2015 International Conference on Advances in Computer Engineering and Applications (ICACEA)

IMS Engineering College, Ghaziabad, India

A Review on Analysis of EEG Signals

Er. Jasjeet Kaur Er. Amanpreet Kaur

Research Scholar, Electronics and Communication Assistant Professor, Electronics and Communication
Engineering Department, Thapar University Engineering Department, Thapar University
Patiala Patiala

Abstract: Electroencephalography (EEG) enlighten about bipolar. In mono polar recording the electrode potential at an
the state of the brain i.e. about the electrical bustle going on in the active electrode is measured w.r.t some reference point such as
brain. The electrical activity measured as voltage at different earlobe. The bipolar recording is simply a voltage difference
points of brain act as basis of EEG. These signals are generally between two active electrodes. The mono polar recording is a
time-varying and non-stationary in nature. These signals can be
more popular method and is widely used.
scrutinized using various signal processing techniques. In this
paper, few statistical approaches to analyze EEG data are The signal extracted from electrodes is called raw EEG
conversed. signal. The raw EEG signal may include some non-cerebral
signals called artefacts. These are actually contaminations in
Keywords: EEG signals, analysis, methods. the signal. The EEG signals is contaminated by much other
kind of signals for example: - movement related potentials, eye
I. INTRODUCTION blinks, movement of facial muscles etcetera. Sometimes
signals like ECG, EMG, EOG also gets mixed up EEG signals.
The electric current produced by a normal human brain has These are called biomedical artefacts. These are the most
order of a few microvolts. These voltage fluctuations are the difficult to remove as most of the times these resemble the
result of ionic current that flows between brain and the actual EEG signal. Another category of artefacts are the
neurons. The spontaneous activity of brain is observed for environmental artefacts like line noise, pulse, electrode
about 20-40 minutes. This results in generation of the EEG stabilization etc. The elimination of these artefacts is very
signals. The electric potential picked up by a single neuron is important from the clinical point of view as minute mistake in
very small and thus it is impossible to detect. So the EEG the interpretation of the signal may turn fatal to the patient.
determines the summation of the synchronous activity of a Improvement in technology can easily decrease externally
large number of neurons present in the brain. These neurons generated artefacts, such as line noise, but biological artefact
have similar spatial orientation to get determined. Also, the signals must be removed after the recoding process.
Brain Computer Interface is one of the most anticipative edges Thus, a raw EEG signal undergoes various processes in
technology between humans and machines in today’s era. order to become readable. The signal is filtered and various
Electroencephalogram-based BCI systems has turned out to be techniques to obtain actual signal are performed. The filtering
a raging field in the investigation of brain engineering and of these signals is one area and analysis of these signals is
reintegration represented by another area. The feature extraction is generally
performed using statistical methods. And, after performing
There are four major types of EEG waves viz. alpha, beta, feature extraction, the signal is finally classified using support
theta and delta waves. The nature and occurrence of these vector machines or the neural networks. A few statistical
waves helps in identification of a number of diseases such as approaches for the analysis of EEG signals are discussed in
epilepsy, insomnia and many more. For instance, if high theta next section.
activity is present in an adult when awake then it must be
suffering from some abnormal pathological conditions. The II. STATISTICAL APPROACHES FOR EEG ANALYSIS
age and mental state also plays a vital role in the brain activity
of an individual. This can be judged from the example that an The statistical methods may be used as pre-processing or
alpha wave in an infant is completely normal but in fully post-processing for an EEG analysis. Sometimes a
grown adults, it might be a sign of some disease. The presence combination of one or more methods is used to achieve better
of spikes may be a sign of epileptic seizure or stroke. accuracy. Statistical analysis may include linear or non-linear
methods to fulfil its purpose. A number of linear and non-
The recording of EEG signals is simply done by placing linear methods have been applied to analyse the EEG signals.
electrodes on the scalp of the subject. The amplitude of EEG But some methods cannot be categorized as either of two
signal may vary from 10 to 100 micro volts with a frequency methods. This section gives a brief account of such methods
range of 1-100 Hz. The recordings may be mono polar or for example, wavelet transform, PCA and ICA.

