Biomedical Signal Processing and Control 9 (2014) 1–5

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Classification of ictal and seizure-free EEG signals using fractional

linear prediction
Varun Joshi a , Ram Bilas Pachori a,∗ , Antony Vijesh b
a
Discipline of Electrical Engineering, Indian Institute of Technology Indore, Indore 452017, India
b
Discipline of Mathematics, Indian Institute of Technology Indore, Indore 452017, India

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, we present a new method for electroencephalogram (EEG) signal classification based on
Received 22 April 2013 fractional-order calculus. The method, termed fractional linear prediction (FLP), is used to model ictal and
Received in revised form 29 July 2013 seizure-free EEG signals. It is found that the modeling error energy is substantially higher for ictal EEG
Accepted 20 August 2013
signals compared to seizure-free EEG signals. Moreover, it is known that ictal EEG signals have higher
energy than seizure-free EEG signals. These two parameters are then given as inputs to train a support
Keywords:
vector machine (SVM). The trained SVM is then used to classify a set of EEG signals into ictal and seizure-
Electroencephalogram (EEG) signal
free categories. It is found that the proposed method gives a classification accuracy of 95.33% when the
Fractional linear prediction (FLP)
Support vector machine (SVM)
SVM is trained with the radial basis function (RBF) kernel.
Epileptic seizure classification © 2013 Elsevier Ltd. All rights reserved.

1. Introduction multi-wavelet transform [14], the smoothed pseudo-Wigner–Ville

Epileptic seizures are the result of abnormally excessive or
synchronous neural activity in the brain. One of the widely used
methods to assess brain activity is through the Electroencephalo-
gram (EEG) signals. Detection of epileptic seizures using the EEG
signals is important for the diagnosis of epilepsy [1].
During epileptic seizures major changes occur in a patient’s
EEG signal due to synchronous electrical activity of the neurons.
One of the definite characteristics of seizure EEG signal is the
occurrence of spikes and sharp waves [2]. Detection of seizures
using EEG signals is required in both diagnostics and therapy. The
parameters extracted from EEG signals can be used as valuable
diagnostic features for automatic detection of epileptic seizure
[3]. Spectral parameters based on the Fourier transform are com-
monly used features for detection and classification of epileptic
seizure EEG signals [4,5]. However, the underlying assumption
of the Fourier transform based analysis is that the signal being
analyzed is stationary. Previous studies have shown that the fre-
quency components of EEG signal change over time i.e., the EEG
signal is a non-stationary process [6–10]. Several time–frequency
domain based methods have been developed for detection of
epileptic seizure from EEG signals. These methods include the short
time Fourier transform [11], the wavelettransform [12,13], the
∗ Corresponding author. Tel.: +91 7324240716.
E-mail address: pachori@iiti.ac.in (R.B. Pachori).
distribution [15], and the multifractal analysis and wavelet trans-
form [16,17]. The improved generalized fractal dimension has
been used for discriminating ictal EEG signals [18]. Recently,
empirical mode decomposition (EMD) based methods for clas-
sification of ictal EEG signals have also been reported in
literature [19–24].
Autoregressive models are also used to find the seizure locations
by estimating the power spectrum of epileptic EEG signals [25]. The
success of epileptic seizure detection using linear prediction error
energy [2] motivated us to use fractional linear prediction for EEG
signal modeling.
The purpose of this paper is to classify a given set of EEG sig-
nals into ictal and seizure-free categories. A new technique for
EEG signal classification is presented which is based on fractional-
order calculus. The EEG signal is passed through a fractional linear
prediction (FLP) filter. Coefficients of the filter are calculated by a
least squares analysis to get the best possible model of the sig-
nal. A prediction error is defined as the difference between the
modeled signal and the actual signal. Since the fractional linear
prediction has a low-pass nature, it cannot accurately model the
sharp changes that occur in ictal EEG signals thus increasing the
prediction error. The prediction error energy for a set having both
ictal and seizure-free EEG signals is calculated. The prediction error
energy and the signal energy of each signal are given as parame-
ters to train a support vector machine (SVM). Then a new set of
error and signal energy values is given as input to the SVM. The
SVM subsequently classifies the points of the new set into ictal and
seizure-free categories.

