Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 46 (2015) 1476 – 1482

International Conference on Information and Communication Technologies (ICICT 2014)

Wavelet sub band entropy based feature extraction method for BCI
Siva Sankar Aa,1,∗, Suparna S Naira,1,∗, Venu S Dharana,1,∗, Praveen Sankarana
a National Institute of Technology Calicut, Calicut 673 601, Kerala, India

Abstract
The study and analysis of the electrical activity of the brain is valuable in understanding the human mental state, intentions and will.
This aids the development of Brain Computer Interface (BCI), facilitating communication between the human brain and computer,
by converting the brain waves (EEG waves - Electroencephalography) into control signals. These control signals can then be used
to trigger an external device, thereby enabling a seamless communication with the intelligent system. It unfolds various avenues for
research and further applications in the realm of prosthetic device control, development of thought controlled intelligent systems
and other complex interfaces. This will also be an aid to persons with disabilities or various other amputations. In this work, a
novel feature extraction algorithm is proposed for extracting event related potentials from the EEG signals using wavelet sub band
entropy which can be used in BCI applications. An attention index is defined which gives a measure of the amount of concentration
or attention the subject has, upon focussing on a particular event or thought.

©c 2015
2014 The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of organizing committee of the International Conference on Information and Communication
Peer-review under
Technologies responsibility
(ICICT 2014). of organizing committee of the International Conference on Information and Communication
Technologies (ICICT 2014)
Keywords: BCI, Wavelet sub bands, Entropy, Feature extraction, Evoked potential, EEG

1. Introduction

The idea of electricity being generated by the body goes back as far as 1798, when the Italian anatomy professor,
Luigi Galvani claimed to have found out electricity in the muscle of a frog 1 . With the advent of modern day technology
of amplification and data processing, the study of these signals opened new avenues of knowledge. Currently rapid
development have been observed in this field of bio-signal processing. Motor cortex signal extracted from a group of
neurons are captured and used to actuate external devices in many cases. Research groups led by Richard Andersen,
John Donoghue, Phillip Kennedy, Miguel Nicolelis and Andrew Schwartz have achieved remarkable feats in this
area 2 . Using EEG to interface human thoughts with a machine is one among the primary areas of ongoing research.
Students of McMaster University has succeeded in bringing up an EEG controlled wheelchair prototype using real
time BCI and fast machine learning algorithm 3 . Research in prosthesis field involving the development of a bionic
limb has also progressed to a notable level. The DEKA Arm is a DARPA funded project aiming at the development
of functionality for the disabled which is still under development 4 . Fully developed products that provide support

∗ Corresponding authors Tel.: +0-904-816-1122

1 These authors contributed equally to this work
E-mail address: [email protected]

1877-0509 © 2015 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of organizing committee of the International Conference on Information and Communication Technologies (ICICT 2014)
doi:10.1016/j.procs.2015.02.067
 A. Siva Sankar et al. / Procedia Computer Science 46 (2015) 1476 – 1482 1477

for amputees are also becoming prominent. The i-limb ultra prosthetic hand is a powered device that can sufficiently
mimic the grasping operation of the human hand 5 .
An effective feature extraction algorithm is imperative to be able to successfully classify the various mental activi-
ties of the person for Brain Computer Interface applications. The EEG signal can be thought of as a superposition of
the electrical activities pertaining to a large number of tasks. These brain activities are oscillations that gets synchro-
nized by stimulation or while performing a particular task. The transition from a disordered to an ordered state results
in event related oscillations and the spectral content of these oscillations change in a particular fashion. Analysis of
these synchronizations can be efficiently done with the help of entropy based approaches since it can give a qualitative
measure of disorderness in the states of brain oscillations. This approach has been successfully applied in medical
studies, such as in epileptic seizures.

