issn: 0975-766x coden: ijptfi available online through … · 2016-07-24 · harikumar rajaguru1,...
TRANSCRIPT
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11904
ISSN: 0975-766X CODEN: IJPTFI
Available Online through Research Article
www.ijptonline.com COMPARISON OF SUPPORT VECTOR MACHINE WITH PARTICLE SWARM
OPTIMIZATION TECHNIQUE FOR EPILEPSY CLASSIFICATION FROM EEG Harikumar Rajaguru
1, Sunil Kumar Prabhakar*
2, Manjusha.M
3
1,2,3 Department of ECE, Bannari Amman Institute of Technology, Sathyamangalam, India.
Email: [email protected]
Received on 29-04-2016 Accepted on 24-05-2016
Abstract
In this work we have evaluated the Performance of Support Vector machine and Particle Swarm Optimization technique
for robust Classification of Epilepsy from EEG signals. Epilepsy is a brain disorder in which the normal neural activity is
hampered because of unusual firing of neurons. Millions of people around the world are affected by epilepsy. EEG
analysis of brain waves provides a good way to diagnose epilepsy. We have used EEG dataset of twenty epileptic patients
in this work. The EEG signals are sampled and artifacts are removed. To deal with the curse of higher dimensionality,
Power spectral density estimation is implemented here. Finally both Support vector machine with various kernels and
particle swarm optimization technique are used as post classifiers and their performance is compared. High Performance
index and Quality Value of 98.24% and 24.19 is achieved with RBF SVM when compared to value of 95.11% and
22.1697 with the PSO technique. Among SVM Kernels, the RBF with gamma value 10 performs best followed by
polynomial kernel. Hence from above investigation we find that SVM is good at Classifying epilepsy risk level than the
PSO technique. This method can be employed as cost effective way of Epilepsy diagnosis which can provide better life
for many patients.
Keywords: Epilepsy; EEG; Power Spectral Density; Support Vector machine; SVM kernels; Particle Swarm
Optimization; Performance Index; Quality Value.
I. Introduction
Epilepsy is a serious brain disorder which involves interruption of the normal activity of the human brain [1][2]. The
person experiences seizures which can be life threatening. During the epileptic attack, the neurons present in brain begins
to fire rapidly and losses their coordination [3]. This results in loss of consciousness and convulsions [4]. The symptoms of
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11905
epileptic seizure may vary from person to person and depend on the type of epileptic attack. Epilepsy is broadly classified
into two types, partial and generalized epilepsy [5].
Partial epileptic seizure involves only a part of brain while generalize seizures involves the hampering of the entire brain
activity. The early diagnosis of epilepsy is vital. The Electroencephalogram is the most widely used technique for brain
wave analysis [6]. It is a non invasive diagnostic technique with no harmful effects to the patient with little discomfort.
EEG collects the electrical activity of neurons [7]. Epilepsy leads to change in the normal EEG wave forms. A set of 21
electrodes are placed over the scalp which picks up the EEG signals form brain [8]. The brain waves are classified into four
types namely alpha, beta, gamma, theta based upon the frequency and amplitude [9].
Their normal rhythm and characteristics is altered during epileptic attack. These brain signals are analyzed for seizure
detection. This paper is further organized as follows. Section II reviews the materials used for the study and describes the
methods proposed for epilepsy risk level classification. Section III combines the result of two techniques namely SVM and
PSO and discussion is done. Section IV describes our conclusion.
II. Materials and Methods
A. EEG Data Acquisition
To investigate the performance of Support Vector Machine and Particle Swarm Optimization for epilepsy risk level
classification, we have collected EEG signals from twenty epileptic patients. These patients were under treatment and
examination in the Neurological Department of Sri Ramakrishna Hospital, Coimbatore. The paper record of EEG signals is
obtained using 16 channel EEG monitoring system. The electrode placement on the scalp follows standard 10-20 system,
adopted by American EEG Society. The raw EEG signals directly collected from the patients are contaminated by noise
and artifacts. Some of the main causes of artifacts are from movement of patients, respiratory movements, ocular
movement, EMG, ECG electrode contact impedance and interference from electronic instruments in the room [10]. The
artifact free EEG signals are selected with the help of Neurologist and scanned using Umax 6696 scanner having a
resolution of 600dpi. The EEG records obtained are filtered between 0.5Hz to 50 Hz using a band pass filter having five
poles. These signals were continuous for thirty minutes, and they are divided into two second epochs which are sufficient
to detect any changes in brain activity. The highest frequency component of EEG signal is 50Hz hence the sampling
frequency is set as 200Hz . Each and every patient has three epochs and each epoch has got 16 channel values.
