
Research Article


Detection of Epileptic Seizure in EEG Recordings by Spectral Method and Statistical Analysis 

D.K. Ravish,
S. Shenbaga Devi,
S.G. Krishnamoorthy
and
M.R. Karthikeyan



ABSTRACT

Electroencephalographs are records of brain electrical activity.
It is an indispensable tool for diagnosing epileptic seizure. Manually reviewing
EEG recordings for detection of epilepsy pattern is a timeconsuming process.
It is, therefore, necessary to automate the epileptic seizure detection. Further
analyzing after the detection using large data set is a good supplement to the
wide range of algorithms currently used for analysis. Seizure evolution is typically
a dynamic and non stationary process and the EEG signals are composed of many
frequency bands. The objective of this work was to determine features that differentiated
epileptic seizure from a normal activity. Subjects suffering from a commonly
occurring generalized epileptic seizure EEG segments and non seizure EEG segments
are used for the study. The Spectral and Statistical methods are applied to
the signal and the features are extracted. Spectral power, Standard deviation,
Variance, Root Mean Square and Measure of Spread are the features that differentiate
abnormal activity from normal activity whereas Median, Mode, Skewness, kurtosis
are not able to differentiate. This study gives a method to detection of seizure
and analysis in offline. It can be further extended to the real time. The algorithm
is tested with two different databases covering children and as well as adult
data sets.





Received: October 02, 2012;
Accepted: January 05, 2013;
Published: February 21, 2013