2. Methods

2.1. Data acquisition and analysis

Data acquisition of the raw EEG data was performed using NeuroSky MindwaveR , a single channel EEG acquisi-
tion device. The device incorporates a sensor located on the forehead, a reference that is clipped to the earlobe and
a chip that does the pre processing of the raw EEG data acquired from the electrodes. It provides single channel raw
EEG data along with Delta (0.1Hz to 3Hz), Theta (4Hz to 7Hz), Alpha (8Hz to 12Hz), Low Beta (12Hz to 15Hz),
Mid range Beta (16Hz to 20Hz) and High Beta (21Hz to 30Hz) values. Raw EEG data with an ADC resolution of 12
bits and a sampling frequency of 512 Hz is provided 6 .
NI LabVIEW R provides an excellent platform for graphical programming. Signal processing and analysis was
done on this Application specific Development Environment(ADE). It supports rapid prototyping and incremental
development of applications, from measurement and automation to real time embedded and general purpose applica-
tions. An unmanaged dynamic linked library based on C is provided by NeuroSky to access the virtual COM port
emulated by the headset. The interface with LabVIEW was done based on a LabVIEW - NeuroSky Driver for access-
ing the full functionality of the system level driver provided by NeuroSkyR . A window of 512 raw EEG samples were
read from the COM port at a time and processed in LabVIEWR .

Fig. 1. Raw EEG signal acquired from NeuroSky Mindwave

R
, a 512 window sample

2.2. Evoked Potentials

EEG consists of brain activities that span several frequency ranges, which are normally in a disordered state.
In the event of a particular brain activity such as attention, meditation, motor imagery or other perceptions, these
disordered waves synchronize or gets ordered. This transition from a disordered to an ordered state, results in a
resonance phenomenon due to the global synchronization of the EEG, giving rise to event related oscillations in
several frequency ranges 7 8 . Başar et al (1980) put forth this hypothesis according to which EEG consists of the
activity of an ensemble of generators producing oscillations in several frequency ranges 9 . The ERPs are reflected as
changes in the temporal domain in EEG during a particular event.
1478 A. Siva Sankar et al. / Procedia Computer Science 46 (2015) 1476 – 1482

2.3. Wavelet Decomposition

The theory of wavelets and filter banks have far reaching applications in the analysis of non - stationary signals. It
is a powerful tool for feature extraction, signal denoising and compression. Using Fourier analysis, one can express a
signal as the sum of an infinite series of sines and cosines. The major drawback in Fourier analysis is that it has no time
resolution. The Fourier basis functions are localized in frequency alone, and not in time. Small frequency changes will
hence produce changes everywhere in the time domain. Short Time Fourier Transform approach demands a constant
time window which in turn affects the frequency resolution. Wavelets are local in both frequency and time, which is
an advantage in many cases. Wavelet multi-resolution analysis is done in order to extract the various sub bands of the
raw EEG signal for the analysis of event related potentials.

2.3.1. Wavelet Multi - resolution analysis

Multi-resolution analysis (MRA) is an efficient tool in analyzing the information content in signals at various
resolutions or scales. It gives a hierarchical network of approximation and detail spaces. The idea of vector spaces is
the underlying concept to this representation. A vector space of higher and lower resolutions exists for each scale of
decomposition. The higher resolution finally yields the actual signal. The scaling function of wavelet is the basis of
each of the decomposed vector spaces.

2.3.2. Mathematical formulation of Mallat’s Fast Wavelet Transform algorithm

MRA defines a sequence of nested spaces of functions, {V j }. Mathematically, this can be represented as:

· · · ⊂ V j ⊂ V j−1 · · · ⊂ V2 ⊂ V1 ⊂ V0 ⊂ V−1 ⊂ V−2 · · · (1)

The lower the index, smoother the functions that belong to the space. At each decomposition level, the approximation
function becomes coarser. This sequence will, at the end, cover the space of the finite energy functions, L2 (R). For
x(t) ∈ L2 (R) and V j = span{φ j (t − k)} at j + 1th level of decomposition,
 
x(Aj+1) (t) = x(t), φ j+1 (t − k) (2)

The approximated signal at each decomposition level is obtained by taking the inner product of the original signal
with the scaling basis function corresponding to that scale. The two scale difference equation, which is central to
Discrete Wavelet Transform (DWT) connects the finer and the coarser scales 10 11 .