1746-8094/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.bspc.2013.08.006
2 V. Joshi et al. / Biomedical Signal Processing and Control 9 (2014) 1–5

2. Fractional linear prediction

Motivated by the effectiveness of fractional order modeling

techniques over linear prediction (LP) techniques for speech signals
[26], we propose to use this method for modeling EEG signals. The
aim is to have a more accurate representation of seizure-free EEG
data and hence better discrimination between ictal and seizure-free
categories. Recently, the EEG signal modeling based on LP tech-
niques has been studied for epileptic seizure detection [2]. It should
be noted that many real life signals and other phenomena have been
shown to have inherently fractional order dynamics and hence frac-
tional calculus based techniques are more suitable for modeling
these signals with greater accuracy [27,28]. Due to these reasons
we expect FLP to be a more accurate representation compared to
traditional LP technique.
There are many ways to define the fractional derivative in the
literature. The most commonly used Riemann–Liouville definition
Fig. 1. Flow chart of the proposed method.
of the fractional derivative of order p of a function x(t) can be
expressed as follows [26,29]:
coefficients by the column vector f. Then we need to solve the fol-
 t
dp x(t) 1 dm x() lowing equation to get  k :
= d (1)
dt p  (m − p) dt m (t − )p−m+1 −1
0 f = (∧T ∧) ∧T x (8)
where m − 1 is an integer, m − 1 < p ≤ m, and the Euler’s Gamma  
where ∧ = 1 2 . . .Q .
function  (z) is defined as:
 ∞
3. Support vector machine
 (z) = e−x xz−1 dx (2)
0
To classify ictal and seizure-free EEG signals using prediction
For performing numerical simulations on a computer the error energy and signal energy we use an SVM. The basic principle
Grünwald–Letnikov approximation of the fractional derivative is of an SVM is most easily understood for a two-dimensional case
commonly used. This is defined as follows: [30]. Here we require to classify a series of data points into two
[
t/h]   different classes of data. These two classes can be represented by
p −p q p A and B. The SVM method provides a boundary H between the two
D x(t) = lim h (−1) x(t − qh) (3)
h→0 q classes such that the margin is maximized. This means that the
q=0
distance between the boundary and the nearest data point in each
Now, similar to representation of output signal as a linear combi- class is maximal. The nearest data points are named as support
nation of derivatives of input signal with integer order in traditional vectors. 
continuous-time linear system, we can express a predicted EEG sig- Given a training sample set S = (xi , yi ), i = 1. . .l , where each
nal as a linear combination of its fractional derivatives as shown sample xi ∈ Rd belongs to a class y : ∈ {+1, − 1}. The boundary can be
[28]: expressed as follows [31]:

Q
ω · x + a = 0, (9)
x̂(n) = k Dpk x(n) (4)
where ω is a weight vector and a is a bias. Then the decision function
k=1
can be used to classify in to two different classes as follows:
Note that here a negative value of pk corresponds to fractional
integral of order pk . We may recast the above equation using Eq. g(x) = sign(ω · x + a) (10)
(3) as follows: In order to obtain the optimal plane we need to

Q
Q
minimize
1
||ω||2 ,
x̂(n) = k Dpk x(n) = k ˚k (n) (5) 2 (11)
k=1 k=1 subject to yi [(ω · xi ) + a] − 1≥0, i = 1. . .l
where  k are the required FLP parameters.
We may rewrite the optimization problem by the use of
These parameters can be determined by minimizing the energy
Lagrange multipliers ˇi ≥ 0 as follows [31]:Minimize
of prediction error. The prediction error is defined by:

l
−1 
l
e(n) = x(n) − x̂(n) (6) M(ω, a, ˇ) = ˇi ˇi ˇj yi yj (xi · xj ), (12)
2
and its energy is given by: i=1 i,j=1


