Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication

(ICCMC)

Performance Comparison of IST and Multi Scale

Principal Component Analysis in the
EEG Signal Processing
Dr.B.Krishna Kumar1 and Dr.K.V.S.V.R.Prasad2
1
Professor, Dept.of ECE, Methodist College of Engineering and Technology, Hyderabad, TS, INDIA
e-mail: saisantu2004@yahoo.co.in
2
Prof & Head Dept. of ECE, D.M.S.S.V.H Engg.College, Machilipatnam, Andhra Pradesh
e-mail: kvsvr@yahoo.com
ABSTRACT

The removal of Ocular Artifacts (OA) in artifacts are the outstanding enemies of high quality
Electroencephalogram (EEG) data is one of the key EEG signals. These artifacts basically fall into two
challenges in the analysis of brain recordings. Brain major categories: technical and physiological
activity produces electroencephalogram signals, which
artifacts. The technical artifacts are often found in
consists of some of vital signs of neurological disorders
power line noise 50/60Hz results from poor electrode
such as epilepsy, tumor cerebrovascular lesions and the
problems associated with the trauma. These signals can
application on the scalp and transducer’s artifacts.
be acquired by placing the electrodes on the scalp at The physiological artifacts are often due to ocular,
specified positions and exists in order of 1-5μv, whose heart and muscular activity; are the EOG, ECG and
frequency range is DC-64 Hz. Acquisition of these EMG artifacts respectively [5].
signals mainly suffers from different unwanted
signals (artifacts or noise) resulting in less signal To date, many methods were presented to remove
information for identification. EOG artifacts. Some methods were based on
regression in the time domain or frequency domain
In this paper, two algorithms are proposed namely,
[6-8] for removing eye blink artifacts. However, they
Multi Scale Principal Component Analysis (MSPCA)
always need a reliable reference channel. Moreover,
and Iterative Soft Thresholding (IST) using wavelets in
removing the Ocular Artifacts (OA) present in the EEG
EOG Reference channel often contains brain signals
signals. This paper discusses not only the performance which will be also removed inevitably from the EEG
comparison of two algorithms on statistical parameters by the procedure of regression. Therefore, the
of EEG signals such as Signal to Noise Ratio, (SNR), methods based on regression may not be the best way
SNRI or Noise Figure (NF) and Absolute Average Error to remove EOG artifacts.
(AAE) but also estimated the run time of each
algorithm i.e., computational time of each algorithm. Principal Component Analysis is among the most
popular methods for extracting the information from
Keywords: EEG, EOG, Wavelets, Denoising,
the data and has found applications in a wide range of
Thresholding, PCA, Multi scale PCA(MSPCA), Ocular
artifacts and Wavelet transform, Iterative Soft Thresholding disciplines [9]. PCA was introduced by Pearson in
1901[10] and developed by Hotelling [11]in the year
1933.Using Multi Scale Principal Component
1. INTRODUCTION Analysis the EEG signal is decomposed into 6 levels.
Later, estimation of the statistical parameters of EEG
Electrooculography (EOG) artifacts inevitably and
signals is done. Finally the performance comparison
frequently interfere with the electroencephalogram
of these two algorithms is made.
(EEG) signals. Eye-movement and eye-blink artifacts
are the main sources of ocular artifacts. The In this paper, IST[12] using wavelets algorithm [13]
frequency components of EEG signals are in the the EEG signal is decomposed into a level 4, which
order of 1-5μV, and their frequency content differs gives us approximate and detail coefficients. The
among the different neurological rhythms, such as, detail coefficients which are having more noise
alpha, beta, delta and theta rhythms [1-4]. However, information are processed with IST and later

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 536

 Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication
(ICCMC)

estimated the statistical parameters such as SNR, clipping, etc [12]. The observed signal y can be
SNRI and AAE [13]. written as