INTRODUCTION
Epilepsy is a common chronic neurological disorder, affecting almost 50 million
people worldwide (WHO, 2007). Epileptic seizures are
paroxysmal brain dysfunction caused by excessive neuronal discharge (Fix,
1995). It is associated with some altered state of consciousness, recurrent
and sudden malfunction of the brain. EEG used to be a firstline diagnosis test.
The diagnosis of epilepsy also achieved by different examinations, such as Positron
Emission Tomography, Magnetic Resonance Imaging, Computed Tomography and Electroencephalogram
(EEG). Of these, EEG most is important and economical one which gives high temporal
resolution. A lot of useful information in EEG can be extracted by signal processing
methods (Gevins et al., 1995, Mansouri
et al., 2012). Some of the information is helpful for diagnosis and
treatment of epilepsy patients (Wang and Xu, 2009).
The research work on epileptic EEG processing mainly focuses on epileptic events
detection and seizure prediction. In the epileptic EEG, the presence of epileptiform
activities, such as spikes, slow rhythm and highfrequency epileptiform oscillations
confirms the diagnosis of epilepsy (Padmasai et al.,
2010). Traditionally, EEGs are scanned for epileptic spikes by experienced
physicians. With the development of EEG acquisition system, longterm EEG collection
can be achieved.
Approximately, 1% of the world’s population is affected by epilepsy and
25% of epilepsy patients cannot be treated sufficiently by any available therapy
80 people per day develops the condition (WHO, 2007;
Liang et al., 2010). Quality of life of a person
may be severely affected by epilepsy because of both psychological and social
reasons. If an automatic seizuredetection system is available, it could reduce
the time required by a neurologist to perform an offline diagnosis by reviewing
electroencephalogram data. It could be used to produce an online warning signal
to alert healthcare professionals (Liang et al., 2010).
Seizure evolution is typically a dynamic and non stationary process and the
signals are composed of multiple frequencies. Visual and conventional frequencybased
direct spectral method has limited application (Tzallas
et al., 2009). An algorithm proposed applying wavelet packet analysis
and determined dominant frequency bands during electro convulsive therapy (Zandi
et al., 2007). Wavelet analysis method to isolate EEG bands had shown
good performance (Tafreshi et al 2006). A method
proposed based on the standard clinical sub bands of EEG (Yucel
and Ozguler, 2008).A simplified method of feature extraction and classification
proposed by the method based on energy ,entropy and kurtosis were considered
for discrimination of various categories of EEG signals (Pal
and Panda, 2010). The statistical method combined with a simplified classification
algorithm was proposed to discriminate epileptic EEG signal (Choe
et al., 2010).
The main objective of the present proposed work is to detect the seizures in EEG signal and extracting the features for further analysis. The proposed method uses an algorithm based on combined spectral and statistical methods. The combined analysis of identified features from various age group and large dataset could reveal better results. This work is to automate the detection process and to help the doctors in analyzing the seizure EEG signal. MATERIALS AND METHODS
We worked with the database collected from (Andrzejak et
al., 2001) the Epilepsy Centre at the University of Bonn, Germany (http://epileptologiebonn.de).
In this study, two sets of EEG, B and E, each containing (normal and seizure
signal) 100 single channel EEG segments of 23.6 seconds duration with a sampling
rate of 173 Hz, are considered. The spectral bandwidth of the data set from
0.5 Hz to 85 Hz is used for detection and analysis. EEG time series data made
available online by the Children's Hospital Boston, consisting of EEG recording
of subjects with intractable seizures are also considered (http://physionet.org/physiobank/database/chbmit/).
Data sets available are filtered already using hardware filters 0.570 Hz and
50 Hz notch filters. Recordings are from 22 subjects (5 males, ages 322 and
17 females, ages 1.518). Mean (Standard Deviation) ages of the patients are
9.81 (5.75 years). The International 1020 system of EEG electrode positions
and nomenclature are used for 23 channels. These recordings are FP1F7, F7T7,
T7P7, P7O1, FP1F3, F3C3, C3P3, P3O1, FP2F4, F4C4, C4P4, P4O2, FP2F8,
F8T8, T8P8, P8O2, FZCZ, CZPZ, P7T7, T7FT9, FT9FT10, FT10T8 and T8P8.
All signals are sampled at 256 samples per second with 16bit resolution and
EEG data recorded is exactly one hour of digitized signal duration, some are
of two hour duration and some other 4 h duration. From this set, 50 segments
of normal EEG are considered as non seizure conditions and 50 segments of EEG
with seizure are considered as seizure cases for the present work. The conditions
namely non seizure and seizure and the duration are specified in the database
itself.
The data used in this investigation have been collected from two databases
as mentioned above. The EEG from Epilepsy Centre at the University of Bonn,
Germany (http://epileptologiebonn.de)
is used for finding spectral, statistical features and to classify normal and
seizure. For the purpose of EEG time series seizure detection and analysis,
the second data set from the database of Children’s Hospital Boston is
used (http:// physionet.org/physiobank).
To present the results, the algorithm is tested using single channel (C_{z}P_{z})
50 normal and 50 seizure EEG time series segments. The seizure cases known as
generalized as given in database are used as gold standard (Andrzejak
et al., 2001), (http://
physionet.org/physiobank/database/chbmit/). Figure 1 shows
the block diagram of the model for a proposed method.
Pre processing: The EEG data are containing many artifacts, such as
power line noise and movement. The recordings of single channel EEG segments
collected from the databases (Andrzejak et al., 2001
and http://epileptologiebonn.de
and Children's Hospital Boston) are first taken into the Lab VIEW platform and
EEG signal baseline wanders is corrected and the signal amplitude is quantified
to micro volts. The EEG signal is filtered using a digital low pass Finite Impulse
Response (FIR) filter with Hamming window technique to remove power line noise
along with outof band noise. The order of the filter is 40 and cut off frequency
is 32 Hz. Flatness without a ripple in the pass band is desirable in the analysis
of EEG signals which leads to the use of FIR filter (Lessard,
2006). Filtered EEG segments are chosen for seizure detection and analysis.
A normal EEG and seizure EEG segments are shown in Fig. 3a
and 4a.
Outline of spectral method: EEG is a non stationary signal and hence
applying linear methods to compute direct spectrum results in less resolution
(Blanco et al., 1998).

Fig. 1: 
Schematic drawing of the proposed method 

Fig. 2: 
Box plot for 25 non seizure and 25 seizure segments 
For the preprocessed normal and seizure EEG segments, averaged spectrum and
averaged power measure are calculated by the following procedure. Spectral analysis
is estimation of power from the observation of the signal over time. The Power
spectrum of the signal is computed using Fast Fourier Transform (FFT) for every
twosecond window with an overlap of one second of the signal (Djuric
and Kay, 1999; Rangayyan, 2002). The equation for
FFT is given in Eq. 1. The computation of Fourier transform
by definition:
We know that:

Fig. 3(ab): 
Normal EEG segment and its averaged spectrum, LED is in ‘OFF’
condition (a) Normal EEG segment and (b) Average spectrum 
For one value of ‘k’ observe that the multiplication of x(n) and
is done for ‘N’ times, since n = 0 to N1. That is there are ‘N’
complex multiplications for one value of k. Since, ‘K’ also has ‘N’
values (since k = 0, 1, …, N1).
RESULTS Spectral based detection: Consider sequence
that is x(n) ,n = 0, 1,…, N1.