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11906
This leads to large dimension of EEG data set. The proposed methodology involves reducing the dimensionality of the
preprocessed EEG signals using Power Spectral Analysis. Further Support Vector Machine using various kernels and
particle Swarm optimization technique is applied to spectral values.
The performance parameters and plot for both the techniques are compared. The overall methodology is given below in
Figure 1.
Figure 1. System Overview..
B. Power Spectral density as a DimensionalityReduction Technique.
Each sample of preprocessed EGG signal has got 400 values which contribute to large data space. In our work we have
considered power spectral values as feature. Other dimensionality reduction techniques like principal component analysis,
singular value decomposition and Independent Component Analysis fails in case of complex and non linear data [11][12].
We have opted spectral analysis for dimensionality reduction. For each sample data spectral value is determined. These
values provide the frequency information present in the data. For this purpose rectangular window is applied. The power
spectral density of a signal x(t) is given by
P(f) = Re2 [X(f)] + Im
2 [X(f)]
where X(f) is the Fourier transform of signal x(t) and P(f) is power spectral density of x(t) [13]. P (f) gives the information
present in EEG signals in frequency domain. Only the maximum spectral values are selected for further analysis [14].
This reduces the large data space. After the feature estimation both support vector machine and particle swarm
optimization are performed on them.
Raw EEG Data Collection
Preprocessing
Dimensionality Reduction
(PSD)
PSO SVM
Performance Comparison of
Epilepsy risk level detection
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11907
C. Support Vector Machine
Support Vector Machine was introduced by Vapnik in 1963. Since then it is widely used for pattern recognition,
Classification and regression [15]. SVM works on minimizing the structural risk rather than minimizing objective function
[16].
Support vector machine doesn’t require any prior knowledge of data.Hence the performance on SVM on outside data is
better when compared to other classification algorithms. A hyper plane is constructed which acts as decision surface. The
decision boundary must separate the negative and positive values such a way that the margin of separation is high . SVM is
also known as large margin Classifier [17]. Greater the margin of separation higher will be the generalization ability of
SVM classifier.
SVM can act on both the linear and non linear data well. In case of linear data simple hyper plane is enough to classify the
data accurately. The decision boundary is defined by function.
f(x) = wTx +b
The margin of separation is given by
b = 2/ || w ||
In our case, the EEG data is non linear which necessities the need for non linear Classifier. Here the kernel trick of non-
Linear SVM is employed. The kernel function maps the non linear data into high dimension state where they can be easily
separated by hyperplanes [18].
Most commonly used kernel functions are linear, Polynomial, Radial Basis function and sigmoid. In our work, we have
implemented non linear SVM using linear, Polynomial with order 2 and RBF with varying value of gamma.
The functions for the above kernels are given below [17].
Linear function: K (X, Y) = (XTY + 1)
Polynomial function: K (X, Y) = (XTY + 1)
d
Radial Basis Function: K (X, Y) = exp {γ || X – Y|| 2 }
Linear kernel is the simplest one. While polynomial kernel is slightly modified one with user defined d and c values. These
values determine the classification efficiency. Radial Basis kernel is the most commonly used which gives better accuracy
for non linear data.
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11908
The value of sigma determines the spreading of the function. In our work we have considered two values of sigma 5 and 10
respectively.
The following steps are done to implement SVM for epilepsy risk level classification.
1. For the Power spectral values k-Means clustering is carried out with different centroid values.
2. The centroid values obtained for each cluster is mapped into higher dimension using kernel function to achieve a
proper shape.
3. Finally linear separation is done with the help of SVM.
The accuracy of this algorithm is manipulated using various performance measures which are discussed in the next
section.
D. Particle Swarm Optimization
Particle Swarm Optimization technique is developed by James Kennedy and Russell Eberhart in 1995 for continuous non
linear functions [19]. It is inspired by bird flocking and fish schooling concept. It also has roots related to Evolutionary
computing and Swarm theory [20]. The birds while searching for food in the nature exchanges their position and velocity.