The noisy EEG signal is convolved with a low and y = Hx + n

high pass filter whose impulse response is determined Where x is the signal of interest which one would
by the wavelet chosen. The output of each filter like to estimate, n is additive noise, and H is a matrix
produces the same number of samples as the original representing the observation processes. For example,
signal, so both outputs are down sampled by 2 if one observes a blurred version of x then H will be a
resulting in the Approximate and detail coefficients convolution matrix. The estimation of x from y can
each with half the number of points as that of the be viewed as a linear inverse problem. A standard
original signal. approach to solve linear inverse problems is to define
a suitable objective function J(x) and to find the
The coefficients represent a correlation between the
signal of interest and wavelet chosen at different signal x minimizing J(x)[12]
scales and during translation. Because all of the
coefficients are preserved, the original signal or any Generally, the chosen objective function [12] is the
level of decomposition can be reconstructed using a sum of two terms:
filter scheme similar to decomposition shown in
Figure 1. The process is reversed and now the J (x)=D(y,Hx)+OR(x)
coefficients are up sampled (interpolated), filtered,
and summed. Where D(y, Hx) measures the discrepancy between y
and x and R(x) is a regularization term (or penalty
function). The parameter λ is called the regularization
parameter and is used to adjust the trade-o between
Down sample the two terms; λ should be a positive value. On one
hand, we want to find a signal x so that Hx is very
Low pass filter 2 similar to y; that is, one would like to find a signal x
Approximate coefficients which is consistent with the observed data y. For
D(y; Hx), we will use the mean square error, namely

Signal Down sample 2

D(y,Hx) = y - Hx 2

High pass filter 2 The v

2
represents the sum of squares vector 'v'
Detailed Coefficients 2

Figure.1.Block diagram to show DWT decomposition 2.1 Iterative Threshold Algorithm

The EEG signal is decomposed into 4 levels using The objective function minimization, previous case
dB10,Sym8 and Haar wavelets and corresponding
approximate and detail coefficients are shown in the J(x) = y - Hx
2
+λ x
2 1
Fig.2.From derived coefficients high frequency
components will be distributed in approximate
The J(x) can be minimized by combining
coefficients and low frequency components presents Majorization-Minimization approach
in Detail coefficients cD4, cD3, cD2, cD.
To find a majorizer of J(x), coincides J(x) at x k and
2. ITERATIVE SOFT THRESHOLDING which is easily minimized. Adding non-negative
function to J(x)
This method describes an approach for the restoration
of degraded signals using sparsity. This approach, Gk (x)= J(x)+(x - x k )t (αI - Ht H)(x - x k )
which has become quite popular, is useful for
numerous problems in signal processing: denoising, By design Gk(x) coincides J(x) at xk. one needs to
de convolution, interpolation, super-resolution, de minimize Gk(x) to get xk+1.

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 537

 Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication
(ICCMC)

From the above equation, we can get reveals that the presented method of modeling gives

2
the good performance among all the previous
Gk (x)= y - Hx 2 +(x - x k )t (αI- HtH)(x - x k ) techniques.

Rewriting the equation using soft thresholding, A4

2 20
1
Gk (x)= α x k + Ht (y - Hx k )- x + λ x 1 + K 0
α 2
-20
0 20 40 60 80
D1
Where, K is constant w.r.t x. Minimizing G k(x), 2
equivalent to Minimizing (1/ D )Gk ( x) so xk+1, 0
-2
obtained by minimizing: 0 100 200 300 400 500 600
D2
2
2
1 λ 0
xk + Ht (y - Hx k )- X + x1
D 2 D -2
0 50 100 150 200 250 300
D3
2
Therefore minimizing is achieved by the following
0
soft-thresholding equation [12] -2
0 50 100 150
§ 1 λ ·
Xk+1 = soft ¨ (x k + Ht (y - Hx k )- X), ¸
D4
© D 2D ¹ 5
0
-5
Where α≥ maxeig (HtH). This gives the Iterated Soft 0 20 40 60 80
Threshold Algorithm.

The iterative soft Thresholding algorithm is used for

a)Approximate and detail coefficients using dB10
denoising the EEG signal. The model consists of
three main steps i.e, normalization, wavelet
decomposition and applying the IST for the detail
coefficients. In this paper EEG signals are taken from
Physionet database [15].

Data can be normalized to increase the accuracy of

the application as follow:

X- X
Xnorm =
std(X)

Normalizing of EEG signal is very important because

of possible variations in signal acquisition from trial
to trial, hence the normalization of data is necessary
for valid results.EEG data is decomposed into 4
levels using dB10, Sym8 and Haar wavelets as shown
in Figure 2(a,b and c) and all the detail coefficients
are scaled by IST algorithm[13], finally by applying
inverse wavelet transform results the enhanced EEG
signal. Algorithmic flow result is shown in Figure 3

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 538

 Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication
(ICCMC)

