 Open Access
 Total Downloads : 1743
 Authors : Miss. Priyanka G. Bhosale, Prof. Dr. S.T.Patil
 Paper ID : IJERTV2IS1017
 Volume & Issue : Volume 02, Issue 01 (January 2013)
 Published (First Online): 30012013
 ISSN (Online) : 22780181
 Publisher Name : IJERT
 License: This work is licensed under a Creative Commons Attribution 4.0 International License
Classification of EEG Signals Using Wavelet Transform and Hybrid Classifier for Parkinson’s Disease Detection
Miss. Priyanka G. Bhosale
Department of Computer Science and Engineering and Information Technology.
Vishwakarma Institute of Technology, Pune, India.
Prof. Dr. S.T.Patil.
Department of Computer Science and Engineering and Information Technology.
Vishwakarma Institute of Technology, Pune, India.
Abstract Feature extraction and classification of Electroencephalograph (EEG) signals for normal and abnormal person is a challenge for engineers and scientists. Various signals processing techniques have already been proposed for classification of nonlinear and nonstationary signals like EEG. In this work, Support Vector Machine (SVM) and Multilayerperceptron (MLP) based classifier was employed to detect Parkinsons Disease from background electroencephalograph signals. Signals features are extracted using the Discrete Fourier transform (DFT). And the stastical features are calculated using the percentage power than output of this stage are different frequency bands and which are the input to the classifiers. The proposed classifier show the promising classification accuracy.
Keywords Electroencephalograph (EEG), Support Vector Machine(SVM), Multilayerperceptron (MLP), Discrete Fourier Transform (DfT), Parkinsons Disease (PD).

INTRODUCTION
The growth in the population age brings the growth of number of diseases. And one of the major group of diseases that affect elderly people is that neurodegenerative diseases.
The Parkinson Disease (PD) is disorder of certain nerve cells in the part of the brain which produces dopamine. These nerve cells break down, dopamine levels drop and brain signals which are responsible for the moment become abnormal. PD usually begins in the middle or late life (after age 50). It progresses gradually for 1015 years. This results in more and more disability. Patients suffering from PD present more clinical abnormalities of moment like resting tremor, rigidity, bradykinesia and postural instability.
Electroencephalograph (EEG) is the recording of electrical activity along the scalp, produced by the firing of neurons within the brain. In clinical context EEG refers to recording of the brains spontaneous electrical activity over a short period of time, as recorded from multiple electrodes placed on the scalp. In this paper the main diagnostic application of EEG is in the case of Parkinsons disease.[1]
Brain patterns form wave shapes that are commonly sinusoidal; usually they are measured from peak to peak and normally range from 0.5to 100 ÂµV in amplitude. Brain waves
has been classified into four basic groups or bands depending on the frequency range (as shown in Fig1).[2]

beta (>13 Hz),

, alpha (813 Hz),

, theta (48 Hz),
– , delta (0.54 Hz).
Fig.1.Brain wave samples with dominant frequencies belonging to , , and bands respectively (upper to lower).
Two basics steps must be performing to analysis of EEG signals: Feature Extraction, & Signal Classification. Feature extraction can be calculated based on statically characteristics or syntax description components of domain, frequency, time and timefrequency domain can be used to extract features from EEG signals. The fourier analysis is extremely useful for data analysis, as it breaks down the signal into constituent sinusoidal of different frequencies. For sampled vector data, Fourier analysis is performed using the discrete fourier transform (DFT). The fast Fourier transform (FFT) is an efficient algorithm for computing the DFT of sequence. It is not a separate transform.[ 2.2.3]
In past few years many research groups focused their work on classifying EEG records to desired mental task classes. Several algorithms has been investigated by purpose or increasing the classification rate and accuracy.
In this work an algorithm based on Discrete Fourier Transform (DFT) and Hybrid classifier (combination of SVM & MLP) is used to detect the PD signals from EEG signals. The dataset of 10 subjects which contain normal as well as Parkinsons disease persons are included.
The rest of the paper is organized as follows section

contain methodology, section 3 contains Results and
discussion section 4 contains Future Work, and in section 5 contain conclusion.


METHODOLOGY
In the signal classification system, First the signals are enrolled in the system, this stage is called as the training stage and the unknown signals are identified in the testing stage.
Fig.2 System Architecture.
The steps involved in electroencephalograph (EEG) signal classification and detection of the Parkinsons Disease patient detection are as follows:

Feature Extraction: The input to this subsystem is the EEG signal data. Then by applying the Discrete Fourier Transform, the frequency components of that signal are obtained. Then the unwanted frequencies (i.e. the frequencies which are greater than 30 Hz) are removed. Then these frequency components are input to the next step.