This inspired the formation of PSO methodology. Each Particle in the swarm is assumed to have same behavior and
characteristic [21]. Each particle explores the search space for optimal solution and keeps track of their position and
velocity which is eventually shared among particles for social interaction. The velocity function alters the position towards
the solution.
The velocity and position update equation is given below [22].
Vi (k +1) = w *Vi + C1* rand1 (…) *( Pbesti – Xi ( k)) + C2*rand2 (…)*(Gbest– Xi (k))
Xi (k + 1) = Xi (k) + Vi (k)
Where Vi (k) and Xi (k) represent the velocity and position of particle i at iteration k, Pbest is the personal best position of
particle and Gbest is the global best solution obtained by any particle in the population, w is the inertial weight which
controls the effect of previous velocity on the current velocity, rand1 and rand2 are uniformly distributed random numbers
in the range 0 to 1, C1 is known as coefficient of self recognition and C2 is called as coefficient of social recognition.
Larger inertial value gives greater global exploration and smaller one leads to local exploration.
The algorithm of particle swarm optimization is given below [23].
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11909
1. Initialize all the particles in the population with random position and velocity vector.
2. Formulate fitness function for each particle.
3. Calculate the best position achieved by each particle , if current is better than previous update pbest with current
value.
4. Calculate the best position achieved by any particle in the population if current is better than previous update gbest.
With current value.
5. Update the velocity and position of each particle using (6) and (7).
6. Terminate algorithm if desired criteria (maximum iteration or good fitness function) is achieved else go to step 3.
Until the desired target is achieved the velocity and new position is calculated iteratively. PSO algorithm is applied to
spectral values in our work. The values of C1 and C2 are set as 2 which are normally used in almost all applications. The
results of PSO applied to PSD values are tabulated below in section III. Further the comparison between different SVM
kernels and PSO technique for epilepsy risk level classification is done.
III. Results and Discussion
To evaluate the performance of different SVM classifiers and particle Swarm Optimization we determine two parameters
namely performance index and Quality value [24]. With RBF SVM we have got a higher performance index value when
compared to other SVM and PSO. For RBF SVM with gamma 10 the quality value obtained is 24.1995 and the
performance index value is 98.2486%. Following RBF SVM the Quality value achieved with Polynomial SVM of order 2
is high as 23.4378. With PSO technique the Quality value got is 22.1697 and Performance index is 97.5696. The false
alarm values for different SVM are equal to zero but in PSO we got false alarm rate of 2.2449 %. The formulas to calculate
various performance parameters [2], [17] are given below.
100
PC MC FA
PIPC
where PC – Perfect Classification, MC – Missed Classification, FA – False Alarm,
The Sensitivity, Specificity and Accuracy measures are stated by the following
100
PCSensitivity
PC FA
100
PCSpecificity
PC MC
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11910
100
2
Sensitivity SpecificityAccuracy
With SVM we got sensitivity of 100% for all three kernels linear, polynomial and RBF whereas with Particle swarm
optimization the value is 97.5696 %. Another important parameter called Quality factor is employed to measure the
robustness of classifiers.
The Quality factor is a value which shows the overall performance of the classifier. The Quality factor value depends on
four factors. The factors are false alarm detected, the perfect classification percentage , the risk levels that are missed and
the delay value encountered [2][5][17][24]. The formula is given below.
( 0.2)*( * 6* )
V
fa dly dct msd
CQ
R T P P
Where, C is scaling constant which is set to 10 as this scales the value of Qv obtained into a range which is readable.
Rfa - the number of false alarm per set,
Pdct - the percentage of perfect classification
Pmsd - the percentage of perfect risk level missed
Tdly - the average delay of the onset classification in seconds
A higher quality value means the classifier is efficient. Table I describes the performance parameters of linear SVM,
Polynomial SVM, RBF SVM and PSO methodology. Figure 1 shows the ROC of SVM kernels and PSO. It is evident that
SVM is more stable than PSO. Figure 2 shows the specificity and sensitivity measures. Figure 3 shows the plot between
Quality value and Time delay of both classifiers. The Quality value comparison and Average detection rate of SVM and
PSO is shown is Figure 4 and Figure 5 respectively.
