Epileptic seizure detection using multiwavelet transform based approximate entropy and artificial neural networks

Page 1

Ee

LD

a

ARRA

KEEMAA

1

afrcsopcrbd

iuEeicd

0d

Journal of Neuroscience Methods 193 (2010) 156–163

Contents lists available at ScienceDirect

Journal of Neuroscience Methods

journa l homepage: www.e lsev ier .com/ locate / jneumeth

pileptic seizure detection using multiwavelet transform based approximatentropy and artificial neural networks

ing Guo ∗, Daniel Rivero, Alejandro Pazosepartment of Information Technologies and Communications, University of La Coruna, Campus Elvina, 15071 A Coruna, Spain

r t i c l e i n f o

rticle history:eceived 3 July 2010eceived in revised form 23 August 2010ccepted 24 August 2010

eywords:lectroencephalogram (EEG)pileptic seizure detection

a b s t r a c t

Epilepsy is the most prevalent neurological disorder in humans after stroke. Recurrent seizure is the maincharacteristic of the epilepsy. Electroencephalogram (EEG) is the recording of brain electrical activity andit contains valuable information related to the different physiological states of the brain. Thus, EEG is con-sidered an indispensable tool for diagnosing epilepsy in clinic applications. Since epileptic seizures occurirregularly and unpredictably, automatic seizure detection in EEG recordings is highly required. Multi-wavelets, which contain several scaling and wavelet functions, offer orthogonality, symmetry and shortsupport simultaneously, which is not possible for scalar wavelet. With these properties, recently multi-

ultiwavelet transform (MWT)pproximate entropy (ApEn)rtificial neural network (ANN)

wavelets have become promising in signal processing applications. Approximate entropy is a measurethat quantifies the complexity or irregularity of the signal. This paper presents a novel method for auto-matic epileptic seizure detection, which uses approximate entropy features derived from multiwavelettransform and combines with an artificial neural network to classify the EEG signals regarding the exis-tence or absence of seizure. To the best knowledge of the authors, there exists no similar work in theliterature. A well-known public dataset was used to evaluate the proposed method. The high accuracy

t clas

obtained for two differen

. Introduction

Epilepsy is the most prevalent neurological disorder in humansfter stroke. About 40 or 50 million people in the world sufferrom epilepsy (Kandel et al., 2000). Epilepsy is characterized byecurrent seizure in which abnormal electrical activity in the brainauses altered perception or behavior. Patients experience variedymptoms during seizures depending on the location and extentf the affected brain tissue. Epileptic seizures may cause negativehysical, psychological and social consequences, including loss ofonsciousness, injuries and sudden death. Until now, the occur-ence of epileptic seizure is unpredictable and the mechanismsehind the seizure are little understood. Thus, efforts towards itsiagnosis and treatment are of great importance.

Electroencephalogram (EEG) is the recording of electrical activ-ty of the brain. EEG recordings contain valuable information fornderstanding epilepsy. The detection of seizures occurring in theEGs is an important component in the diagnosis and treatment of

pilepsy. However, usually massive amounts of data are includedn EEGs and visual inspection for discriminating EEGs is a timeonsuming and costly process. Thus, developing automatic seizureetection methods is of great significance for reviewing EEGs.

∗ Corresponding author. Tel.: +34 981 167000x1302; fax: +34 981 167160.E-mail address: [email protected] (L. Guo).

165-0270/$ – see front matter © 2010 Elsevier B.V. All rights reserved.oi:10.1016/j.jneumeth.2010.08.030

sification problems verified the success of the method.© 2010 Elsevier B.V. All rights reserved.

Research on seizure detection began in the 1970s and variousmethods addressing this problem have been presented. Mohseniet al. (2006) applied short time Fourier transform analysis of EEGsignals and extracted features based on the pseudo Wigner–Villeand the smoothed-pseudo Wigner–Ville distribution. Then thosefeatures were used as inputs to an artificial neural network(ANN) for classification. Kalayci and Ozdamar (1995) used wavelettransform to capture characteristic features of the EEG signals andthen combined with ANN to get satisfying classification result.Nigam and Graupe (2004) described a method for automateddetection of epileptic seizures from EEG signals using a multistagenon-linear preprocessing filter for extracting two features: relativespike amplitude and spike occurrence frequency. Then they fedthose features to a diagnostic artificial neural network. In thework of Jahankhani et al. (2006), the EEGs were decomposed withwavelet transform into different sub-bands and some statisticalinformation were extracted from the wavelet coefficients. Radialbasis function network (RBF) and multi-layer perceptron network(MLP) were utilized as classifiers. Subasi (2005, 2007) decomposedthe EEG signals into time–frequency representations using discretewavelet transform (DWT). Some features, such as the mean of the

absolute value, average power, standard deviation, and ratio of theabsolute mean value derived from the wavelet coefficients werecalculated and applied to different classifiers, such as feed-forwarderror back-propagation artificial neural network (FEBANN),dynamic wavelet network (DWN), dynamic fuzzy neural network