978-1-4673-6911-4/15/$31.00©2015 IEEE

957

Authorized licensed use limited to: VIT University. Downloaded on May 18,2022 at 05:32:15 UTC from IEEE Xplore. Restrictions apply.
 2015 International Conference on Advances in Computer Engineering and Applications (ICACEA)
IMS Engineering College, Ghaziabad, India

1).Linear Methods Other than GARCH method, some other methods such as
The first in this category consists of prediction methods. Burg Method, Durbin Recursion and Yule-Walker method are
In the prediction methods a set of parameters is determined used as non-linear, non-parametric methods. Short data
that would help to figure out the signal generation system. The sequences are efficiently analyzed using this method.
linear prediction method mainly works in the time domain.
Mathematically, a linear prediction model can be represented 3).Short Time Fourier Transform
as This method deals with the signal in spectral domain. The
method provides a bridge between the Fourier analysis and
wavelet transform.
(1)
The FT is not an appropriate method to be used when it
Where , m = 1, 2, . . . , p, are linear parameters, n comes to non-stationary signals. This is because FT does not
represents the discrete sample time, and x(n) is the noise input. provide simultaneous time-frequency analysis and cannot
The discrete samples of time are normalized to unity. The represent discontinuities at the corners. Thus the FFT
given equation represents the autoregressive (AR) modeling of algorithm is applied on such signals.
an EEG signal. The AR modeling method is the most
commonly used linear prediction method. There are some But a better approach is STFT, in which the signal is
other methods such as adaptive filters, moving average filters, divided into small segments and the signal within this segment
weighted moving average filter, calculating randomness of is assumed to be stationary. Statistically,
signals and many more. However, these methods are not
always guaranteed to be linear.

The linear prediction method was efficiently used in

analysis, storage and transmission of EEG data [2]. This (3)
method gives a better view of background EEG activity. The Where,
autoregressive moving average (ARMA) method and multi x(t) is the signal,
variate AR (MVAR) approach are used as linear methods to y(t) is the window function,
analyze EEG signals in time domain. And the L.H.S represents the STFT of the signal.

2).Non-linear Methods The segmentation of signal is done using the window

Much like linear models, the output in non-linear models function that has equal width as the segment of the signal. The
is depends upon the output of previous models but non- choice of window function depends upon the minimum
linearly. A simple mathematical method to represent a non- frequency separation required to resolve two amplitude
linear model can be given by the following equation frequency components [7]. However, the size of window
function also require a good trade-off between the frequency
and time resolution as wider window function results in good
(2) frequency but poor time resolution and vice-versa. The
solution to this problem is overcome by wavelet transform [5].
Where, z(n) represents the input to the signal and x(n)
represents the output. Also, the functions h(.) and g(.) denote 4).Wavelet Transform
some kind of nonlinearity in the signal. Generally, these Wavelet transform has good localization properties in
functions represent non-linearity in mean and variance of time as well as frequency domain. The conversion of signal
signal, respectively. from time domain to the frequency domain gives a better
understanding of the signal [3]. The wavelet transform is
This method is popularly known as generalized preferred to be implemented for analyzing EEG signals
autoregressive conditional heteroskedasticity (GARCH) because of its dual property i.e. it can be used for discrete
method [4] and was originally used for time-varying volatility. (Discrete WT) and analog (Continuous WT).
The drawback of GARCH method is that the sign symmetries
are not entertained i.e., the inverse problem [1]. However, a The DWT is appropriate for analysis and the synthesis of
number of alternatives are present to deal with this problem. In the signal whereas the CWT is more suitable to identify
the field of EEG signals localization of sources deal with the diseases. Wavelet Transform is a powerful tool to analyze EEG
issue and the method is called low-resolution electromagnetic signals as there is an option to choose from many mother
tomography (LORETA) [6]. wavelets [10, 11]. Haar and Daubechies are the examples for
the same. The choice of mother wavelet also affects the
accuracy of the received output.

