International Journal of Scientific & Engineering Research Volume 3, Issue 12, December-2012 1

ISSN 2229-5518

CLASSIFICATION AND CLUSTERING OF BRAIN SEIZURE ACTIVITY USING WAVELET TRANS- FORM AND RADIAL BASIS NEURAL NETWORK

Shweta Kumari

Abstract— Electroencephalogram (EEG) is the record of the brain electrical activity and it contains valuable information related to the dif- ferent physiological and pathological states of the brain. Epilepsy is known to be the most prevalent neurological disorder in humans and seizure discharge is the main characteristics of the epilepsy. EEG is an important clinical tool for the diagnosis and monitoring of seizures. Epileptic seizure occurs irregularly and unpredictably manner due to temporary electrical disturbance of the brain. The aim of this project is Epileptic seizure detection in multichannel EEG. This paper presents a novel method for automatic epileptic seizure detection, useing recur- rent rates derived from discrete wavelet transform in combination with a radial basis function neural network for classification and clustering of the pattern feature of EEG signals. The output of the neural network aids in finding existence or absence of seizures in the EEG data. We have used discrete wavelet transform (DW T) of Daubechie’s wavelet order 4 to decompose the EEG signal at different levels in extracting approximation and detail coefficients. W e have evaluated a unique dynamical parameter (recurrence rate) from the wavelet co-coefficients of EEG of different subjects (normal and epileptic). The recurrence rate has been used in a radial basis function neural network for seizure detection. The performance of the network has been evaluated in terms of the accuracy, specificity and sensitivity detecting in unknown EEG time series.

Index Terms— Discrete Wavelet Transform (DW T), EEG, Radial basis function Neural Network (RBFNN), Recurrence Quantification

Analysis (RQA), Recurrent Rate

1 INTRODUCTION

An epileptic seizure (“a fit”) is a common neurological disorder that has been with us ever since ancient times and approx 50 million people in the world badly affected by epilepsy. It may occur usually following physical or metabolic insult resulting in sudden surge of electrical activity in the brain [1]. Clinically an epileptic seizure is an intermittent, usually unprovoked, stereo- typical, disturbance of consciousness, behavoiur, emotion, mo- tor function or sensation that is result of cortical neuronal dis- charge [7].
Patient experience different types of symptom during sei- zures and it’s depends on the location and extend of the affected brain tissue. The classification of seizure has been standardized by the International League Against Epilepsy (ILAE). Partial seizures affects small part of the brain and generalized seizures affects all parts of the brain [2].
Electroencephalograph (EEG) is the recording of brain activi- ty and its signals contains valuable information for understand- ing of epileptic seizures. In recent years, classification of EEG signals increases based on the machine learning techniques ap- plications, which assist physicians to diagnose the epileptic sei- zure. Although the epileptic seizures are unpredictable, so for

————————————————

• Shweta kumari is done masters degree program in biomedical engineering in VIT University, vellore, India, PH-09664682286 E-mail: shwetasu- man31@gmail.com

automatic detection of epileptic discharges, this can be used to predict the onset of seizure [6].
Discrete wavelet transforms (DWT) is a versatile signal processing tool which helps to decompose the EEG signals in different levels of approximation and detailed coefficient based on frequency sub bands, which produces information about the brain activities [4]. Here, we have used DWT because wavelet transform has mother wavelets with a finite start and finish as compare to fourier transform with sines and cosines function, which have infinite in mathematical terms to generate the coef- ficients. We can say that mother wavelet has “compact sup- port”. The importance of having compact support is that when you fit it to the signal you get a localized result rather than ge- neralized [6]. There are hundreds of mother wavelets available but Daubechies wavelet level 4 (db4) is more appropriate to detect changes in EEG signals.
Recurrence analysis is a method to measure the Recurrent Rate (RR) which is the density of recurrence points on a Recur- rence Plot (RP). It is a fundamental property of dynamical sys- tems, which can be exploited to characterize the system’s beha- viour in phase space [8].
Recurrence quantification analysis is a sensitive tool for de- tecting any dynamic changes, and it can be easily affected by settings like embedding dimension, time delay and mainly af- fected by threshold values. These measures are mostly based on the recurrence point density and the diagonal line structures of the RPs. It is a simplest application to measure the recurrent rate or per cent recurrences [13].
An Artificial Neural Network (ANN) is a powerful data- modeling tool that can capture and represent complex in- put/output relationships. The neural network technology mate- rialized from the desire to develop an artificial system that could perform intelligent and complex tasks similar to those