Page 2

cience

(stosnIfrrooutttsEtAc

ot2(coasat(oSdt

aAtae1dotm

cehnaois

ptseaitu

L. Guo et al. / Journal of Neuros

DFNN) and mixture of expert system (ME), for epileptic EEG clas-ification. In the work of Srinivasan et al. (2005), five features fromime-domain and frequency-domain were employed individuallyr jointly for classifying EEG signals. The high classification resultshowed that one feature combination with the Elman recurrenteural network exhibited excellent discrimination performance.

n Güler et al. (2005)’s work, Layapunov exponents were extractedrom EEGs with Jacobi matrices and then applied as inputs toecurrent neural networks (RNNs) to obtain good classificationesults. Übeyli (2006b) classified the EEG signals by combinationf Lyapunov exponents and fuzzy similarity index. Fuzzy sets werebtained from the feature sets (Lyapunov exponents) of the signalsnder study. The results demonstrated that the similarity betweenhe fuzzy sets of the studied signals indicated the variabilities inhe EEG signals. Thus, the fuzzy similarity index could discriminatehe different EEGs. In the research of Kannathal et al. (2005),everal entropy measures were investigated for discriminatingEG signals. The classification ability of the entropy measures wasested through an adaptive neuro-fuzzy inference system (ANFIS).bout 92% of accuracy was obtained for normal and epileptic EEGslassification.

Based on the literature review, there have been numerous worksn applying wavelet transform for epileptic EEG signal classifica-ion and recognition (Kalayci and Ozdamar, 1995; Jahankhani et al.,006; Subasi, 2006, 2007). Among those work, only scalar waveletswavelets generated by one scaling function) were used to pro-ess the EEG signals. Multiwavelets, which consist of more thanne scaling and wavelet functions, recently have attracted morettention because of their significant characteristics. Multiwaveletsimultaneously possess orthogonality, short support, symmetry,nd a high order of approximation through vanishing moments,hat all of them are important for signal processing applicationStrela, 1996). Multiwavelets have shown superior performancever scalar wavelets in image classification (Ramakrishnan andelvan, 2006), image compression (Cotronei et al., 2000) andenoising (Hsung et al., 2005). In this work, for the first time, mul-iwavelet is applied in epileptic seizure detection.

Choosing suitable features that can best represent the char-cteristics of the EEG signals is important for seizure detection.pproximate entropy (ApEn) is a statistical parameter to measure

he regularity of a time series data and it has been widely used in thenalysis of physiological signals, such as estimation of regularity inpileptic seizure time series data (Radhakrishnan and Gangadhar,998). Diambra et al. (1999) have shown that the value of the ApEnrops abruptly due to the synchronous discharge of large groupsf neurons during an epileptic activity. Thus, it is a suitable fea-ure to characterize the EEGs. In this work, ApEn derived from the

ultiwavelet transform is used to discriminate EEG signals.Artificial neural networks have been used as the most common

lassifier for classifying EEGs (Nigam and Graupe, 2004; Jahankhanit al., 2006; Güler et al., 2005; Übeyli, 2006a). ANN is a parallelighly inter-connected structure consisting of a number of simple,on-linear processing elements. ANN can perform computations atvery high speed if implemented on a dedicated hardware. Becausef its adaptive nature, it can adapt itself to learn the knowledge ofnput signals. Thus, an ANN model is also chosen as the classifierystem for this research.

In this paper, a novel epileptic seizure detection method is pro-osed. The method consists of three steps. Initially, multiwaveletransform is used to decompose the EEG signal to several sub-ignals. Then, the approximate entropy feature is extracted from

ach sub-signal. Finally, the extracted features are used as input ton artificial neural network, which discriminates the EEGs accord-ng to the specified classification problems. To the knowledge ofhe authors, there is no other work in the literature related tosing approximate entropy based on multiwavelet transform as the

Methods 193 (2010) 156–163 157

input to an ANN for automatic epileptic seizure detection in EEGs.A dataset containing 500 EEG segments is employed. The proposedmethod is evaluated through two classification problems, while dif-ferent selection of EEGs from the whole dataset is required for eachclassification problem. The obtained high accuracies indicate thegood classification performance of the proposed method.

The rest of the paper is organized as follows. In Section 2, thebasics of multiwavelet transform theory, approximate entropy andartificial neural network are described. In Section 3, the datasetused in this work and the proposed methodology are described indetail. Then, in Section 4, the evaluation procedure and the exper-imental results are presented, followed by further discussion onthe obtained results. Finally, some conclusions and future work aregiven in Section 5.

