PARAMETRIC PERSON IDENTIFICATION FROM

THE EEG USING COMPUTATIONAL GEOMETRY

M. P o u ~ o s , (M.
~ )Rangoussi,(2)V. Chrissikopoulos,(l) A. E v a n g e l o ~ , ( ~ )
(')Dept. of Informatics,
University of Piraeus,
80, Karaoli and Dimitrioy str.,
Piraeus, GR- 18534,

(')Dept. of Electronics,
Tech. Educ. Inst. of Piraeus,
250, Thivon str.,
Aigaleo, GR-12244,

(3)Dept. of Exp. Physiology,

School of Medicine,
Univ. of Ioannina,
Ioannina, GR-45110,

GREECE
tel: $30 14112463

GREECE
tel: $30 1 5381225,

E-mai1:marios p@usa.net

E-mail:maria(Qem.t e i p i r . g r

GREECE
tel: -t30 651 97577
fax: $30 651 32045

ABSTRACT
Person identification based on features extracted
parametrically from the EEG spectrum is investigated
in this work. The method proposed utilizes computational geometry algorithms (convex polygon intersections), appropriately modified, in order to classify unknown EEGs. The signal processing step includes EEG
spectral analysis for feature extraction, by fitting a linear
model of the AR type on the alpha rhythm EEG signal.
The correct classification scores obtained on real EEG
data experiments (91% in the worst case) are promising in that they corroborate existing evidence that EEG
carries genetically specific information and is therefore
appropriate as a basis for person identification methods.

1. INTRODUCTION
The existence of genetic information in the EEG has
been investigated since as early as the 1930s, [a],[4].
However, it was not until in the 1960s that a direct connection was established between a person's EEG (especially the alpha and beta rhythms) and his/her genetic
information or code, thanks to the pioneering work of
Vogel with twins [18],as well as that of other researchers
in the same area, [3], [SI, [ll]. These reports showed that
EEG waveform morphology from monozygotic twins is
strikingly similar and demonstrates high inter-twin correlation for a variety of measures, [9], [13], [7],
[17].Such
results suggest that individual variation in EEG is determined genetically, although some inconsistency has
been observed and the strength of the association between dizygotic twins is not as high as it might be expected, when spectral analysis methods are employed to
quantify the EEG, [15], [13], [14].
Taken together, these findings have implied that the
EEG spectrum is determined by a polygenic mechanism,
such that the EEG appears to be an emergenic trait, or
one that has Y , a phenotypic characteristic that results

0-78 03-5682-919 9/$10.000 1999 I E E E .

from the (nonadditive) interaction of a configuration of

several independent genes, [9].
Earlier studies aimed in extracting genetic information among twins, monozygotic or dizygotic, and among
family members or shared features among pathological
cases (e.g., epileptics), [6]. In contrast, the main objective of this work is to show the existence of an one-toone correspondence between the genetic information of
a healthy adult person and certain features of his/her
EEG and, furthermore, to develop a system for person
identification based on the individuality of the EEG.
The proposed method constitutes an alternative to
the person identification method proposed by the authors in [la], where a different signal preprocessing step
is used. More specifically, the modification proposed
here is the parametric spectral processing of the EEG
for feature extraction purposes, by fitting a linear model
of the AR type on the signal. The target of this work
is twofold: to reach conclusions concerning the effectiveness of the present study regarding healthy person identification via EEG and to improve our previous approach
in terms of computational complexity.

2. T H E PROPOSED METHOD
The proposed method for processing of the EEG
recordings, feature extraction and classification can be
analyzed in the following steps.
Step 1. Isolation of the a rhythm EEG signal :
The recorded signal is spectrally analyzed using
Fourier Transform and the band between 7.5 Hz
and 12.5 Hz (alpha rhythm) is retained for further
processing. This frequency band is further segmented into five (5) subbands of 1 Hz width each.
At the resolution used, each subband contains 180
frequency points. The corresponding per subband
discrete signals ~ ; ( n ) i, = 1 , . . . , 5 are regenerated
via an inverse Fourier Transform.