IJSER © 2012 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 3, Issue 12, December-2012 2

ISSN 2229-5518

performed by the human brain. The main advantages of artifi- cial neural network are adaptive learning and ability to learn how to do tasks based on initial experience or the data given for training. Self-organization in which an ANN can create its own organization and also represent the information it receives dur- ing learning time, real time operation which enables ANN computations to be carried out in parallel, fault tolerance due to which network capabilities may be retained even with major network damage [9]. There are many types of networks used, among which one is Radial Basis Function Neural Network (RBFNN). Its learning capacity is fast compare to other multi- layer feed forward neural network. The Probabilistic Neural Network (PNN) classification comes under the RBNN, which we are implementing in our project. It is a form of multilayered feed forward network with four layers; input layer, pattern layer, summation layer, output layer. PNN is closely related to Parzen window Probability Density Function (PDF) estimator [12]. The main objective of this work is to diagnose the epileptic seizure using EEG signals. Discrete wavelet transform helps to decompose the signal in different levels of approximation and detail coefficients. Recurrence quantification analysis is used for measuring the recurrent rate of wavelet decomposition levels. The recurrence rate of each sub bands has been used as an input in PNN for seizure detection. The performance of the network has been evaluated in terms of the accuracy, specificity and sen- sitivity detecting in unknown EEG time series.

2 LITERATURE REVIEW

Many researchers have various techniques including uncon- ventional approaches such as engineering diagnostic tech- niques, for determining patient‘s condition. A review of litera- ture includes, automatic analysis of EEG waveforms for de- tecting the epileptic seizure began in 1970s. Since then many researches has been conducted on detection of epileptic sei- zure by EEG signal processing involving various methods and algorithms. Further on [1] Gotman in (1983) measured inter- channel differences in onset times to study seizure propaga- tion. In this work, he had analyzed a technique for automated seizures detection. [2] Osorio et al. (1998) have used a measure called seizure intensity they achieved perfect detection of sei- zure and the average detection latency was 2.1 seconds eva- luated on a database of 125 patients, but they used same data for training and validation. [3] Nigam and Graupe (2004), de- scribed a method for automated non-linear detection of epi- leptic seizure from EEG signals using a multistage non-linear preprocessing filter for extracting two features namely, occur- rence frequency and relative spike amplitude, then they feed those feature to diagnostic artificial neural network for classi- fication. [4] Abdulhamit subasi et al (2005) have used fast Fourier transform and autoregressive model which have max- imum likelihood estimation to optimize the epileptic and nor- mal EEG signal which was used as an input of artificial neural network and accuracy rate was obtained grater then 92%. [5] Jahankhani et al. (2006) perform the wavelet transform de- composition of EEG signals into different sub-bands and some
statistical information were extracted from the wavelet coeffi- cient, which is used as a input of the radial basis function net- work and multilayer perceptron for the classification of epilep- tic seizure. [6] Ling Guo et al. (2010) have used a multi- wavelet transform to decompose the EEG signals to several sub signals then the approximate entropy feature was ex- tracted from each sub signals and finally extracted feature was classified with the help of artificial neural network. [7] Umut Orhan et al. (2011) implemented a new method for feature extraction called probability distribution based on Equal Fre- quency Discretization (EFD) to be used in the detection of epi- leptic seizure from EEG signals

3 PROCEDURE

A block diagram of the key stages of this work is shown in Figure 3.1. The methodology followed a traditional machine learning approach: (1) publically available data were used; (2) the data were appropriately preprocessed; (3) features were selected and extracted; (4) a classifier was trained; and (5) the performance of the system is evaluated. The entire process was iterated until acceptable performance results were achieved.

Fig. 1. Schematic illustration of the proposed method.

3.1 DATA SELECTION AND RECORDING

We have used the publicly available data described in Andrze- jak et al. [7]. The complete data set consists of five sets denoted as A-E and each one containing 100 single-channel EEG seg- ments of 23.6 s duration. The sets were selected from EEG records after purifying artifacts. These segments were selected and cut out from continuous multi-channel 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 that were carried out on five healthy volunteers. Volunteers were relaxed in awake state with eyes open (A) and eyes closed (B), respectively. Sets C–E originated from EEG archive of presurgical diagnosis. Segments in set D were recorded from within 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 only activity meas- ured during seizure free intervals, set E only contained seizure activity. Here segments were selected from all recording sites exhibiting ictal activity. 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). In this study, I used two dataset

IJSER © 2012 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 3, Issue 12, December-2012 3

ISSN 2229-5518

(A and E) of the complete dataset.