N−1 Subject to ˇi ≥ 0, and
∈ = (e(n))2 (7) n
n=0 ˇi yi = 0.
i=1
where N is the number of samples in the signal. Then, the obtained decision function can be given as follows:
The aim is to determine  k while minimizing the prediction
l 
error energy ∈. The above equations are written more conveniently 
using matrix notation. Denote the sequence corresponding to frac- g(x) = sign ˇi yi (xi · x) + a . (13)
tional integral ˚k (n) by N × 1 column vectors k and the required i=1
 V. Joshi et al. / Biomedical Signal Processing and Control 9 (2014) 1–5 3

120

100

80

60

40
Amplitude

20

-20

-40
Original Signal
Modeled Signal
-60
Prediction Error

-80
0 100 200 300 400 500 600 700 800
Sample Number

Fig. 2. FLP modeling of seizure-free EEG signal.

If the separation into two classes is not possible by a linear (2) Polynomial kernel: The polynomial kernel can be defined as
boundary then a hyperplane needs to be created to do linear follows [32]:
separation in higher dimensions. This is achieved by using a trans- l
formation T(x) that maps the data from input space to feature space. M(x, xi ) = (xT xi + 1) (17)
If a kernel function
where l is the order of the polynomial.
M(x, y) = T (x) · T (y) (14) (3) Radial basis function (RBF) kernel: The RBF kernel can be
defined as follows [32]:
is used to perform the transformation, then the basic form of SVM
2 /2 2
can be expressed as follows: M(x, xi ) = e−||x−xi || (18)

l 
 where  controls the width of RBF function.
g(x) = sign ˇi yi M(x, xi ) + a (15)
i=1 4. Proposed method
In this paper we have used the following kernel functions:
(1) Linear kernel: The linear kernel can be defined as follows [32]: The EEG signal is passed through an FLP filter. The filter then
calculates the FLP coefficients of the signal using a least-squares
approach. The FLP coefficients are used to model the signal accord-
M(x, xi ) = xT xi (16) ing to Eq. (4). The difference between the actual signal and the

1000

500

0
Amplitude

-500

-1000

-1500
Original Signal
Modeled Signal
Prediction Error
-2000
0 100 200 300 400 500 600 700 800
Sample Number

Fig. 3. FLP modeling of ictal EEG signal.