I005

Step 2. Model fitting - parameter estimation :

A linear rational model of the ARMA type is fitted to the alpha rhythm EEG signals x,(n), which
is treated as a superposition of a signal component
(deterministic and periodic) plus additive noise
(random), due to imperfections of the recording
process. Standard parameter estimation procedures can be used for the estimation of the model
order and model paremeters, using second or third
order correlations of x l ( n ) (see, e.g., [16]).
At the present stage of our work we considered
purely AR models. The reason for that was to
keep the complexity of the proposed method low,
as it is known that for MA parameter estimation one should either resort to non-linear minimization or employ third order correlations (cumulants) of the data. EEG signals can be adequately handled via second order correlations,
since there is no evidence of non-causality present.
Therefore, the following overdetermined set of equations, based on autocorrelation alone, was formed
for each subband signal x t ( n ) and solved for the
AR parameters ap = [l a(1) a(2) . . . a(p)lT,
CZ,x(T)

+
T

as

eB

Figure 1: T h e convex polygon of an

vector of 10 coefficients.

AR parameter

a6

x6-

a7

P
Q(2)

CZ,r(T - 2)

0,

(1)

t=l

where q + 1 5

a10

5 M+p+

as

es

q , or, in matrix form,

C Z ,. ~aP = O,,

where p , q are the corresponding AR and MA part

moder orders, M > 0 accounts for the overdeterminancy present in the linear equation system and
matrix C Z ,consists
~
of the corresponding autocorrelation lags and is of dimensions M + p x p + 1.
In our case, model orders were set to q = 0 (AR
model) and p = 10. The later choice was made
experimentally; it should be noted that it is in
accordance with existing studies ([6]), where AR
orders ranging from p = 8 to p = 12 are proposed
for EEG recordings longer that 1 min.
The estimated AR parameters ap are used as
features describing the EEG signal ~ ~ ( n ) .

Step 3. Transformation of the EEG signal into a convex polygon:

The convex hull (or convex surrounding polygon)
of a set of points on the 2-D plane is computed.
These points are the values of the coefficients of
the AR model which were extracted in Step 2,
plotted against their index numbers [ l . ..p]. The
convex polygon, therefore, includes the values of
t,he coefficients of tshe AR model, (see Figure 1).
An appropriately modified version of standard computational geometry algorithms, 1121, is employed
for the computation of the convex hull of these
points. Thus each EEG recording produces a number of convex polygons equal to the number of
subbands.
Step 4. Intersection of convex polygons

1006

Figure 2: Intersection of two convex polygons, each

obtained from a n AR vector.
In this step the intersection of two convex polygons, such as those obtained in Step 3, is computed using the polygon intersection algorithm of
[lo] (see Figure 2).

Step 5 . Extraction of the characteristic polygon

A set of convex polygons, obtained as in Step 3
from a corresponding set of EEG recordings of a
given person, are iteratively intersected using the
Step 4 algorithm. This iterative procedure has
empirically been seen to converge to a final characteristic convex polygon, which is the common
section of the whole set of polygons and it represents this set. Convergence is achieved when the
surface change between two successive intersection
steps is less than a threshold surface, which is set
to a percentage of the whole surface. This polygon is assumed to characterize the individual and
it is the proposed feature set for the classification/identification step to follow.
Step 6. Classification :
First, for each one of a group of individuals and
for each frequency band, a characteristic polygon
is computed, as the common section polygon produced iteratively according to Steps 1, 2, 3 , 4 and
5 . Then, each candidate EEG is subjected to the
transformation and feature extraction Steps 1, 2

named X. The final pool of EEG recordings thus contained (4 x 45

75 x 1=) 255 recordings. All recordings were taken using a digital electroencephalograph
with the PHY-100 Stellate software. Subjects were a t
rest, with closed eyes. Voltage difference (in mVolts)
was recorded between leads 0 2 and CZ (one channel).
All EEG recordings lasted for three ( 3 ) continuous minutes, thus producing a 23040 samples long record each,
at the 128 Hz sampling rate used. A 1 - 30 Hz bandpass
filter was then applied on the EEG signals to retain only
useful information. Further processing was carried out
in Matlab, on a Pentium PC.