2. State of the art

2.1. Multiwavelet transform (MWT)

Multiwavelet opens a new chapter to wavelet theory in recentyears. Multiwavelet is considered as the generalization of scalarwavelet. However, some important differences exist between thesetwo types of wavelet. Specifically, scalar wavelets only have asingle scaling and wavelet function, whereas multiwavelets havetwo or more scaling and wavelet functions. For notational con-venience, the set of scaling functions are written as the vectornotion˚(t) = [�1(t),�2(t), . . .,�r(t)]T, where˚(t) is called multiscal-ing function and T denotes the vector transpose. Correspondingly,multiwavelet function is defined from the set of wavelet functionsas � (t) = [ 1(t), 2(t), . . ., r(t)]T, where r > 1 is an integer. Whenr = 1, � (t) is called a scalar wavelet, or simply wavelet. Although rcan be arbitrarily large, most existing multiwavelets to date haver = 2 (Martin and Bell, 2001). The dilation and wavelet equations formultiwavelet take the following forms, respectively:

˚(t) =√

2∑

k

Gk˚(2t − k) (1)

� (t) =√

2∑

k

Hk˚(2t − k) (2)

The pair {Gk, Hk} is called a multiwavelet filter bank (this is oftenabbreviated to multifilter bank). Gk and Hk are matrix low-pass filterand matrix high-pass filter, respectively. They are all r × r matricesfor each integer k. The matrix elements in these filters provide moredegrees of freedom than a scalar wavelet. This makes multiwaveletshave some significant properties, such as short support, orthogo-nality, symmetry, and vanishing moments. These properties are allimportant in signal processing (Strela et al., 1999).

Similar to the scalar wavelet transform, the multiwavelet trans-form (MWT) is also based on the idea of multiresolution analysis.During a single level decomposition using scalar wavelet trans-form, the input data is decomposed into two sub-bands, whichrepresent the output of low-pass and high-pass filters individually.While, MWT decompostion produces two low-pass sub-bands andtwo high-pass sub-bands for each level decomposition. In MWTeach multifilter bank contains two channels, so that there willbe two sets of scaling coefficients and two sets of wavelet coeffi-cients for each level decomposition. Since multiple iterations over

the low-pass data are desired, the scaling coefficients for the twochannels are stored together. Likewise, the wavelet coefficientsfor the two channels are also stored together. Successive itera-tions are performed on the two sets of scaling coefficients fromthe previous stage. The MWT decomposition algorithm is given as

Page 3

158 L. Guo et al. / Journal of Neuroscience

(

c

d

Klioiafsvvvrcfi

(rpPgficocidttt

2

icueg(

(

and bias values according to Levenberg–Marquardt optimization.

Fig. 1. MWT decomposition.

Jafari-Khouzani and Soltanian-Zadeh, 2003):

j−1,k =∑

n

Gncj,2k+n (3)

j−1,k =∑

n

Hndj,2k+n (4)

The procedure of MWT decomposition is shown in Fig. 1 (Jafari-houzani and Soltanian-Zadeh, 2003). G and H represent matrix

ow-pass and high-pass filters, respectively. [c1,0,k, c2,0,k]T is thenput signal vector, which form the initial expansion coefficientsf the given multiwavelet system. After the input signal vectors filtered through G and H matrix filters, four output streamsre generated. Each of these streams is then downsampled by aactor of two. [c1,−1,k, c2,−1,k]T represents the low-pass sub-bandignal vector, which is multiscaling coefficients of the input signalector. [d1,−1,k, d2,−1,k]T represents the high-pass sub-band signalector, which is the multiwavelet coefficients of the input signalector. For the sucessive decomposition, the same procedure isepeated on the first level multiscaling coefficients vector [c1,−1,k,2,−1,k]T to get the second level multiscaling and multiwavelet coef-cients.

Unlike scalar wavelets in which Mallat’s pyramid algorithmsMallat, 1989) can be employed directly, the application of MWTequires that the input signal to be first vectorized, namely pre-rocessing or prefiltering in multiwavelet (Strela et al., 1999).reprocessing generate multiple (vector) input streams from aiven scalar source stream, and form the initial expansion coef-cients of a given multiwavelet system. Since preprocessing is arucial step for the success of MWT in applications, various meth-ds have been developed. The repeated-row method is the mostommonly used, which is repeating the scalar input signal to get annput vector (Strela et al., 1999). This preprocessing method intro-uces oversampling of the original data by a factor of two. Sincehe repeated-row preprocessing has proved to be superior in fea-ure extraction (Soltanian-Zadeh et al., 2004), it is also adopted inhis work.

.2. Approximate entropy (ApEn)

Approximate entropy is a measure that quantifies the complex-ty or irregularity of a time series. Larger value of ApEn means moreomplexity and irregularity of the signal. Recently, ApEn has beensed for the detection of epilepsy (Kannathal et al., 2005). A robuststimate of ApEn can be obtained by using short, noisy datasets. Theeneral procedure of estimating ApEn is described in the following

Sabeti et al., 2009):

1) Let the original signal containing N data points be X(n) = [x(1),x(2), . . ., x(N)].

Methods 193 (2010) 156–163

(2) m-vectors X(1), . . ., X(N − m + 1) are defined according to

X(i) = [x(i), x(i+ 1), . . . , x(i+m− 1)] i = 1,2, . . . , N −m+ 1

(5)

These vectors represent m consecutive x values, starting withthe i-th point.