4 V. Joshi et al. / Biomedical Signal Processing and Control 9 (2014) 1–5

modeled signal is defined as the prediction error. The energy of Table 1

Classification accuracy for different kernel functions and EEG data sets.
prediction error is estimated once the signal modeling is com-
plete. The energy of the signal is also calculated and this procedure Kernel function Set Description Accuracy Average
is repeated for the entire set of EEG signals. It should be noted (parameter) accuracy
that the impulsive nature or sharp changes in the ictal EEG sig- Set F Seizure-free 48/50(96%) 88.67%
nals will require high order of FLP in order to model the signal. Linear Set N Seizure-free 50/50 (100%)
The requirement of FLP order will be low for seizure-free EEG sig- Set S Ictal 35/50 (70%)
Set F Seizure-free 45/50 (90%) 94.67%
nals due to absence of impulses or sharp changes. For the same Polynomial
Set N Seizure-free 50/50 (100%)
order modeling error in the seizure-free EEG signals will be less (l = 5)
Set S Ictal 47/50 (94%)
compared to ictal EEG signals. The modeling error together with Set F Seizure-free 46/50(92%) 95.33%
RBF
signal energy helps us to develop a classification system in order Set N Seizure-free 49/50 (98%)
( = 0.02)
Set S Ictal 48/50 (96%)
to classify the ictal and seizure-free EEG signals. Next, we choose
50% of the signals each from the ictal category and the seizure-free
category and use their prediction error energy and signal energy
energy for each signal was calculated. The error energy and signal
as features to train a SVM. Finally, the rest of the prediction error
energy were given as inputs to train a SVM. For training 50% of the
energy and signal energy data is used for classification of the EEG
data was used, the remaining 50% data was kept for classification.
signals into ictal and seizure-free categories. We vary the kernel
The SVM can be trained using different Kernel functions and after
functions and their parameter values used for training the SVM to
trial and error it was found that the maximum classification accu-
get the highest accuracy. The performance of the method is evalu-
racy of 95.33% was obtained for radial basis function (RBF) kernel
ated through SVM classification plots and by calculating accuracy
with  = 0.02. The classification accuracy results for different ker-
(Acc), sensitivity (SEN), and specificity (SPE) values for the set of
nel functions for each set of data are summarized in Table 1. The
classified data. A flow chart of the proposed method is shown in
results show a substantial increase in accuracy as we go from the
Fig. 1.
linear kernel to the RBF kernel function.
The classification test performance of the SVM-classifier can
5. Results be determined by computation of sensitivity (SEN) and specificity
(SPE) along with accuracy (Acc). They are defined as:
To verify our proposition we did simulations on the EEG dataset
available publicly in Ref. [33]. The dataset consists of five subsets true positives
SEN = × 100 (19)
(denoted as Z, O, N, F, and S) each containing 100 EEG signals, each total positives
one having 23.6 s duration. In this study, we have used only the
subsets F, N, and S to perform simulations. The signals in the subset true negatives
F and N have been measured in seizure-free intervals from five SPE = × 100 (20)
total negatives
patients. Subset F is measured from the epileptogenic zone and
N from the hippocampal formation of the opposite hemisphere of
currectly classified
the brain. The subset S contains seizure activity, selected from all Acc = × 100 (21)
total
recording sites exhibiting ictal activity. The sampling frequency of
the EEG signals in the dataset is 173.61 Hz. These values were calculated for different kernel functions and
In this work, we have performed modeling on the first 800 sam- are presented in Table 2. The modeling of seizure-free and ictal
ples of each signal. First, each of the signals was passed through EEG data for a sample signal is shown in Figs. 2 and 3 respectively.
a fractional linear prediction filter and the optimal coefficients The classification of data into ictal and seizure-free classes for RBF
were estimated. Next, using these coefficients the prediction error kernel is shown in Fig. 4. It is clear from the Fig. 4 that the proposed

8
x 10
4.5

3.5

3
Signal Energy

2.5

1.5

1
Hyperplane
0.5 Seizure-free
Ictal

0
0 2 4 6 8 10 12 14 16 18
5
Error Energy x 10

Fig. 4. SVM classification plot.