~~

3.2

Figure 3: Examples illustrating the 4 Possible Polygon intersection results. In all 4 cases, t,he same
P (bold line) is intersected
characteristic polygon
- -with 4 different test polygons X (thin line).

and 3. A convex polygon is thus produced which

represents the test EEG. This test convex polygon is intersected (Step 4) with the characteristic
polygons of each one of the group of individuals.
These intersections are computed for all frequency
bands. Let P denote a characteristic polygon of
one of the individuals in the group and X denote
the convex polygon of a test individual to be classified. There are four possible cases that can arise
from the intersection of P with X :

n X =P,
: P nX =X ,

case 1 : P
case 2

case 3 :

case 4 :

X = A,
X = 0.

Figure 3 illustrates these four cases, where PI is a

sub-polygon of P , produced in the third case. For
the test individual to be classified as a specific one
of the individuals of the group, the test convex
polygon should produce a case 1 result when
intersected with the characteristic polygons of the
specific individual, i n all frequency subbands. Otherwise the test individual is considered not to be
the specific individual.

3. EXPERIMENTAL PART
3.1

Data acquisition

A number of forty five (45) EEG recordings were taken

from each one of four (4) subjects named A, B, c and D,
In addition, one EEG recording was taken from each one
of seventy five (75) different subjects, to form a group

Application of t h e proposed method

In order to safely include alpha rhythm in the spectral

portion retained, the frequency band between 7.5 and
12.5 Hz was selected from each recording. This part
of the spectrum was next partitioned into five (5) subbands of 1 Hz width each. The discrete time signal that
corresponds to this part of the spectrum only was then
obtained via an inverse Fourier transform and a linear
model of the AR type was fitted on the data. Model
order p = 10 was empirically selected. The coefficients
of the AR model were plotted in the 2-D plane against
their index numbers 1 to 10 and their convex hull was
then computed, as in Step 3 above. For each one the t,he
five (5) frequency bands, four (4) characteristic polygons
were iteratively obtained as in Step 5 above. These are
representing individuals A, B, C and D, respectively. At
this point it was observed that each characteristic polygon would converge to a steady form on the 2-D plaen,
after approximately 10 iterations of the intersection St,ep
4. The rest 35 out of the total 45 polygons produced no
significant change of shape.

3.3

Results and discussion

Following Steps 1 to 5 of the proposed method and using

10 EEGs from each individual A, B, C and D, four characteristic polygons are obtained, PA, PB,
Pc and PD.
For the actual classification test, a test set of 215 EEGs
is formed using EEGs different from those used for the
characteristic polygon computations. The test set contains 35 EEGs from each of subjects A, B, C and D plus
75 EEGs from other subjects. Every characteristic polygon is intersected with each one of the test EEGs and
the same procedure is repeated for all five subbands.
The results are obtained as described in Step 6 above.
Correct classification scores are shown in Table 1.
As it can be seen in Table 1, an average correct classification score of 95% is obtained along the diagonal
(133/140), with diagonal entry scores ranging per individual from 91% (D) to 97% (A and C). This score represents positive classification answers (count a success
when an EEG of a certain individual is correctly classified as belonging to that individual). Negative classification tests were also performed on the same set of test
data: each one of the test EEGs was checked against the
characteristic polygons of all ot,her classes and a success
(EEG does not
was counted for each legative
to that
Test EEGs Of
Others were
mat,ched against the characteristic polygons of all four
I 007

n classified as: -+

11

1 C 1

Total

fl

[3] Buchbaum, M. S., Gershon, E.S., Genetic factors

in EEG, Sleep and evoked potentials, In: H u m a n
consiousness and its transformations, Ed. Davidson, Phenomen Press, 1978.
[4] Davis, H., Davis, P.A., Action potentials of the
brain in normal persons and in normal states of
cerebral act-ivit,y,Archives of Neurological Psychiatry, vol. 36, pp. 1214-1224, 1936.

!,

Table 1: Correct classification scores obtained in the

four subject classification test.

classes A, B, C and D. A correct negative classification

score of 86.5% (623/720) was obtained, which is comparable to the positive score of 95%.
Note that entries in a given row of Table 1 do not
sum up to 35 or to 75, respectively. This is because, the
way this test was set up, a positive answer (test EEG
belongs to a particular subject) was not guaranteed to
be unique, i.e., a test EEG might produce more than
one positive answers. Moreover, a positive answer was
not guaranteed a t all.
A comparison of the results shown in Table 1 to
those reported in [la], shows that the positive classification score has remained a t the same level (around 95%)
while the negative classification score has here decreased
to 86% from 96%. However, thanks to its paramatric
nature, the approach taken is seen to bear significant
improvement over that of [la], in terms of lower complexity and lower number of iterations for characteristic
polygon convergence.

