A Novel Signal Modeling Approach For Classification of Seizure and Seizure-Free EEG Signals

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

A Novel Signal Modeling Approach for


Classification of Seizure and Seizure-free EEG
Signals
Anubha Gupta, Senior Member, IEEE, Pushpendra Singh, Member, IEEE, Mandar Karlekar

Abstract—This paper presents a signal modeling based new changes in mental state [12]. The alpha wave, for example, is
methodology of automatic seizure detection in EEG signals. The observed to be reduced in children, elderly [13], and in patients
proposed method consists of three stages. First, a multirate with dementia, schizophrenia, stroke, and epilepsy [14].
filterbank structure is proposed that is constructed using the
basis vectors of discrete cosine transform (DCT). The proposed A visual inspection and analysis of EEG signals is an
filterbank decomposes EEG signals into its respective brain empirical science. It requires a large amount of time and neu-
rhythms: delta, theta, alpha, beta, and gamma. Second, these rological knowledge [15] that makes the process of epilepsy
brain rhythms are statistically modeled with the class of self- detection tedious and subjective [16] and may lead to incorrect
similar Gaussian random processes, namely, fractional Brownian diagnosis. In order to eliminate these limitations, a computer
motion (fBm) and fractional Gaussian noises (fGn). The statistics
of these processes are modeled using a single parameter called based fully- or semi-automated analysis of EEG signals is
the Hurst exponent. In the last stage, the value of Hurst exponent required. These computer assisted methods extract features
and autoregressive moving average (ARMA) parameters are used from EEG signals in order to classify them as inter-ictal
as features to design a binary SVM classifier to classify pre-ictal, (seizure-free interval of an epileptic patient) or ictal (seizure),
inter-ictal (epileptic with seizure free interval), and ictal (seizure) respectively.
EEG segments. The performance of the classifier is assessed via
extensive analysis on two widely used dataset and is observed Fourier transform (FT) based spectral features are frequently
to provide good accuracy on both the dataset. Thus, this paper used for the classification of EEG signals. However, direct
proposes a novel signal model for EEG data that best captures application of FT may not provide correct spectral information
the attributes of these signals and hence, allows to boost the of EEG signals because 1) one may analyze each separate
classification accuracy of seizure and seizure-free epochs. segment of EEG with FT assuming that to be a deterministic
Index Terms—Fractional Gaussian noise, epileptic EEG data, signal, while EEG signals form a random process; 2) FT is not
DCT, Seizure Detection suited for the analysis of short data segments; and 3) short-
time FT leads to side-lobe leakage effects due to windowing
I. I NTRODUCTION [17]. FT can be used to compute the power spectral density of
EEG assuming it to be a stationary random process, but EEG
A seizure is an event that causes an abrupt surge of electrical is a non-stationary random process [18]. Thus, FT should be
activity in the brain, where epilepsy is the disease involv- used carefully for the analysis of EEG signals.
ing recurrent unprovoked seizures. Worldwide, approximately Several FT based spectral methods are developed to process
50 million people are affected by epilepsy [1] and this is EEG signals. For example, [19] uses the autoregressive model
one of the most common neurological diseases. Electroen- for EEG signals but it is unsuitable for non-stationary signals
cephalograph (EEG) is a recording of the electrical activity in addition to the need for finding the correct order of the
of the brain generated by cortical neurons. It is, typically, system [20]. In [21], Welch method is used to estimate the
a noninvasive diagnostic method for the determination and periodogram for the analysis and classification of EEG signals.
treatment of epileptic seizures [2]. An EEG signal is band- However, periodogram is a biased estimator of frequency
limited in frequency (0.1–60 Hz), modeled and classified into spectrum [20].
five rhythmic waves, namely, delta (0.1–4 Hz), theta (4–8 Hz), A number of methods have been developed for the classi-
alpha (8–12 Hz), beta (12–30 Hz), and gamma (30–60 Hz) fication of epileptic data [22]–[33]. In [34], wavelet packet is
waves that capture different brain activities [3]. The recording used to detect epileptic seizure. Modular energy and modular
and analysis of EEG signal is useful in diagnosing epileptic entropy are used as features in [35] for the classification of
seizures [2], [4]–[6], EEG signal classification [7] for the EEG signals. In [36], feature-waves are extracted using non-
reliable operation of brain-computer interfaces, detection and linear distribution methods. In [37], approximate entropy and
monitoring of brain injury [8], classification of sleep stages [9], sample entropy have been used to identify epileptic signals.
[10], detection of abnormal brain states [11], and detection of However, the method is heavily dependent on the record length
and the results are greatly affected by system noise [38]. The
Anubha Gupta is with the Deptt. of ECE, IIIT-Delhi, India. email: largest Lypunov exponent and correlation dimension provide
anubha@iiitd.ac.in. Lab: http://sbilab.iiitd.edu.in/ useful clinical information in order to segregate seizure data
Pushpendra Singh is with Bennett University, Noida, U.P., India. email: from the healthy [39]. However, it is observed that lower
spushp@gmail.com.
Mandar Karlekar was with the Georgia Tech. University, USA at the time of sampling rates and variable length data segments tend to distort
completion of this work. email: mandarkarlekar03@gmail.com the estimation of correlation dimension [40].

