International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017

RESEARCH ARTICLE OPEN ACCESS

Fractal Dimension and Higher Order Statistics Based Features for

Classification of Different Epileptic States
Kaushik Das1, Asif Ahmed2
1,2
(Electronics and Communication Engineering, North-Eastern Hill University, Shillong)

Abstract:
Here we have presented a method for the classification of different types of electroencephalogram (EEG)
signals in the empirical mode decomposition (EMD) domain. Here we have used a EEG dataset which is available
online, in the dataset out of five subsets we have considered three subsets forming normal, interictal and ictal
states. Here we have used some of the statistical moments like variance, skewness and kurtosis and we have also
used fractal dimension and sample entropy on the intrinsic mode functions (IMFs) which are obtained by doing
EMD on the main EEG signals. All the obtained features are feed to a support vector machine for classification of
normal and ictal states as well as interictal and ictal states. The mentioned method gives a classification accuracy
of 100% in almost all the cases for classification of these states.

Keywords — Electroencephalogram (EEG), Support vector machine (SVM), Epileptic seizure, Empirical
mode decomposition (EMD)

signal analysis have been proposed in [4], time

domain and frequency domain features along with
I. INTRODUCTION
artificial neural network (ANN) have been used for
The human brain is a highly complex system and normal and epileptic EEG signal analysis [5]. EMD
consists of billons of neurons and these are is a technique now researchers are using widely,
responsible for maintaining brains electrical charge. The measure namely, area of analytic IMF in the
EEG is an electrophysiological monitoring method complex plane has been used in order to
that contains the information about the human brain discriminate between seizure and normal EEG
activity. The EEG signal can be obtained by placing signals [6]. Computation of amplitude modulation
the sensors on the scalp or using the intracranial and frequency modulation bandwidths have been
electrodes. The EEG signals can be used for various used for the classification of seizure and normal
purposes like emotion recognition [1], brain– EEG signals [7].
computer interfaces [2], etc. Seizure is an Here firstly we will perform EMD on the EEG
neurological disorder of the brain and this occurs signals to get the IMFs, after that we will perform
when the human brain cell fires electrical impulses higher order statistical moments like variance,
three times more than that of normal one, a pattern skewness and kurtosis as well as some of the
of repeated seizure is called epilepsy. Epilepsy can features like fractal dimension and sample entropy
be caused by alcohol, drugs, brain injury and can on the IMFs and these will be used as a input to the
also be genetically. During seizure period some of support vector machine for classification. The paper
the symptoms like rapid eye blinking, shaking of is organised as follows: The EEG dataset in section
body, clenching of teeth are observed [3]. There are II, methods in section III, which includes EMD,
lots of method developed in the literature for EEG fractal dimension, sample entropy, higher order
signal analysis and classification are based on time statistical moments and support vector machine.
domain, frequency domain and time-frequency The results and discussion is in section IV and
domain. The spikes detection methods for EEG finally section V concludes the paper.

II. DATASET

ISSN: 2395-1303 http://www.ijetjournal.org Page 13

 International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017

2) Formation of upper and lower envelopes

In this paper the dataset we have used is available and el (t ) respectively, by connecting the maxima
online in [7],[8]. The dataset consists of five subsets and minima with cubic spline interpolation.
e (t ) + el (t )
(namely Z, O, N, F, and S), each containing 100 3) Calculate the local mean as a(t ) = m
single-channel EEG signals, each of having 23.6 s 4) Extract the detail h1 (t ) = x(t ) − a(t ) 2
duration. These signals have been taken out from 5) Decide whether h1 (t ) is an IMF or not by
continuous multichannel EEG recording after visual checking the two basic conditions as described
inspection of artefacts. The subsets Z and O have above.
been recorded extracranially, whereas the subsets 6) Repeat steps (1) to (4) and end when an IMF h1 (t )
N, F, and S have been recorded intracranial. The is obtained.
subsets Z and O have been acquired from surface As soon as the first IMF is derived,
EEG recordings of five healthy volunteers with define c1 (t ) = h1 (t ) , which is the smallest temporal
eyes open and closed, respectively. Subsets N and F scale in x(t). To determine the rest of the IMFs,
consist of EEG recorded during seizure free generate the residue r1(t) = x(t)−c1(t) . The residue can
intervals from epileptogenic zone and hippocampal be treated as the new signal and we can repeat the
formation of the opposite hemisphere, respectively. above explained process until the final residue is a
subset E contains signals with seizure activity. The constant or a function from which no more IMFs
signals are recorded in a digital format at a can be derived. At the end of the decomposition, the
sampling rate of 173.61 Hz. Out of five subsets we original signal x(t) is represented as
have taken three subsets named Z, F and S forming M
normal, interictal and ictal states respectively. Here
we have considered two cases first case is
∑
x(t) = cm(t) +rM (t)
m=1
(1)