A4
20 Normalizing Wavelet 4 level
0 Transform Decomposition
-20
0 20 40 60 80
D1
2
0 Detail Coefficients Approximate
-2 coefficients
0 100 200 300 400 500 600
D2
2
0 IST IDWT
-2
0 50 100 150 200 250 300
D3
2
0
-2 Clean
0 20 40 60 80 100 120 140 EEG
D4
5
0
-5
0 20 40 60 80 FIGURE 3: IST Using Wavelets Algorithm Flow

b) Approximate and detail coefficients using sym8

The noisy EEG signal and modeled EEG signal using
IST algorithm is shown in the Figure 4.
A4
20
0 x 104 Collected EEG signal
2
-20
0 10 20 30 40 50 60 70 0
D1
2 -2
0 100 200 300 400 500 600 700 800 900 1000
0 Original EEG signal
-2 5
0 100 200 300 400 500
D2 0
2
-5
0 0 100 200 300 400 500 600 700 800 900 1000
-2 Noisy EEG signal
0 50 100 150 200 250 5
D3 0
5
0 -5
0 100 200 300 400 500 600 700 800 900 1000
-5 Denoised EEG signal
0 20 40 60 80 100 120 140 5
D4
5 0
0 -5
-5 0 100 200 300 400 500 600 700 800 900 1000
0 10 20 30 40 50 60 70

Figure.4: Noisy EEG Signal and Modeled signal

c) Approximate and detail coefficients using Haar

FIGURE 2

Noisy
EEG

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 539

 Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication
(ICCMC)

3. RESULTS OF IST Table.2. statistical measures of EEG signal using

MSPCA
The results obtained from IST using wavelets
algorithm [13] are provided in the Table.1 Wavelet /Threshold AAE(dB) SNR(dB) SNRI(dB)
Sym8 (soft) 23.58 34.95 25.34
Table.1. statistical measures of EEG signal using IST Sym8 (hard) 19.26 24.52 48.00
Haar (soft) 24.89 20.29 10.75
Wavelet/Thresho AAE SNR SNRI or
4. MULTISCALE PCA ld type (dB) (dB) NF(dB)
Sym8 (soft) 23.58 34.95 25.34
PCA is a technique which is generally used for
Sym8 (hard) 19.26 24.52 48.00
reducing the dimensionality of multivariate datasets
i.e. reducing the number of dimensions, without Sym8 (IST) 18.26 38.27 50.20
much loss of information. Considering a vector of n Haar (soft) 24.89 20.29 10.75
random variables x for which the covariance matrix Haar (hard) 18.18 25.10 46.08
is Σ, the principal components (PCs) can be defined Haar (IST) 17.22 39.40 52.30
by z = Ax Where z is the vector of n PCs and A is dB10(soft) 14.88 28.89 34.58
the n by n orthogonal matrix with rows that are the dB10(hard) 19.98 23.74 47.69
eigenvectors of Σ. The Eigen values of Σ are dB10(IST) 14.44 29.43 49.43
proportional to the fraction of the total variance Haar (hard) 18.18 25.10 46.08
accounted for the corresponding eigenvectors, so the dB10 (soft) 14.88 28.89 34.58
PCs explaining most of the variance in the original
dB10 (hard) 19.98 23.74 47.69
variables can be identified. If, as is usually the case,
some of the original variables are correlated, a small MSPCA 13.72 37.43 46.91
subset of the PCs describes a large proportion of the
variance of the original data.
From Table.1 and Table.2 it is clearly observed that
Multi scale PCA (MSPCA)[14,16] combines the both the proposed algorithms i.e., IST and MSPCA
ability of PCA to extract the cross correlation or are providing better results over the conventional
relationship between the variables, with that of Thresholding techniques namely soft and hard
orthonormal wavelets to separate deterministic Thresholding.
features from stochastic processes and approximately The algorithmic flow of MSPCA [14] is shown in the
de correlate the autocorrelation among the Figure.5.
measurements. To combine the benefits of PCA and
wavelets, the measurements for each variable For Each Column Data Matrix
(column) are decomposed to its wavelet coefficients
using the same orthonormal wavelet for each
variable. This results in transformation of the data Compute Wavelet Decomposition
matrix, X into a matrix, WX, where W is an n ´ n
orthonormal matrix representing the orthonormal
wavelet transformation operator containing the filter Compute Covariance Matrix for
coefficients.
Selected Scales
MSPCA removes the artifacts like EMG and ocular
artifacts. To remove the artifacts PCA is done twice
on the same data by selecting only first four principal Compute PCA
components,and the results were given in figures 5
for two different data sets taken from reference
[15].The results of MSPCA [14, 16] are provided in Select Important PC’S
the Table.2.

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 540

 Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication
(ICCMC)

Algorithm Computational Time(sec)

IST 1.391266
Select Wavelet coefficients using MSPCA 2.694674
threshold
6. CONCLUSIONS

From the above Table.1, Table.2 and Table.3 it is

For all scales together compute PCA clear that IST not only providing better statistical
parameters of EEG signal but also take less time in
computing them when compared to MSPCA.