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

Empirical mode decomposition (EMD) is a recent technique limitations of the method in the ‘Discussion and Limitations’
which has caught the eye of researchers in nonlinear signal Section. Some conclusions are presented in the last Section.
analysis. The mean frequency obtained from the intrinsic mode
functions (IMFs) provides considerable distinction between the
inter-ictal and ictal groups. In [41], higher order statistics are II. M ATERIALS
calculated for different IMFs to distinguish between healthy A. Description of Dataset-1
and epileptic subjects. Amplitude modulation bandwidth and
frequency modulation bandwidth of IMFs are used as features This study uses EEG database which was recorded at the
for classification of EEG signals in [42]. However, EMD is University of Bonn, Germany for detailed study of electrical
computationally expensive. Also, the number of IMFs differs activity of the brain [54]. The complete database consists of
for every set of data. five sets (A–E) of EEG signals and each set contains 100
This paper presents a novel signal-modeling based classifier EEG segments of 23.6 seconds duration each [54]. SET A and
for seizure detection in EEG signals. The contributions of this B consist of data collected from five healthy volunteers with
paper are as follows: relaxed open and closed eyes, respectively, using international
10–20 system scalp surface electrodes placement scheme. The
1) A new 3-level multirate filterbank structure is proposed other three sets C, D and E are obtained from five epileptic
that is based on discrete cosine transform (DCT) basis patients undergoing pre-surgical evaluations using intracranial
vectors for the extraction of delta, theta, alpha, beta, and EEG recordings. The dataset C and D have been measured
gamma brain rhythms; in inter-ictal periods (seizure-free intervals) from five patients
2) A new statistical modeling of brain rhythms is proposed in the epileptogenic zone (D) and from the hippocampal
via a class of self-similar random processes, namely, formation of the opposite hemisphere of the brain (C). SET
fractional Brownian motion (fBm) and fractional Gaus- E consists of data recorded during a ictal (seizure) event.
sian noise (fGn). Delta brain waves are seen to belong to Each of these sets has 100 single-channel EEG recordings
1st order fBm processes. Rest of the brain waves, i.e., with the sampling rate of 173.6 Hz, obtained from a 12-bit
theta, alpha, gamma, and beta are modeled as higher A/D converter, resulting in 86.8 Hz bandwidth. The duration
order fGn processes. Further, this is to note that fBm of EEG duration, of each channel in all five sets, is of
are non-stationary random processes, while fGn are the 23.6 seconds and leads to 4097 non-overlapping samples in
stationary random processes. Thus, delta brain waves each channel. More information about the dataset is available
exhibit non-stationarity, while the rest of the brain waves in [54].
are stationary random processes;
3) Autoregressive moving average (ARMA) parameters
along with H values have been used as features to design B. Description of Dataset-2
a 10-fold Support Vector Machine (SVM) based classi-
fier that classifies EEG segments into pre-ictal, inter-ictal In order to test the robustness of the proposed classifier, it
(seizure free interval), and ictal (seizure) signals. has been tested on one more dataset collected from another
4) Classification accuracy has been tested on two widely source. This second dataset consists of segmented EEG time
used dataset to establish the robustness of the proposed series recordings collected at Neurology and Sleep Centre,
methodology irrespective of the dataset. Hauz Khas, New Delhi from ten epilepsy patients. The dataset
was acquired using Grass Telefactor Comet AS40 amplifica-
To the best of our knowledge, this is the first work that tion system at a sampling rate of 200 Hz with gold plated scalp
presents such a higher order fractional Gaussian noise based EEG electrodes placed according to 10-20 electrode placement
signal-modeling approach for EEG signal classification. Al- system. Signals were passed through band-pass filter (0.5 Hz
though attempts have been made previously on modeling of to 70 Hz) and then segmented into pre-ictal, interictal, and
EEG signals via fBm [43], the modeling has been restricted to ictal stages by a team of experts. There are three folders that
1st -order modeling. This study has extended this modeling to are named according to the epileptic seizure stages and each
the rhythms of EEG signals and it is shown that delta rhythms folder contains fifty MAT-files of EEG time series. Each MAT-
belong to 1st -order fBm processes; theta, alpha, and gamma file consists of non-overlapping 1024 samples of one EEG time
rhythms belong to 1st -order fGn, while beta rhythms belong series data that corresponds to 5.12 seconds of time duration.
to 2nd -order fGn. In this paper, the pre-ictal data is labeled as SET X, inter-
While classification between ictal-free and ictal dataset ictal data as SET Y, and ictal data as Z, where all these data
has been studied by many authors [44]–[48], this study has are originating from the same person. This dataset is publicly
considered all possible cases of pre-ictal, inter-ictal, and ictal, available for download [55] and many studies related to seizure
where inter-ictal and ictal classification is more important diagnosis, detection and classification [56]–[59] have been
because it may predict the coming ictal activity. performed by the authors of this dataset. More details on this
This paper is organized into five sections. The ‘Materials’ data can be referred from these papers [56]–[59].
Section describes dataset used. The ‘Methods’ Section presents
modeling of EEG signals as fBm and fGn processes, feature This study considers all possible cases of classification
identification, and design of classifier. Results are presented between pre-ictal, inter-ictal, and ictal EEG segments for both
in ‘Results’ section, while these are discussed along with the the above dataset.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

III. M ETHODS This study has extended this method to estimate H for higher
First, EEG signals and corresponding brain rhythms are order (mth -order with m ≥ 1) dfGn processes. The detailed
modeled statistically as fBm and fGn processes. Next, sta- algorithm is provided in Appendix as Algorithm-1.
tistically significant features are identified and collected based Both the public dataset considered in these paper have
on the above modeling to build the classifier. been captured without any extrinsic stimulus. As discussed
above, it is appropriate to model given EEG signals and
their constituent brain rhythms via dfBm and dfGn random
A. Statistical Modeling of EEG Signals processes. This is to note that although EEG signals have
It is known that EEG signals are random processes [18] [60] been modeled previosuly using fractal dimension assuming
that are non-stationary [18]. In general, EEG signals consist of deterministic self-similarity [62]–[67], there is no attempt to
an intrinsic stimuli related response, also called as an ongoing model these signals or constituent brain rhythms as random
brain activity, and an extrinsic stimuli related response. The processes.
intrinsic stimuli related response has been known to be a Using the ML estimation method presented in Algorithm-1,
zero mean signal [61]. Since an EEG signal is a cumulative the values of Hurst exponent H averaged over all the segments
response of millions of neurons in the brain, these signals of SET A, B, C, D, E, X, Y, and Z have been computed and
may be considered as Gaussian random processes by applying presented in Table-I.
Central Limit theorem (CLT). Thus, in the absence of any
extrinsic stimuli, EEG signals can be assumed to be zero-mean TABLE I: Average Hurst exponent H of EEG dataset
Gaussian non-stationary random processes.
Dataset SET Hurst Exponent H
Fractional Brownian motion comprise an important class 1 A 0.9159
of self-similar random processes that are zero-mean Gaussian 1 B 1.2701
non-stationary processes with self-similarity index H and with 1 C 0.9042
stationary increments (shown mathematically in (6)). The 1 D 0.9755
1 E 1.6509
parameter H is called the self-similarity index or the Hurst 2 X 1.4709
exponent and helps in writing explicit expressions of the 2 Y 1.1375
statistics of these random processes. 2 Z 1.5015
Depending on the value of H, the order of an fBm process
is specified. For example, an fBm process with H ∈ (0, 1) is It is observed that H is positive and is either between 0 to 1
called as 1st -order fBm [17], with H ∈ (1, 2) is called as 2nd - or between 1 to 2 for all the SETs. This implies that all these
order fBm, and with H ∈ (m − 1, m) is called as mth -order EEG signals belong to 1st -order or 2nd -order dfBm processes
fBm process [50], [51]. A discrete-time fractional Brownian and hence, are subsequently modeled accordingly.
motion (dfBm) is defined by considering uniformly sampled
continuous-time fBm process [50].
Another important subclass of self-similar processes are B. Extraction of EEG rhythms using Multirate Filterbank
the stationary Gaussian random processes called as fGn that It is desired to extract the following brain waves from the
are formed by computing the normalized increment of fBm given EEG signals: Delta 0.1-4Hz, Theta 4-8Hz, Alpha 8-
processes [50]. For example, Wiener process is a fractional 12Hz, Beta 12-30Hz, and Gamma 30-60Hz. In [51], [68],
Brownian motion with H = 1/2 and its normalized increment it has been shown that the basis vectors of DCT form the
provides a zero-mean stationary white Gaussian noise, i.e., an eigenvectors of the covariance matrices of dfBm processes in
fGn noise. Correspondingly, for a discrete-time fBm process, the asymptotic sense. Further, it has been shown in [69] that
discrete-time fGn is computed by considering the difference if a multirate filterbank is designed using the eigenvectors of
of dfBm. For example, a 1st -order dfGn (1-dfGn) process is the covariance matrix of a dfBm process, the corresponding
defined as the 1st -order difference of dfBm [50] and 2nd -order subband outputs are statistically uncorrelated with respect to
difference of a dfBm process provides the dfGn (2-dfGn). each other at all time instants. Since this study has been able
In order to distinguish dfBm, 1-dfGn, and 2-dfGn based on to model EEG signals as dfBm processes (Table-I) and that it
the value of Hurst exponent H, following ranges of H have is desired to extract brain rhythms that are uncorrelated with
been considered: 0<H<1 for dfBm, -1<H<0 for 1-dfGn, and respect to each other, DCT basis vectors are utilized to design
-2<H<-1 for 2-dfGn, i.e., it is assumed that the spectrum a multirate filterbank structure [70] to derive brain rhythms
of all these random processes is proportional to 1/|f |2H+1 , from the given EEG signals.
irrespective of whether they are dfBm or dfGn processes of The block diagram (Figure 1) of the proposed multirate
any order. filterbank provides all the five brain rhythms. Here, input EEG
In [53], Lundahl et al. [53] used maximum likelihood signal is first passed through a 2-channel DCT-based uniformly
(ML) method for the estimation of Hurst exponent of 1- decimated filterbank that decomposes the input EEG signal of
dfBm process. Since ML can be applied on stationary random 0-60Hz into 0-30Hz signal and 30-60 Hz (Gamma rhythm)
processes and 1-dfGn is a stationary process, Lundahl et al. signal. This Gamma rhythm is obtained from the highpass
[53] computed H of the successive difference of the given analysis subband output.
dfBm process, i.e., the 1-dfGn. The Hurst exponent of the The lowpass output of first level (0-30Hz signal) is passed
1-dfBm is then computed as H1−df Bm = H1−df Gn + 1. through a 5-channel uniformly decimated DCT filterbank. The