consisting of normal and ictal states and second where M is the number of IMFs, c (t ) is the mth
m
case is consisting of interictal and ictal states.
IMF and rM (t ) is the final residue. Each IMF in (1)
is assumed to yield a meaningful local frequency,
III. METHODS and different IMFs do not exhibit the same
A. EMPIRICAL MODE DECOMPOSITION (EMD) frequency at the same time. Then, (1) can be written
as
It is a signal processing technique that
M
decomposes a signal so called intrinsic mode
functions. The EMD method does not require any
x(t∑) =
m=1
Am (t)cos[ϕm (t)] (2)
condition about whether a signal is stationary or
The plot of subsets Z, F and S are shown in figure
not. The main aim of the EMD method is to
1. and their empirical mode decomposition are
decompose a signal x(t) into a numbers of intrinsic
shown in fig. 2, 3 and in 4.
mode functions (IMFs). Each IMF satisfies two
basic conditions: 1) the number of maxima-minima B. FRACTAL DIMENSION (FD)
and the number of zero crossings must be the same The term “fractal dimension” is given by
or differ at most by one; 2) at any instant, the mean Mandelbrot, on the basis of fractal geometry. It has
value of the envelope formed by the local maxima been found that fractal dimension is a promising
and the envelope formed by the local minima is parameter in distinguishing the non linear and non
zero. stationary property of EEG signal [16]. In this
The EMD algorithm of the signal x(t) can be paper, the most used Higuchi’s algorithm [17] is
explained as follows [7],[9]. used for the fractal dimension calculation of EEG
1) Detect the extrema (maxima and minima) of the sequences.
dataset x(t). Let us consider a discrete time sequence X
e m ( t ) containing N data points, X = {x(1), x(2), . . . ,

ISSN: 2395-1303 http://www.ijetjournal.org Page 14

 International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017

x(N)}, the reformed time series xrn with n discrete Sample Entropy is more immune to noise and does
time interval between points is given as [17]: not depend on the data series length. A lower value
of Sample Entropy means an increased matching in
N −r 
xrn = {x(r) + x(r + n) + x(r + 2n) + ... + x(r +  n)} (3) the time series data. Sample Entropy can be given
 n  by the following equation
Here, symbol  y  represents the greatest integer SaEn(m, r, k) = − ln[Bm+1(r) / Bm(r ) ] (7)
which is less than or equal to y; r = 1, 2, . . . ,n and
it means the initial time value.
The average length J m (n) of each subsequence xrn
is defined as
b 

∑ x(r + in) − x(r − (i − 1)n) ( N − 1)

i =1
J m ( n) = (4)
b  n
where b is computed as

b = (N − r) / n (5)
Therefore, the fractal dimension of any time
sequence X can be calculated as
FD = log( J m (n)) / log(1/ n) (6)
In this paper we have taken n=5.

Fig. 2. Empirical Mode Decomposition of two ictal

EEG signals.

Fig. 1. Plot of two normal (Z), interictal (F) and

ictal (S) EEG signals.
C. SAMPLE ENTROPY (SaEn)
Sample entropy is an alteration of the
approximate entropy used for measuring the
complexity of a time-series data. Sample Entropy is
given by the negative natural logarithm of the
conditional probability means any two sequences
which are same for m points will remain same at
the next point where r is identified as similarity Fig. 3. Empirical Mode Decomposition of two
criterion and m is the length of data segment interictal EEG signals.
[11],[13]. Compared to Approximate entropy,

