Stable EEG Features

V. Stefanidis, G. Anogiannakis, A. Evangelou, and M. Poulos

Abstract The aim of this chapter is to identify stable points and stationary wavelets
in EEG signals. Generally an EEG signal is a very complex nonstationary signal. It
is very difficult to recognize specific EEG features such as Biometric patterns and
Pathological changes. Using a repeated autocorrelation procedure and symmetry
features of EEG time series on real EEG Time Series Data, we experimentally
investigate stable points in EEG signals. Also we investigate standing waves shafts
around these stable points, which reveals the existence of stationary wavelets in
EEG signals.

Keywords EEG signal • Data mining • Stationarity • Time series • Autocorre-

lation coefficients • LVQ neural network

1 Introduction

Time-frequency analysis of electroencephalogram (EEG) through several mental

tasks received significant consideration. As EEG is nonstationary, time-frequency
analysis is crucial to analyze brain states during different mental tasks [1]. In partic-
ular, it has been exposed that large-scale patterns of matched neuronal EEG activity
are ever varying and thus exhibit a substantial variability over time. Therefore, until
now, analysis of the EEG signal has been based mostly on statistical data processing
in order to acquire stable and reliable characteristics. The assumption underlying

V. Stefanidis • A. Evangelou
Laboratory of Physiology, Medical School, University of Ioannina, Greece
e-mail: [email protected]; [email protected]
G. Anogiannakis
Laboratory of Physiology, Medical School, Aristotle University, Thessaloniki, Greece
e-mail: [email protected]
M. Poulos ()
Laboratory of Information Technologies, Faculty of Information Science and Informatics,
Ionian University, Corfu, Greece
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 349

A. Migdalas, A. Karakitsiou (eds.), Optimization, Control, and Applications
in the Information Age, Springer Proceedings in Mathematics & Statistics 130,
DOI 10.1007/978-3-319-18567-5_18

[email protected]
350 V. Stefanidis et al.

such statistical analyses is the “stationarity” of the registered signal [1]. However,
in studies [1, 2], the EEG sources are considered quasi-stationary. In study [3] is
introduced that the length of EEG of which is so short that signal within it can be
treated as stationary or quasi-stationary [3].
The basic aim of the research study is to examine the stationary lengths of each
EEG in order to be corroborated the hypothesis of the previous study [3]. Thus, for
this implementation a novel EEG signal analysis is introduced.
It is known that the time-frequency information of EEG signal can be used as a
feature for classification in brain–computer interface (BCI) applications [4–14] or
for Diagnostic Purpose [7].
There are two alternative approaches to examining stationarity, the parametric
and the nonparametric. Parametric approaches are widely used by those undertaking
research in the time domain, such as economists, who are building certain assump-
tions about the nature of their data.
Nonparametric approaches are most commonly seen by researchers working in
the frequency domain, such as electrical engineers, who often treat the system as an
unknown entity and cannot make any inferences or reckoning based on the nature
of the data.
Nonparametric tests are not based on the knowledge that the population data
are normally distributed. By making no assumptions about the nature of the data,
non parametric tests are more widely applicable than parametric tests which often
require normality in the data.
While more widely applicable, the downside is that nonparametric tests are also
less powerful than parametric tests merely because the assumptions underlying their
use are fewer and weaker than those associated with parametric tests.
In this study, the symmetrical features of EEG are investigated using a well-fitted
autocorrelation coefficients (ACC) procedure. For this the schedule of this algorithm
is developed in the following three steps:
1. The ACC calculation using symmetrical features
2. Polynomial procedure on ACC
3. Graph Interpolation Procedure

2 The Method

2.1 The ACC Algorithm

This study is based on the hypothesis that the shape of a segment of a Time Series
may be described by the degree of asymmetry around a characteristic point. The
degree of asymmetry of a segment is obtained via the Pearson criterion [9] and is
described by the following equation:

X − Mo
S= (1)
s

[email protected]
Stable EEG Features 351

where S is the degree of asymmetry, X is the mean value of a time series segment,
Mo is the value of the characteristic time series (data) point, which is received as
the maximum value, s is the standard deviation of a time series segment. The degree
of asymmetry may be characterized as a necessary characteristic coefficient of the
time series segment in our case because this depicts a total geometry picture of
the segment. In this stage, the time series segment xt is subjected to power-spectral-
density analysis of each time series overlap segment, and computed using frequency
estimation of the standard periodogram, as follows:

1 T
Sx (f ) = | ∑ xt ej−2π tf |2 (2)
T t=1

The samples xt are replaced by the values of the periodogram given by |fn | where,
which can be computed using the fast Fourier transform (FFT) algorithm [10] thus,
according to the Inverse Function of formula 1, which is given as follows:

fn = (Sn−1 (Sn f )) (3)

This approach has also been applied to several cases, e.g., EEG study [4–14].
Using Eqs. (2) and (3), Eq. (1) is transformed into 4:
√
T T − 2(|f̂ | − |fg |)
Sk = 7 (4)
T/2 T/2
2T ∑k=1 fn2 − 4(∑k=1 fn )2

where fg = max(fn ) for 1  g  (int[T/2]).

Thereafter, we considered set D of sequences, which consists of the following
coefficients:

{D} = {D̂1 , D̂2 , . . . , D̂k−1 }

where

D̂1 = {S1 , S2 }, D̂2 = {S1 , S2 , S3 , }

D̂3 = {S1 , S2 , S3 , S4 }, . . . , D̂k−1 = {S1 , S2 , S3 , . . . , Sk }

Then, the ACC of the proposed method are computed as follows:

Ĉ = [Cm ] (5)

where m = 3, . . . , k + 2 is calculated via the condition of Eq. (6) which expresses the
corresponding weighing function using a Power Spectral Density window:

T
wm = int( ), where wm  3 (6)
m

[email protected]
352 V. Stefanidis et al.

Then, the ACC are given by

6
m=1 (D̂k D̂k ) − 4 | ∑m=1 (D̂k ) |
2k ∑k+2 T k+2 2
Sk = (7)
m(m − 2)

2.2 Polynomial Procedure on ACC

After that, we have to find the coefficients of the sixth degree polynomial p(x) (see
[4–14], as a justification of this option of 6 degrees of freedom for the fitting), in
order to achieve best fitting of p(x(i)) to C(i), where x(i) are the horizontal axis
elements (time or points) and C(i) are the ACC. The produced polynomial is known
as “Interpolation Polynomial.” The result p is a row vector of length n+1 containing
the polynomial coefficients in descending powers:

p(x) = p1 xn + p2 xn−1 + . . . + pn x + pn+1 (8)

2.3 Interpolation Procedure

Then, in order to find exactly the graph which best fits the coefficients of the
Interpolation Polynomial, for each element of the horizontal axis (x) we compute
the corresponding element on the vertical axis (y) according to following equation:

y = p1 xn + p2 xn−1 + . . . + pn x + pn+1 (9)

where pi is the previously computed Interpolation Polynomial. This is achieved

using delta error estimation, in which delta is an estimate of the standard deviation
of the error in predicting a future observation at x by p(x) [see Eq. (8)]

3 Experimental Part

As input for our experiments we used EEG signals which we received from people
who were in the process of relaxation. Three men and a woman participated in this
procedure. The duration of the EEGs was approximately 46 s (46,000 ms). The input
files approximately consist of 23,000 points (1 point/2 ms).
For the implementation of the experiments we made a function at Matlab,
which had an input of two parameters. The first parameter (k) is the number of
windows (samples) we will use from the EEG section, the second parameter (w)
is the size of the window (sample length). In case the number of windows is more
than one (k > 1), we have overlapping of used points (for instance 5000. . . 5020,
5001. . . 5021, etc). We executed the experiments with several combinations of the

[email protected]
Stable EEG Features 353

two parameters and we received several results which we present below. We used
mostly samples (w) of 20 points. From such a sample, of 20 points length, 28 ACC
{D} are isolated.
As we said in the previous chapter, we have to compute the coefficients of the
sixth degree “Interpolation Polynomial” p(x) [Eq. (8)], in order to achieve best
fitting of p(x(i)) to C(i), where x(i) are the horizontal axis elements (time or points)
and C(i) are the ACC. To achieve that we used the Matlab polyfit function:

p = polyfit(x, s, n) (10)

Figure 1 we can see a plot of these coefficients.

Then, in order to find exactly the graph which best fits the coefficients of the
Interpolation Polynomial, for each element of the horizontal axis (x) we compute
the corresponding element on the vertical axis (y) according to following equation:

y = p1 xn + p2 xn−1 + . . . + pn x + pn+1

where pi is the previously computed Interpolation Polynomial.