(3) Denote the distance between X(i) and X(j) by d[X(i), X(j)],defined as the maximum absolute difference between theirrespective scalar components, i.e., the maximum norm

d[X(i), X(j)] = maxk=1,2,...,m

|x(i+ k − 1) − x(j + k − 1)| (6)

(4) For a given X(i), find the number of (j = 1, . . ., N − m + 1, j /= i)so that d[X(i), X(j)] ≤ rr, denoted as Nm(i). Then, for i = 1, . . .,N − m + 1

Cmrr (i) = Nm(i)N −m+ 1

(7)

(5) Cmrr (i) measures the frequency of patterns similar to the onegiven by the window of length m within a tolerance rr.

(6) Compute the natural logarithm of eachCmrr (i) and average it overi

�m(rr) = 1N −m+ 1

N−m+1∑

i=1

lnCmrr (i) (8)

(7) Increase the dimension to m + 1. Repeat steps (2)–(6) to obtainCm+1rr (i) and �m+1(rr).

Finally, ApEn is computed based on the following formula:

ApEn(m, rr,N) = �m(rr) − �m+1(rr) (9)

Before computing ApEn value of the signal with length N, twoparameters must be specified: m, the embedding dimension and rr,a tolerance window.

2.3. Artificial neural network (ANN)

An artificial neural network is an information-processing sys-tem that is based on simulation of the human cognition process.It consists of many computational neural units connected toeach other. In ANNs, knowledge about the problem is distributedthrough the connection weights of links between neurons. The neu-ral network has to be trained to adjust the connection weights andbiases in order to produce the desired mapping. ANNs are widelyused in the biomedical field for modeling, data analysis, and diag-nostic recognition. The ANN’s capability of learning from examples,the ability to reproduce arbitrary non-linear functions of input,and the highly parallel and regular structure make them especiallysuitable for pattern recognition problems (Basheer and Hajmeer,2000).

The training algorithm is an important part of the ANN model. Agood topology can be inefficient if trained by an inappropriate algo-rithm. A suitable training algorithm has a short training process,while achieving better accuracy. One of the most common train-ing algorithms is Bayesian regularization back-propagation, whichis also used in current work. This algorithm updates the weight

It minimizes a combination of squared errors and weights and thendetermines the correct combination so as to produce a network thatgeneralizes well. The process is also called Bayesian regularization(Mathworks, 2002).

Page 4

L. Guo et al. / Journal of Neuroscience Methods 193 (2010) 156–163 159

(Z, O, N

3

3

wadsiostw(pemcotrutEi

fmsis

Fig. 2. Example EEG segments from each of the five sets

. Materials and methods

.1. Dataset description

The data described by Andrzejak et al. (2001) is used in currentork. The whole dataset consists of five sets (denoted as Z, O, N, F

nd S), each containing 100 single-channel EEG segments of 23.6 suration, with a sampling rate of 173.6 Hz. These segments wereelected and cut out from continuous multi-channel EEG record-ngs after visual inspection for artifacts (e.g., due to muscle activityr eye movements). Sets Z and O consisted of segments taken fromurface EEG recordings that were carried out on five healthy volun-eers using a standardized electrode placement scheme. Volunteersere relaxed in an awake state with eyes open (Z) and eyes closed

O), respectively. Sets N, F and S originated from an EEG archive ofresurgical diagnosis. Segments in set F were recorded from thepileptogenic zone, and those in set N from the hippocampal for-ation of the opposite hemisphere of the brain. While sets N and F

ontained only activity measured during seizure free intervals, set Snly contained seizure activity. All EEG signals were recorded withhe same 128-channel amplifier system, using an average commoneference. The data were digitized at 173.61 samples per secondsing 12-bit resolution and they have the spectral bandwidth ofhe acquisition system, which varies from 0.5 Hz to 85 Hz. TypicalEG segments (one from each of the five described sets) are shownn Fig. 2.

In this work, two different classification problems are createdrom the above dataset in order to verify the performance of our

ethod. In the first problem, two sets are examined, normal andeizure, the normal class includes only set Z while the seizure classncludes set S. The notation of the problem is simplified as Z–S. In theecond problem, all the EEGs from the dataset are used and they are

Fig. 3. Block diagram of the proposed method.

, F, and S). From top to bottom: segment Z to segment S.

classified into two different classes: sets Z, O, N and F are includedin the non-seizure class and set S in the seizure class, which thenotation is simplified as ZONF–S. The second classification problemis more close to the clinical applications.

3.2. Method

In this work, a novel method based on multiwavelet trans-form and approximate entropy is proposed for classifying the EEGdata into normal/non-seizure and epileptic. The raw EEG data isfirstly decomposed into several sub-signals through MWT, thenthe approximate entropy feature is extracted for each sub-signalto form a feature vector. Finally, the constructed feature vector isput as input to an artificial neural network to classify the EEG intonormal/non-seizure and seizure. The block diagram of the proposedapproach is shown in Fig. 3.