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

f31[-n] 3
delta rhythm
0-4Hz
f32[-n] 3
0-12Hz theta rhythm
4-8Hz
f21[-n] 5 5 f21[n] f33[-n] 3
alpha rhythm
8-12Hz
f22[-n] 5 5 f22[n]

f23[-n] 5 5 f23[n]

f24[-n] 5 5 f24[n]
Input EEG signal 0-30Hz
0-60Hz signal
f11[-n] 2 f25[-n] 5 5 f25[n]
Beta rhythm
12-30Hz
f12[-n] 2
Gamma rhythm
Level-1 30-60Hz Level-2 Level-3

Fig. 1: DCT basis vector based 3-level multirate filterbank structure for the extraction of EEG brain rhythms

higher three frequency bands are combined using the corre- the system parameters. Since all the subband signals
sponding synthesis branches to obtain beta band. In the end, except for the Delta band are WSS random processes,
a 3-channel uniformly decimated DCT filterbank is utilized to ARMA modeling is carried out for all these subbands
obtain delta, theta, and alpha brain rhythms. The filters used (except for the delta band). A (2,1) order ARMA system
in each of the three levels of this composite filterbank are the is assumed and the parameters are estimated using the
DCT basis vectors given in Appendix and the filter responses prediction error minimization method. Delta rhythms are
are shown in the supplementary material. largely observed to belong to 1st order dfBm process.
This implies that the first successive difference of these
C. Modeling of EEG Brain Rhythms rhythms will lead to stationary 1st -order fGn process.
Hence, ARMA parameters of delta rhythms have been
Next, these brain rhythms are modeled based on their statis-
computed from the the 1st successive difference signal
tical properties. In [69], it has been shown that a dfBm process
xd [n] computed as:
when passed through a multirate filterbank with eigenvectors
of its covariance matrix as filters leads to stationary processes x[n] = xd [n] − xd [n − 1]. (2)
in all the subbands except for the lowpass branch. This implies
that except for the delta band, other brain rhythms should E. Design of Classifier
comprise zero-mean, Gaussian, stationary random processes,
The complete method from EEG brain rhythm modeling to
i.e., the fGn processes. In order to ascertain this fact, the Hurst
building the classifier is shown in Figure-2.
exponent H of brain rhythms for all the four sets SET C, E,
Y, and Z using Algorithm-A outlined in Appendix.
Features
Input DCT Based
Hurst Exponent Class 1
EEG Filterbank—
D. Feature Identification and SVM
signal EEG
ARMA Classifier
Following features have been chosen for building the clas- 0-60Hz Rhythms Class 2
Parameters
Extraction
sifier: Extraction
1) Hurst exponent: The H value has been computed for all
brain rhythms and is chosen as a feature. Fig. 2: Block diagram of the Method
2) Autoregressive moving average (ARMA) model: ARMA
models are mathematical models for predicting the be- A binary SVM classifier is built using the features listed
havior of a wide sense stationary (WSS) random process above for all possible SET comparisons: A vs B, A vs C, A
using its past and present samples [17]. This model vs D, A vs E, B vs C, B vs D, B vs E, C vs D, C vs E, D vs E,
assumes that a given random process has been produced AB vs CD, AB vs E, CD vs E, ABCD vs E, X vs Y, X vs Z, Y
by passing white noise through a pole-zero filter of order vs Z, and XY vs Z. Results are computed with an RBF (radial
(p,q) where p denotes the number of poles and q denotes basis function) kernel based 10-fold SVM classifier (repeated
the number of zeros used in the filter. For example, the 10 times).
ARMA model equation of order (2,1) is given as below: The value of H and ARMA parameters are used as features
in each subband. In the delta band, ARMA modeling has been
y[n]+a1 y[n−1]+a2 y[n−2] = w[n]+b1 w[n−1], (1)
done after computing the first successive difference. This leads
where y[n] denotes the output WSS random process, to 100 feature vectors of dimension four for all subbands for
w[n] is the input white noise, and a1 , a2 , and b1 are sets A to E and 50 feature vectors for sets X to Z.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

Performance is quantified using standard measures, namely, Table-II: Overall performance of classifier for all 18 test cases (14 cases of
sensitivity, specificity, and accuracy. dataset-1, and 4 cases of dataset-2)
Cases Sensitivity Specificity Accuracy
TP A vs. B 0.9710 0.9520 0.9615
Sensitivity(sen) = × 100% (3) A vs. C 0.9700 0.9600 0.9650
TP + FN A vs. D 0.9790 0.9890 0.9840
TN A vs. E 0.9540 0.9430 0.9485
Specificity(spec) = × 100% (4) B vs. C 0.9960 0.9900 0.9930
TN + FP
B vs. D 1.0000 0.9900 0.9950
TP + TN
Accuracy(acc) = × 100% (5) B vs. E 0.9930 0.9870 0.9900
TP + TN + FP + FN C vs. D 0.7620 0.7490 0.7555
C vs. E 0.9800 0.9700 0.9750
where TP, TN, FP and FN stand for true positive (e.g., inter- D vs. E 0.9650 0.9620 0.9635
ictal classified as inter-ictal), true negative (e.g., ictal classified AB vs. CD 0.9745 0.9795 0.9770
as ictal), false positive (e.g., ictal classified as inter-ictal) and AB vs. E 0.9740 0.9715 0.9727
false negative (e.g., inter-ictal classified as ictal), respectively. CD vs. E 0.9685 0.9700 0.9692
ABCD vs. E 0.9797 0.9760 0.9779
X vs. Y 0.7600 0.7320 0.7460
IV. R ESULTS X vs. Z 0.7880 0.8060 0.7970
Y vs. Z 0.9720 0.9580 0.9650
The average Hurst exponent value for each of the brain XY vs. Z 0.8930 0.9430 0.9180
rhythms is computed (Table-A1, Appendix). It is noted that
the delta band has positive H value for most of the dataset. TABLE-III: Classifier Performance: Inter-ictal vs. ictal (in %)
This confirms that delta rhythm is a non-stationary zero-mean, Dataset-1: Inter-ictal (D) vs. ictal (E)
Gaussian fBm process that absorbs all the non-stationarity of (%) Alpha Beta Gamma Delta Theta Overall
Sen 81 82 88 95 90 97
EEG signals. Further, it is noted that the H value of rest of Spec 79 77 89 91 88 97
the brain rhythms are negative. This implies that these brain Acc 80 79 89 93 89 97
rhythms constitute stationary dfGn processes. These observa- Dataset 2: Inter-ictal (Y) vs. ictal (Z)
tions are in consonance with the theoretical results of [69] and (%) Alpha Beta Gamma Delta Theta Overall
hence, provides validation to the modeling proposed in this Sen 72 67 86 97 95 97
Spec 60 74 90 98 94 96
paper. Further, it is noted that theta, alpha, and gamma rhythms Acc 65 70 88 97 94 97
belong to 1st -order dfGn processes with H ∈ (−1, 0) and beta
rhythms belong to 2nd -order dfGn with H ∈ (−2, −1).
Next, the mean values of ARMA parameters a1 , a2 , and
indeed useful. Further, ARMA parameters are also found
b1 are estimated for both the dataset (Table-A2 and Table-
to be statistically significant features for the required SET
A3, Appendix). The statistical significance of the features
comparisons. Out of the three ARMA parameters, about two
extracted, on the binary SET classifications for all cases, was
parameters are observed to be significant in different brain
also carried out.
rhythms (except for the beta rhythm) for all SET comparisons.
In general, an independent sample t-test is used to test the
In general, H fails to distinguish between brain activities in
statistical significance of the features with Null Hypothesis
the alpha rhythm, while ARMA parameters fail in the beta
that the mean of the feature across the two SET is same. This
rhythm.
test requires the feature to be normally distributed. In case
Two SET comparisons C Vs D and X Vs Y are observed
of failure of the normality assumption, Mann Whitney U test
to have low accuracy results of around 75%. These results
(Wilcoxon Rank Sum Test) can be conducted on the median
are understandable because these SET comparisons have only
of the feature. Since all the features did not follow the normal
a few significant features, 3 significant features, respectively
distribution, Wilcoxon Rank Sum Test at 5% significance level
(Table-A4, Appendix). SETs C Vs D comparison is between
(p < 0.05) is used. In addition, Bonferroni-Holm correction
two inter-ictal SETs. Although these SETs are captured from
for multiple comparisons was used to control the family-wise
different regions of the brain, these have been captured simul-
error rate at 5% significance level (p < 0.05). These results
taneously from the same subjects and hence, may not have
are shown for all the features on all possible SET comparisons
much differences in the brain activity.
in Table-A4, Appendix.
SET B could be classified from SETs C, D, and E with
Next, classifier performance for all possible SET compar-
approximately 100% accuracy. This shows that the designed
isons is computed (Table-II). Since inter-ictal versus ictal is
features are able to easily distinguish pre-ictal (SET B) from
generally more difficult classification problem, results of a
inter-ictal (C and D) and ictal (E) brain activity.
classifier trained on the features of only one of the brain
Between inter-ictal and ictal comparisons (D Vs E and Y
rhythms for both the dataset are also computed (Table-III).
vs Z), it is observed that results of sensitivity, specificity,
and accuracy of the classifier built using only delta band
V. D ISCUSSION AND L IMITATIONS features are above 95% (Table-III). This shows very good
The statistical significance testing on features indicates that differentiation ability of delta rhythms. In fact, the slow wave
parameter H is a statistically significant feature for beta, abnormalities associated with brain dysfunction are found
gamma, delta, and theta rhythms for almost all the SET mainly in the delta rhythm.
comparisons (Table-A4). This shows that this modeling is Likewise, good classification results of above 90% are