For every value x of the horizontal axis (time or points), the polyval function of
Matlab returns the value of a polynomial of degree n evaluated at x:

p = polyval(p, x) (11)

In Fig. 2 we can see the graph produced by the procedure described above.
The diagram we see in Fig. 2 corresponds to the representation of the polynomial
[Eq. (9)] which is the result of the execution of the function with k = 1 and w =
20 for a “man” subject. In Fig. 3 we can see a similar representation of the same
polynomial of a woman and another man EEG.
If we increase the number of samples k = 10 we get the representation of Fig. 4.
The representations we get, as we said before, have overlapped points.

Fig. 1 Coefficients of
interpolation polynomial

[email protected]
354 V. Stefanidis et al.

Fig. 2 Polynomial of degree

n evaluated at x [Eq. (9)]
graphical representation

Fig. 3 Polynomial [Eq. (9)] representation for a woman and a man, respectively

In Fig. 4 also, we can clearly distinguished the points we have stationarity. We

can easily distinguish these stable points (points with approximate zero amplitude
of motion) among which standing wave shafts are formed.
At the female signal we were forced to increase the width of the sample in order
to find the points for which stationarity appears. In Fig. 5, in the first diagram k is
equal to 10 (k = 10), w is equal to 20 (w = 20) and as signal section, we received,
randomly, the points from 5,000 up to 5,500. In the first diagram the stable points
are not easily distinguished. In the second diagram of the same figure k is equal to
10, w is equal to 20, and as a signal section, we received the points from 3,000 up
to 5,500 (we increased the section length). In this figure the stable points as well as
the shafts of standing waves are easily distinguished (Fig. 5).
A standing wave pattern is not actually a wave, but rather a pattern of a wave.
Thus, it does not consist of crests and troughs, but rather nodes and antinodes. In
other words, shafts are formed. The pattern is the result of the interference of two
waves to produce these nodes and antinodes. The parameters of this pattern are
shown in Fig. 6. In Fig. 7 we can see 6 EEG’s per person which depicted with a

[email protected]
Stable EEG Features 355

Fig. 4 Polynomial [Eq. (9]

representation for a man
(k = 10,w = 20), stable points

Fig. 5 Polynomial [Eq. (9)] representation for a woman (5,000–5,500 and 3,000–5,500 points)

Fig. 6 Static waves pattern

parameters (wavelength and
amplitude of oscillation)

color and (total 5 persons). Each EEG is submitted (auto repeat 20 times). In this
figure we can have a more cautious approach of standing waves.

4 Conclusions

According to above findings we conclude that stable points are presented clearly
during many intersections of the polynomials. These findings are consistent with
the theory [15] in which a common stable point may be shown to exist then

[email protected]
356 V. Stefanidis et al.

Fig. 7 Static waves in 6 EEG’s per person

the commuting polynomials have a common stable point and the community
polynomials yield a certain homomorphisms [16]. However, the homomorphisms
are continuous functions that preserve topological properties [17] and these are very
significant for pattern recognition reasons such as biometric EEG features [18] and
for diagnostic purposes [19].
Furthermore, the Data Mining of Time Series using Autocorrelation Coefficients
and symmetry features of EEG time series is addressed in this work. A repeated
autocorrelation procedure was performed on real EEG Time Series Data, in an
attempt to experimentally investigate and establish the connection between Time
Series data and hidden information relating to the properties of stationary Time
Series. These results are in agreement with previously proposed research methodol-
ogy/methods, exhibiting a Time-Series Analysis of real EEG, which carries genetic
information, as well as demonstrating the potential relevance of our approach for
Stationary Identification as a tool in Time Series Analysis.
Results show that the proposed algorithm can provide an optimal time-frequency
resolution using autocorrelation procedure in order to highlight stable EEG points
around a trust region. Although it is generally accepted that the EEG signals are
not stationary, we could isolate points which are stable and we can distinguish
stationarity.
Our future research will be focused on the following points:
• The verification of these results with more EEG data is the next step of the
proposed research.
• More extensive experimentation is needed in order to obtain statistically signifi-
cant results and, thus, check and verify assumptions on a Real Data set about the
existence of a one-to-one correspondence between the Time Series and symmetric
spectral features.
• The wavelet pattern recognition in order to create personal wavelet is the next aim.
• The comparison of the verification of the unique feature of each wavelet using
neural network is asked in the third researcher step.
• Finally, the connection of the findings and the possible biological-biometric
features are considered as further target of this research.