3.2.1. MWT decompositionThe basic multiwavelet transform theory has been explained

in Section 2.1. In current work, three famous multiwaveletsare empolyed, which are Gernoimo–Hardin–Massopust (GHM),Chui–Lian (CL) and SA4. All of these multiwavelets contain twoscaling and two wavelet functions, which are illustrated inFigs. 4, 5 and 6, respectively (for review details of GHM, CL andSA4, see Geronimo et al., 1994; Chui and Lian, 1996; Shen et al.,2000).

The raw scalar EEG data is firstly processed through therepeated-row preprocessing to construct an input vector. Then theinput vector is decomposed by multiwavelet up to 4 levels. The mul-tiwavelet decomposition is similar to scalar wavelet, but has somedifference. Since the low-pass and high-pass filters in MWT decom-position contain two channels (see Fig. 1), after the input vector isdecomposed with four levels, 10 sub-signals are generated and each

sub-signal represents the original signal information in differentfrequency sub-bands. Whereas, usual scalar wavelet transform onlygenerated 5 sub-signals for representing the signal information indifferent frequency sub-bands. Based on the advantages of multi-wavelet containing two scaling and wavelet functions, a feature of

Page 5

160 L. Guo et al. / Journal of Neuroscience

Fig. 4. GHM two scaling functions (top) and two wavelet functions (bottom).

Fig. 5. CL two scaling functions (top) and two wavelet functions (bottom).

Fig. 6. SA4 two scaling functions (top) and two wavelet functions (bottom).

Methods 193 (2010) 156–163

the signal through the scalar wavelet transform can be decomposedinto two components that makes the features more comprehensive.Thus, incorporating the two components of multiwavelet can offera more efficient feature extraction than the scalar wavelet.

3.2.2. ApEn feature extractionApproximate entropy is a statistic parameter that measures the

regularity or predictability of a specific time series. It is also knownthat ApEn possesses good characteristics such as robustness inthe characterization of the epileptic patterns. Therefore, in currentstudy ApEn is chosen as the features to discriminate EEGs. ApEnvalue for each sub-signal of the EEG data decomposed with MWTis calculated to form a feature vector. Before computing ApEn, twoimportant parameters, which are embedding dimension (m) andtolerance window (rr), have to be defined. Based on the sugges-tions by Pincus (1991), the values m and rr are set optimally to 2 and0.15 times the standard deviation of the data, respectively. After theApEn values were obtained for 10 sub-signals derived from MWTof the EEG, they construct a feature vector that is used as input toan artificial neural network for classifying EEGs.

3.2.3. Artificial neural network (ANN) classificationMulti-layer perceptron neural network (MLPNN) is the most

widely used ANN structure for classification problems. It has fea-tures such as the ability to learn and generalize, smaller trainingdata requirement, fast operation, and easy implementation. In thiswork, a three-layer MLPNN with Bayesian regularization back-propagation training algorithm is used to classify EEGs based onthe previous obtained ApEn feature vector.

3.3. Statistical parameters

The evaluation of the proposed method on classification prob-lems is determined by computing the statistical parameters ofsensitivity, specificity and classification accuracy. The definitionsof these parameters are as:

• Sensitivity: number of correctly detected positive patterns/totalnumber of actual positive patterns. A positive pattern indicates adetected seizure.

• Specificity: number of correctly detected negative patterns/totalnumber of actual negative patterns. A negative pattern indicatesa detected normal/non-seizure.

• Classification accuracy: number of correctly classified pat-terns/total number of patterns.

4. Results and discussion

4.1. Results

The two classification problems described above are used toevaluate the proposed method. The classification implementationprocedure is: EEG signal under study is firstly preprocessed withrepeated-row method to form an input signal vector, then the inputsignal vector is decomposed into 10 sub-signals through multi-wavelet with 4-level decomposition. After that, the approximateentropy feature is calculated for each sub-signal to form a featurevector with a dimension of 10. Finally, the generated feature vectoris fed into an MLPNN to classify normal/non-seizure and seizureEEGs. The proposed method is carried out with GHM, CL and SA4multiwavelets individually. For addressing the specific problems,

several different MLPNN structures have been tried. The networkstructure that gave the best results for the classification problemsof this work is: one input layer with 10 neurons (equals to theapproximate entropy feature vector dimension), and one hiddenlayer with 10 neurons and one output layer with 1 neuron. The

Page 6

L. Guo et al. / Journal of Neuroscience Methods 193 (2010) 156–163 161

Table 1The performance evaluation parameters of three multiwavelets for Z–S classificationproblem.

Multiwavelet Specificity (%) Sensitivity (%) Classification accuracy (%)

uattMHfwasrmtaci

TswwcamsowsdiaemS

muwabcddibtdcs

TTc

multiwavelets have more than 99% correct classification while thebest result of three scalar wavelets is only 97.15% with Db6. ForZONF–S classification problem, multiwavelet SA4 showed the bestperformance with 98.27% accuracy. All three multiwavelets havemore than 95% discriminate accuracy while the best result of three

GHM 99.20 100 99.85 ± 0.15CL 98.13 99.92 99.23 ± 0.70SA4 98.77 100 99.38 ± 0.61

nits in the hidden layer are sigmoid units with hyperbolic tangents transfer function, while the output is linear. The target value ofhe MLPNN output layer was defined as 0 or 1, which 0 representshe normal/non-seizure EEG and 1 represents the seizure EEG. The

LPNN was implemented by using MATLAB software version 7.9.alf of the data were randomly selected for training, while the rest

or testing. The training dataset was used to train the MLPNN net-ork, whereas the testing data was used to verify the accuracy

nd the effectiveness of the trained MLPNN for specific EEG clas-ification problems. One hundred different training-test sets wereandomly created for each classification problem. The test perfor-ance of the proposed method was determined by computation of

he three statistical parameters described in Section 3.3, which theverage results of one hundred executions for three multiwaveletsorresponding to Z–S and ZONF–S classification problems are givenn Tables 1 and 2, respectively.