Fig. 4(ab): 
Seizure EEG segment and its averaged power spectrum, LED is
in ‘ON’ condition (a) Seizure EGG segment and (b) Averaged spectrum 
To find the periodogram of data x[n],we split up the Npoint data record into
Mpoint segments x_{i}[n] that overlap with each other by segments of
length L, such that ith segment is given by the sequence x_{i}[n]. The
sequence that
overlaps successive sequences by D samples. Then = L+D (K1), here N is the
total number of samples in the entire EEG time series data, L is the length
of the segment and D is the overlapping samples. The lth sequence is
denoted by:
where, x_{i} [n] is the signal to be analysed and [n] the weighting
or temporal window of null value outside the observation interval. This temporal
product is transformed into the frequency domain by a convolution product of
the Fourier transforms of the sequence and window. The Hamming window technique
offers a very simple means of (linear) Zero phases, stable, simpler and easier
for our application (Sadati et al., 2006). EEG
signals are transformed from a time domain to a frequency domain. Here, we used
Welch Overlapping Averaged Spectral (WOSA) method. The averaged spectrum is
calculated using Hamming window considering 50% overlap. The averaged power
is thus computed by using Welch Overlapped Spectral Averaging (WOSA) method.
Hamming window function is used in the averaging process covering from delta
(δ) to alpha (α).
Table 1: 
Spectral power computed for 25 normal and 25 seizure segments. 

SD: Standard deviation 
The spectral power of the ith segment is:
Here,
is the averaged periodogram of the data samples x[n] weighted by a Hamming window
w[n].The spectrum obtained is given by:
where, M = 0, 1, 2 …K segments. Now spectral power is computed using relation:
Here,
is the averaged periodogram of the data sample x[n] weighted by hamming window.
The frequency band from 0.5 Hz to 14 Hz is considered for further processing
as this band covers δ, θ and α. The seizure activity has generally
been considered to be associated with these bands. The researchers also point
out that frequency dynamics characterized by an activity of seizure are originally
at alpha and are slowing down to about delta (Zandi et
al., 2007; Tafreshi et al., 2006). From
the averaged spectrum, the average power for the frequency range from 0.5 Hz
to 14 Hz is thus calculated. This power value is shown in the upper right corner
in Fig. 3b. This value of average power will be helpful in
finding the presence of seizure. The average power value is computed for each
8 second segments of EEG data. This is performed for 25 normal EEG segments
and 25 EEG segments with seizure condition. The average power value for the
25 segments found out to be 39.45 dB with Standard Deviation (SD) of 2.15 and
the same for seizure condition is 55.80 dB with SD of 4.37. These two values
show a large difference in average power value and during seizure, the value
is much higher than normal. This concept is applied further to detect epileptic
seizure from EEG. The seizure is detected by looking for elevation of power
during the seizure compared to normal. The average spectral powers for the 25
normal and 25 seizure segments (Andrzejak et al., 2001)
are tabulated in Table 1. The results are presented graphically
using box plot as shown in Fig. 2. The averaged power spectral
value of 25 normal and 25 epileptic EEG signals are analyzed to fix the threshold
value (Zandi et al., 2007; Rangayyan,
2002). The threshold value is fixed as 48 dB for detecting the seizure based
on the non overlap region in the box plot. If there is a seizure, the elevation
of power in the spectrum is more.
Totally, the database has 100 segments each for normal and seizure conditions.
Out of these 25 cases from each are analyzed for spectral variations. For the
remaining 75 segments of EEG of each normal and seizure data, the average power
value is calculated for the 0.514 Hz range as before and when the average power
in the 8 second segment is less than the threshold of 48 dB, a green LED in
the upper right corner of panel is shown in the off condition and the plot of
average power spectrum is shown (Fig. 3b). When the average
power is greater than 48 dB, it is detected as seizure in the 8 second segment
and the LED will be made on (Fig. 4b). The average spectral
powers for the remaining 75 normal and 75 seizure segments are calculated for
data set segments (Andrzejak et al., 2001).
Spectral method for children hospital EEG time series data set: We also
extended our work with EEG timeseries data collected from the children’s
hospital Boston. We only used single channel (CZPZ) data for development and
testing of the seizure detection algorithm (Greene et
al., 2008; Hoeve et al., 2001). Non seizure
EEG time series signal of a subject as shown in Fig. 5a and
a seizure signal of another subject is shown in Fig. 6a. EEG
signals are transformed from a time domain to a frequency domain.
The selected peak points are shown in Fig. 5c and Fig.
6c for normal and seizure conditions, respectively. This peak point is mapped
to the time series EEG and this point as the centre point, a window has to be
chosen (Fig. 5d, 6d). Here, we used Welch
Overlapping Averaged Spectral (WOSA) method. The averaged spectrum is calculated
using Hamming window considering two seconds segments with an overlap of one
second. The relative spectral power is computed from the Eq. 3.
The relative spectral power peak value is detected (Fig. 5b,
6b). According to the researchers, the minimum seizure duration
is 10 seconds (Zandi et al., 2007; Bruce,
2006). Taking this into consideration a window of 8 second data are selected
for detection and analysis (Fig. 5e, 6e).
Figure 5f and 6f show the detected results
by the algorithm and this well coincides with the diagnosis results given by
the doctor in the database.
This box plot shows that there is a clear demarcation in the spectral power
values during seizure and normal EEG and the value considered for this work
is as follows; if power less than 48 dB no seizure.
Table 2: 
Spectral powers computed for 150 non seizure and 150 seizure
segments 