φ j+1,l (t) = ck,l φ j (t − k) (3)
k

Where, the scaling function coefficients 10

 
ck,l = φ j+1,l (t), φ j,k (t) (4)

φ j,k (t) = φ(2− j t − k) (5)

φ j+1,l (t) = φ(2−( j+1) t − l) (6)

Let τ = 2− j t − k
    k + τ − 2l  
φ j+1,l (t), φ j,k (t) = φ , φ(τ) (7)
2
 
 
τ k
= φ − l− , φ(τ) (8)
2 2
= cl− k (9)
2
 A. Siva Sankar et al. / Procedia Computer Science 46 (2015) 1476 – 1482 1479

The above equation yields,


φ j+1,l (t) = cl− k φ j (t − k) (10)
2
k

Substituting in 2, we get:

x(Aj+1) (l) = cl− k xAj (k) (11)
2
k

This can be interpreted as the convolution of the given signal, with the ck values, followed by down sampling yielding
the decomposed space. Using a similar approach in the detail space, we have:

x(Dj+1) (l) = dl− k xDj (k) (12)
2
k

where,
 
dk,l = ψ j+1,l (t), φ j,k (t) (13)

ψ(t) is the wavelet function and dk is the wavelet function coefficients. This is the basic idea behind Mallat’s Fast
Wavelet Transform Algorithm 12 . The choice of the wavelet depends largely on the signal under consideration. The
wavelet should have high approximation ability.

2.4. Wavelet entropy based approach

The Entropy, H for a discrete random variable X is defined as 13 :


H(X) = − P(X = ai )logb P(X = ai ) (14)
i

Where ai are the possible values of the random variable X. It reflects the degree of disorderness that the random
variable possesses. Considering the non stationary characteristic of EEG signal and the event related potential, multi-
resolution analysis of the signal has been done, which extracts signal in various frequency bands.
Wavelet packet decomposition performed on the raw EEG data with 5 levels yielded 64 sub bands, using Daubechies
wavelet with 8 taps (db8). The energy in each of these bands, Ek is calculated as:


8
Ek = x2j (15)
j=1

where, k is the index of the sub band and j is the number of data samples in each sub band.
The total energy of all the sub bands, Etotal are calculated as :


64
Etotal = Ek (16)
k=1

where, k is the index of the sub band.

Probability distribution for each level, pk given by :
Ek
pk = (17)
Etotal
The wavelet sub band entropy, WS k is given by :

WS k = −pk log2 (pk ) (18)

1480 A. Siva Sankar et al. / Procedia Computer Science 46 (2015) 1476 – 1482

Upon focussing or concentrating, the distribution of wavelet entropy along the sub bands was found to change in
a particular fashion as long as the subject maintained that level of concentration. Sub bands W10 to W15 were found
to have a significant change in the distribution pattern during the attention or concentration state. These 6 sub bands
were extracted and the sub band entropy sum, WS sum was calculated for these bands as:

k=15
WS sum = WS k (19)
k=10

The backward difference of the sub band entropy sum, ∇WS sum was calculated to define an index to measure the
level of attention.

∇WS sum = WS ksum − WS k−1

sum (20)

2.5. Attention Index

A fraction of the backward difference of the sub band entropy sum was cumulatively added if it goes below a
particular threshold and subtracted if it overshoots that threshold. This defines the Attention Index, A.
⎧
⎪
⎨An−1 +  , | ∇WS sum | ≤ η
An = ⎪⎩A −  , | ∇WS (21)
n−1 sum | > η

where, n = 1, 2, 3, ... is the nth window,  = k∇WS sum , k = 0.2, η is threshold value, 0.15 and A0 = 0.
The algorithm is summarized in the block diagram shown in figure 2.

Since the value of WS k lies between 0 and 12 , WS sum has a lower bound given by 0 and the upper bound is
determined by the number of wavelet sub bands under consideration. W10 to W15 has been used for calculating WS sum
i.e. number of sub bands(n) considered is six. WS sum is bounded according to the relation:
n
0 ≤ WS sum ≤ (22)
2

Fig. 2. General block diagram of the algorithm

2.6. Testing

The proposed algorithm for feature extraction of EEG can be used for performing various tests pertaining to cogni-
tive analysis. An experiment to analyze the workload or amount of cognitive stress was carried out on various subjects.
 A. Siva Sankar et al. / Procedia Computer Science 46 (2015) 1476 – 1482 1481

The subject was asked to concentrate or focus on a particular object or thought and the attention index was analyzed.
The measure of cognitive workload performed by the subject (here, the attention or concentration level) was used as
a parameter for moving an object (a wireless toy car).

