7/30/2019 01020517

1/4

2001 Proceedings of the 23rd Annual EMBS International Conference, October 25-28, Istanbul, Turkey

A NEW APPROACH FOR DIAGNOSING EPILEPSY BY USING W AVELETTRANSFORM AND NEURAL NETWORKSM.Akin', M.A.Arserim',M.K.Kiymik*, I.Turkoglu3

'Dep. of Electric and Electro nics Engineering, Dicle University, Diyarbakir, Turkey*Dep.of Electric and Ele ctronic s Engineering, SI. niversity, Kah ramanmaras, Turkey'Electronics and Com puter Dept., Tech nical Education Faculty, Firat University, 23 1 19, Elazig, Turkey.Abstract: Today, epilepsy keeps its importance as a majorbrain disorder. However, although some devices such asmagnetic resonance (MR), brain tomography (BT) areused to diagnose the structural disorders of brain, forobserving some special illnesses especially such asepilepsy, EEG is routinely used for observing the epilepticseizures, in neurology clinics. In our study, we aimed toclassify the EEG signals and diagnose the epilepticseizures directly by using wavelet transform and anartificial neural network model.

EEG signals are separated into 6, 8, cx, and p spectralcomponents by using wavelet transform. These spectralcomponents are applied to the inputs of the neuralnetwork. Then, neural network is trained to give threeoutputs to signify the health situation of the patientsKeywords: wavelet, neural network, epilepsy,EEG

I. INTRODUCTIONIn m edicine, EEG keeps its im portan ce for identifying thephysiological, and the psychological situations of the humanand the functional activity of the brain. In neurology c linicsEEG device is used efficiently for observing the braindisorders.According to the spectral components, and the amplitudesof these spectral components, which EEG consists, differentinterpretations can be made about the recorded waveform(the patient is healthy or not). The m ost important frequencycomponent of the human's EEG is a wave (approximatelybetween 8-12Hz), and a wave is sometimes called as thenatural frequency of the brain (1). This wave appears whenthe eyes are closed and one beg ins to rest. In epilepsy cases,however, when the epileptic seizures occurs, 6, 8 waves,which have lower f reque ncies, and higher m agnitudes withrespect to a waves, should be seen (6,8 aves has 0-4Hz, 4-8Hz frequency ranges, respectively). In ad dition, brainproduces desynchronize waves, which have higher frequency,lower m agnitude, called p waves (frequency range is higherthan 13Hz). Therefore, for diagnosin g the brain disord ers,these spectral components must b e analyzed carefully.When the EEG waveform is observed, it is seen that EEGwaveform is a non-stationary signal. For this reason, whenthe frequency components of the EEG is extracted by usingthe Short Time Fourier Transform (STFT) and the wavelettransform, including stft, should be useful than the otherspectrum analyzing methods (AR, A R M , FFT etc).Furthermore, viewing the results of the wavelet transfo rm intime domain should be useful to make add itional comm ents.

After these processes, if we think that the person whodiagnoses the illnesses is a doctor, use of an artificial neuralnetwork (ANN) should be offered. Because, by using theartificial neural netwo rk should minim ize the errors d one bydocto rs when they diagnose the illnessIn our study, EEG data sets are collected by a system, whichhas been used in our previou!; studies. From the EEG datasets, obtained 6, 8, a, nd p waves are extracted by usingwavelet transform. After all, according to these waves anartificial neural network trained, and developed to diagnosethe epileptic cases.U. MATERTIALSAND METHODS

A. Obtaining The EEG D ata SetsIn our previous studies, a d ata accusation and pro cessingunit (PCI-MIO-16-E%) is used to record the EEG data tomake com puter-based analysis. Record ings have been madeas 202 samples during 6 seconds. The accusation unit has a12 bits analog to digital converter (AD 7572, % 0.02sensitivity, 0.014 ms conversion time) to discritisize the EEGwaveform. The EEG recording unit is shown in fig. 1.

B. Wavelet T ransformIf a signal does not change much ove r time, we would callit as a stationary signal. Fourier transform could be applied tothe sta tionary signals easily and go od result can be taken.However, like EEG, a plenty of signals contain non-stationaryor transitory characteristics, and Fourie r Transformis not suited properly to detect the no n-stationary signals.