SD: Standard deviation 
Table 3: 
Detection results 

TP: True positive, FN: False negative, FP: False positive,
TN: True negative 
If power is greater than 48 dB, seizure condition.
To show the performance of the method for full data set 150 seizure EEG and 150 non seizure EEG segments the result obtained are tabulated in the Table 2 .The spectral analysis results are plotted graphically in the Fig. 7. Sensitivity, specificity and accuracy measures: The EEG segments are executed for detecting seizure. Table 3 shows the test result of the spectral method. The sensitivity (S_{n}) of seizure detection is the probability that the detection is positive when the EEG segments are with the seizure. The specificity (S_{p}) is defined as the probability that the seizure detection result says a non seizure segment, when in fact, they are seizure free: The three measures sensitivity, specificity and accuracy are used as evaluation criteria to check the performance of the proposed detection method. The result of detection method is shown in Table 3. There are 150 normal EEG segments and 150 seizure EEG segments for the both databases put together. The sensitivity of the proposed algorithm is 93% specificity is 97% and overall accuracy is 95%.
Statistical analysis: Statistical methods provide information on the
amplitude variation of EEG signal. Statistical parameters used are Mean, Mode,
Variance, Standard Deviation, Measure of Spread, Skewness, kurtosis and RMS
of EEG for understanding normal and seizure features. The unit of amplitude
variation of EEG signal is in micro volts. The mean, standard deviation, RMS
and Measure of Spread have the same units as the EEG signal amplitude which
is in micro volts whereas unit of variance is the square of the micro volts.

Fig. 5: 
Normal time series and its spectral power, LED is in ‘OFF’
condition (a) EEG normal time series, (b) Relative power in egg segment,
(d) EEG normal time series, (e) 8 sec epoch and (f) Averaged spectrum 

Fig. 6: 
Seizure time series and its Spectral power, LED is in ‘ON’
condition (a) EEG normal time series, (b) Relative power in egg segment,
(d) EEG normal time series, (e) 8 sec epoch and (f) Averaged spectrum 