4. CONCLUSIONS
Person identification based on features extracted parametrically from the EEG spectrum is investigated in this
work. The method proposed utilizes computational geometry algorithms (convex polygon intersections) in order to classify unknown EEGs. It provides an alternative approach to the one taken in [12], where features
extracted non-parametrically were used. While attaining the same level of correct positive classification scores
as before, the present approach results in considerable
computational savings, thanks to drastic reduction of
the complexity both in the feature extraction and the
classification steps. The classification scores obtained
(91% in the worst case) are promising in that they corroborate existing evidence that EEG signal carries genetically specific information and is therefore appropriate as a basis for person identification methods.

References

[5] Graham, R., An efficient algorithm for determining the convex hull of a finite planar set, Information processing Letters, vol. l , pp. 132-133, 1972.
[GI Hazarika N., Tsoi, A., Sergejew, A . , . Nonlinear
Considerations in EEG signal Classification, IEEE
Transactions o n Signal Processing, vol. 45, no. 4,
pp. 829- 836, 1997.
[7] Juel-Nielsen, N., Harvand, B., The electroenceplalogram in univular twins brought up apart ,
Acta genetica, vol. 8, pp. 57-64, Basel, 1958.
[8] Lennox, W., Gibbs, E., Gibbs, F., The brainpattern, an hereditary trait, T h e Journal of Heredity, vol. 36, pp. 233-243, 1990.
[9] Lykken, D.T., Research with twins: the concepts
of emergenesis, Phychophgsiology, vol. 19, pp. 361373, 1982.

[lo]

0 Rourke, J., Chien, C.B., Olson, T., Naddor,

D., A new linear algorithm for intersecting convex
polygons, Computer Graphics and Image Processing, vol. 19, pp. 384-391, 1982.

[ll] Plomin, R., The role of inheritance in behavior,

Science, vol. 248, pp. 183-188, 1990.

[12] Poulos, M., Rangousi, M., Kafetzopoulos, V., Person identification via the EEG using computational geometry algorithms, Proc. Intl. Conf. EUSIPCO98, Greece, Sept. 1998.

[13] Stassen, H.H., Bomben, G., Propping, P., Genetic aspects of the EEG: an investigation into the
within-pair similarity of monozygotic and dizygotic
twins with a new method of analysis, Electroencephalography and Clinical Neurophysiology, vol.
66, pp. 489-501, 1987.
[14] Eischen, S.E., Luckritz, J.Y., Polish, J., Spectral
analysis of EEG from families, Biological Psychology, vol. 41, pp. 61-68, 1995.
[15] Sviderskaya, N.E.,
Korolkova, T.A., Genetic features of the spatial organization the human cerebral
cortex, Neuroscience and Behavioral Physiology,
vol. 25, no. 5, 1995.
[I61 Swami, A . , Mendel, J.M., AR identification using cumulants, Proc. Workshop o n Nigher- Order
Spectral Analysis, Vail, C O , USA, pp. 1-6, 1998.

Vogt, P., Schalt, E., Vogel, F., A genetic study

of the human low-voltage electroencephalogram,
H u m a n Genetics, vol. 90, pp. 99-112, 1992.

[17] Varner, J., Potter, R., Rohrbaugh, J., A procedure for automatic classification of EEG genetic
variants, Processing of biological signals, 30.9-3,
Annual nternational Conference of the IEEE Engineering in Medicine and Biology Society, vol. 13,
no. 1, 1991.

[2] Berger, H., Das Elektrenkephalogramm des Menschen, Nova A c t a Leopoldina, Bd. 6 . Nr. 38, 1938.

[18] Vogel, F., The genetic basis of the normal EEG,

H u m a n Genetics, vol. 10, pp. 91-114, 1970.

[1] Anoklin, A., Steinlein, O., Fisher, C., Mao, Y.,

I008