Fig. 1. Data acquisition system

0-7803-721 1-5/01/$17.000 001 IEEE 1596

7/30/2019 01020517

2/4

In an effort to correct this deficiency, Dennis Gabor (1946)adapted the Fourier transform to analyze only a sm all sectionof the signal at a time, which is called as Short Time FourierTransform . One of the major features of stft is mapping thesignal in two-dime nsional function of time and frequency.The -WaveletTransform deco mposes a signal onto a set ofbasic functions called wavelets. These basic functions areobtained by dilations, contractions and shifts of a uniquefunction called the wa velet prototype.In order to the input signal x(t), Wavelet Transform should beseparated as Continuous Wavelet Transform (CWT) andDiscrete Wavelet Transform (DWT). We can identify theCWT as in (1);

CW T(a,b)=j x(t).Y *a,b(t).dt (1)where * denotes the com plex conjugate, a R' represents thescale param eter, b E R' represents the translation, and Y ,,(t)is obtained by scaling the prototype wavelet Y(t) at a time b,and scale a as in (2);

Generally in wavelet applications, orthogonal dyadicfunctions are chosen as the mother wavelet. This transform isoften discritisized in a and b on a dyadic grid with the timeremaining continuou s. The m other wavelet, commonly used,is (3);Yj ,&t )=2-"2y(2- ' t - ) (3)

where {\Yj,k(t),j,keZ} for L2(R)C . Artificial Neural Network

Neural networks are used as a powerful means inengineering area after the development especially, incomputer technology. The fundamental characteristic of theneural network s is an adaptive, non-algo rithmic and parallel-distributed memory [11.Artificial neural networks are modeled by inspiring frombiological neural system and have a more simple structure.Many neural networks were developed for resembling severalknown characteristics of biological neural networks such aslearning and reacting. Some characteristics, however, arerealized with an engineering approach instead ofneuropsychological one [2].III. EXPERIMENTAL,TUDY

In this study, first EEG waveforms have been recorded bya data acquisition and processing unit. One of the recordedEEG waveform is shown below. Then, the wavelettransforms of the recorded EEG waveforms are taken byusing daubechies wavelets. Recorded EEG waveforms are

first divided into low and high wav elet coefficients, and theselow and high wavelet coefficients are divided in to their sub -high and sub low coefficients. Therefore, 6, 8, a, and pwavelets of the original EEG waveform are obtained.0.2 I0

-020 2 4 60.027 I0 2 4 60.01 ,

-0.01 I0 2 4 60.05, 1

-0.05 I0 2 4 60.2 I 1

-0.2 10 2 4 6t ime(sec )

Fig. 2. SimulatedEEG waveform and ts spectral components due to wavelettransformThe results of Wavelet Transform of the different EEG's areshown in figure 2,3 , and 4.

In these figures first the EEG waveform has been given.Then the sub-spectral components depending each EEG aregiven. The 6, 8; a, and /3 waves are viewed in the figure bythe following windows. And figures 2,3,4 show the EEGwaveforms as simulation, healthy and epileptic respectively.Classification is based on the partition of every section ofthe space formed by EEG wavelet signals and determinationof a p artitioning function related with those sections; in caseof the igno rance of the mathematical forms of the partitioningfunctions, first a learning activity should be realized.Learning activity provides the determ ination of the realvalues of these functions with the aid of the examples fromevery class (training set) [3]. Since the classifiers are basedon de ciding by learning, they lead to more successful resultswith respect to the traditional (non-learnin g) methods [4].Back propagation network is a multi-layer feed forwardnetworks. It is an artificial neural network between the inp utand an output layer, of which m ore than one layer is used. Inthese immediate layers called as hidden layer, there areprocessing elements, which don't receive input and give

1597

7/30/2019 01020517

3/4

output without any means. The general layout of a multi-layer neural network classifier, shown in fig. 5 . is given [ 5 ] .output classes

-0.5 10.5

0-0.50.2

0

0 2 4 6

0 2 4 6

- 0 .20.5 0 2 4 6

I- 0 . 5 I i

0 2 4 6time(sec)

Fig. 3. Epileptic EEG waveform and its spectral components due to wavelettransform

0 . 5

- 0 . 5 10 2 4 60.1 1 I

-0.1 ' I0 2 4 60 . 1

-0.1 I I0 2 4 60 . 2 1

- 0 . 2 ! I0 2 4 6t i rne(sec)

do dl d* 4Wavelet SignalsFigS. Multi layer feed forward neural network classifier.