Fig. 7: 
Box plot for 150 non seizure and 150 seizure segments 
Skewness and kurtosis have no units; it is a pure number like a Zscore.
The Arithmetic Mean is the standard average, often simply called the mean.
For a given EEG signal, the mean value of the EEG data at any point in time,
n is the average value of its sample functions at a fixed time (Bruce,
2006). The Mean of the signal is calculated by using the Equation:
The Mode is the value that occurs most frequently in a data set or a probability distribution. The variance is the mean of the squared differences between individual data points and the mean of the array. Variance is calculated by using the Eq. 10:
The standard deviation is the square root of the variance. A Measure of Spread
tells us a whether data sample is spread out or scattered. We can use the range
to measure the spread of a sample. We get a good measure of spread by summing
the squares of the deviations from the mean. The Root Mean square (abbreviated
RMS), also known as the quadratic mean, is a statistical measure of the magnitude
of a varying quantity. RMS is calculated by Eq. 11:
Range is used to explain the variability in an EEG segment samples. It is used in connection with a measure of central tendency, such as the mean, median, to provide overall description of a set of EEG samples that is Measure of Spread MOS. Quartiles tell us about the measure of spread. The quartiles are computed by breaking data set into quarters. The measures available in quartiles are first quartile (Q1), second quartile (Q2) and third quartile (Q3). A common measure of expressing a quartile is an inter quarter range (Q3Q1). Hence, for EEG data segment the inter quartile range is: Skewness describes asymmetry from the normal distribution in a set of statistical data, as data becomes more symmetrical as its value approaches zero. Normally distributed data, by definition has little skewness and on other hand positively skewed or right sided skewed data has positive and negatively skewed or left sided skewed has negative value. Skewness can be calculated by the Eq. 13: Kurtosis is a statistical measure used to describe the distribution of observed data around the mean. It is the degree to which a data set is peaked. Kurtosis can be calculated mathematically by the Eq. 14: These statistical values are computed for already detected normal and seizure segments of 8 second duration each. The results obtained for variance, root mean square, standard deviation, measure of spread and inter quartile range for the normal and seizure EEG signals are plotted in the Fig. 8.
In the analysis, all parameters are computed based on the equations given above.
The statistical parameters for the 150 normal and 150 seizure segments (Andrzejak
et al., 2001 and Epilepsy Centre at the University of Bonn, Germany
http://epileptologiebonn.de)
obtained are tabulated in Table 4. The box plot in Fig.
9 shows a distribution of statistical features computed for 150 normal and
150 seizure segments. From the figures it can be clearly seen that wide difference
is there in all parameters for normal and seizure conditions.
The algorithm is tested on a test data (http://www.vis.caltech.edu/~rodri/data.htm).
The spectral method is applied on test data and the instant at which seizure
occurs are identified (power spectral value of 51 dB). This correlates with
the gold standard for member.

Fig. 8(al): 
Statistical parameters for normal and seizure segment (a)
Normal EEG segment, (b) Variance, (c) RMS, (d) SD, (e) MOS, (f) IQR, (g)
Seizure EEG segment, (h) Variance, (i) RMS, (j) SD, (k) MOS and (l) IQR 
Table 4: 
Statistical parameters computed for 150 normal and 150 seizure
segments 

RMS: Root mean square, SD: Standard deviation, MOS: Measure
of spread and IQR: Inter quartile range 
The statistical parameters for the seizure durations are computed and they fall in the ranges specified for seizure condition (variance of 26963, RMS of 163.24, SD of 163.59 and MOS of 733.50 and Range of 218.82). This shows the efficiency of the algorithm in detecting the seizure. Here Variance, Standard deviation, RMS and Measure of spread shows some good variation during seizure period. Other parameters like Mean, Mode, Skewness and kurtosis do not show much variation in the seizure period. So the parameters variances, Standard deviation, RMS, MOS are considered for analysis. RESULTS AND DISCUSSION
In this study, a method to detect epileptic seizure in an automatic manner
is proposed. Power in the EEG signal is relatively increased during the seizure.
Taking this into consideration, first method uses the spectral technique for
detection and analysis. For this, EEG signal from (Andrzejak
et al., 2001) 100 normal and 100 segments with epileptic seizure
are tested and analyzed. All 100 normal cases are identified correctly but 95
out of 100 epileptic seizure conditions are detected correctly by the algorithm.
The results clearly show that functional differences during normal and seizure
activities. . The discussion of our study in comparison to researches that deal
with detection of epileptic seizure from EEG recordings of same database (Andrzejak
et al., 2001) is given in Table 5. The result obtained
using our method with average accuracy of 97.5%.
The algorithm is further tested with another database (http://epileptologiebonn.de).
The power spectral values for 50 seizures and 50 non seizures EEG timeseries
have been tested. Of the 50 epileptic seizures, 45 are identified correctly
and the remaining five segments are not detected.