[email protected]
Stable EEG Features 357

References

1. Kaplan, A.Y., Fingelkurts, A.A., Fingelkurts, A.A., Borisov, S.V., Darkhovsky, B.S.: Nonsta-
tionary nature of the brain activity as revealed by EEG/MEG: methodological, practical and
conceptual challenges. Signal Process. 85(11), 2190–2212 (2005)
2. Zygierewicz, J., Mazurkiewicz, J., Durka, P.J., Franaszczuk, P.J., Crone, N.E.: Estimation
of short-time cross-correlation between frequency bands of event related EEG. J. Neurosci.
Methods 157(2), 294–302 (2006)
3. Wang, C., Xu, J., Lou, W., Zhao, S.: Dynamic information flow analysis in vascular dementia
patients during the performance of a visual oddball task. Neurosci. Lett. 580, 108–113 (2014)
4. Poulos, M., Rangoussi, M., Alexandris, N.: Neural network based person identification using
EEG features. In: Proceedings of 1999 IEEE International Conference on Acoustics, Speech,
and Signal Processing, 1999, pp. 1117–1120. IEEE (1999)
5. Poulos, M., Rangoussi, M., Chrissikopoulos, V., Evangelou, A.: Parametric person identifi-
cation from the EEG using computational geometry. In: Proceedings of ICECS’99. The 6th
IEEE International Conference on Electronics, Circuits and Systems, 1999, pp. 1005–1008.
IEEE (1999)
6. Poulos, M., Rangoussi, M., Chrissikopoulos, V., Evangelou, A.: Person identification based on
parametric processing of the EEG. In: Proceedings of ICECS’99. The 6th IEEE International
Conference on Electronics, Circuits and Systems, 1999, pp. 283–286. IEEE (1999)
7. Poulos, M., Georgiacodis, F., Chrissikopoulos, V., Evagelou, A.: Diagnostic test for the
discrimination between interictal epileptic and non-epileptic pathological EEG events using
auto-cross-correlation methods. Am. J. Electroneurodiagnostic Technol. 43, 228 (2003)
8. Poulos, M., Rangoussi, M., Chrissikopoulos, V., Evangelou, A., Georgiacodis, F.: Comparative
analysis of the computational geometry and neural network classification methods for person
identification purposes via the EEG: part 1. J. Discret. Math. Sci. Cryptogr. 7, 319–347 (2004)
9. Poulos, M., Papavlasopoulos, S.: Automatic stationary detection of time series using auto-
correlation coefficients and LVQ - Neural network. In: Fourth International Conference on
Information, Intelligence, Systems and Applications (IISA), 2013, pp. 1–4. IEEE (2013)
10. Mendonca, M.W.: Multilevel Optimization: convergence theory, algorithms and application
to derivative-free optimization. Ph.D. thesis, Facultés Universitaires Notre-Dame de la Paix,
Namur, Belgium (2009)
11. McArdle, J.J., Hamagami, F.: Modeling incomplete longitudinal and cross-sectional data using
latent growth structural models. Exp. Aging Res. 18, 145–166 (1992)
12. Felzer, T.: On the possibility of developing a brain-computer interface (bci). Technical
University of Darmstadt, Department of Computer Science, Darmstadt (2001)
13. Palaniappan, R.: Brain computer interface design using band powers extracted during mental
tasks. In: Conference Proceedings of 2nd International IEEE EMBS Conference on Neural
Engineering, 2005, pp. 321–324. IEEE (2005)
14. Marcel, S., Millán, J.R.: Person authentication using brainwaves (EEG) and maximum a
posteriori model adaptation. IEEE Trans. Pattern Anal. Mach. Intell. 29, 743–752 (2007)
15. Boyce, W.M.: Commuting functions with no common fixed point. Trans. Am. Math. Soc. 137,
77–92 (1969)
16. Ritt, J.F.: Permutable rational functions. Trans. Am. Math. Soc. 25, 399–448 (1923)
17. Oxtoby, J.C., Ulam, S.M.: Measure-preserving homeomorphisms and metrical transitivity.
Ann. Math. 42, 874–920 (1941)
18. Poulos, M.: On the use of EEG features towards person identification via neural networks.
Inform. Health Soc. Care 26(1), 35–48 (2001)
19. Poulos, M., Papavlasopoulos, S., Alexandris, N., Vlachos, E.: Comparison between auto-cross-
correlation coefficients and coherence methods applied to the EEG for diagnostic purposes.
Med. Sci. Monit. 10(10), 99–100 (2004)

[email protected]