Then the training characteristics of neural network used inthis study are as follows;Structure:Layer number: 3

The number of neuron on the layers: (4x202) 15 3Training Parameters:

Adaptive learning coefficient: 0.0005Momentum coefficient: 0.95Sum-squared error-sse: 0.0005Activation Function: tangent sigmoid

The variation of system error in according to the learningiteration during the training stage of back propagationnetwork is given in fig. 5. There is not any instability orroughness in training process of the network. This shows theconvenience of the parameters chosen to train the networks.In the second stage, the trained network was tested withEEG wavelet signals. As a result it was seen that byobserving the output vector produced by the network it waspossible to diagnose the disease.Finally several types of EEG recordings that we have usedin the study have tested the developed network. And theresponses of the network to these test signals are shown intable 1.

Fig. 4. Normal EEG waveform and its spectral components due to wavelettransform

1598

7/30/2019 01020517

4/4

Training fo r 1245 Epochslo 2 I

-.._. . .

-. . . . . . -,-_. . . . . .__.I-0 500 1000

2

(U 1.5cm-E.- " 1E2 0.5m

0

Epoch

0 500 1000 1500EpochFig. 6. s.s .e and learning rate versus iteration number

Table 1. Result of he test signalsSignals Diagno sis Recog nitionTest signal 1 Epileptic 97Test signal 2 Healthy 95Test signal 3 Healthy 98Test signal 4 Healthy 97Test signal 5 Healthy 95Test signal 6 Pathologic 93

Rate (%)

IV . CONCLUSIONIn our study, we have tried to find a new solution fordiagnosing the epilepsy. For this aim, the Wavelet Transformof the EEG signals have taken, and the 6, 8,a, nd p sub-frequencies are extracted. Depending on these sub-frequencies an artificial neural network has been developedahd trained. The accuracy of the neural network outputs istoo high (%97 for ep ileptic case, %98 for healthy case, and %93 fo r pathologic case that have been tested). This means that

this neural network identifies the health conditions of thepatients approx imately as 90 of 100. From this point we cansay that an application of this theo retical study will be helpfulfor the neurolog ists when they diagn ose the epilepsy.

Furthermore we want to develop the practical applicationof this study. After all a small model of this system will bevery useful for the patients suffer from epilepsy.

REFERENCES[11 J.E. Dayhoff, Neural Network Architectures, VanNostrand Reinhold, New York, 1990.[2] P. Simpson, Artificial Neural Systems, Pergamon PressCompany, New York, 1990.a1 of Science Institute, 1996, p.147-158.[3] R.O. Duda and P.E. Hart, Pattern Classification andScene Analysis. Stanford Research Institute, 198 9.[4] M.J. Zurada, Introduction to Artificial Neural System s,New York. West Publishing Company, 1992.[5] C.M. ishop, Neural Networks for Pattern Recognition,Claretzdon Pre ss, Oxford, 1996.[6] N.Hazarika, Classification of EEG signals using thewavelet transform, Signal Processing [H.W. Wilson - AST];May 1 997; Vol. 59, ISS: 1;pg. 61[7] M.Akin, M .K.Kiymik, M.A.Arserim, i.Tiirkoglu,Separation of B rain Signals Using FFT and Neural Networks,Biyomut 200 0,Istanbul, Turkey[8] M.Akin, M.K.Kiymik, Application of Periodogram andAR Spectral Analysis to EEG Signals, Journal of MedicalSystem s, Vo1.24 No.4, 20 00[9] Mehta, S.V., Koser, R.W., Venziale, P.J.,Waveletanalysis as a potential tool for seizure detection, Time-Frequency and Time-Scale Analysis, 1994., Proceedings ofthe IEEE-SP International Symposium on , 1994 Page(s):[101 Reuter, M. Analysing epileptic events online by neuralnets, special preprocessing methods includedIntelligent Control and Autom ation, 2000. Proceedings of the3rd World Congress on, Volume: 2,20 00, Page(s): 919 -924v01.2[111 Geva, A.B., Kerem,D.H., Forecasting generalizedepileptic seizures from the EEG signal by wavelet analysisand dynamic unsupervised fuzzy clustering, BiomedicalEngineering, IEEE Transactions on ,Volume: 45 Issue: 10 ,Oct. 19 98 Page(s): 12 05 -1216

584 -587

1599