Fig. 9(ae): 
Box plot obtained for 150 normal and 150 seizure segments 
Table 5: 
Comparison of the results obtained by our method 

It is also tested with 50 non seizure cases 46 are identified correctly and
the remaining four segments are not saying correctly non seizure. These results
are shown in the form of sensitivity, specificity and accuracy.
As expected, during the epileptic seizure activities, the magnitude of the spectral power increased. Box plots of the distributions of the normal EEG segments and the seizure segments are given in Table 2 and Fig. 7. There is an increase in the standard deviation during the seizure. Spectral method is applied here for single channel EEG segments since the seizure cases are generalized. The work can be extended to eight channels so that more insight into the localization of the seizure can be obtained. In the second method, statistical parameters are calculated for EEG signals for the datasets collected from two databases. The results of the investigation yields statistically significant differences in the parameters studied. The results shown in the Table 4 as statistical features namely variance, RMS, Standard Error Mean, Standard Deviation, Measure of Spread and Inter Quartile Range have statistically significant distribution for non seizure and during seizure. For the non seizure condition, the statistical feature values are considerably lower. The Box plots of the distribution of the statistical parameters are shown in Fig. 9. The study of the large set of data with normal and seizure measures are plotted and tabulated using spectral and statistical methods. CONCLUSION
The study presents automated detection technique of epileptic seizure activity
based on spectral and statistical methods. The EEG signals from two databases
are used here (Andrzejak et al., 2001 and http://epileptologiebonn.de).
We tested EEG segments each 150 signals with normal and seizure activity which
yielded a sensitivity of 93%, the specificity of 97% and average accuracy of
95%.
The statistical features are analyzed and it is inferred that discrimination between normal and seizure segments can be performed in a better manner if these are included along with spectral features.