958

Authorized licensed use limited to: VIT University. Downloaded on May 18,2022 at 05:32:15 UTC from IEEE Xplore. Restrictions apply.
 2015 International Conference on Advances in Computer Engineering and Applications (ICACEA)
IMS Engineering College, Ghaziabad, India

And then, sometimes, multi resolution analysis (MRA) is intrinsic mode functions. The main characteristics of IMFs are
performed to achieve good time resolution and poor frequency given as follows:-
resolution at high frequencies and vice-versa.
i) They should have same number of zero-crossings and
5).Principal Component Analysis extrema (or difference should not be more than one).
A different technique that has been utilised to analyse ii) They should be symmetric w.r.t. local mean i.e. mean
EEG signals is principal component analysis that aims at the values of upper and lower envelopes must be equal to zero.
dimensionality reduction of the system. However, it is a
theoretical approach that that applies mathematical operations The signal is thus decomposed into oscillating
to obtain Eigen values but the method is extensively used for components which are further broken into lower and higher
analysis of EEG and removal of artefacts from it. It is a frequency components. Thus, it is not necessary for the signal
powerful tool for analysing and for dimension reduction of to be stationary. Further, the results of EMD when applied to
data without loss of information [8]. The data is linearly support vector machine yield even better results.
transformed in such a way that only orthogonal components
are retained. These orthogonal components provide maximum 8).Fractional Dimension
information about the signal. In particular it allows us to This method is helpful in reducing the complexity of
determine the principal directions in which data varies. These signals by means of their quantization. In this method, a set of
principal directions actually represent the data in the best points represent a bigger set of original values [16].
possible manner. PCA combined with different techniques For a given set of S original values, the fractional
provide great results [12]. One more advantage of PCA is that dimension is given
g as
it can be used for pattern recognition as well as compression of
data without much loss of useful information.
(4)
6). Independent Component Analysis
PCA reduces the data into orthogonal components. But a Where, is the minimum number of circles with
better and improved approached is independent component radius required to cover S.
analysis (ICA).
In this method higher order statistics such as kurtosis Fractional Dimension is more effectively used for short
parameter are used to separate the components. Independent EEG time series [14]. This technique has greater accuracy than
component analysis (ICA) separates EEG data into neural any other method.
activity and artefact; once identified; such components can be
deleted from the data. The orthogonality of the signal is not an
issue in this method. It is dependent on two assumptions: III. CONCLUSION

i) The source signals are independent of each other; The EEG signals have gained a lot of importance in the
ii) The distribution of values in each source signal is non- field of biomedical science in the past few decades. The
Gaussian. advancement in technology and its ever increasing demands
Independent Component Analysis is a strong tool to have encouraged the engineers to ascertain new methods for
separate different data such as artefacts or spatially analyzing these signals. Some of the most widely used
overlapping EEG activities [13]. But still, there are some methods have been discussed in this paper. These methods can
limitations of ICA. Firstly, it can be applied to decompose the be further modified or combined with some other methods to
number of sources equal to the number of scalp electrodes used get more appropriate results. Also, in the later stages, the
to collect the data. The number of factors contributing to the support vector machines or the neural networks can be used for
signal is generally unidentified. Secondly, for a small dataset, the classification of signal.
the temporal independence may not yield satisfying results.
Thirdly, it is not always right to assume that cerebral and IV. REFERENCES
artifactual sources are spatially fixed. This is the principle
assumption of ICA. [1]. D. B. Nelson,"Stationarity and persistence in the GARCH(1,1) model," J.
Econometrics, vol.45, 1990.
7).Empirical Mode Decomposition
[2]. N. Pradhan, D. N. Dutt, "Data compression by linear prediction for storage
Another statistical technique similar to principal and transmission of EEG signals," International Journal of Biomedical
component analysis is the Empirical Mode Decomposition. Computing, vol. 35, no. 3, pp. 207-217, April 1994.
The data is taken in time-series. The principle of this method is
to first decompose the signal into a sum of functions called [3]. S P Ghael, A M Sayeed, R G Baraniuk, "Improved Wavelet Denoising via
Empirical Wiener Filtering," in Proc. of SPIE, San Diego, 1997, pp. 389-399.