Fig.2. EEG wave at various frequencies

Fig.3. The 10-20 international system of electrode placement

2.2 Discrete wavelet transform

Here, we used multilevel one dimensional wavelet analysis using specific wavelet ‘Daubechies’(db) to decompose the sig- nal in different levels of approximation and detailed coeffi- cients. The number of decomposition level was chosen to be 5. Thus, the EEG signals were decomposed into approximation (A1-A5) and detailed (D1-D5). We used Daubechies wavelet of order 4(db4) which helps to smooth the features, more appro- priate to detect changes in EEG signals. The DWT decomposi- tion can be described as;

Where, a(i)(l) and d(i)(l) are the approximation and detail coef- ficients of resolution i respectively.

Fig.4. multilevel one dimension decomposition

2.3 Recurrence Quantification Analysis

It is a method of non linear data analysis. The RQA's parame- ter like embedding dimension (m), time delay (t), threshold, window size and window shift values are obtained to find recurrent rate. The m and t are found based on standard me- thod like false nearest neighbour (for m) and average mutual information (for t), which avoid autocorrelation effects. Thre- shold value is one of the most critical parameter. Even a small change can dramatically affect the result of RQA. Threshold value is obtained with the help of Dr. Marwan recommended principle that is normalize the data and then use a fraction of the standard deviation as the value of threshold parameter. The window size (W) short windows focus on small-scale re- currences, whereas long windows focus on large-scale recur- rences. The window shift value (WS) is the first five numbers after the data series. After finding all the parameter RR was measured;

Here, t-time delay, Pt (l) - diagonal line length

2.4 Radial Basis Function Neural Network

Probabilistic neural network (PNN) process is faster than backpropagation network and it is an inherently parallel struc- ture. Training sample can be easily added and removed in this network without any extensive training. In this work, a four layer neural network is used to classify EEGs based on the previous obtained RR. The evaluation of this proposed me- thod determined by computing the statistical parameters like

IJSER © 2012 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 3, Issue 12, ISSN 2229-5518

specificity, sensitivity and classification of accuracy.

Fig.5. Architecture of Probabilistic Neural Network

3. Experiments

In order to apply the neural networks, we used Matlab R2010a with neural network program. In this project in Matlab platform we implemented several scripts that allowed us to run the experiments.

4 Evaluation formulas

To evaluate our result we used three different formulas: speci- ficity (the capacity of correctly identified negative cases), sen- sitivity (the capacity of correctly identified positive cases), and accuracy (the proportion of correct classified instances). These formulas are mainly used in this project, making easier to compare our results with other works.

5 Result And Discussion

The phenomena of seizure activity can be physiologically un- derstood as an enhanced coherence resulting from the simul- taneous bursts of firing across a mass of neurons. The EEG shows a discharge of seizure. The EEG segment contains a seizure discharge and from the recurrent we classify the ab- sence and presence of seizure activity. First part of the project shows the result of the decomposition of DWT in the levels of approximate and detail coefficient.

Fig.6. Original normal signal and wavelet decomposition approximated and detailed levels 1 to 5

IJSER © 2012 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 3, Issue 12, December-2012 5

ISSN 2229-5518

Fig.7. Original seizures signal and wavelet decomposition approximated and detailed levels 1 to 5

The recurrent rate is the unique dynamical parameter of recur- rence plot. Recurrent rate of seizures and non- seizures data are different because seizures tend to be periodic that’s be- cause of periodic recurrent are small. The recurrent rate of each wavelet decomposition approximate coefficient (CA) and detailed coefficient (CD) levels are calculated and values are tabulated below:

Table 1. Seizure signal decomposed DWT coefficient recur- rent rate.

rent quantification analysis are one of best method for extract- ing the epileptic seizure feature. Radial basis function neural network trained the network more robustly and tolerantly then traditional backpropagation neural network. Here, PNN shows the 100% classification result between seizure and non- seizure EEGs data. The classification of PNN values are tabu- lated below:

Table 3. The performance evaluation parameters of recurrent rate for A–E data set classification

Table 2. Normal signal decomposed DWT coefficient recur- rent rate.

