEG and respiration signals.

Automatic seizure detection using Stockwell transform and boosting algorithm for long-term EEG.

Quantification of effective attenuation coefficients using continuous wavelet transform of photoacoustic signals.

Automatic epileptic seizure detection using scalp EEG and advanced artificial intelligence techniques.

Application of wavelet transform for PDZ domain classification.

The detection of epileptic seizure signals based on fuzzy entropy.

Automatic seizure detection using diffusion distance and BLDA in intracranial EEG.

Image reconstruction of compressed sensing MRI using graph-based redundant wavelet transform.

Ensemble classifier for epileptic seizure detection for imperfect EEG data.

Prediction of Epileptic Seizure by Analysing Time Series EEG Signal Using k-NN Classifier.

Classification of 5-S Epileptic EEG Recordings Using Distribution Entropy and Sample Entropy.

Fetal Heart Sounds Detection Using Wavelet Transform and Fractal Dimension.

Energy efficient acquisition and reconstruction of EEG signals.