From the performance evaluation parameters shown inables 1 and 2, the proposed approach was verified to be quiteuccessful in detecting epileptic seizures in EEGs. For Z–S problem,ith GHM, CL and SA4 multiwavelets, the classification accuraciesere all higher than 99%. Moreover, GHM achieved 99.85% classifi-

ation accuracy with a standard deviation of 0.15%, which is the bestmong the three multiwavelts. For the ZONF–S problem, which isore close to real clinic application, those three multiwavelets also

howed excellent discriminate performance. SA4 outperformed thether two multiwavelets by 98.27% overall classification accuracyith a standard deviation of 0.81%. From the results of the two clas-

ification problems, it is proven that the approximate entropieserived from the multiwavelet transform significantly character-

zed the EEG signals. In addition, through comparing the resultsmong the three multiwavelets on approximate entropy featurextraction, different multiwavelet basis exhibit different perfor-ance for specific problem: GHM is the best for Z–S problem while

A4 is the most suitable for ZONF–S problem.In order to test the efficiency of mulwitwavelets, the perfor-

ance of scalar wavelets on EEG signal classification is also studiednder the same experimental conditions. Three scalar wavelets,hich are Daubechies wavelets with order 2, 4 and 6 (Db2, Db4,

nd Db6) are chosen in the work since Daubechies wavelets haveeen proven as the most effective scalar wavelet in EEG signallassification (Subasi, 2006, 2007). With 4-level scalar waveletecomposition of EEG signals, 5 sub-signals are obtained. Thus, theimension of approximate entropy feature based on scalar wavelet

s 5. The best MLPNN network structure for approximate entropy

ased on scalar wavelet is: one input layer with 5 neurons (equalso the approximate entropy feature vector dimension), and one hid-en layer with 5 neurons and one output layer with 1 neuron. Thelassification accuracy comparison of three multiwavelet and threecalar wavelets for the same two EEG classification problems are

able 2he performance evaluation parameters of three multiwavelets for ZONF–S classifi-ation problem.

Multiwavelet Specificity (%) Sensitivity (%) Classification accuracy (%)

GHM 89.91 98.62 96.69 ± 1.18CL 89.21 96.57 95.15 ± 1.40SA4 95.50 99.00 98.27 ± 0.81

Fig. 7. The classification accuracy comparison of multiwavelets and scalar waveletsfor Z–S problem.

shown in Figs. 7 and 8. It can be seen that multiwavelets offer betterdiscrimination performance than scalar wavelets for both classi-fication problems. Thus, it can be concluded that multiwaveletsoutperform the scalar wavelets in discriminating EEG signals.

4.2. Discussion

In this paper, a novel method for epileptic seizure detectionin EEGs is proposed. The method combines multiwavelet trans-form based approximate entropy with artificial neural network forclassifying the EEGs. This method is evaluated through two differ-ent classification problems originated from medical diagnosis. Thehigh classification accuracies proved the outstanding performanceof the approach. Equipped with specific properties for signal pro-cessing, multiwavelet with repeated-row preprocessing achieveshigher classification accuracy than that of scalar wavelet. For Z–Sclassification problem, multiwavelet GHM obtained the highestaccuracy with 99.85% among the three multiwavelets. All three

Fig. 8. The classification accuracy comparison of multiwavelets and scalar waveletsfor ZONF–S problem.

Page 7

162 L. Guo et al. / Journal of Neuroscience Methods 193 (2010) 156–163

Table 3A comparison of classification accuracy obtained by our method and others’ method for two EEG classification problems.

Researchers Method Classification problem Classification accuracy (%)

Nigam and Graupe (2004) Non-linear preprocessing filter—diagnostic neural network Z–S 97.2Srinivasan et al. (2005) Time- and frequency-domain features—recurrent neural network Z–S 99.6Kannathal et al. (2005) Entropy measures—adaptive neuro-fuzzy inference system Z–S 92.22Polat and Günes (2007) Fast Fourier transform—decision tree Z–S 98.72Subasi (2007) Discrete wavelet transform—mixture of expert model Z–S 95Tzallas et al. (2007) Time–frequency analysis—artificial neural network Z–S 100Guo et al. (2009) Discrete wavelet transform–relative wavelet energy—MLPNN Z–S 95.2

featu

workfeatu

stn

aCsftfcaodfsb(a

sbpdfofZctdahbiiodr

5

esereiBut

This work Multiwavelet transform–approximate entropy

Tzallas et al. (2007) Time–frequency analysis—artificial neural netThis work Multiwavelet transform–approximate entropy

calar wavelets is 94.10% with Db2. Therefore, it can be concludedhat multiwavelet is more appropriate for classifying the EEG sig-als comparing with scalar wavelet.