959

Authorized licensed use limited to: VIT University. Downloaded on May 18,2022 at 05:32:15 UTC from IEEE Xplore. Restrictions apply.
 2015 International Conference on Advances in Computer Engineering and Applications (ICACEA)
IMS Engineering College, Ghaziabad, India

[4]. M. Dacorogna, U. Muller, R. B. Olsen, O. Pictet, "Modelling short-term

volatility with GARCH and HARCH models," Nonlinear Modelling of High
Frequency Financial Time Series, Econometrics , vol. 7, pp. 161-176, May
1998.

[5]. R. M. Rao, A. S. Bopardikar, Wavelet Transforms introduction to theory

and applications.: Addison Wesley, Longman Inc, Reading, 1998.

[6]. R. D. Pascual-Marqui, M. Esslen, K. Kochi, D. Lehmann, "Functional

imaging with low resolution brain electromagnetic tomography (LORETA): a
review," Methods and Findings in Experimental and Clinical Pharmacology,
vol. 24C, pp. 91-95, Feb 2002.

[7]. M. G. Tsipouras, D. I. Fotiadis, A. T. Tzallas, "Epileptic Seizure Detection

in EEGs using time-frequency analysis," Information Technology in
Biomedicine, vol. 13, no. 5, september 2009.

[8]. N. N. Mbeledogu,M. Odoh., M.N. Umeh., "Stock Feature Extraction using

Principal Component Analysis," in International Conference on Computer
Technology and Science (IPCSIT), Hong Kong, 2012.

[9]. A. Galka,O. Yamashita, T. Ozaki, "GARCH modeling of covariance in

dynamical estimation of inverse solutions," Physics Letters A, vol. 333, 2004.

[10]. P. Senthil Kumar, R. Arumuganathan., K. Sivakumar, C. Vimal,

"Removal of artifacts from EEG Signals using Adaptive Filter through
Wavelet Transform," in International Conference on Signal Processing
(ICSP), Beijing, 2008.

[11]. A. Prochazka, J. Kukal, O. Vysata, "Wavelet Transform use for Feature

Extraction and EEG Signal Segments Classification," in International
Symposium on Communications, Control and Signal Processing, St. Julians,
2008.

[12]. R. Kottaimalai, M. P. Rajaskeran, V. Selvam, B. Kannapiran, "EEG

Signal classification using Principal Component Analysis with Neural
Network in Brain Computer Interface Applications," in International
Conference on Emerging Trends in Computing, Communication and
Nanotechnology, Tiruniveli, 2013.

[13]. S. Makeig, M. J. McKeown, A.J. Bell, Te-Won Lee, T. J. Sejnowski

Tzzy-Ping Jung, "Imaging Brain Dynamics using Independent Component
Analysis," Proceedings of IEEE, vol. 89, no. 7, 2001.

[14]. W. Lutzenberger, F. Pulvermüller, N. Birbaumer “ P. Hubert, "Fractal

dimensions of short EEG time series in humans," Neurosci. Lett., vol. 225, no.
2, pp. 77-80, 1997.

[15]. B. B. Mandelbrot, Geometry of nature. Sanfrancisco: W. H. Freeman

and Co,1983.

960

Authorized licensed use limited to: VIT University. Downloaded on May 18,2022 at 05:32:15 UTC from IEEE Xplore. Restrictions apply.