A Different Statistical Approach Aiming at EEG Parameter Investigation
for Brain Machine Interface Use
Maria Claudia F. de Castro and Fabio Gerab
1
Electrical Engineering Department, Centro Universit
´
ario da FEI, S
˜
ao Bernardo do Campo, Brazil
2
Mathematics Department, Centro Universit
´
ario da FEI, S
˜
ao Bernardo do Campo, Brazil
Keywords:
EEG, Power Spectral Density, Frequency Bands, Spatial Feature Selection, Pattern Recognition, Statistical
Analysis.
Abstract:
A lot of effort has been made to investigate EEG features that could better represent signal characteristics.
The results are usually based on the best mean recognition rates and statistical analysis is done only when
different methods are compared. In this work, we propose a new approach that applies multiple rate inter-
comparisons based on large samples aiming at detecting differences among treatments in order to recognize
their importance for the classification rates. Ten frequency band compositions expressed by power spectral
density averages were extracted from 8 EEG channels during 4 motor imageries, and spatial feature selections
were also considered during the recognition process. Classification rate in large samples can be represented
by a normal distribution and, for multiple rate inter-comparisons, the level of significance was corrected based
on the Bonferroni Method. The variables were considered to be independents and the test was performed as
non paired samples in a very conservative approach. The results showed that there are significant differences
among cases of spatial feature selection and thus the considered electrodes are important parameters. On the
other hand, considering or not the Delta and Theta bands along with different arrangements for Gamma band
resulted in no significant difference.
1 INTRODUCTION
Brain Machine Interfaces (BMI) studies have been fo-
cused on the development of rehabilitation systems
and, more recently, on the application to games and
automation systems. One of the ways to implement
these systems is to use motor imageries recorded
through Electroencefalogram (EEG) from the cortex,
using some kind of processing technique to identify
specific patterns related to the intended movement.
After that, these patterns can be translated into con-
trol commands to external devices (Al-Ani and Trad,
2010; Millan et al., 2010).
Many authors declare that motor imageries can
modify the neuronal activities in the primary senso-
rimotor areas in a very similar way as observed when
the movements are really executed. This implies the
predominance in the acquisition of the EEG from the
motor area (C3, C4, Cz) (Hema et al., 2010; Herman
et al., 2008). Other studies are based on multichannel
acquisition (Liu et al., 2005; Higashi et al., 2009).
In addition to the position of the acquired signals,
it is important to extract features from original EEG
signals, which are able to distinguish among men-
tal states. The EEG signal has a frequency spectrum
ranging from 0.1 Hz to 100 Hz which is classified into
five frequency bands. There is no consensus about
the exact band limits. Small differences exist depend-
ing on the author. However, in general, the frequency
bands can be considered as Delta (0.1 - 4 Hz) , Theta
(4 - 8 Hz), Alpha (8 12 Hz) , Beta (12- 28 Hz)
and Gamma (28 - 100 Hz). The task of EEG spec-
tral quantification is particularly challenging consid-
ering the complexity of the dynamics of non station-
ary EEG. It is required to take into account the time
variation of the relevant frequency components (Liu
et al., 2005; Herman et al., 2008; Al-Ani and Trad,
2010; Hema et al., 2010).
Another part of the success of a BMI is dependent
on subject’s training and motivation, making them
able to learn to control the intensities of specific fre-
quency bands, which can be used for the communi-
cation feature (Herman et al., 2008; Al-Ani and Trad,
2010; Hema et al., 2010).
Herman et al. (2008) presented an extensive
comparative study involving different approaches to
244
Claudia F. de Castro M. and Gerab F..
A Different Statistical Approach Aiming at EEG Parameter Investigation for Brain Machine Interface Use.