REFERENCES 
1: WHO, 2007. Epilepsy: Aetiogy, epidemiology and prognosis. World Health Organization, Switzerland.
2: Fix, J.D., 1995. Neuroanotomy. 2nd Edn., Williams and Wilkins, Baltimore, MD., USA., ISBN13: 9780683032499, pp: 393398.
3: Gevins, A., H. Leong, M.E. Smith, J. Le and R. Du, 1995. Mapping cognitive brain function with modern high resolution electroencephalograph. Trends Neurosci., 18: 429436. CrossRef 
4: Mansouri, A., A. Fallah and T.A. Valiante, 2012. Determining surgical candidacy in temporal lobe epilepsy. Epilepsy Res. Treat.,
5: Wang, J. and G. Xu, 2009. Some highlights on epileptic EEG processing. Recent Patents Biomed. Eng., 2: 4857. Direct Link 
6: Padmasai, Y., K. SubbaRa, V. Malini and C.R. Rao, 2010. Linear prediction modeling for the analysis of the epileptic EEG. Proceedings of the International Conference on Advances in Computer Engineering, June 2021, 2010, Bangalore, pp: 69.
7: Liang, S.F., H.C. Wang and W.L. Chang, 2010. Combination of EEG complexity and spectral analysis for epilepsy diagnosis and seizure detection. EURASIP J. Adv. Signal Process., 10.1155/2010/853434
8: Tzallas, A.T., M.G. Tsipouras and D.I. Fotiadis, 2009. Epileptic seizure detection in EEGs using timefrequency analysis. IEEE Trans. Inform. Technol. Biomed., 13: 703710. CrossRef 
9: Zandi, A.S., R. Tafreshi, G.A. Dumont, C.R. Ries, B.A. MacLeod and E. Puil, 2007. Electroconvulsive therapy: A model for seizure detection by a wavelet packet algorithm. Proceedings of the 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, August 2226, 2007, Lyon, France, pp: 19161921.
10: Tafreshi, R., G. Dumont, D. Gross, C.R. Ries, E. Puil and B.A. MacLeod, 2006. Seizure detection by a novel wavelet packet method. Proceedings of the 28th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, August 30September 3, 2006, New York, USA., pp: 61416144.
11: Yucel, Z. and A.B. Ozguler, 2008. Detection of epileptic indicators on clinical subbands of EEG. Proceedings of the 16th European Signal Processing Conference, August 2529, 2008, Lausanne, Switzerland, pp: 15.
12: Pal, P.R. and R. Panda, 2010. Classification of EEG signals for epileptic seizure evaluation. Proceedings of the IEEE Students Technology Symposium, April 34, 2010, Kharagpur, India, pp: 7276.
13: Choe, S.H., Y.G. Chung and S.P. Kim, 2010. Statistical spectral feature extraction for classification of epileptic EEG signals. Proceedings of the International Conference on Machine Learning and Cybernetics, Volume 6, July 1114, 2010, Qingdao, China, pp: 31803185.
14: Andrzejak, R.G., K. Lehnertz, F. Mormann, C. Rieke, P. David and C.E. Elger, 2001. Indications of nonlinear deterministic and finitedimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Phys. Rev. E, Vol. 64. 10.1103/PhysRevE.64.061907
15: Lessard, C.S., 2006. Signal Processing of Random Physiological Signals. Morgan and Claypool Publishers, USA., ISBN13: 9781598290387, pp: 120121.
16: Blanco, S., A. Figliola, R.Q. Quiroga, O.A. Rosso and E. Serrano, 1998. Timefrequency analysis of electroencephalogram series. III. Wavelet packets. Phys. Rev., 57: 932940. Direct Link 
17: Djuric, P.M. and S.M. Kay, 1999. Spectrum Estimation and Modeling. In: Digital Signal Processing Handbook, Madisetti, V.K. and D.B. Williams (Eds.). CRC Press, Boca Raton, USA., pp: 346347.
18: Rangayyan, R.M., 2002. Biomedical Signal Analysis: A CaseStudy Approach. IEEE Press, New York, USA., ISBN13: 9780471208112, pp: 200284, 289290.
19: Greene, B.R., G.B. Boylan, W.P. Marnane, G. Lightbody and S. Connolly, 2008. Automated single channel seizure detection in the neonate. Proceedings of the 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, August 2025, 2008, Vancouver, Canada, pp: 915918.
20: Hoeve, M.J., R.D. Jones, G.J. Carroll and H. Goelz, 2001. Automated detection of epileptic seizures in the EEG. Proceedings of the 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Volume 1, October 2528, 2001, Istanbul, Turkey, pp: 943946.
21: Bruce, E.N., 2006. Biomedical Processing and Signal Modelling. John Wiley and Sons, New York, pp: 354356.
22: Kannathal, N., U.R. Acharya, C.M. Lim and P.K. Sadasivan, 2005. Characterization of EEGA comparative study. Comput. Methods Programs Biomed., 80: 1723. CrossRef 
23: Kannathal, N., M.L. Choo, U.R. Acharya and P.K. Sadasivan, 2005. Entropies for detection of epilepsy in EEG. Comput. Methods Programs Biomed., 80: 187194. CrossRef 
24: Guler, M.L., E.D. Ubeyli and I. Guler, 2005. Recurrent neural networks employing Lyapunov exponents for EEG signals classification. Expert Syst. Appl., 29: 506514. CrossRef 
25: Sadati, N., H.R. Mohseni and A. Maghsoudi, 2006. Epileptic seizure detection using neural fuzzy networks. Proceedings of IEEE International Conference on Fuzzy Systems, July 1621, 2006, Vancouver, BC., Canada, pp: 596600.
26: Subasi, A., 2007. EEG signal classification using wavelet feature extraction and a mixture of expert model. Expert Syst. Appli., 32: 10841093. CrossRef 
27: Tzallas, A.T., M.G. Tsipouras and D.I. Fortiadis, 2007. Automatic seizure detection based on time frequency analysis and artificial neural networks. Comput. Intell. Neurosci., 10.1155/2007/80510
28: Sivasankari, N. and K. Thanushkodi, 2009. Automated epileptic seizure detection in EEG signals using FastICA and neural Network. Int. J. Adv. Soft Comput. Appli., 1: 114. Direct Link 
29: Ubeyli, E.D., 2009. Combined neural network model employing wavelet coefficients for EEG signals classification. Digital Signal Process., 19: 297308. CrossRef 
30: Bedeeuzzaman, M., O. Farooq and Y.U. Khan, 2009. EEG signal classification for seizure analysis using statistical measures. Proc. EIP, 1: 366372.
31: NaghshNilchi, A.R. and M. Aghashahi, 2010. Epilepsy seizure detection using eigensystem spectral estimation and Multiple Layer Perceptron neural network. Biomed. Signal Process. Control, 5: 147157. CrossRef 
32: Song, Y. and P. Lio, 2010. A new approach for epileptic seizure detection: Sample entropy based feature extraction and extreme learning machine. J. Biomed. Sci. Eng., 3: 556567. Direct Link 