Table-I: Performance Comparison of Support Vector Machine and Particle Swarm Optimization.
Parameters
Linear SVM Polynomial
SVM order-2
RBF SVM
with Gamma-5
RBF SVM
with Gamma-
10
PSO
Perfect Classification (%) 91.2337 94.9322 90.5046 98.2486 95.4135
Missed Classification (%) 8.8858 3.9575 6.5920 1.8735 2.2911
False Alarm (%) 0 0 0 0 2.2449
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11911
Weighted delay in Seconds 2.3055 2.1555 2.2871 2.0654 2.0443
Performance Index (%) 93.2337 94.2334 90.5046 98.2486 95.1075
Sensitivity (%) 100 100 100 100 97.5696
Specificity (%) 92.0597 96.0457 92.7735 98.3988 97.7778
Average Detection (%) 96.0885 98.0217 95.1194 99.0968 97.7086
Quality value 21.7863 23.4378 22.4150 24.1995 22.1697
Fig. 1. Sensitivity and Specificity measures of different SVM kernels and PSO technique.
Fig. 2. Quality value and Time delay measures of different SVM kernels and PSO technique.
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11912
Fig.3. Quality value of different SVM kernels and PSO technique.
Fig.4. Average Detection value of different SVM kernels and PSO technique.
IV. Conclusion
The prime aim of our work is to classify epilepsy risk level from EEG signals using different Support Vector Machine
kernels and Particle Swarm Optimization method with high performance index, Quality value and less False alarm and
missed Classification. Initially the EEG waveforms are preprocessed and digitized. Then the Power Spectral density values
for the EEG signals is calculated which are used as feature vectors. Finally SVM with different kernels and Particle Swarm
Optimization technique is applied to PSD values.The results shows that RBF SVM with gamma value 10 outperforms
Particle swam optimization technique and other SVM kernels. SVM kernels produced nil false alarm rates and PSO got
2.2449 for the same. RBF SVM has highest perfect classification rate of 98.2486% with Quality value of 24.1995. Using
this methodology a system to identify epilepsy risk levels in patients can be effectively formulated.
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11913
References
1. Ye Yaun, “Detection of epileptic seizure based on EEG signals”, 2010 3rd International Congress on Image and
Signal Processing, vol. 9, pp. 4209 – 4211, October 2010.
2. Sunil Kumar Prabhakar, Harikumar Rajaguru, “ An Efficient Classification of Epilepsy Risk Levels from EEG Signals
Using Hard Thresholding Computation Applied to Code Converters”, Proceedings of 2015 IEEE Student Symposium
in Biomedical Engineering and Sciences (ISSBES), November 4, 2015, Universiti Teknologi Mara, Malaysia
3. Leon D.Iasemidis etal., “Adaptive Epileptic SeizurePredictionSystem”, IEEE Transactions on Biomedical
Engineering, , vol. 50(5), pp.616-627, May2003.
4. A.S.Gevins, “Analysis of the electromagnetic signals of the human brain: Milestones, Obstacles, and Goals, “IEEE
Trans. Biomed Eng, vol. 31, pp. 833 850, 1984.
5. Dalia Nashat, G.Z. Khalifa, M.G. Mohamed, N. Abd Alraheem, “ Android Application for Epilepsy Detection and
Monitoring” International Journal of Advanced Reaearch in Computer Science and Software Engineering, vol. 5,
September 2015.
6. Sunil Kumar Prabhakar, Harikumar Rajaguru “ICA, LGE and FMI as Dimensionality Reduction Techniques followed
by GMM as Post Classifier for the Classification of Epilepsy Risk Levels from EEG Signals”, 2015 IEEE European
Modelling Symposium.
7. Alison A. Dingle, Richard D. Jones, Grant J. Carroll, W. Richard Fright, “A Multistage System to Detect Epileptifrom
Activity in the EEG”, IEEE Transactions On Biomedical Engineering, vol. 40, No. 12, December 1993
8. Sasikumar Gurumurthy, Vudi Sai Mathi, Rittwika Ghosh, “Analysis ans simulation of brain signal data by EEG signal
processing technique using MATLAB”, International Journal of Engineering and Technology, vol. 5 , No. 3, June-
July 2013.