DOI: 10.5220/0004804602440250
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2014), pages 244-250
ISBN: 978-989-758-011-6
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
spectral signal representation such as power spec-
tral density techniques, atomic decompositions, time-
frequency energy distributions, continuous and dis-
crete wavelet approaches, and also various classifiers
aiming at distinguishing between right and left hand
motor imageries. They used EEG data from 2 dif-
ferent datasets, from a total of eleven subjects. The
EEG signal was acquired during 8 seconds trials, with
bipolar electrodes over C3 and C4 locations based on
international standard 10/20 system, and band-pass
filtered in the range of 0.5 - 30 Hz. The statistical
analysis was based on five-fold cross validation for
comparisons amongst classifiers. Classification ac-
curacy with Tukeys Honestly Significant Difference
criterion defined the best methods to extract spectral
features, and ANOVA was used to evaluate the vari-
ability among subjects. Using these criterions, the
power spectral density approaches demonstrated the
most consistent robustness and effectiveness in ex-
tracting the discriminant characteristics. With regard
to classification methods, the study has shown the su-
periority of Support Vector Machines with Gaussian
kernel. Furthermore, it was concluded by Herman
et al. (2008) that the combination of different EEG
datasets undertaken by subjects, with varying levels
of prior experience and in motor imagery, enables a
higher inter-subject variance.
Hema et al. (2010) investigated the effect of the
power spectrum feature of each of ten sub-bands of
10 Hz width using motor imagery signals of 10 sec-
onds (s) period recorded from C3 and C4 channels
and a neural classifier aiming at distinguishing among
four tasks. The data signal was segmented into 0.5
s segments with an overlap of 0.25 s, and each seg-
ment was filtered using a Chebyshev band pass filter
with a bandwidth of 10 Hz. The sum of the power
spectrum values was calculated and then a logarithmic
transform was applied. For each subject, ten neural
network models, for each of the ten sub bands of 10
Hz bandwidth, were developed. Classification rates
between 89.23% and 94.47% were achieved for sub-
band frequencies from 21 - 40 Hz.
Fitzgibbon et al. (2004) demonstrated that gamma
band power spectrum corresponding to complex men-
tal tasks had great increase relative to a Control condi-
tion, and the spatial distribution of such increase was
task-related. These studies suggested that higher fre-
quency bands (30 - 100 Hz) may contain useful in-
formation for classification between different mental
states and encouraged Liu et al. (2005) to also investi-
gate the effect of 10 Hz width sub-bands into the clas-
sification of different mental tasks. Differently, Liu
et al. (2005) used EEG signals from C3, C4, P3, P4,
O1 and O2. Seven subjects participated in the exper-
iment and five mental tasks were analyzed, including
no specific mental task, mental multiplication, mental
letter composing, geometric figure rotation and visual
counting. Again, a period of 10 s constituted a trial,
and the sum of weighted power spectrum was used
as feature extracted from ten 10 Hz wide sub-bands,
segmented into 1 s segments with an overlap of 0.9 s.
A Fisher’s Linear Discriminant based classifier was
used to distinguish between the mental tasks. In this
case, frequencies ranging from 30 to 100 Hz resulted
in greater classification accuracy, over 85%.
2 METHODOLOGY
Based on the previously discussed studies, and in
order to validate the proposed approach, EEG sig-
nals were systematically recorded at 1000 Hz using a
Bioamplifier plus PowerLab 16/30 configuration from
AdInstruments, according to the approved protocol
(COEP - USJT - No.088/2011). Three able body sub-
jects were requested to sit comfortably during the ex-
periment in which they had to perform 4 motor im-
ageries (right or left hand close, right or left arm flex).
EEG signals were acquired transversally from F, C, P
and O areas (from Fz, Cz, Pz and, Oz to F3, F4, C3,
C4, P3, P4, O1 and O2 respectively). The commands
were randomly given through the computer monitor
summing 45 repetitions for each movement imagina-
tion. After eliminating those corrupted with unusu-
ally excessive noise, or with simultaneous movement,
only 39 samples were used for each motor imagery
for each subject.