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

observed on D Vs E and Y vs Z with theta rhythm features. used as features for classification of ictal and inter-ictal EEG
It has been reported earlier that larger contingents of theta dataset with 10-fold cross validation (repeated 10 times) in
waves are abnormal and are caused by pathological problems SVM classifier and are shown to yield an averaged accuracy
in waking adults [20]. Also, for various brain disorders, theta of 97% on both dataset-1 and dataset-2 (Table-III).
and delta waves become more active [20]. Further, seizure may The overall averaged performance of classifier for a total of
occur at different frequencies but generally occur in the range 18 tests cases (14 cases for dataset-1, and 4 cases for dataset-
of 3-6 Hz and are, therefore, best captured in the delta and 2) is computed (Table-II). The performance of the classifier
theta ranges [20]. Correspondingly, results of this study also built using the proposed methodology is observed to be above
indicate that these rhythms should be analyzed carefully for 95% in 14 combinations. SVM classifier is used with three
building robust classifiers and diagnostic tools. different types of kernels, namely, the Gaussian, polynomial
The accuracy in the alpha band is lower but that is no and linear kernels, and it is observed that the overall averaged
surprise because alpha waves are usually found in the occipital sensitivity, specificity, and accuracy are almost the same with
lobes [20], while the above used dataset is a single channel all these kernels. The LS-SVM classifier was also tried with
EEG dataset from other lobes. This is one of the reasons for all three kernels and results were observed to be almost the
the poor classification results in the alpha band. same as reported in this study with no significant change.
Authors in [71] studied ictal and inter-ictal EEG dataset (C Thus, the proposed method is robust and is almost inde-
versus E) classification using the permutation entropy features pendent of the type of kernels and classifiers. All the results
and SVM. They reported classification accuracy of 88%. reported are obtained from the SVM classifier with Gaussian
Authors in [72] used discrete wavelet transform-based approx- kernel (Table-II and Table-III). The four least classification
imate entropy (ApEn) features with artificial neural network accuracy, obtained amongst 18 cases of both dataset, are 75.5%
to classify the EEG dataset and obtained 100% (A versus E), in C vs. D, 74.6% in X vs. Y, 79.7% in X vs. Z, and 91.8% in
99.2% (B versus E), 100% (C versus E) and 95% (D versus E) XY vs. Z sets (Table-II). Accuracy was observed to be more
accuracy. The least squares support vector machine (LS-SVM) than 95% in all other cases and reaches up to 99.5% in B vs.
model-based methods coefficients are used for classification of D sets classification. The averaged classification accuracy of
dataset (A versus E) and 99.56% accuracy obtained by authors seizure and seizure-free (i.e. ABCD vs. E) sets in dataset-1
in [46]. Authors in [45] used discrete wavelet transform, is 97.79%. Thus, the robustness and efficacy of the proposed
principal component analysis (PCA), independent component method is validated thoroughly on two dataset with 18 cases
analysis (ICA), linear discriminant analysis (LDA) and SVM using the SVM classifier and by employing three different
for classification of EEG dataset (A versus E) and obtained kernels. Results are consistent on both the dataset as well as
98.75% (PCA), 99.50% (ICA), and 100% (LDA) accuracy. are consistent with the theory of EEG signals.
Line length features with artificial neural networks are used Proposed method is computationally efficient. Although
by authors in [44] for classification of EEG dataset and coding was not optimzed for best performance, computation
obtained 99.85% (A versus E) and 97.77% (ABCD versus E) of features on SET A consisting of 100 sample functions of
accuracy. Authors in [42] used EMD and LS-SVM approach 4097 samples took 100.13 seconds on an Intel (R) Core (TM)
to classify EEG dataset (ABCD versus E) and obtained 100% i5-6200U CPU @2.3GHz; RAM 8GB; 64 bit OS Windows
classification accuracy. Permutation entropy and SVM are used 10 on MATLAB2017a. The classifier design for SET A Vs.
in [47] for EEG dataset classification and obtained 93.55% (A E took 14.84 sec including training and testing time. Thus,
versus E), 82.88% (B versus E), 88% (C versus E), and 79.94% the method can be easily implemented for real-time epileptic
(D versus E) classification accuracy. seizure diagnosis.
Recently, authors in [73] studied the same EEG dataset (C
versus E) classification using the features mean frequencies
and RMS bandwidths of EEG rhythms derived form the fast Limitations
Fourier transform (FFT) and reported 100% accuracy. The best
classification accuracy for inter-ictal and ictal EEG dataset-2 The work has some limitations. Firstly, the proposed method
(Y versus Z) classification is 99.6% which was reported in [56] has been used in two dataset where dataset-2 is relatively
using the standard deviation and root-mean-square features small. Thus, training and testing of the classifier built may
with a general regression neural network (GRNN) classifier. not be sufficient to produce consistent and optimal results.
All these studies have used either dataset-1 or dataset-2 for Secondly, although we have assumed the EEG signals to
classification of ictal, inter-ictal, and normal EEG. They have belong to a class of self-similar random processes, this is
derived features from the dataset either directly or by diving an oversimplified assumption owing to two reasons: 1) the
the each channel data into EEG rhythms. self-similarity index is computed using given (short) sample
This study has developed a novel signal-modeling based functions and 2) in general, these may belong to multifractal
classifier approach for seizure detection in EEG signals. First, random processes. Hence, the proposed methodology needs to
a new method of 3-stage DCT based multirate filterbank is be validated with larger dataset and on different classification
proposed to decompose the EEG signals into brain rhythms. problems such as real-time epileptic seizure diagnosis, focal
Next, the theta, alpha and gamma rhythms are modeled as 1st and non-focal EEG signals classification, along with some
order fGn, while the beta rhythms are modeled as 2nd order other brain computer interface (BCI) classification problems.
fGn processes. The Hurst exponent and ARMA parameters are We plan to take up these problems in the near future.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