The results from the current work verify that the multiwavelet isn efficient signal processing technique for the feature extraction.ompared with scalar wavelet, the main reasons of multiwaveletuccess in current research lie in: Firstly, in multiwavelet trans-orm, the repeated-row preprocessing of the scalar input produceshe oversampling of the original signal, which makes the extractedeatures are more discriminative. Secondly, since the multiwaveletontains two or more scaling and wavelet functions, the low-passnd high-pass filters in MWT decomposition are matrices insteadf scalars. Thus, a feature in the usual scalar wavelet transform isecomposed into two or more components in multiwavelet trans-orm, which makes the features more sensitive to differentiate theignals. To the best of our knowledge, multiwavelet has alreadyeen widely used in signal or image compression and denoisingCotronei et al., 2000; Strela et al., 1999), with the first time it ispplied for EEG signal classification.

There are many other methods proposed for the epilepticeizure detection. Table 3 presents a comparison on the resultsetween the method developed in this work and other methodsroposed in the literature. Only methods evaluated in the sameataset are included so that a comparison between the results iseasible. For Z–S classification problem, the accuracy obtained fromur method is the second best presented, which is only 0.15% dif-erence compared with the best result shown in the literature. ForONF–S classification problem, which is a case more close to thelinic expert needs, the result obtained in our work is better thanhat of the Tzallas’s work, in which the authors employed the energyistribution features extracted from the time–frequency plane ton ANN for classifying EEGs. The quality of the proposed methodas been proven from the obtained results. The accuracy achievedy our approach for the epileptic seizure detection is more than sat-

sfactory and also its automated nature makes it suitable to be usedn real clinical applications. A system that may be developed basedn the result of this study can assist the medical professionals toetect seizures quickly and accurately through examining the EEGecordings.

. Conclusion and future work

This study explores the capability of applying approximatentropy derived from multiwavelet transform to classify EEGignals. The original EEG signal is firstly decomposed into sev-ral sub-signals through 4-level multiwavelet transformation withepeated-row preprocessing. For each sub-signal, the approximate

ntropy feature, which measures the regularity or predictabil-ty of the signal, is calculated. Finally, a three-layer MLPNN withayesian regularization back-propagation training algorithm issed for classification. High classification accuracies obtained withhree multiwavlets for two classification problems, derived from a

re—MLPNN Z–S 99.85

ZONF–S 97.73re—MLPNN ZONF–S 98.27

well-known database, verified the success of the proposed method.Through comparing the performance of the multiwavelets withthat of the scalar wavelets on the same classification problems, itcan be concluded that the multiwavelets achieve more satisfactoryresults than the scalar wavelets with reference to the EEG signaldiscrimination.

The multiwavelet transform of a given signal results in morenumerous sub-signals than the scalar wavelet transform (with thesame raw signal). An excessive number of features corresponding tothose sub-signals increases the computation cost. Future researchis underway to devise some soft computing techniques for reducingthe feature dimension. Since the proposed method has shown suc-cess on epileptic EEG classification problems, future work wouldinclude applying the proposed method to more wide range of pat-tern recognition problems which are also important to humans,such as Alzheimer’s and Parkinson’s diseases detection and diag-nosis.

Acknowledgements

Ling Guo was financially supported through a fellowship ofthe Agencia Espanola de Cooperación International (AECI) and theSpanish Ministry of Foreign Affairs.

References

Andrzejak R, Lehnertz K, Mormann F, Rieke C, David P, Elger C. Indications of non-linear deterministic and finite-dimensional structures in time series of brainelectrical activity: dependence on recording region and brain state. PhysicalReview E 2001;64(6):061907-1–8.

Basheer I, Hajmeer M. Artificial neural networks: fundamentals, computing, design,and application. Journal of Microbiological Methods 2000;43(1):3–31.

Chui C, Lian J. A study of orthonormal multi-wavelets. Applied Numerical Mathe-matics 1996;20(3):273–98.

Cotronei M, Lazzaro D, Montefusco L, Puccio L. Image compression through embed-ded multiwavelet transform coding. IEEE Transactions on Image Processing2000;9(2):184–9.

Diambra L, de Figueiredo J, Malta C. Epileptic activity recognition in EEG recording.Physica A: Statistical Mechanics and Its Applications 1999;273(3–4):495–505.

Geronimo J, Hardin D, Massopust P. Fractal functions and wavelet expan-sions based on several scaling functions. Journal of Approximation Theory1994;78(3):373–401.

Güler N, Übeyli E, Güler I. Recurrent neural networks employing Lyapunovexponents for EEG signals classification. Expert Systems with Applications2005;29(3):506–14.