A segment of 2.5 s of each trial was selected and
the Power Spectral Density Averages in ten different
frequency band configurations were computed as fea-
tures. The common part among those arrangements
was: (A) Alpha (8 - 12 Hz), (B1) Beta1 (12 - 16 Hz),
(B2) Beta2 (16 - 20 Hz), (B3) Beta3 (20 - 28 Hz). The
investigation was based on the influence of consider-
ing or not the (D) Delta ( 0 - 4 Hz) and (T) Theta ( 4 -
8 Hz) bands, and five configurations of Gamma band
: (G1) 28 - 32 Hz; (G2) 28 - 64 Hz; (G3) 28 - 100 Hz;
(G4) Gamma1 (28 - 32 Hz), and Gamma2 (32 - 64
Hz); (G5) Gamma1 (28 - 32 Hz), Gamma2 (32 - 64
Hz), and Gamma3 (64 - 100 Hz). The ten treatments
are defined in Table 1. A spatial feature selection,
applying different electrode combinations and aims
to finding the most useful and discriminatory infor-
mation, was also carried out to improve classification
performance.
Fisher’s Linear Discriminant Analysis (Thomaz
and Gillies, 2005) was used as classifier, implement-
ing a few experiments aiming at discriminating be-
ADifferentStatisticalApproachAimingatEEGParameterInvestigationforBrainMachineInterfaceUse
245
Table 1: Band Configurations Treatments.
Treatment Considered
Bands
T1 A, B1, B2, B3, G1
T2 A, B1, B2, B3, G2
T3 A, B1, B2, B3, G3
T4 A, B1, B2, B3, G4
T5 A, B1, B2, B3, G5
T6 D, T, A, B1, B2, B3, G1
T7 D, T, A, B1, B2, B3, G2
T8 D, T, A, B1, B2, B3, G3
T9 D, T, A, B1, B2, B3, G4
T10 D, T, A, B1, B2, B3, G5
tween right and left hands, right and left arms, right
and left limbs, right arm and right hand, left arm
and left hand, and hands and arms (Castro et al.,
2013). Subject’s training and feedback that usually
contribute to increase the classification rates were ab-
sent in this work. The EEG were acquired just af-
ter volunteers received instructions and there was no
feedback and the classification process was done of-
fline.
To determine the importance and the significance
of the different frequency band configurations and
also the spatial feature selection in the classification
process, multiple inter-comparisons, two by two, of
the mean classification rates were made, for each pa-
rameter. This is a well established statistical proce-
dure, however, to the best of our knowledge, it has
not previously been applied in this context.
The result of a classification has a binomial distri-
bution (it is right or wrong). This distribution, with
large samples, according to the central limit theorem,
follows a normal distribution with
p : N( ˆp;
ˆp(1 ˆp)
n
) (1)
where p is the population classification rate under
study, ˆp is the sample classification rate and n is the
sample size.
The sample size is defined as the number of rep-
etitions (always 39), multiplied by the number of
treatments (10 for the frequency band configurations
and 15 for the spatial feature selection) multiplied
by the number of possibilities during classification (2
classes). This originates the sample sizes shown in
Table 2, confirming the previous assumption of large
samples.
For multiple comparisons, two by two, with a 5%
Table 2: Sample sizes.
Parameter 2 classes
Frequency Bands 1170
Spatial Feature Selection 780
level of significance, a correction was made by the
Bonferroni Method (Bland and Altman, 1995), as-
suming a very concervative approach. The new level
of significance is given by:
α
0
=
α
k
2
(2)
where k is the number of treatments. Therefore,
α
0
2
is the focus when performing a bilateral test.
For example, using k = 15 for spatial feature se-
lection results in
α
0
2
= 0.00023809 and Z
α
0
2
= 3.4938.
While, using using k = 10 as for frequency band con-
figurations,
α
0
2
= 0.0006 and Z
α
0
2
= 3.2389.
Based on the normal approximation, as mentioned
before, comparing between two classification rates
the Z score is given by:
Z
res
=
ˆp
1
ˆp
2
ˆ
σ
p
1
;p
2
(3)
where
ˆ
σ
p
1
;p
2
=
r
p
(1 p
)(
1
n
1
+
1
n
2
) (4)
and
p
=
n
1
ˆp
1
+ n
2
ˆp
2
n
1
+ n
2
(5)
Thus, if |Z
res
| > Z
α
0
2
there is a significant differ-
ence between both rates.