Feature Vector Calculation: In this stage feature vector is obtained by applying the percentage power formula on the frequency components which are the output of the previous step. The output of this stage are different frequency bands and these are the inputs to the classifier.

Feature Classification: In this step there are two different classifiers are used to classify the signals, these are Multilayer Perceptron Backpropogation (MLPBP), and Support vector Machine Classifier (SVM). The input to these classifiers are different frequency bands. And the output of each classifier is either person is normal or having Parkinsons Disease.
Discrete Fourier transform (DFT) and Feature Vector Calculation:
Fourier analysis is extremely useful for data analysis, as it breaks down a signal into constituent sinusoids of different frequencies. For sampled vector data, Fourier analysis is performed using the discrete Fourier transform (DFT). The fast Fourier transform (FFT) is an efficient algorithm for computing the DFT of a sequence; it is not a separate transform. It is particularly used in area such as signal processing, where its uses range from filtering and frequency analysis to power spectrum estimation []. Computation using DFT of each level gives an indication to the frequencies that the bands contained in.
The DFT is used for periodic, discretetime or digital signals x[n]. The DFT for a signal with period N is
These equations say that the coefficients X[k] represent the periodic discretetime signal x[n]. Notice, that only N samples of the time signal x[n] are used to compute X[k] (for k = . .
.), and also only N of the coefficients X[k] are used in the inverse transform.

A periodic discretetime signal has a periodic spectrum (just like any other discrete time signal), i.e., X[k + N] = X[k].

The X[k] terms are evenly spaced samples of the continuous spectrum
X1(ej) of the signal x1[n] = x[n] for 0 n < N; x1[n] = 0 for n < 0, n N
Properties of Discrete Fourier Transform:

Linearity & Periodicity:
Linearity : let {X0 , X1 XN1}and {Y0 , Y1 YN1} be the two sets of discrete samples with corresponding DFT is given X(m) and Y(m) then the DFT of samples set { X0+ Y0, X1+ Y1, XN1 + YN1} is given by X(m)+Y(m).
Periodicit: We have evaluated DFT at m=0,1,. N1. There
after (M>N) is shows periodicity.

DFT symmetry:
If samples are real, then extracting in frequency domain X(0), X(N1)seems counter intuitive. Because from N bits of information in one domain (time), we are deriving 2N bits of information in frequency domain.
The suggests that there is some relationship between X(0)., X(N1) as per DFT symmetry property, X(Nm= X*(m)m = 0,1,.. N=1 where symbol * indicates complex conjugative.

DFT phase shifting:
Dft phase shifting property states that, for a periodic sequence with periodicity N i.e. X(m)=X(m+iN), I as an integer, an offset in sequence manifests itself as phase shift in the frequency domain. In other words , if we decide to sample X(n) starting at n equal to some integer k, s opposed to n=0, the DFT of those time shifted samples.
Feature Classification:
BCI Classifiers Problem :
Some problem related to BCI classifier were reported by [7] which can be summarized as follows

Course of Dimensionality and small training set: Dimensionality means that features come from various channels in different time intervals. So 0that when size of training dataset is small as compare to size of feature vector then classifier gives poor results.

BiasVariance tread off:
It is due to the noise of BCI system i.e. biasness in mapping from & variation of training set that cause error. As a solution biasness & variation must be removed.
By using classifier but the problem is that if we use stable classifier it has high biasness & low variance and if we use unstable classifier it has low biasness & high variance.
Support Vector Machine
An SVM also uses a discriminate hyper plane to identify classes. However, concerning SVM, the selected hyper plane is the one that maximizes the margins, i.e. the distance from the nearest training points. Maximizing the margins is known to increase the generalization capabilities. An SVM uses a regularization parameter C that enables accommodation to outliers and allows errors on the training set.
Fig 3. SVM and the optimal hyperplane for generalization.
Such an SVM enables classification using linear decision boundaries, and is known as linear SVM. This classifier has been applied, always with success, to a relatively large number of synchronous BCI problems. However, it is possible to create nonlinear decision boundaries, with only a low increase of the classifier's complexity, by using the kernel trick". It consists in implicitly mapping the data to another space, generally of much higher dimensionality, using a kernel
function K(x; y). The kernel generally used in BCI research is the Gaussian or Radial Basis Function (RBF) kernel:
The corresponding SVM is known as Gaussian SVM or RBF SVM.
Multilayer perceptron (MLP)
A Mulitlayer Perceptron is a feed forward artificial neural network model that maps sets of input data onto a set of appropriate output. An MLP consists of multiple layers of nodes in a directed graph, with each layer fully connected to the next one. Except for the input nodes, each node is a neuron (or processing element) with a nonlinear activation function. MLP utilizes a supervised learning technique called backpropogation for training the network. MLP is a modification of the standard linear perceptron, which can distinguish data that is not linearly seperable.
Fig.4 FeedForward Network The basic neural network consists of 3 layers.