Reconstruct the data matrix from the

7. REFERENCES
selected and threshold scores at each
scale [1]. N.V. Thakor et al. “Multi Resolution Wavelet Analysis of
Evoked Potentials”, IEEE Transactions on Biomedical
Engineering, Vol. 40, No 11, pp. 1085-
1093,November,1993.
FIGURE.5: MSPCA Algorithm Flow
[2]J.Malmivuo and R. Plonsey, "Bioelectromagnetism: Principles
The noisy EEG signal and modeled EEG signal using and
Applications of Bioelectric and Biomagnetic Fields", New York:
Multi Scale Principal Component Algorithm is Oxford University Press, New York, Ch.28,1995.
shown in the Figure 6.
[3] D. Nov´ak, L. Lhotsk´a, V. Eck, and M. Sorf, "EEG and VEP
Signal Processing. Cybernetics", Faculty of Electrical
Engineering, pp.50–53, 2004.

[4] L. S¨ornmo and P. Laguna, "Bioelectrical Signal Processing in

Cardiac and Neurological Applications", Elsevier, Academic Press,
Original EEG signal data1 2005.
5
[5]. S. Ventakaramanan, P. Prabhat, S.R Choudhury, H.B Nemade,
0 and J.S. Sahambi “Biomedical Instrumentation Based On
Electrooculogram (EOG) Signal Processing And Application To A
-50 100 200 300 400 500 Hospital Alarm System”, Indian Institute Of Technology (IIT)
EOG Corrupted EEG signal data1 Guwahati, Proceedings of IEEE ICISEP, pp.535-539,2000.
5 [6] Schlogl A, Keinrath C, Zimmermann D et al. ”A fully
automated correction method of EOG artifacts in EEG recordings”
0 Clinical Neurophysiology.118, pp. 98–104, 2007.
-50 [7] Jung T-P, Makeig S, Humphries C, Lee T-W, McKeown
100 200 300 400 500 MJ,Iragui V and Sejnowski TJ. “Removing
Recovered EEG signal data1 Electroencephalographic artifacts by blind source separation”
5
Psychophysiology.37, pp.163 –178,2000.
0 [8] Gratton G, Coles MG and Donchin E “A new method for off-
line removal of oucular artifact”, Electroencephalography Clinical
-50 100 200 300 400 500 Neurophysiology 55(4), pp.484– 486, 1983.
Original and Recovered EEG signal-Lead1
data1
5 data2 [9] M.Aminghafari,N. cheze, J.M Poggi, ”Multivariate denoising
0 using wavelets and principal component Analysis,” computational
statistics & Data Analysis, vol. 50 pp. 2381-2398,2006.
-50 100 200 300 400 500
[10] Pearson, K, "On Lines and Planes of Closest Fit to Systems of
Points in Space", Phil. Magazine, vol.6, issue. 2, pp.559–572,
FIGURE.6: The noisy EEG signal and modeled EEG 1901.
signal using MSPCA [11] Hotelling. H," Analysis of a Complex of Statistical Variables
Into Principal Components", J. Educ. Psychol., 24, pp.417–441,
5. RESULTS pp.498–520, 1933.
Computational Efficiency of IST is better than
[12] Ivan W.Selesnick, "Sparse Signal Restoration",
MSPCA is provided in the Table.3. pp.1-16, April6, 2010.
Table.3. Computational Efficiency of IST and
MSPCA [13]B.Krishna Kumar, Dr.K.V.S.V.R.Prasad
Dr.K.Kishan Rao and J. Sheshagiri Babu, “Modeling of EEG
Signals using Wavelets with Iterative Soft Thresholding”, i-

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 541

 Proceedings of the IEEE 2017 International Conference on Computing Methodologies and Communication
(ICCMC)

manager’s Journal on Future Engineering & Technology, Vol.6,

No.4, pp.37-42, May - July 2011.

[14]B.Krishna Kumar, Dr.K.V.S.V.R.Prasad, Dr.K.Kishan Rao

and Narsimha Baddiri, "Analysis of EEG Signals Using Multi-
Scale Principal Component Analysis", IEEE Conference, Cape
Institute of Technology, Levingapuram, Kanyakumari, pp.359-363,
December15-18, 2011.

[15]http://physionet.ph.biu.ac.il/physiobank/database

[16] www.mathworks.com

978-1-5090-4890-8/17/$31.00 ©2017 IEEE 542