3 RESULTS
The best mean classification rate for each experiment
and each subject is presented in Figure 1. Inside each
bar, there is the indication of the best electrode com-
bination followed by the specific treatment resulted
from the best classification value. Sometimes more
than one combination is shown as in the case of Right
versus left Hand Classification for Subjects 2 and 3. It
is clear that the results are subject dependent and that
the training and feedback could contribute to reduce
classification rates close to aleatory values. Results
from subject 1 outperformed those obtained by the
others, probably due to some abilities to imagine. The
figure also shows a lack of a standard pattern related
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
246
Figure 1: Best Mean Classification Rates for each experiment based on the investigated parameter influence.
to the studied parameters. For each classification ex-
periment the best parameter combination was differ-
ent. However, there was a predominance in the use of
the F, P and O areas related to cognition and vision,
respectively, over the C area, related to movement.
This area appeared more in combination with others
and for subject 2 and 3. A great variation related to the
treatments can also be noticed, with a small advantage
in the treatments that included Delta and Theta bands
(from T6 to T10) over their absence.
Applying the statistical methods for subject 1
data, obtained from the experiment Right versus Left
Hands Classification Rates, the Z
res
scores for each
inter-comparison based on Frequency Bands Config-
uration and Spatial Feature Selection, can be seen
in Figures 2 (a) and (b), respectively. To facilitate
the analysis, the numbers were arranged in a decres-
cent order from the top left corner to the bottom right
one. The cells marked in light orange are those with
|Z
res
| > Z
α
0
2
, indicating a significant difference be-
tween both rates. The cells in white indicate that
there are no significant differences among them dur-
ing the comparison process. Thus, when analyzing
both figures, one can observe that, when aiming at dis-
criminating between Right and Left Hands, the results
based on different frequency band arrangements (2
(a)) had significant different classification rates only
if the treatment T10 was used in comparison with
the results obtained when treatment T1, T2 and T3
were used (see Table 1). On the other hand, the re-
sults based on the different electrode combinations (2
(b)) had significant different classification rates when
P and PO electrodes were used in respect to FC, F,
FCO, C, FCPO, O, CO, FO, and FPO electrode com-
binations. A subset was formed by the results when
CP was used, showing a significant difference from
those obtained by using FC, F, FCO, and C options.
Following the sequence, the use of FP resulted in clas-
sification rates different from those reached by using
FC, F, and FCO, while CPO results differentiated only
from those when FC was used.
In general, the feature frequency band, configured
into ten treatments did not show to make great dif-
ferences during classification processes. For subject
1, treatments T1 and T6, that differ by the absence
or not of the bands Delta and Theta, showed classifi-
cation rates significantly different when compared to
the other treatments, being the worst ones in exper-
iments Left Arm versus Left Hand and Arms versus
Hands. Regarding subject 2, for the experiment aim-
ing at discriminating between Right and Left Hands,
the treatment T6 achieved significantly different re-
sults when compared to T2, T3 and T4, while the
treatment T9 also produced results different from T3.
In other words, the highest classification rates (those
obtained with T2 and T3) were significantly differ-
ent from the lowest ones (obtained by using T6 and
T9). The intermediate cases did not present signifi-
cant differences. The same situation was verified for
subject 3 to distinguish between Right and Left Arm.
The treatment T1 obtained classification rates signif-
icantly higher than those obtained when used T2 and
T8.
On the other hand, in all experiments involv-
ing spatial feature selection, the use of a different
combination of electrodes always presented situations
where there were significant differences, as shown in
Figure 2 (b). Even for subjects 2 and 3, in experiments
where the classification rates were close to aleatory
values, as to differentiate between Arms and Hands
(Figure 3) or Right and Left Limbs (Figure 4), it was
possible to detect significant differences. The higher
differences were always between the electrode combi-
nation that produced the best mean classification rates
against those combinations that produced the worst
results. However, it can be noticed that intermediate
groups also showed significant differences.