VI. C ONCLUSIONS A discrete-time fractional Brownian motion (dfBm) is de-


This paper has presented a new methodology of modeling fined by considering uniformly sampled continuous-time fBm
EEG signals as fractional Brownian motion random processes process [50]. Since a dfBm process is a non-stationary random
to describe EEG data that allows to enhance classification process, its auto-covariance is a function of two time indices
accuracy of seizure and seizure-free epochs. DCT basis vectors and is given as below [50]:
are used as filters to design a 3-level multirate filterbank. 2
σH
This composite multirate filterbank structure is utilized to rH [n1 , n2 ] = (|n1 |2H − |n1 − n2 |2H + |n2 |2H ), (7)
2
decompose EEG signal into its constituent brain rhythms
that are modeled as discrete-time fBm and higher order fGn 2 1
where σH = var(ZH [1]) = .
random processes. Γ(2H + 1)|sin(πH)|
The Hurst exponent H characterizing these brain rhythms
and ARMA parameters are used as features in an SVM B. Discrete-time fGn Processes
classifier to classify ictal and inter-ictal EEG segments. The
classifier is tested and validated successfully on data available A 1st -order dfGn (1-dfGn) process is defined as the 1st -
from two public dataset. Good results on both the dataset order difference of dfBm [50] as below:
underscore the robustness of the proposed signal modeling 4
and classification method. The classifier provides an overall Y1 [n] = ZH [n + 1] − ZH [n]. (8)
accuracy of 97% in both dataset for ictal and inter-ictal EEG
The covariance sequence of a 1-dfGn process can be written
sets classification. The classification accuracy of seizure and
easily using (7) and (8) as below:
seizure-free (i.e. ABCD vs. E) sets in dataset-1 is 98%. The
proposed method has been validated with two dataset with 18 2
σH
different classification cases, produced good consistent results rY1 [n] = (|n + 1|2H − 2|n|2H + |n − 1|2H ). (9)
2
in terms of averaged performance parameters, namely, sensi-
tivity, specificity and accuracy. It is also found to be robust A 1-dfGn process is a wide-sense stationary random process
and almost independent of the type of kernels (Gaussian, with power spectral density falling as 1/|f |2H−1 where H ∈
polynomial and linear kernels) and classifiers (SVM and LS- (0, 1) [50]. In other words, the spectrum of 1-dfGn can be said
SVM) used in this study. to be proportional to 1/|f |2H+1 with H ∈ (−1, 0).
In summary, this paper has made a significant contribution Similarly, 2nd -order difference of a dfBm process provides
in modeling of EEG signal rhythms based on higher order the dfGn (2-dfGn) defined as below [50]:
fractional Gaussian noise processes and extraction of rhythms 4
via multilevel DCT based multirate filterbank. EEG signal Y2 [n] = ZH [n + 2] − 2ZH [n + 1] + ZH [n]. (10)
classification on different dataset establish the robustness of
Correspondingly, the covariance of 2-dfGn process is given
the proposed methodology.
by:
To demonstrate the efficacy of the proposed method, in this
study, we have considered two dataset of EEG data classifica- 2 
σH
tion. The directions of future research would be (1) to apply rY2 [n] = − |n + 2|2H + 4|n + 1|2H
2
the proposed method to a large dataset for seizure and seizure- 
− 6|n|2H + 4|n − 1|2H − |n − 2|2H . (11)
free EEG signals classification, (2) to use the proposed method
for classification of focal and non-focal EEG signals, along
The spectrum of a 2-dfGn is directly proportional to
with some other nonstationary signal classification problems.
1/|f |2H−2 with H ∈ (0, 1) or it can be said to be directly
proportional to 1/|f |2H+1 with H ∈ (−2, −1). In general, an
A PPENDIX mth -order dfGn is defined as
A. Discrete-time fBm Processes k  
4 X k
Fractional Brownian motion are denoted symbolically as Yk [n] = (−1)k−j ZH [n + j] (12)
j
j=0
d
ZH (at) = aH ZH (t), (6)
with the covariance given by
where a is a scaling constant greater than zero and H is called k
2
 
the self-similarity index or the Hurst exponent. This parameter σH X 2k
rYk [n] = (−1)k (−1)j |n + j|2H . (13)
helps in writing explicit expressions of the statistics of the 2 k+j
j=−k
random process ZH (t). The equality in (1) holds for all finite
order distributions in the statistical sense [49]. In order to distinguish dfBm, 1-dfGn, and 2-dfGn based on
Depending on the value of H, the order of an fBm process is the value of Hurst exponent H, we considered 0<H<1 for
specified. These processes are non-stationary with an averaged dfBm, -1<H<0 for 1-dfGn, and -2<H<-1 for 2-dfGn, i.e.,
power spectral density (PSD) proportional to 1/|f |2β , where we assumed that the spectrum of all these random processes
β = 2H + 1 and f is frequency (in Hertz) [52]. Hence, these is proportional to 1/|f |2H+1 , irrespective of whether they are
processes are also called as 1/f processes. dfBm or dfGn processes of any order.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

TABLE-A1: Hurst exponent H of all brain rhythms


Brain Dataset 1 Dataset 2
Rhythms SET A SET B SET C SET D SET E SET X SET Y SET Z
Alpha -0.5098 -0.5004 -0.5016 -0.5059 -0.5141 -0.4980 -0.5094 -0.5067
Beta -1.4637 -1.3926 -1.4163 -1.4002 -1.4020 -1.3408 -1.3750 -1.3785
Gamma -0.5418 -0.3561 -0.2556 -0.1958 -0.2154 -0.0978 -0.2217 -0.1339
Delta 0.1811 -0.1702 0.5339 0.4706 -0.1584 0.2565 0.1844 0.5596
Theta -0.7013 -0.7148 -0.5455 -0.5678 -0.6944 -0.6454 -0.6909 -0.5247