9. Anders Isaksson, Arne Wennberg, Lars H. Zetterberg, “Computer analysis of EEG sisgnals with Parametric Models”,
Proceedings Of The IEEE, vol. 69, No. 4, April 1981.
10. Wanli Ma, Dat Tran, Trung Le, Hong Lin, Shang-Ming Zhou, “Using EEG Artifacts for BCI Applications”, 2014
International Joint Conference on neural Networks, July 2014, Beijing, China.
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11914
11. Arunasakthi. K, KamatchiPriya. L, “A Review On Linear and Non – Liner dimensionality Reduction techniques,
Machine Learning an Applications: An International Journal( MLAIJ) vol. 1, No. 1, September 2014.
12. Tian Siva Tian, Gareth M.James, “Interpretable Dimesion Reduction for Classifiying Function Data”,omputional
Statistics and Data Analysis. vol. 57, Issue. 1, pp. 282 – 296, January 2013.
13. Khalid M. Aamir, Mohammad A. Maud, “Efficient Spectral analysis of Quasi stationary Time Series”, International
Journal of Computer, Electrical, Automation, Control and Information Engineering, vol.1, No.2, 2007.
14. N.Faud, M.N.Taib, R.Jailani, M.E.Marwan, “ Brainwave classification for Brain Balancing Index (BBI) via 3D EEG
Model using KNN technique, International Journal of Computer, Electrical, Automation, Control and Information
Engineering, vol. 8, No. 8, 2014
15. Christopher J.C.Burges, “A Tutorial on Support Vector machines for Pattern Recognition”, Data Mining and
Knowlwdge Discovery, vol . 2, pp. 121-167, 1988.
16. Ming Guang Zhang, Wei Wu Yan, Zhan Ting Yuan, “Study of nonlinear system Identification based on Support
Vector Machine”, Proceedings of the third International Conference on Machine Learning and Cybernetics, Shanghai,
26-29 August 2004.
17. A.Keerthi Vasan, R.Harikumar, M.Logesh Kumar, “Performance Analysis of Support Vector Machine (SVM) for
Optimization of Fuzzy based Epilepsy Risk Level Classifications using different Types of kernel Functions from
EEG Signal parameters”, Proceedings of the International MultiConference of Engineers and Computer Scientists
2009,vol 1, IMECS 2009, March 18-20, 2009, Hong Kong.
18. V.Vapnik, “ The Nature of Statistical learning Theory”, Springer - Verlag, New York, 1996.
19. J. Kennedy, R. Eberbert, “ Particle swarm Optimization”, Proceedings of IEEE Interntional Conference on Neural
Networks , Piscataway, NJ, vol.4, pp. 1942-1948, 1995.
20. R. Eberhart, Y. Shi, “Evolving Artificial Neural Networks”, Proceedings of the 198 International Conference on
Neural Networks and Brain, pp. 5-13, 1998.
21. Ismail Babaoglu, Oguz Findik, Erkan Ulker, Nazif Aygul, “A Novel Hybrid Classification method with Particle
Swarm Optimization and K- Nearest Neighbour algorithm for diagnosis of Coronary Artery disease using exercise
stress test data”, Interntional Journal of Innovative computing Information and Control, vol. 8, No. 5(B), may 2012.
Sunil Kumar Prabhakar* et al. International Journal Of Pharmacy & Technology
IJPT| June-2016 | Vol. 8 | Issue No.2 | 11904-11915 Page 11915
22. R.Harikumar, T. Vijayakumar, R. Kasthuri, “Analysis of PSO in Calculation of Epileptic Risk level in EEG”,
Interntional Journal of Soft Computing and Engineering , vol.3, Issue.1, March 2013.
23. Y. Shi, Particle Swarm Optimization, IEEE Neural Networks Society, pp. 8-13, 2007.
24. Sunil Kumar Prabhakar, Harikumar Rajaguru, “Application of Linear Graph Embedding as a Dimensionality
Reduction Technique and Sparse Representation Classifier as a Post Classifier for the Classification of Epilepsy Risk
Levels from EEG Signals”, Proceedings of the International Conference on Graphic and Image Processing (ICGIP),
October 23-25, 2015, Singapore.
Corresponding Author:
Sunil Kumar Prabhakar*,
Email: [email protected]