In this project, we used a DWT, RQA and PNN. The current work verifies that the discrete wavelet transform and recur-

5. Conclusion

In this project, a novel method for epileptic seizure detection in EEGs is proposed. Study explores the capacity of applying recurrent rate derived from discrete wavelet transform to clas- sify EEG signals. The EEG signals are decomposed into 5 sub- signals through 5-level DWT. For each approximate and de- tailed coefficient recurrent rate feature obtained, which is used as an input of PNN classifier. We used data set A and E for the classification of healthy segments and epileptic seizure seg- ments and the proposed method shows 100% classification accuracies result.

Acknowledgment

I would like to thank Dr. N. Pradhan, Senior Professor, Psy- chopharmacology department, National Institute of Mental Health and Neuro Sciences (NIMHANS), Bangalore, Karnata- ka, who has given me the much sought for opportunity to work in this prime research institute, and for guiding me with immense patience. And I am sincerely grateful to Prof. Chai- tanya Srinivas L.V, Dept. of Biomedical Engineering, VIT Uni- versity, Vellore, Tamil Nadu, India for his affectionate guid- ance and moral support throughout my project work.

Future scope

Seizure prediction with the help of EEG is however, still far from maturity and the uses of nonlinear techniques such as discrete wavelet transform and recurrent quantification analy- sis discussed in this paper should be considered exploratory. These explorations are essential, and they must continue to be pursed if there is eventually to evolve a body of techniques that can confidently be used to study signals from the brain. The methodology employed in this project is well defined for stationary time series generated by a low dimensional dy- namical system moving around an attractor. Since the epilep- tic discharge is stationary for a brief period of time this tech- nique is successful in estimating the onset of seizure dis- charge. But this methods fails in investing event related brain potential (ERP) because they are nonstationary by definition. Event related brain potential are characteristic changes in the EEG of a subject during and short after stimulus (surprising moment).

IJSER © 2012 http://www.ijser.org

International Journal of Scientific & Engineering Research Volume 3, Issue 12, December-2012 6

ISSN 2229-5518

Refrences

[1] J. Gotman, "Seizure recognition and analysis," Electroencepha- logram Clin Neurophysiol Suppl, vol. 37, pp. 133-45, 1985

[2] I. Osorio, M. G. Frei, and S. B. Wilkinson, "Real-time automated

detection and quantitative analysis of seizures and short-term prediction of clinical onset,"Epilepsia, vol. 39, pp. 615-27, 1998

[3] Nigam V, Graupe D. A neural-network-based detection of epi- lepsy. Neurological Research 2004;26(1):55–60

[4] Jahankhani P, Kodogiannis V, Revett K.EEGsignal classification using wavelet feature extraction and neural networks. In: IEEE John Vincent Atanasoff 2006 InternationalSymposium on Mod- ern Computing (JVA’06); 2006. p. 52–7

[5] Ling Guo*, Daniel Rivero, Alejandro Pazos, “Epileptic seizure

detection using multiwavelet transform based approximate en-

tropy and artificial neural networks” Journal of Neuroscience

Methods 193 (2010) 156–163

[6] Origin lab, Multi-scale wavelet decomposition, http://www.originlab.com/www/helponline/origin/en/Cate gory/MultiScale_Wavelet_Decomposition.html

[7] Umut Orhan & Mahmut Hekim & Mahmut Ozer, ‚‘‘Epileptic

Seizure Detection Using Probability Distribution Based On Equal Frequency Discretization” DOI 10.1007/s10916-011-9689- y

[8] N. Marwan, M.H. Trauth, M. Vuille, J. Kurths, Comparing mod- ern and Pleistocene ENSO-like influences in NW Argentina us- ing nonlinear time series analysis methods, Clim. Dynam. 21 (3–

4) (2003) 317–326

[9] Andrew Blais and David Mertz, -An Introduction to Neural Networks – Pattern Learning with Back Propagation Algo- rithm/, Gnosis Software, Inc., July 2001

[10] R. Esteller, Detection of seizure onset in epileptic patients from intracranial EEG signals, Ph.D dissertation, Georgia Institute of Technology, April 2000

[11] L. D. Iasemidis, “Epileptic seizure prediction and control,” IEEE Trans. Biomed. Eng., vol. 50, no. 5, pp. 549–556, May 2003.

[12] R. J. Howlett and L. C. Jain, Radial Basis Function Networks

1:Recent Developments in Theory and Applications. Berlin, Germany:Springer, 2001.

[13] N. Marwan, J. Kurths, Nonlinear analysis of bivariate data with

cross recurrence plots, Phys. Lett. A 302 (5–6) (2002) 299–307.

IJSER © 2012 http://www.ijser.org