 V. Joshi et al. / Biomedical Signal Processing and Control 9 (2014) 1–5
Table 2
Sensitivity, specificity, and accuracy values for different kernel functions.
Kernel function Sensitivity Specificity Accuracy
Linear 70% 98% 88.67%
Polynomial 94% 95% 94.67%
RBF 96% 95% 95.33%
method can be used as a diagnostic tool for detecting ictal EEG
signals.
In order to evaluate the performance of the proposed method
for classification of ictal and seizure-free EEG signals, a comparison
with the proposed method in [2] is done. The method proposed in
[2] has provided average classification accuracy of 94% for classifi-
cation of ictal and seizure-free EEG signals, whereas our proposed
method provides higher classification accuracy which is 95.33% for
classification of ictal and seizure-free EEG signals. We have com-
pared our method for classification of ictal and seizure-free EEG
signals with the method proposed in Ref. [2] with same number of
EEG signals of the same dataset.
6. Conclusion
FLP is a powerful and effective method for modeling of EEG sig-
nals. The prediction error energy arising out of this modeling and
the energy of signal are used as features to classify ictal and seizure-
free EEG signals. The classification of EEG data using error energy
and signal energy as parameters to the SVM has proved to be suc-
cessful with a maximum classification accuracy of 95.33%. Hence,
FLP promises to become an important tool for biomedical signal
processing applications. Improvements in classification accuracy
may be possible by using other Kernel functions such as Morlet
wavelet, Mexican hat etc.
References
[1] L.D. Iasemidis, D.-S. Shiau, W. Chaovalitwongse, J.C. Sackellares, P.M. Pardalos,
J.C. Principe, P.R. Carney, A. Prasad, B. Veeramani, K. Tsakalis, Adaptive epilep-
tic seizure prediction system, IEEE Trans. Biomed. Eng. 50 (May (5)) (2003)
616–627.
[2] S. Altunay, Z. Telatar, O. Erogul, Epileptic EEG detection using the linear pre-
diction error energy, Expert Systems with Applications 37 (August (8)) (2010)
5661–5665.
[3] U.R. Acharya, S.V. Sree, G. Swapna, R.J. Martis, J.S. Suri, Automated EEG analysis
of epilepsy: a review, Knowledge-Based Systems 45 (June) (2013) 147–165.
[4] V. Srinivasan, C. Eswaran, N. Sriraam, Artificial neural network based epilep-
tic detection using time-domain and frequency-domain features, Journal of
Medical Systems 29 (December (6)) (2005) 647–660.
[5] K. Polat, S. Güneş, Classification of epileptiform EEG using a hybrid system based
on decision tree classifier and fast Fourier transform, Applied Mathematics and
Computation 187 (April (2)) (2007) 1017–1026.
[6] B. Boashash, M. Mesbah, P. Colditz, Time frequency detection of EEG abnor-
malities, in Time–Frequency Signal Analysis and Processing: A Comprehensive
Reference, Elsevier, Oxford, UK, 2003, pp. 663–670, ch.15, article 15.5.
[7] R.B. Pachori, P. Sircar, EEG signal analysis using FB expansion and second-order
linear TVAR process, Signal Processing 88 (February (2)) (2008) 415–420.
[8] H. Adeli, Z. Zhou, N. Dadmehr, Analysis of EEG records in an epileptic patient
using wavelet transform, Journal of Neuroscience Methods 123 (February (1))
(2003) 69–87.
5
[9] Y.U. Khan, J. Gotman, Wavelet based automatic seizure detection in intrace-
rebral electroencephalogram, Clinical Neurophysiology 114 (May (5)) (2003)
898–908.
[10] A.T. Tzallas, M.G. Tsipouras, D.I. Fotiadis, Automatic seizure detection based on
time-frequency analysis and artificial neural networks, Computational Intelli-
gence and Neuroscience 2007 (2007) 13, Article ID 80510.
[11] R. Schuyler, A. White, K. Staley, K.J. Cios, Epileptic seizure detection, IEEE Engi-
neering in Medicine and Biology Magazine 26 (March/April (2)) (2007) 74–81.
[12] S. Ghosh-Dastidar, H. Adeli, N. Dadmehr, Mixed-band wavelet-chaos-neural
network methodology for epilepsy and epileptic seizure detection, IEEE Trans-
actions on Biomedical Engineering 54 (September (9)) (2007) 1545–1551.