ISSN: 2395-1303 http://www.ijetjournal.org Page 15

 International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017

where B m ( r ) is the probability of matching two where µ 4 is a moment of fourth order around the
sequences for m points and B m +1 ( r ) is the mean.
probability that two sequence will match in the next
E. SUPPORT VECTOR MACHINE (SVM)
point i.e. for m+1 points.
In order to classify the normal and ictal as well as
interictal and ictal EEG signals, all the parameter
values obtained are given as a input to a support
vector machine. Suppose we have two classes and
an unknown feature vector which is to be classified,
then our goal is to design a hyperplane that
classifies all the training vectors in two classes, but
we may have different hyperplane so best choice
will be the hyperplane that leaves the maximum
margin for both classes [12]. The decision function
of a SVM is given by
k
l ( x) = sgn(∑ α i ri Ψ ( x, xi ) + b) (11)
j =1

where Ψ defines the kernel type, k is the number of

support vectors, xi is the input data, ri is the target
class of training dataset and b is the bias term. In
this paper we have used radial basis function kernel.
Fig. 4. Empirical Mode Decomposition of two IV. RESULTS AND DISCUSSION
normal EEG signals. Here we have calculated variance, skewness,
D. HIGHER ORDER STATISTICS OF IMFS kurtosis, fractal dimension and sample entropy for
Here we have used higher order statistical different IMFs obtained as a result of EMD and
moments like variance, skewness and kurtosis. before that we have segmented each IMF into 8
Generally variance is a measure of how far a set of parts. Here we have considered two cases first one
numbers are spread out from their mean, Skewness is consisting of normal and ictal EEG signals and
tells us about asymmetry of a probability the second one is consisting of interictal and ictal
distribution function and Kurtosis tells us about the EEG signals. The class of discrimination ability of
thickness or weight of the tails of a distribution above mentioned parameters for both the cases are
[14]. performed using Kruskal-Wallis statistical test and
Let us consider a random variable X has the mean, the p-values obtained as a result of this test are
µ = A(X), then the variance Var(X) of X is given by: shown in table I and II.
TABLE I
Var( X ) = A[( X − µ)2 ] (8) p-values obtained as a result of Kruskal-Wallis statistical test
Skewness is given by the following formula: for normal and ictal signal
µ
δ1 = 33 (9) IMF SaEn Kurtosis Skewness Variance FD
σ IMF1 8.6e-19 2.1e-38 0.0606 1.5e-77 4.3e-39
where σ is the standard deviation and µ3 is a IMF2 2.7e-15 3.9e-25 0.0783 1.7e-78 1.1e-45
IMF3 0.0811 9.0e-05 0.0349 4.0e-80 1.7e-39
moment of third order around the mean. IMF4 0.0342 0.0156 0.0140 9.1e-80 3.9e-51
Kurtosis is given by the following formula:
µ
δ 2 = 44 (10) The calculation of the sample entropy is done by
σ keeping pattern length parameter m as 1 and
tolerance parameter r as .2, after that we have

ISSN: 2395-1303 http://www.ijetjournal.org Page 16

 International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017

designed a classifier in order to classify the ictal Fig. 6. Box-plot for the comparison of normal and
and normal signal as well as interictal and ictal ictal signals using kurtosis values.
signal and to get the classification accuracy we
need the training and the test data, for that we have
selected 60% of the dataset for training and
remaining data are for testing and the results are
shown in table III.
TABLE II
p-values obtained as a result of Kruskal-Wallis statistical test
for interictal and ictal signal

IMF SaEn Kurtosis Skewness Variance FD

IMF1 7.5e-35 1.3e-61 0.0357 4.1e-79 5.0e-28
IMF2 1.8e-21 1.6e-67 0.0378 8.0e-78 2.8e-54
IMF3 0.0811 2.3e-37 0.0973 1.9e-77 1.2e-17 Fig. 7. Box-plot for the comparison of normal and
IMF4 0.2642 4.6e-04 0.4603 5.9e-60 4.8e-12 ictal signals using skewness values.
TABLE III
Classification performances of different cases of EEG signals
for the first four IMFs
IMF Ictal (S) vs Normal Ictal (S) vs Int. ictal
(Z) (F)
Accuracy(%) Accuracy(%)
IMF1 100% 100%
IMF2 100% 100%
IMF3 100% 100%
IMF4 100% 99.88%

Fig. 8. Box-plot for the comparison of normal and

ictal signals using variance values.

Fig. 5. Box-plot for the comparison of normal and

ictal signals using sample entropy values. Fig. 9. Box-plot for the comparison of normal and
ictal signals using fractal dimension values.

ISSN: 2395-1303 http://www.ijetjournal.org Page 17

 International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017

Fig. 10. Box-plot for the comparison of interictal Fig. 14. Box-plot for the comparison of interictal
and ictal signals using sample entropy values. and ictal signals using fractal dimension values.
V. CONCLUSION
In this paper, at the beginning we have applied
empirical mode decomposition on the EEG signals
to obtain IMFs, the parameters extracted from the
IMFs have been used in order to distinguish
between normal (Z) and ictal (S) as well as
interictal (F) and ictal (S). Here we have calculated
above mentioned parameters up to fourth IMF, and
got an average accuracy of 100% in order to
Fig. 11. Box-plot for the comparison of interictal discriminate between normal and ictal, also got an
and ictal signals using kurtosis values. average accuracy of 99.97% in order to
discriminate between interictal and ictal EEG
signals.