TABLE-A2: ARMA parameters of Dataset-1


Brain A B C D E
Rhythms a1 a2 b1 a1 a2 b1 a1 a2 b1 a1 a2 b1 a1 a2 b1
Alpha -0.3613 -0.1176 -0.4363 -0.1537 -0.1478 -0.2437 0.0405 -0.1176 0.0096 0.0665 -0.1245 0.0395 -0.2252 -0.0801 -0.2946
Beta -0.3557 0.4860 -1.000 -0.4386 0.3924 -1.000 -0.4065 0.3815 -1.000 -0.4219 0.3660 -1.000 -0.4269 0.4523 -1.000
Gamma -0.6305 0.4593 -0.4970 -1.0211 0.6331 -0.8697 -0.7616 0.2763 -0.3313 -0.7095 0.3161 -0.1678 -0.4602 0.4227 0.3154
Delta* -0.2614 0.0923 -0.6832 0.5215 0.3852 -0.1305 -0.5511 0.3866 -0.3590 -0.4727 0.4335 -0.3328 -0.3820 0.3712 -0.8746
Theta 0.3717 0.1892 0.0284 0.7206 0.3221 0.3257 -0.5217 0.3042 -0.4603 -0.4749 0.3286 -0.4390 -0.2666 0.1741 -0.5228
* ARMA parameters of Delta band are calculated after computing its first difference.
TABLE-A3: ARMA parameters of Dataset-2
Brain X Y Z
Rhythms a1 a2 b1 a1 a2 b1 a1 a2 b1
Alpha 0.1684 -0.1809 0.1610 0.3785 -0.1759 0.3739 -0.0804 -0.1297 -0.1195
Beta -0.4375 0.2929 -1.000 -0.4347 0.2754 -1.000 -0.4319 0.3480 -1.000
Gamma -1.150 0.5426 -0.4325 -1.0874 0.4636 -0.7115 -0.5720 0.2981 0.1265
Delta* -0.1746 0.4084 -0.0594 -0.0219 0.4933 -0.0162 -0.6995 0.1797 -0.6414
Theta -0.0573 0.3510 -0.0136 0.0709 0.4086 0.0356 -0.5751 0.1471 -0.5940
* ARMA parameters of Delta band are calculated after computing its first difference.

Table-A4: Statistical significance testing of features


Band feature A Vs B A Vs C A Vs D A Vs E B Vs C B Vs D B Vs E C Vs D C Vs E D Vs E X Vs Y X Vs Z Y Vs Z
H             
a1             
Alpha
a2             
b1             
H             
a1             
Beta
a2             
b1             
H             
a1             
Gamma
a2             
b1             
H             
a1             
Delta
a2             
b1             
H             
a1             
Theta
a2             
b1             
: Passed the test; : failed the test;
Bonferroni-Holm correction test has been used for multiple comparisons to control the family-wise error rate at level p≤0.05.

C. Estimation of Hurst Exponent for Higher Order Self- (pdf) as below:


Similar Processes
−(yT K−1 y)
1
In this study, we have extended the maximum likelihood p(y; H) = e 2 . (14)
(ML) method of estimating Hurst exponent of 1-dfBm process (2π)N/2 |K|1/2
[53] to estimate H for higher order (mth -order with m ≥ 1)
where K = E[yyT ] is the covariance matrix.
dfGn processes with detailed steps as below.
2) In order to estimate H, the logarithm of p(y; H) given
Algorithm-A: Step-1: below needs to be maximized (ML criterion).
1) Construct a vector y={Ym [0], Ym [1], ..., Ym [N − 1]} of −N 1 (yT K−1 y)
an m-dfGn sample function. The vector y is zero-mean log p(y; H) = log (2π) − log |K| − .
2 2 2
Gaussian distributed with probability density function (15)

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

3) The covariance matrix K of an m-dfGn process is formed [7] D. Garrett, D. A. Peterson, C. W. Anderson, and M. H. Thaut, “Compar-
using (13) where K can be written as ison of linear, nonlinear, and feature selection methods for EEG signal
classification,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 11, no. 2,
2
K = σH K0 (16) pp. 141–144, 2003.
[8] H. C. Shin, S. Tong, S. Yamashita, X. Jia, R. G. Geocadin, and N. V.
2
and σH is as given in (7). Thakor, “Quantitative EEG assessment of brain injury and hypothermic
neuroprotection after cardiac arrest,” in Engineering in Medicine and
On substituting (16) into (15), we obtain Biology Society, 2006. EMBS’06. 28th Annual International Conference
of the IEEE. IEEE, 2006, pp. 6229–6232.
−N N 2
log p(y; H) = log (2π) − log σH [9] A. J. Lim, and W. D. Winters, “A practical method for automatic real-
2 2 time EEG sleep state analysis,” IEEE Trans. Biomed. Eng., vol. BME-27,
1 (yT K0−1 y) no. 4, pp. 212–220, 1980.
− log |K0 | − 2 . (17) [10] M. Diykh, Y. Li, and P. Wen, “EEG sleep stages classification based
2 2σH on time domain features and structural graph similarity,” IEEE Trans.
2 Neural Syst. Rehabil. Eng., vol. 24, no. 11, pp. 1159–1168, 2016.
4) Equation (17) is maximized over σH by taking the [11] J. C. Principe and S.-H. Park, “An expert system architecture for abnor-
2
derivative with respect to σH and equating it to zero. mal EEG discrimination,” Engineering in Medicine and Biology Society,
This gives 1990., Proceedings of the Twelfth Annual International Conference of
2 yT K0−1 y the IEEE. IEEE, pp. 1376–1377, 1990.
σ̂H = . (18) [12] A. Myrden, T. Chau, “A passive EEG-BCI for single-trial detection of
N changes in mental state,” IEEE Trans. Neural Syst. Rehabil. Eng., vol.
5) On substituting (18) into (17), we obtain the function 25, no. 4, pp. 345–356, 2017.
[13] A. Tuunainen, U. Nousiainen, A. Pilke, E. Mervaala, J. Partanen,
that is to be maximized over H as below and P. Riekkinen, “Spectral EEG during short-term discontinuation of
antiepileptic medication in partial epilepsy,” Epilepsia, vol. 36, no. 8,
Ĥ = max (−N log (yT K0−1 y) − log |K0 |) − m. pp. 817–823, 1995.
0<H<1
(19) [14] B. Clemens, G. Szigeti, and Z. Barta, “EEG frequency profiles of