Input layer: The input layer consists of source nodes. This layer captures the features pattern for classification. The number of nodes in this layer depends upon the dimension of feature vector used at the input. In proposed system there are four input neurons which are delta band, theta band, alpha band and beta band obtained from feature extraction stage.

Hidden layer: This layer lies between the input and output layer. The number of hidden layers can be one or more. Each hidden layers have a specific number of nodes (neurons) called as hidden nodes or hidden neurons. The hidden nodes can be varying to get the desired performance. These hidden neurons play a significant role in performing higher order computations. The output of this layer is supplied to the next layer. In this system four hidden neurons are present.

Output layer: The output layer is the end layer of neural network. It results the output after features is passed through neural network. The set of outputs in output layer decides the overall response of the neural network for a supplied input features. The given system give output is person is normal or abnormal.
Strengths and weaknesses of Classifiers:[7][8]

SVM : Strengths:

Provides maximized margin (degree of seperat) in training data.

Due to use of the kernel SVM is flexible in threshold attributes.

Best in classification and always provides single unique solution rather than multiple solution.

Generally best overall accuracy as compare to KNN and ANN.

It did not required training again and again.

Best for those BCI features which contains noise & outlier.

Low error rate.

BCI EEG data contain high dimensionality therefore SVM gives good results in case of high dimensionality and smaller training set
Weakness:

In testing phase SVM is slow.

Contain high complexity = O(n2).

Wastage of memory.

If dataset is large then no solution to train it.


MLP : Strengths:

MLP d not use assumption or guess for decision making, it is capable to provide decision directly from trained data.

Capable to learn complex decision boundaries.

Good in classification performance but not good for discriminating the non class data.

Best in signal classification.

Capable of reduce signal noise.
Weakness :

High complexity i.e. O(n2).

Only provides solution against linearly separated data.

Sensitive to overtraining especially with such noisy and nonstationary EEG data.

Recognition.