4 DISCUSSION
This work applied a well established statistical
method over multiple inter-comparisons of the mean
classification rates based on different arrangements of
the frequency bands and spatial feature selection aim-
ing at investigating the significance of the variation of
ADifferentStatisticalApproachAimingatEEGParameterInvestigationforBrainMachineInterfaceUse
247
(a)
(b)
Figure 2: Z
res
scores for each inter-comparison using subject 1 data, for the experiment Right versus Left Hands, based on (a)
Frequency Bands Configurations (b) Spatial Features Selection.
each parameter for the classification process.
At first, following some statistical analysis fea-
tures, the assumption that both variables are indepen-
dent (treatments based on frequency band arrange-
ments and spatial feature selection based on different
electrode combinations) is a conservative approach,
due to the possibility of non definition of differences
for the same significant level. In other words, a higher
difference between the compared situations is neces-
sary when considering the hypothesis of dependency
between such variables.
The Bonferroni Method was used to correct the
level of significance for multiple inter-comparisons.
Again, it is a conservative approach because it is the
most restrictive treatment in respect to the statistically
significant differences.
The sample sizes used in this work were large
enough to make the approximation by the normal dis-
tribution useful and the continuity correction not nec-
essary.
The methodology adopted following this ap-
proach was able to indicate the significant differences
in classification results when various arrangements of
the same variable were applied, identifying and justi-
fying its importance and the need for further investi-
gation.
The results showed that the principal frequency
bands are the Alpha and Beta and sometimes the
lower part of the Gamma Bands, which are the same
ones usually used in other works. The use or not of
the other bands, and the way that they are used do
not produce significant classification rate differences.
In other words, they do not present useful information
related to motor imageries. This is in accordance with
Hema et al. (2010) that found the range from 21 - 40
Hz as the principal frequency band for motor imagery,
while the higher frequencies, as showed by Liu et al.
(2005) are related to complex cognitive activities.
On the other hand, different electrode combina-
tions based on the spatial features selection made sig-
nificant difference. This feature is related to the cere-
bral area and thus with the brain function. In al-
most all the performed experiments, the principal ar-
eas which contributed to the best results were those
related to cognitive processes, sensation, and vision
(F, P and O). The motor area (C) had a very restricted
contribution for motor imagery based on the protocol
used in this work. This area showed significance only
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
248
Figure 3: Z
res
scores for each inter-comparison using subject 2 data, for the experiment Arms versus Hands using different
electrode combinations.
Figure 4: Z
res
scores for each inter-comparison using subject 3 data, for the experiment Right versus Left Limbs using different
electrode combinations.
for subjects 2 and 3 to distinguish between Right and
Left Limbs. These results showed that the imagina-
tion may be more a cognitive and sensorial activity
than a motor one. However, these findings were in
disagreement with other authors who used only the
motor area saying that motor imagination cause sim-
ilar changes in the motor cortex as movement execu-
tion (Hema et al., 2010; Herman et al., 2008).
It is important to mention here that the direction
of the used signal was transversally acquired from the
scalp and the power spectral density was extracted
over the 2.5 s period entirely. Hema et al. (2010) and
Liu et al. (2005) used a segmented data, and both of
them, as well as Herman et al. (2008) used a total pe-
riod of time at least more than 3 times longer than
the 2.5 s used in this work. A longer period of time
increases the resolution of the feature and provides
more information. These differences may be respon-
sible for the inconsistencies and for the lower values
noticed in some classification rates. Thus, this point
needs further investigation.
The differences in the results of different subjects,
specially those close to aleatory values presented by
subjects 2 and 3 (Figure 1), indicate that more inves-
tigation is needed and that a training process, together
with some feedback to the subject, could contribute to
the improvement of classification results. This is due
to the possibility of active participation of the subject
to modulate the signal and adapt to the system crite-
rion. The results showed here were acquired without
this process. Nevertheless, the methodology proposed
was able to identify significant differences. The in-
ter subject variability is described by Herman et al.
(2008) and others as a regular behavior. Thus, an
adaptive system could be more appropriated for on-
line use.