idiopathic generalised epilepsy syndromes,” Epilepsy Res., vol. 42,
The estimated value of H using (19) for 1-dfGn will no. 2–3, pp. 105–115, 2000.
be in the range -1<H<0, while that of 2-dfGn will be [15] C. Borel, and D. Hanley, “Neurological intensive care unit monitoring,”
-2<H<-1. Critical Care Clinics Symposium on Neurological Intensive Care. PA:
Saunders, 1985, pp. 223–239.
[16] R. Agarwal, J. Gotman, D. Flanagan, and B. Rosenblatt, “Automatic
D. DCT basis based filters EEG analysis during long-term monitoring in the ICU,” Electroen-
cephalogr. Clin. Neurophysiol., vol. 107, no. 1, pp. 44–58, 1998.
The filters used in each of the three levels of this composite [17] D. G. Manolakis, V. K. Ingle, and S. M. Kogon, Statistical and adap-
filterbank of Fig-1 are the DCT basis vectors given as below: tive signal processing: spectral estimation, signal modeling, adaptive
filtering, and array processing. McGraw-Hill Boston, 2000.
1 [18] L. Rankine, N. Stevenson, M. Mesbah, and B. Boashash, “A nonstation-
fk1 [n] = √ ; k = 1, 2, 3; 0 ≤ n ≤ Q − 1
Q ary model of newborn EEG,” IEEE Trans. Biomed. Eng., vol. 54, no. 1,
r pp. 19–28, 2007.
2 π(2n + 1)(l − 1) [19] G. Tognola, P. Ravazzani, T. Locatelli, F. Minicucci, F. Grandori, and
fkl [n] = cos ; (20)
Q 2Q G. Comi, “A parametric method for the analysis of temporal and spatial
variability in the interictal EEG signal,” in Engineering in Medicine
k = 1, 2, 3; 2 ≤ l ≤ Q; 0 ≤ n ≤ Q − 1 and Biology Society, 1994. Engineering Advances: New Opportunities
for Biomedical Engineers. Proceedings of the 16th Annual International
where k is the level number (1, 2, or 3), Q is the downsampling Conference of the IEEE. IEEE, 1994, pp. 1234–1235.
factor (or the length of filters in each stage), and l denotes the [20] S. Tong, and N. V. Thakor, Quantitative EEG analysis methods and
analysis (or synthesis) filter number at each stage with l = 1 clinical applications. Artech House, 2009.
[21] M. A. Naderi, and H. Mahdavi-Nasab, “Analysis and classification of
as the lowpass filter, l = Q denotes the highpass filter, and EEG signals using spectral analysis and recurrent neural networks,”
2 < l < Q − 1 denote the bandpass filters. Filter responses Biomedical Engineering (ICBME), 2010 17th Iranian Conference of.
are shown in the supplementary material. IEEE, pp. 1–4, 2010.
[22] P. Singh, R. B. Pachori, “Classification of focal and nonfocal EEG
signals using features derived from Fourier-based rhythms,” J. Mech.
R EFERENCES Med. Biol., Vol. 17, No. 7, 1740002, pp. 1–16, 2017.
[1] L. D. Iasemidis, “Epileptic seizure prediction and control,” IEEE Trans. [23] M. Sharma, R. B. Pachori, and U. R. Acharya, “A new approach to
Biomed. Eng., vol. 50, no. 5, pp. 549–558, 2003. characterize epileptic seizures using analytic time-frequency flexible
[2] J. Corsini, L. Shoker, S. Sanei, and G. Alarcon, “Epileptic seizure wavelet transform and fractal dimension,” Pattern Recognit. Lett, vol.
predictability from scalp EEG incorporating constrained blind source 94, pp. 172–179, 2017.
separation,” IEEE Trans. Biomed. Eng., vol. 53, no. 5, pp. 790–799, [24] A. K. Tiwari, R. B. Pachori, V. Kanhangad, and B. K. Panigrahi,
2006. “Automated diagnosis of epilepsy using key-point based local binary
[3] M. Steriade, P. Gloor, R. R. Llinás, F. H. Lopes da Silva, and pattern of EEG signals,” IEEE J. Biomed. Health Inform., vol. 21, no.
M. M. Mesulam, “Basic mechanisms of cerebral rhythmic activities,” 4, pp. 888–896, 2017.
Electroencephalogr Clin Neurophysiol, vol. 76, no. 6, pp. 481–508, [25] A. Bhattacharyya, R. B. Pachori, A. Upadhyay, and U. R. Acharya,
1990. “Tunable-Q wavelet transform based multiscale entropy measure for
[4] L. Mingyang, C. Wanzhong, and Z. Tao, “Automatic epileptic EEG automated classification of epileptic EEG signals,” Applied Sciences,
detection using DT-CWT-based non-linear features,” Biomed Signal vol. 7 (4), 385, pp. 01–18, 2017.
Process Control, vol. 34, pp. 114–125, 2017. [26] D. Bhati, M. Sharma, R. B. Pachori, and V. M. Gadre, “Time-frequency
[5] J. Abeg Kumar, and B. Haider, “Local pattern transformation based localized three-band biorthogonal wavelet filter bank using semidefinite
feature extraction techniques for classification of epileptic EEG signals,” relaxation and nonlinear least squares with epileptic seizure EEG signal
Biomed Signal Process Control, vol. 34, pp. 81–92 2017. classification,” Digital Signal Processing, vol. 62, pp. 259–273, 2017.
[6] P. Shivnarayan, and P. Trilochan, “Detection of epileptic seizure using [27] R. Sharma and R. B. Pachori, “Classification of epileptic seizures in EEG
kraskov entropy applied on tunable-q wavelet transform of EEG signals,” signals based on phase space representation of intrinsic mode functions,”
Biomed Signal Process Control, vol. 34, pp. 74–80 2017. Expert Syst Appl., vol. 42, no. 3, pp. 1106–1117, 2015.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2018.2818123, IEEE
Transactions on Neural Systems and Rehabilitation Engineering