REFERENCES
[1] P. C. Petrantonakis and L. J. Hadjileontiadis, “Emotion
recognition from EEG using higher order crossings,”
IEEE Trans. Inf. Technol. Biomed., vol. 14, no. 2, pp.
186–197, Mar. 2010.
[2] Y. Li et al., “An EEG-based BCI system for 2-D cursor
control by combining mu/beta rhythm and p300
potential,” IEEE Trans. Biomed. Eng., vol. 57, no. 10, pp.
2495–2505, Oct. 2010.
Fig. 12. Box-plot for the comparison of interictal [3] Ling Guo, Daniel Rivero, Julian Dorado, Juan R.
and ictal signals using skewness values. Rabunal, Alejandro Pazos, “Automatic Epileptic Seizure
Detection in EEGs based on Line Length Feature and
Artificial Neural Network,” Journal of Neuroscience
Methods, Vol. 191, 2010, pp. 101-109.
[4] Adeli, H., Zhou, Z., Dadmehr, N., “Analysis of EEG
records in an epileptic patient using wavelet transform,”
J. Neurosci. Methods,vol . 123, pp.69-87, 2003.
[5] Srinivasan, V., Eswaran, C., Sriraam, N., “Artificial
neural network based epileptic detection using time-
domain and frequency-domain features,” J. Med. Syst.,
vol. 29, pp. 647-660, 2005.
[6] R. B. Pachori and V. Bajaj, “Analysis of normal and
epileptic seizure EEG signals using empirical mode
decomposition,” Comput. Methods Progr. Biomed., vol.
104, no. 3, pp. 373–381, 2011.
[7] V. Bajaj and R. Pachori, “Classification of seizure and
Fig. 13. Box-plot for the comparison of interictal non-seizure EEG signals using empirical mode
and ictal signals using variance values. decomposition,” IEEE Trans. Inf. Technol. Biomed., vol.
16, no. 6, pp. 1135–1142, Nov. 2012.
[8] EEG Time Series Download Page 2012 [Online].
Available: http://epileptologie-
bonn.de/cms/front_content.php?idcat=193&lang3
[9] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih,
Q. Zheng, N. C.Yen, C. C.Tung, andH.H.Liu, “The
empirical mode decomposition and Hilbert spectrum for
nonlinear and non-stationary time series analysis,” Proc.
R. Soc. Lond. A, vol. 454, pp. 903–995, 1998.
[10] S. M. Pincus, "Approximate entropy as a measure of
system complexity," Proceedings of the National
Academy of Sciences, vol. 88, pp. 2297-2301, 1991.
[11] J. S. Richman, D. E. Lake, and J. R. Moorman, "Sample
Entropy," vol. 384, pp. 172-184, 2004.

ISSN: 2395-1303 http://www.ijetjournal.org Page 18

 International Journal of Engineering and Techniques - Volume 3 Issue 6, Nov - Dec 2017
[12] C.-W. Hsu, C.-C. Chang, and C.-J. Lin, A Practical Guide
to Support Vector Classification. [Online]. Available:
http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf
[13] Vijith V S, Jisu Elsa Jacob, et al., “Epileptic seizure
detection using nonlinear analysis of EEG,” IEEE Int.
Conf. on Inventive Computation Technol., vol. 3, pp. 1-6,
2016.
[14] Mohamed Bedeeuzzaman.V, Omar Farooq, Yusuf
Uzzaman Khan, “Automatic Seizure Detection Using
Higher Order Moments,” IEEE Int. Conf. on Recent
Trends in Information, Telecommunication and
Computing, pp. 159-163, 2010.
[15] George B. Moody, Approximate Entropy. [Online].
Available:
http://physionet.caregroup.harvard.edu/physiotools/ApEn/
[16] B. B. Mandelbrot, The Fractal Geometry of Nature. San
Francisco, CA, USA: Freeman, 1982.
[17] T. Higuchi, “Approach to an irregular time series on the
basis of the fractal theory,” Phys. D, vol. 31, no. 2, pp.
277–283, Jun. 1988.

ISSN: 2395-1303 http://www.ijetjournal.org Page 19