3. Results

The entropy based feature extraction was found effective in giving a measure of the attention level a subject has
when he focusses on a particular thought. The signal processing algorithm can be implemented in a small DSP without
much resources, thereby making it suitable for wearable systems. The algorithm in its crude stage works under the
assumption that the neurons when fired in unity for a particular thought in mind, the entropy changes. This change in
entropy is found and is quantified and can be even stored for future reference in case the same pattern occurs again.
The algorithm is essentially immune to certain artifacts like popular noise source which has a particular fixed energy
pattern (for example, a 50Hz sinusoid). As the entropy of these patterns are constant, the backward difference stage in
the algorithm removes those noise patterns. The particular usefulness can be a problem if the user expects to extract
non transient features since they are removed in the algorithm stage.

Fig. 3. Entropy and energy plots of the sub bands (bar plot)

In figure 3, it is evident that the wavelet sub band entropies and the corresponding sub band energies are prominent
in the lower frequency levels. Significant changes in the sub band entropies were observed only from level 10 to level
15 while performing the test for attention index. The initial levels were discarded as it signifies the eye blink artifact,
which should not be considered in our study.

Fig. 4. Backward difference of the sub band entropy sum and the attention index

In figure 4, the backward difference of the wavelet sub band entropy sum for the sub bands from 10th to 15th level
has been plotted in NI LabVIEWR , with the subject asked to concentrate on a particular thought. The initial portion
1482 A. Siva Sankar et al. / Procedia Computer Science 46 (2015) 1476 – 1482

of the plot represents the backward difference when the subject concentrated and the fluctuation is very less, which
can be seen from the plot. The corresponding attention index is also seen to be increasing. The last part of the plot
shows a sudden spike in the backward difference and a corresponding rapid downward slope in the attention index,
indicating that the subject has lost focus in the thought.

4. Conclusion

In this work, a novel feature extraction algorithm for extracting EEG signals using wavelet sub band entropy sum
has been proposed. This algorithm in conjunction with a proper ANN can create powerful feature extraction and
classifying methods. The power lies in the extraction of the transient features which can be stored as a template for
future reference. Features representing ERPs based on entropy has the advantage of representing the current mental
state of the subject with precision. Wavelet sub band entropy can help in applications that require accurate results
from less number of input channels. Choosing relevant sub band entropies pertaining to specific ERPs for analysis
can aid many BCI platforms to meet their requirements.

References

1. Galvanis frog. http://galvanisfrog.com/; .

2. Brain Computer Interfacing. https://www.classle.net/book/brain-computer-interfacing; 2009.
3. Razvi, A.L.S.D.Z.. EEG-controlled wheelchair - McMaster University. https://sites.google.com/site/engfinalprojects/;
2012.
4. The DEKA Arm - A DARPA funded project . http://www.dekaresearch.com/deka_arm.shtml; 2009.
5. Touch Bionics - a world-leading prosthetic technology and supporting service provider. http://www.touchbionics.com; .
6. NeuroSky Mindwave datasheet. http://store.neurosky.com/products/mindwave-1; .
7. Başar, E.. EEG-brain dynamics: Relation between EEG and brain evoked potentials 1980;.
8. Quiroga, R.Q., Rosso, O.A., Başar, E., Schürmann, E.. Wavelet entropy in event-related potentials: a new method shows ordering of EEG
oscillations. Biological Cybernetics 2001;84:291–299.
9. Jayasree, T., Devaraj, D., Sukanesh, R.. Classification of Transients using Wavelet Based Entropy and Radial Basis Neural Networks.
International Journal of Computer and Electrical Engineering 2009;1(5).
10. Mallat, S.. A Theory for Multiresolution Signal Decomposition : The Wavelet Representation. IEEE Transactions on Pattern Analysis and
Machine Intelligence 1989;11(7).
11. Vetterli, M., Kovačević, J.. Wavelets and Subband Coding. Prentice Hall PTR; 2007.
12. Mallat, S.. A wavelet tour to signal processing. Academic Press; 3rd ed.; 2008.
13. Shannon, C.E.. A Mathematical Theory of Communication. The Bell System Technical Journal 1948;27:379–423,623–656.