10

[28] T. S. Kumar, V. Kanhangad, and R. B. Pachori, “Classification of seizure [52] P. Flandrin, “On the spectrum of fractional Brownian motions,” IEEE
and seizure-free EEG signals using local binary patterns,” Biomed Signal Trans. Inf. Theory, vol. 35, no. 1, pp. 197–199, 1989.
Process Control, vol. 15, pp. 33–40, January 2015. [53] T. Lundahl, W. J. Ohley, S. M. Kay, and R. Siffert, “Fractional Brownian
[29] R. B. Pachori, and S. Patidar, “Epileptic seizure classification in EEG motion: A maximum likelihood estimator and its application to image
signals using second-order difference plot of intrinsic mode functions,” texture,” IEEE Trans. Med. Imag., vol. 5, no. 3, pp. 152–161, 1986.
Comput Methods Programs Biomed., vol. 113, no. 2, pp. 494–502, 2014. [54] R. G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, and C. E.
[30] V. Joshi, R. B. Pachori, and A. Vijesh, “Classification of ictal and Elger, “Indications of nonlinear deterministic and finite-dimensional
seizure-free EEG signals using fractional linear prediction,” Biomed structures in time series of brain electrical activity: Dependence on
Signal Process Control, vol. 9, pp. 1–5, January 2014. recording region and brain state,” Physical Review E, vol. 64, no. 6,
[31] R. B. Pachori, and V. Bajaj, “Analysis of normal and epileptic seizure p. 061907, 2001.
EEG signals using empirical mode decomposition,” Comput Methods [55] https://www.researchgate.net/publication/308719109 EEG Epilepsy
Programs Biomed., vol. 104, no. 3, pp. 373–381, December 2011. Datasets.
[32] R. B. Pachori, “Discrimination between ictal and seizure-free EEG [56] P. Swami, T. K. Gandhi, B. K. Panigrahi, M. Tripathi, and S. Anand, “A
signals using empirical mode decomposition,” Research Letters in Signal novel robust diagnostic model to detect seizures in electroencephalog-
Processing, 2008. raphy,” Expert Syst Appl., vol. 56, pp. 116–130, 2016.
[33] U. R. Acharya, S. V. Sree, G. Swapna, R. J. Martis, and J. S. Suri, [57] T. Gandhi, P. Chakraborty, G. G. Roy, and B. K. Panigrahi, “Discrete
“Automated EEG analysis of epilepsy: a review,” Knowledge-Based harmony search based expert model for epileptic seizure detection in
Systems, 45, 147–165, 2013. electroencephalography,” Expert Syst Appl., vol. 39, no. 4, pp. 4055–
[34] A. S. Zandi, M. Javidan, G. A. Dumont, and R. Tafreshi, “Automated 4062, 2012.
real-time epileptic seizure detection in scalp EEG recordings using an [58] T. Gandhi, B. K. Panigrahi, and S. Anand, “A comparative study of
algorithm based on wavelet packet transform,” IEEE Trans. Biomed. wavelet families for EEG signal classification,” Neurocomputing, vol. 74,
Eng., vol. 57, no. 7, pp. 1639–1651, 2010. no. 17, pp. 3051–3057, 2011.
[35] P. P. Acharjee, and C. Shahnaz, “Multiclass epileptic seizure classifi- [59] T. Gandhi, B. K. Panigrahi, M. Bhatia, and S. Anand, “Expert model
cation using time-frequency analysis of EEG signals,” Electrical and for detection of epileptic activity in EEG signature,” Expert Syst Appl.,
Computer Engineering (ICECE), 2010 7th International Conference of, vol. 37, no.4, pp. 3513–3520, 2010.
pp. 260–263, 2012. [60] H.-W. Steinberg, T. Gasser, and J. Franke, “Fitting autoregressive models
[36] Q. Shuren, and J. Zhong, “Extraction of features in EEG signals with the to EEG time series: An empirical comparison of estimates of the order,”
non-stationary signal analysis technology,” in Engineering in Medicine IEEE Trans. Acoust., Speech, Signal Process., vol. 33, no. 1, pp. 143–
and Biology Society, 2004. IEMBS’04. 26th Annual International Con- 150, 1985.
ference of the IEEE, vol. 1. IEEE, pp. 349–352, 2004. [61] T. W. Picton, S. Bentin, P. Berg, E. Donchin, S. A. Hillyard, R. Johnson,
[37] Y. Song, J. Crowcroft, and J. Zhang, “Epileptic EEG signal analysis G. A. Miller, W. Ritter, D. S. Ruchkin, M. D. Rugg, and M. J. Taylor,
and identification based on nonlinear features,” in Bioinformatics and “Guidelines for using human event-related potentials to study cognition:
Biomedicine (BIBM), 2012 IEEE International Conference on. IEEE, Recording standards and publication criteria,” Psychophysiology, vol. 37,
pp. 1–6, 2012. no. 2, pp. 127–152, 2000.
[38] S. M. Pincus, “Approximate entropy as a measure of system complexity,” [62] D. Easwaramoorthy and R. Uthayakumar, “Analysis of EEG signals
Proc Natl Acad Sci USA, vol. 88, no. 6, pp. 2297–2301, 1991. using advanced generalized fractal dimensions,” in Computing Commu-
[39] H. Adeli, S. Ghosh-Dastidar, and N. Dadmehr, “A wavelet-chaos nication and Networking Technologies (ICCCNT), 2010 International
methodology for analysis of EEGs and EEG subbands to detect seizure Conference on. IEEE, pp. 1–6, 2010.
and epilepsy,” IEEE Trans. Biomed. Eng., vol. 54, no. 2, pp. 205–211, [63] A. Accardo, M. Affinito, M. Carrozzi, and F. Bouquet, “Use of the fractal
2007. dimension for the analysis of electroencephalographic time series,”
[40] H. Jing, and M. Takigawa, “Low sampling rate induces high correlation Biological Cybernetics, vol. 77, no. 5, pp. 339–350, 1997.
dimension on electroencephalograms from healthy subjects,” Psychiatry [64] V. Cabukovski, N. d. M. Rudolf, and N. Mahmood, “Measuring the
Clin Neurosci., vol. 54, no. 4, pp. 407–412, 2000. fractal dimension of EEG signals: selection and adaptation of method
[41] S. M. S. Alam, and M. I. H. Bhuiyan, “Detection of seizure and epilepsy for real-time analysis,” WIT Transactions on Biomedicine and Health,
using higher-order statistics in the EMD domain,” IEEE J. Biomed. vol. 1, pp. 285–292, 1993.
Health Inform., vol. 17, no. 2, pp. 312–318, 2013. [65] A. Petrosian, “Kolmogorov complexity of finite sequences and recog-
[42] V. Bajaj, and R. B. Pachori, “Classification of seizure and non-seizure nition of different preictal EEG patterns,” in Computer-Based Medical
EEG signals using empirical mode decomposition,” IEEE Trans Inf Systems, 1995., Proceedings of the Eighth IEEE Symposium on. IEEE,
Technol Biomed, vol. 16, no. 6, pp. 1135–1142, 2012. pp. 212–217, 1995.
[43] T. Bojić, A. Vuckovic, and A. Kalauzi, “Modeling EEG fractal dimen- [66] N. Pradhan and D. Narayana Dutt, “Use of running fractal dimension
sion changes in wake and drowsy states in humans–a preliminary study,” for the analysis of changing patterns in electroencephalograms,” Comput.
J. Theor. Biol., vol. 262, no. 2, pp. 214–222, 2010. Biol. Med, vol. 23, no. 5, pp. 381–388, 1993.
[44] L. Guo, D. Rivero, J. Dorado, J. R. Rabuñal, and A. Pazos, “Automatic [67] R. Esteller, G. Vachtsevanos, J. Echauz, T. Henry, P. Pennell, C. Epstein,
epileptic seizure detection in EEGs based on line length feature and R. Bakay, C. Bowen, and B. Litt, “Fractal dimension characterizes
artificial neural networks,” Journal of neuroscience methods, vol. 191, seizure onset in epileptic patients,” Acoustics, Speech, and Signal
no. 1, pp. 101–109, 2010. Processing, 1999. Proceedings., 1999 IEEE International Conference
[45] A. Subasi, and M. I. Gursoy, “EEG signal classification using PCA, on, vol. 4. IEEE, pp. 2343–2346, 1999.
ICA, LDA and support vector machines,” Expert Syst Appl., vol. 37, [68] A. Gupta, and S. Joshi, “Connection between DCT and discrete-time
no. 12, pp. 8659–8666, 2010. fractional brownian motion,” in Data Compression Conference (DCC),
[46] E. D. Übeyli, “Least squares support vector machine employing model- 2016. IEEE, 2016, pp. 598–598.
based methods coefficients for analysis of EEG signals,” Expert Syst [69] A. Gupta, and S. Joshi, “Some studies on the structure of covariance
Appl., vol. 37, no. 1, pp. 233–239, 2010. matrix of discrete-time fBm,” IEEE Trans. Sig. Process, vol. 56, no. 10,
[47] N. Nicolaou, and J. Georgiou, “Detection of epileptic electroencephalo- pp. 4635–4650, 2008.
gram based on permutation entropy and support vector machines,” [70] M. Karlekar, and A. Gupta, “Stochastic modeling of EEG rhythms
Expert Syst Appl., vol. 39, no. 1, pp. 202–209, 2012. with fractional Gaussian noise,” in Signal Processing Conference (EU-
[48] S. Mihandoost, and M. C. Amirani, “EEG signal analysis using spectral SIPCO), 2014 Proceedings of the 22nd European. IEEE, 2014, pp.
correlation function & GARCH model,” Signal, Image and Video 2520–2524.
Processing, vol. 9, no. 6, pp. 1461–1472, 2015. [71] N. Nicolaou and J. Georgiou, “Detection of epileptic electroencephalo-
[49] B. B. Mandelbrot, and J. W. Van Ness, “Fractional Brownian motions, gram based on permutation entropy and support vector machine,” Expert
fractional noises and applications,” SIAM review, vol. 10, no. 4, pp. Syst. Appl., vol. 39, no. 1, p. 202–209, 2012.
422–437, 1968. [72] Y. Kumar, M. L. Dewal, and R. S. Anand, “Epileptic seizures detection
[50] E. Perrin, R. Harba, C. Berzin-Joseph, I. Iribarren, and A. Bonami, “nth- in EEG using DWT-based ApEn and artificial neural network,” Signal
order fractional Brownian motion and fractional Gaussian noises,” IEEE Image Video Process., vol. 8, no. 7, pp. 1323–1334, 2014.
Trans. Sig. Process, vol. 49, no. 5, pp. 1049–1059, 2001. [73] P. Singh, S. D. Joshi, R. K. Patney, and K. saha, “Fourier-based feature
[51] A. Gupta and S. D. Joshi, “DCT and eigenvectors of covariance of extraction for classification of EEG signals using EEG rhythms,” Circuits
1st and 2nd order discrete fractional Brownian motion,” arXiv preprint Syst Signal Process, vol. 35, no. 10, pp. 3700–3715, 2016.
arXiv:1302.5556, 2013.

1534-4320 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like