Dr.B.

Krishna Kumar
Professor, Methodist College of Engineering and
Technology
Abids, Hyderabad
• Hans Berger ,in 1929,was the first to measure
ElectroEncephaloGraphy(EEG)Signals.

• EEG refers to the recording of the brain’s

spontaneous electrical activity over a period of
time, with the help of electrodes.
• Recording electrodes are arranged on the
scalp according to 10/20 international
system.

• At the time of recording, EEG signals are

often contaminated by Ocular Artifacts(OA)
such as eye movements and eye blinks.
• The voltage levels of OA are higher than that of
EEG signals. The contamination of EEG signals
due to OA’s, makes them to have low signal to
noise ratio.
• EEG is used to diagnose many diseases and
illnesses related to brain such as epilepsy,
tumor, carbovascular lesions and problems
related to trauma.
• EEG signals falls in the frequency range of
dc to 64Hz

• The voltage levels of EEG signals are in the

range of 1-5µv
 EEG DATA
• To collect EEG data, we follow 10/20
international system
To carry out the research work, the data sets were collected from
Physionet data base.
The details of the two data sets collected from Physionet data base are
provided below.
The sampling frequency is 100Hz and sampling interval 0.01sec

• Row Signal Gain Base Units

• 1 EEG Fpz-Cz 164.66080402 0 uV
• 2 EEG Pz-Oz 184.087078652 -184 uV
• 3 EOG horizontal 34.3294918806 -17 uV
• 4 Rasporo-nasal 32767.5 0
• 5 EMG Submental 9359.46872322 -5550 uV-mrs
• 6 Temp body 648.038569642 -52939 degC
• 7 Event marker 43482.8188624 -24754

• To convert from raw units to the physical units, subtract

'base' and divide by 'gain'.
 Types of Artifacts

• Processing of EEG signals is to remove the

unwanted bio-medical signals such as:

• Ocular Artifacts (Eye included artifacts )

• EMG (Electromyogram) artifacts(Muscle

Activation induced artifacts)

• Power line noise(50Hz)

 Objective

• The main objective is to remove the ocular

artifacts present in the EEG signals and

obtain the cleaned EEG signal and there by

estimating Signal to Noise Ratio (SNR)

HOW TO ACHIEVE THE OBJECTIVE
In EEG signal processing,

The observed signal ‘y’ can be modeled as:

Y=Hx+n
Where,
x is signal of interest

n is additive noise

H is matrix representing the observation process

• Since the EEG data is contaminated by
ocular artifacts (OA’s), hence the Signal
to Noise Ratio is very low.

• The Signal to Noise Ratio is low

because OA’s are 10 to 100 times
stronger than that of EEG signal.
• For the analysis of EEG signal, the SNR
should be very high and the algorithm
employed should provide low NF.

• Work has to be carried out to improve

the above mentioned factors and later
estimated.
 METHODOLOGY

The collected EEG signal is normalized for

using the following formula.

X- X
Xnorm =
std(X)
Where,
• x is the original EEG signal

• X is the mean of original EEG

• std(x) is the standard deviation of x

The observed signal ‘y’ (noisy EEG
signal)can be modeled as;
Y=Hx+n
Where,
x is signal of interest
n is additive noise; Here the noise is EOG
signal
H is matrix representing the observation
process
• Wavelet decomposition is to be done to
noisy EEG signal using different
wavelets separately, which results in
Approximate and Detail coefficients.
60
SNR using soft thresholding and IST

50 48.5

39.4
40 38.27
Soft
34.95
dB

29.43
30 28.89 IST

% Improvement
20.29
20

10 8.67

1.83
0
Sym8 Haar dB10
45
SNR using hard thresholding and IST
40 38.27 39.4
35.93 36.29
35

29.43
30
Hard
24.52 25.1
25 23.74
dB

IST
20 19.33

15 %
Improvement
10

0
Sym8 Haar dB10
BLIND SOURCE SEPARATION
TECHNIQUES
•ICA is one of the Blind Source Separation
methods

•ICA separates the mixture of independent signals

•FastICA is an extended version of ICA

•FastICA also separates the mixture of

independent source signals at a faster rate; hence
computational efficiency of it is better than that of
ICA
• One can extract a set of components or signals
from a set of measurements or signals by using
FastICA.
• The basic assumption is that the components
are mixed either linearly or non-linearly, and
along with the components of the mixing
system or matrix is assumed to be unknown.
• Another assumption is that the sources are
mutually independent.
• The choice of linear mixing model is:
X=AS
Where,
X is a matrix represents the observed signal or
data
Estimation of S from X is done with the help of
matrix A using the following formula:

-1
S= A X
•The EEG and EOG signals were collected from
physionet database.

•These signals are then linearly mixed producing

mixed signals as M1 and M2.

•M1=0.8 *A+0.2 * B
•M2=0.2 *A+0.8 *B
• M1 and M2 are called as mixed signals.

• Using these two mixed signals, one can construct the

observed matrix X and is shown below.

• X=[M1;M2]

• These two mixed signals are applied to FastICA

algorithm, which produces two independent
components namely IC1 and IC2
• Using these two independent components,
one can estimate the RMSV and RMSD
values.
PRINCIPAL COMPONENT
ANALYSIS
• PCA can be applied to EEG data that contains a
large number of measured variables to develop
into smaller number of artificial variables
called principal components.
• Obtaining a smaller number of variables from a
large number of measured variables is to
reduce the redundancy in the measured
variables.
• Here, the redundancy means some of the
variables in the measured data are correlated
with one another, because they are measuring
the same construct.
• The main idea of PCA is to reduce the
dimensionality of the data set, as the data set
consists of large number of interrelated
variables, and trying to retain as much as
possible variation present in the data set.
• In summary, the reduction of dimensionality of
dataset is achieved by transforming it into a
new set of variables called principal
components.
• These principal components are uncorrelated,
orthogonal and ordered in such a way that the
first few components retains most variation
present in all of the original variables.
• Two data sets namely EEG data set1 and
EEG data set2, were collected from
physionet data base.

• The size of each data set is 1x1000

• The two data sets were down sampled by a

factor of 2.This reduces the size of the each
data set to 1x500.
• For PCA to work properly, one has to find
the mean of each data set. This produces a
data set whose mean is zero. The mean is
calculated using the following formula.
n
 Xi
X = i=1
n
• Later, the mean of each data set is
subtracted from the corresponding data set
and then divide by the variance of
corresponding data set.
• The variance of corresponding data sets is
calculated using the following formula:

n 2
 (Xi-X)
 2
= i=1
n-1
• A noisy signal(EOG), whose variance is of 0.4 and
length 500, is added to the mentioned data sets;
which results in two noisy data sets.
• The two noisy data sets are of size 1x500 are
converted into a column vector.
• The size of column vector will be equal to 500x2.
• This column vector will be treated as noisy EEG
signal.
• The function known as wavelet Multiscale
2

PCA is applied to the noisy EEG signal for

wavelet decomposition to a level 6 and also
obtain principal components.

• Level of decomposition is explained in the

next slide.
50
45.79 SNR using soft thresholding and MSPCA
45
40 37.43 37.43 37.43
34.95
35
28.89 Soft
30
MSPCA
25 22.82
dB

20.29 % Improvement
20
15
10 6.63
5
0
Sym8 Haar dB10
40 37.43 37.43
SNR using hard thresholding and MSPCA
34.49 37.43 36.57
35 32.94
30
24.52 25.1 23.74
25 Hard
20
dB

MSPCA
15
10 % Improvement

5
0
Sym8 Haar dB10