Guo L, Rivero D, Seoane J, Pazos A. Classification of EEG signals using relative waveletenergy and artificial neural networks. In: Proceedings of the first ACM/SIGEVOSummit on Genetic and Evolutionary Computation; 2009. p. 177–84.

Hsung T, Lun D, Ho K. Optimizing the multiwavelet shrinkage denoising. IEEE Trans-actions on Signal Processing 2005;53(1):240–51.

Jafari-Khouzani K, Soltanian-Zadeh H. Multiwavelet grading of pathological imagesof prostate. IEEE Transactions on Biomedical Engineering 2003;50(6):697–704.

Jahankhani P, Kodogiannis V, Revett K. EEG signal classification using wavelet featureextraction and neural networks. In: IEEE John Vincent Atanasoff 2006 Interna-

tional Symposium on Modern Computing (JVA’06); 2006. p. 52–7.

Kalayci T, Ozdamar O. Wavelet preprocessing for automated neural networkdetection of EEG spikes. IEEE Engineering in Medicine and Biology Magazine1995;14(2):160–6.

Kandel E, Schwartz J, Jessell T. Principles of neural science. New York: McGraw-Hill,Health Professions Division; 2000.

Page 8

cience

K

M

M

MM

N

P

P

R

R

S

time–frequency analysis and artificial neural networks. Computational Intel-

L. Guo et al. / Journal of Neuros

annathal N, Choo M, Acharya U, Sadasivan P. Entropies for detection of epilepsy inEEG. Computer Methods and Programs in Biomedicine 2005;80(3):187–94.

allat S. A theory for multiresolution signal decomposition: the wavelet rep-resentation. IEEE Transactions on Pattern Analysis and Machine Intelligence1989;11(7):674–93.

artin M, Bell A. New image compression techniques using multiwavelets and mul-tiwavelet packets. IEEE Transactions on Image Processing 2001;10(4):500–10.

athworks. MATLAB V.6.5.0 Help Files; 2002.ohseni H, Maghsoudi A, Kadbi M, Hashemi J, Ashourvan A. Automatc detection of

epileptic seizure using time–frequency distributions. In: IET 3rd internationalconference on advances in Medical, Signal and Information Processing, MEDSIP2006; 2006. p. 1–4.

igam V, Graupe D. A neural-network-based detection of epilepsy. NeurologicalResearch 2004;26(1):55–60.

incus S. Approximate entropy as a measure of system complexity. Proceed-ings of the National Academy of Sciences of the United States of America1991;88(6):2297–301.

olat K, Günes S. Classification of epileptiform EEG using a hybrid system basedon decision tree classifier and fast Fourier transform. Applied Mathematics andComputation 2007;187(2):1017–26.

adhakrishnan N, Gangadhar B. Estimating regularity in epileptic seizure time-seriesdata: a complexity-measure approach. IEEE Engineering in Medicine and Biology1998;17(3):89–94.

amakrishnan S, Selvan S. Classification of brain tissues using multiwavelet trans-formation and probabilistic neural network. International Journal of Simulation:Systems, Science and Technology 2006;7(9):9–25.

abeti M, Katebi S, Boostani R. Entropy and complexity measures for EEG signalclassification of schizophrenic and control participants. Artificial Intelligence inMedicine 2009;47(3):263–74.

Methods 193 (2010) 156–163 163

Shen L, Tan H, Tham J. Symmetric–antisymmetric orthonormal multiwaveletsand related scalar wavelets. Applied and Computational Harmonic Analysis2000;8(3):258–79.

Soltanian-Zadeh H, Rafiee-Rad F, Pourabdollah-Nejad D. Comparison ofmultiwavelet, wavelet, Haralick, and shape features for microcalcifica-tion classification in mammograms. Pattern Recognition 2004;37(10):1973–86.

Srinivasan V, Eswaran C, Sriraam N. Artificial neural network based epileptic detec-tion using time-domain and frequency-domain features. Journal of MedicalSystems 2005;29(6):647–60.

Strela V. Multiwavelets: theory and application. Ph.D. thesis; 1996.Strela V, Heller P, Strang G, Topiwala P, Heil C. The application of multi-

wavelet filterbanks to image processing. IEEE Transactions on Image Processing1999;8(4):548–63.

Subasi A. Epileptic seizure detection using dynamic wavelet network. Expert Sys-tems with Applications 2005;29(2):343–55.

Subasi A. Automatic detection of epileptic seizure using dynamic fuzzy neural net-works. Expert Systems with Applications 2006;31(2):320–8.

Subasi A. EEG signal classification using wavelet feature extraction and amixture of expert model. Expert Systems with Applications 2007;32(4):1084–93.

Tzallas A, Tsipouras M, Fotiadis D. Automatic seizure detection based on

ligence and Neuroscience 2007;13, article ID 80510.Übeyli E. Analysis of EEG signals using Lyapunov exponents. Neural Network World

2006a;16(3):257–73.Übeyli E. Fuzzy similarity index employing Lyapunov exponents for discrimination

of EEG signals. Neural Network World 2006b;16(5):421–31.