[13] H. Ocak, Optimal classification of epileptic seizures in EEG using wavelet anal-
ysis and genetic algorithm, Signal Processing 7 (July) (2008) 1858–1867.
[14] L. Guo, D. Rivero, A. Pazos, Epileptic seizure detection using multiwavelet trans-
form based approximate entropy and artificial neural networks, Journal of
Neuroscience Methods 193 (October (1)) (2010) 156–163.
[15] A.T. Tzallas, M.G. Tsipouras, D.I. Fotiadis, Epileptic seizure detection in EEGs
using time–frequency analysis, IEEE Transactions on Information Technology
in Biomedicine 13 (September (5)) (2009) 703–710.
[16] R. Uthayakumar, D. Easwaramoorthy, Epileptic seizure detection in EEG signals
using multifractal analysis and wavelet transform, Fractals 21 (June (2)) (2013).
[17] R. Uthayakumar, D. Easwaramoorthy, Multifractal-wavelet based denoising in
the classification of healthy and epileptic EEG signals, Fluctuation and Noise
Letters 11 (December (4)) (2012).
[18] D. Easwaramoorthy, R. Uthayakumar, Improved generalized fractal dimensions
in the discrimination between healthy and epileptic EEG signals, Journal of
Computational Science 01 (March) (2011) 31–38.
[19] V. Bajaj, R.B. Pachori, Classification of seizure and nonseizure EEG signals using
empirical mode decomposition, IEEE Transactions on Information Technology
in Biomedicine 6 (November) (2012) 1135–1142.
[20] R.B. Pachori, V. Bajaj, Analysis of normal and epileptic seizure EEG signals
using empirical mode decomposition, Computer Methods and Programs in
Biomedicine 3 (December) (2011) 373–381.
[21] R.B. Pachori, Discrimination between ictal and seizure-free EEG signals using
empirical mode decomposition, Research Letters in Signal Processing 2008
(2008) 5, Article ID 293056.
[22] R.J. Oweis, E.W. Abdulhay, Seizure classification in EEG signals utilizing Hilbert-
Huang transform, BioMedical Engineering OnLine 10 (May) (2011).
[23] R.J. Martis, U.R. Acharya, J.H. Tan, A. Petznick, R. Yanti, C.K. Chua, E.Y. Ng, L. Tong,
Application of empirical mode decomposition (EMD) for automated detec-
tion of epilepsy using EEG signals, International Journal of Neural Systems 22
(December (6)) (2012).
[24] V. Bajaj, R.B. Pachori, Epileptic seizure detection based on the instantaneous
area of analytic intrinsic mode functions of EEG signals, Biomedical Engineering
Letters 1 (March) (2013) 17–21.
[25] A. Subasi, E. Erçelebi, A. Alkan, E. Koklukaya, Comparison of subspace-based
methods with AR parametric methods in epileptic seizure detection, Comput-
ers in Biology and Medicine 36 (February (2)) (2006) 195–208.
[26] K. Assaleh, W.M. Ahmad, Modeling of speech signals using fractional calculus,
in: Proc. 9th International Symposium on Signal Processing and Its Applications,
12–15 February 2007, Sharjah, UAE, 2007, pp. 1–4.
[27] S. Das, I. Pan, Fractional Order Signal Processing: Introductory Concepts and
Applications, Springer, Heidelberg, 2012.
[28] H. Sheng, Y.Q. Chen, T.S. Qiu, Fractional Processes and Fractional-Order Signal
Processing: Techniques and Applications, Springer, London, 2012.
[29] I. Podlubny, Fractional Differential Equations, Academic Press, San Diego, 1999.
[30] V.N. Vapnik, Statistical Learning Theory, John Wiley & Sons, New York, USA,
1998.
[31] J. Cheng, D. Yu, Y. Yang, A fault diagnosis approach for gears based on IMF AR
model and SVM, EURASIP Journal on Advances in Signal Processing 2008 (2008)
7, Article ID 647135.
[32] A.H. Khandoker, D.T.H. Lai, R.K. Begg, M. Palaniswami, Wavelet-based feature
extraction for support vector machines for screening balance impairments in
the elderly, IEEE Transactions on Neural Systems and Rehabilitation Engineer-
ing 15 (December (4)) (2007) 587–597.
[33] R.G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, C.E. Elger, Indi-
cations of nonlinear deterministic and finite-dimensional structures in time
series of brain electrical activity: dependence on recording region and brain
state, Physical Review E 64 (6) (2001), Article ID 061907.