RESULTS AND DISCUSSION
In this study we used EEG signals of Normal and Parkinsons disease patients, in order to perform an abnormal patient detection and to see in which stage it belongs to. EEG recording were divided into sub band frequencies such as , , and by using DFT. Then a set of statically features was extracted from frequencies (14 Hz, stage 4), (48 Hz,
stage 3), (812 Hz, stage 2), (1230 Hz, stage 1). After normalization EEG signals were decomposed using DFT and statical feature extracted from subbands and these features are classified using the SVM and MLPBP classifier.
The objective of this application was to develop classifier that is able to identify any input combination as belonging to either one of the two classes: Normal or Parkinsons.
Dataset:
The dataset used for this system is the EEG signal recorded from 10 persons which contain normal and Parkinsons disease persons, by using the 8 channels. It means the EEG signals from the persons brain are recorded using the 8 electrodes which are placed over the cap. This dataset contain the 8 channel data for each person and each channel contains the 4096 samples with its amplitude values.
Experimental setups:
In this system Discrete Fourier transform is used to processing the input signals. The outputs of the Discrete Fourier transform are the frequency components. The unwanted frequency components are removed i.e. the frequencies which are greater than 30HZ are removed, because above that frequency the person is said to be ormal. Then these clean frequency components are given as an input to the percentage power to calculate the feature vector. The feature vector size is 1*8 for each channel. The value 8 means the four frequency bands and each having its low and high value. And these feature vectors are given as an input to the classification stage.
category 
No of traine d signal s 
No of teste d signa ls 
Corre ctly classif ied by MLP BP 
Correctl y Classifie d by SVM 
MLP BP accurac y in % 
SVM accur acy in % 
Training phase which include seizure and non seizure signals 
10 
10 
10 
10 
100 
100 
Training phase which include seizure and non seizure signals 
10 
10 
9 
10 
90 
100 
Table 1 Classification accuracy of EEG signal by using SVM and MLPBP by considering %power features
Classi fier 
Computa tion Complex ity 
Accuracy 
Speed 
Memo ry Need 
BCI Applic ablity 
SVM 
Initially high i.e. O(n2) but becomes less due to kernel tricks 
Best 
Classification = Fast
= slow Overall luckily very fast in BCI application. 
Large 
Yes 
MLP 
O(n2) 
Low learning accuracy 
Good in classification, slow in discrimination 
?? 
Yes applica ble to all BCI 
Classi fier 
Computa tion Complex ity 
Accuracy 
Speed 
Memo ry Need 
BCI Applic ablity 
SVM 
Initially high i.e. O(n2) but becomes less due to kernel tricks 
Best 
Classification = Fast
= slow Overall luckily very fast in BCI application. 
Large 
Yes 
MLP 
O(n2) 
Low learning accuracy 
Good in classification, slow in discrimination 
?? 
Yes applica ble to all BCI 
Performance analysis of classifiers:
Table 2. Performance Analysis of SVM And MLPBP Additionally, because the problem involves classification
into two classes, sensitivity, specificity, positive predicative value and negative predicative value were used as a performance measures.[13]
Positive Predictive Value (PPV) = TP / (TP + FP) Negative Predictive Value (NPV) = TN / (TN + FN) Sensitivity= TP / (TP + FN)
Specificity= TN / (TN + FP)
Classification PPV NPV Sensitivity Specificity Method
MLPBP 100% 88.889% 50% 100%
SVM 100% 100% 100% 100%
Table 3 The values of statistical parameters of MLPBP and SVM model for EEG signal Classifiers
4. FUTURE WORK
In the proposed project we tested the result for detection of the Parkinsons disease by using hybrid classifier which is the combination of support vector machine classifier and multilayer perceptron classifier. Then we can compare the results obtained by this work with the other classifiers
In this work 8 channel signals are used for the human computer interface. The process would become easier and more accurate if 16 channels are used for analysis and implementation.
5 CONCLUSION
Brain computer Interface is the simplest method to interface with any real input and output device using Electroencephalograph (EEG) signals based on the frequency, the signal can be classified into different bands of frequency. Each band will predict different conditions with the help of these frequencies control of interfaced device can be made easy and automated.
Instead of using one classification algorithm, we are combining the two classifier such as Support vector machine (SVM) & Multilayer Perceptron (MLP) because the SVM has best training accuracy and the MLP has best testing accuracy than other.
As Compare to other Classifier SVM and MLP are gives best classification accuracy. SVM gives 100% accuracy in treaning as well as testing, while MLPBP gives 100% accuracy in training phase while 90% in testing.
6 REFERENCES.
[1]. P.V.RamaRaju, V.M.Rao, Relevance of wavelet transform for EEG signals. [2]. Cesar Seijas, Antonino Caralli and Sergio Villazana,Estimation of Brain Activity usind Support vector Machines, Proceeding of the 3rd International IEEE EMBS Conference on Neural Engineering Kohala Coast, Hawaii, USA, May 25,2007.
[3]. Dr.D.S.Bormane and Prof.S.T.Patil,Pearl ensemble classifier Decision. [4]. R.Panda, P.S.Kobragade, P.D. Jambhule,Classification of EEG signals using Wavelet Transform and SVM For Epileptic seizure detection, 2010 IEEE International conference on system in medicine and Biology
[5]. Aleas Prochazka, Jaromar Kukal Oldrich Vysata,Wavelet Transform use for feature extraction & EEG signal classification.
[6]. ChiaJung Chang, Time frequency analysis and wavelet transform tutorial, Time frequency analysis for biomedical engineering, National Taiwan University. [7]. F.Lotte, M.Congedo, A.Lecuyer,F.Lamarche and B.Arnaldi, A Review of Classification Algorithms for EEGbased BrainComputer Interfaces.Publication,2007
[8]. Ijaz Ali Shoukat , Mohsin Iftikhar, Suggested Hybrid Approach for Robust Classification of EEG data for Brain Computer Interface, In Proc. of International Conf. on Bioinformatics & Computational Biology, BIOCOMP 2010, July 1215, Vol. 2, 2010, Las Vegas Nevada, USA.
[9]. JR. Panda, 2P, S. Khobragade, 3P, D. Jambhule, 4S. N Jengthe,P.R.Pal 5T. K. Gandhi, Classification of EEG Signal Using Wavelet Transform and Support Vector Machine for Epileptic Seizure Diction,Proceedings of 201 0 International Conference on Systems in Medicine and Biology 1618 December 2010, IIT Kharagpur, India.
[10]. Zahari, Nooritawati, Ihsan.,Classification of Parkinsons Disease based on Multilayer Perceptron neural Network [11]. A.Khorshidtalab, M.J.E. Salami, EEG Signal Classifocation for RealTime BrainComputer Interface Application: A Review, 2011 4th International Conference on Machatronics (ICOM), 1719 May 2011, Kuala Lumpur, Malaysia. [12]. M.M.Shaker, EEG Waves Classifier using Wavelet Transform and Fourier transform, International Journal of Biological and Medical Sciences 1:2 2006.[13]. Abdulhamit Subasi, M. Ismail Gursoy, EEG signal classification using PCA, ICA, LDA and support vector machines Expert Systems with Applications 37 (2010) 86598666