ADifferentStatisticalApproachAimingatEEGParameterInvestigationforBrainMachineInterfaceUse
249
5 CONCLUSIONS
The numerical analysis, by itself, sometimes can
lead us to wrong conclusions, without any meaning.
A higher or lower classification rate might be im-
portant but requires further investigation about their
significance. This work applied a standard statisti-
cal approach, useful for situations of multiple inter-
comparisons of classification rates, but not previously
applied in the context of EEG feature analysis for
a BCI system. Six motor imaginaries were imple-
mented: Right versus Left Hands, Right versus Left
Arms, Right versus Left Limbs, Right Arm versus
Right Hand, Left Arm versus Left Hand, and Arms
versus Hands. A total of 8 EEG potential difference
was used, and the data was transversally acquired.
The method was able to highlight the cases where
there were significant differences between the com-
pared arrangements. The method also indicated that
the use of Delta, Theta and Gamma Bands (above 32
Hz) did not produce significant differences in the clas-
sification rates. On the other hand, the location of the
signal, represented by the several combinations of the
electrodes, achieved significant different results, and
the principal areas were F, P and O excluding the com-
monly used motor area (C). Thus, this variable needs
further investigation as part of a BCI system.
ACKNOWLEDGEMENTS
The authors would like to thank FEI and FAPESP for
the support and Prof. Esther L. Colombini and Dr.
Murilo F. Martins for reviewing this paper and mak-
ing enlightening suggestions.
REFERENCES
Al-Ani, T. and Trad, D. (2010). Signal Processing and Clas-
sification Approaches for Brain-computer Interface.
In Somerset, V. S., editor, Intelligent and Biosensors,
chapter 2. InTech.
Bland, J. M. and Altman, D. G. (1995). Multiple
significance tests: the bonferroni method. BMJ,
310(6973):170.
Castro, M. C. F., Galhianne, J. P. O. P., and Colombini,
E. L. (2013). EEG Motor Imagery Classification of
Upper Limb Movements. In International Conference
on Bioinspired Systems and Signal Processing., pages
314–317.
Fitzgibbon, S., Pope, K., Mackenzie, L., Clark, C.,
and Wiloughby, J. (2004). Cognitive tasks aug-
ment gamma EEG power. Clinical Neurophysiology,
115(8):1802–1809.
Hema, C. R., Paulraj, M. P., Yaacob, S., Adom, A. H., and
Nagarajan, R. (2010). An Analysis of the Effect of
EEG Frequency Bands on the Classification of Mo-
tor Imagery Signals. Biomedical Soft Computing and
Human Sciences, 16(1):121–126.
Herman, P., Prasad, G., McGinnity, T. M., and Coyle, D.
(2008). Comparative analysis of spectral approaches
to feature extraction for EEG-based motor imagery
classification. IEEE Transactions on Neural Systems
and Rehabilitation Engineering, 16(4):317–26.
Higashi, H., Tanaka, T., and Funase, A. (2009). Clas-
sification of single trial EEG during imagined hand
movement by rhythmic component extraction. In 31st
Annual International Conference of the IEEE EMBS,
pages 2482–5.
Liu, H., Wang, J., Zheng, C., and He, P. (2005). Study on
the effect of different frequency bands of EEG signals
on mental tasks classification. In Annual International
Conference of the IEEE Engineering in Medicine and
Biology Society., volume 5, pages 5369–5372.
Millan, J. D. R., Rupp, R., Muller-Putz, G. R., Murray-
Smith, R., Giugliemma, C., Tangermann, M., Vidau-
rre, C., Cincotti, F., Kubler, A., Leeb, R., Neuper, C.,
Muller, K. R., and Mattia, D. (2010). Combining
Brain-Computer Interfaces and Assistive Technolo-
gies: State-of-the-Art and Challenges. Frontiers in
Neuroscience, 4:161–193.
Thomaz, C. and Gillies, D. (2005). Approach for limited
sample size problems - with application to face recog-
nition. In XVIII Brazilian Symposium on Computer
Graphics and Image Processing (SIBGRAPI05), vol-
ume 12, pages 89–96.
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
250