TELKOM
NIKA
, Vol. 11, No. 5, May 2013, pp. 2381
~ 2386
ISSN: 2302-4
046
2381
Re
cei
v
ed
Jan
uary 10, 201
3
;
Revi
sed Ma
rch 5, 2
013;
Acce
pted Ma
rch 1
5
, 2013
Study on Fisher Analysis of Electroencephalo
graph
Data
Yuan Shi*, Qi Wei, Ruijie
Liu, Yuli Ge
Dali
an Institute
of Science a
n
d
T
e
chnolo
g
y
Dali
an Lvs
h
u
n
Econom
ic Dev
e
lo
pment Z
o
n
e
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 2008
80
41@
q
q
.com
A
b
st
r
a
ct
Objective
in
th
is p
aper, w
e
have
do
ne
F
i
s
her
discri
m
i
n
a
n
t an
alysis
to
Electro
ence
p
hal
ogra
m
(EEG) data of
experi
m
ent o
b
jects w
h
ich
a
r
e record
ed
i
m
perso
nal
ly, co
me
up w
i
th a
relativ
e
ly acc
u
rate
meth
od
use
d
i
n
featur
e extr
action
an
d cl
a
ssificati
o
n
d
e
ci
sions. T
h
e
pr
esent stu
d
y is
the gr
oun
dw
ork
ana
lysis for
other a
n
a
l
ysis i
n
EEG study. Me
thods
In acc
o
rdanc
e w
i
th th
e strengt
h of
wa
ve
, th
e
hea
d
electro
des are divid
ed into
fo
ur
spec
ies.
In use
of
p
a
rt of
21 e
l
ectro
des
EEG dat
a of 6
3
pe
opl
e, w
e
h
a
v
e
don
e F
i
sher
di
scrimina
n
t an
al
ysis to EEG data of six
o
b
je
cts. EEG data process
i
ng
an
d statistic an
al
ysi
s
ado
pted i
n
d
e
p
end
ently d
e
sig
ned EEG an
aly
s
is toolb
o
x a
n
d
the progr
a
m
o
f
correlatio
n
an
alysis. Res
u
lts I
n
use of part o
f
EEG data of 63 peo
ple,
w
e
have do
n
e
F
i
sher discr
imina
n
t ana
lys
i
s, the electro
d
e
classificati
on a
ccuracy rates i
s
82.
3%. Con
c
lusio
n
s F
i
she
r
discrimin
ant has hi
gher pr
e
d
ictio
n
accura
cy,
EEG features (
m
a
i
nly
w
a
ve) extract more
a
ccurate. F
i
sh
er discri
m
i
n
a
n
t w
oul
d b
e
b
e
tter
app
lie
d to th
e
feature extracti
on an
d class
i
fi
cation d
e
cisi
on
s of EEG data.
Ke
y
w
ords
:
EEG, Electroence
pha
logr
a
m
,· F
i
sher discri
m
in
a
n
t,
rhythm
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The ordi
na
ry EEG (electroen
cep
halo
-
g
r
aph
)
examin
ation, also
known as bra
i
nwave
discrimi
nant
analysi
s
, aim
s
to find
out
wheth
e
r
brai
n waves
are
normal, an
d
to assi
st in
the
diagn
osi
s
of
brain
le
sion
s.
Tra
d
itional
el
ectro
e
n
c
ep
ha
logra
m
di
scri
minant a
naly
s
is is cond
uct
e
d
throug
h inte
rpreting
multi-cha
nnel EEG
in the re
co
rding p
ape
r b
y
EEG experts, that is, visual
insp
ectio
n
is ado
pted to
und
erstand
and
ev
alua
te the EEG.
The
esse
n
c
e
of expe
rts’
experie
nce-b
a
se
d method
is to use
s
experts’ exp
e
ri
e
n
ce to re
mov
e
the sign
al interferen
ce a
nd
artifact, extra
c
t feature
s
from the EEG sign
als
a
c
cording to sig
nal
s’ frequ
en
cy, amplitude, a
n
d
pha
se info
rm
ation, and m
a
ke
categ
o
ry
descri
p
tion
s for extracte
d feature
s
by
using
acce
p
t
ed
experie
nce, thus
analy
s
is
and eval
uatio
n of EEG ca
n be
compl
e
ted [1]. So far, this method
is
still widely applied in cli
n
i
cs.
Visual methods can
capture pat
hol
ogical wavef
o
rm to a certain
extent, and
e
v
en identify b
r
ain l
e
si
on l
o
cation
s.
Ho
wever, EEG i
s
ch
ara
c
te
rize
d by bei
ng
n
o
n
-
stationa
ry an
d non
-line
a
r, and
the
visual meth
o
d
dep
end
s l
a
rgely o
n
th
e EEG analy
s
ts’
kno
w
le
dge a
nd experi
e
n
c
e [2]. These two a
s
pe
cts
require ne
w method
s to
make brea
kthrou
ghs
in EEG rese
arch. Fish
er
discrimi
nant
analysi
s
is a
dopted in EE
G re
sea
r
ch, whi
c
h will g
r
eatly
facilitate the
extraction
an
d cla
s
sificatio
n
of
EEG
si
gnal info
rmat
ion, thus it will help i
n
EEG
examination
and qu
antitative analysi
s
, providing a
n
effective analytical tool for E
E
G che
c
ks [3
].
2. Rese
arch
Metho
d
2.1. Subjects
The subje
c
ts
are 2
8
men
a
nd 35
wom
e
n
,
aged from
2
0
to 60 yea
r
s
old, with an
a
v
erage
age of 3
6
.7 years. All su
bj
ects
are
heal
thy, with
no reco
rd of
se
ri
ous
nervo
us
system di
se
a
s
e
and p
s
ychiatric d
r
ug
use,
whi
c
h m
ean
s these
pe
opl
e are individ
uals
ch
osen
from the
normal
popul
ation.
2.2. Establis
hing Mathe
m
atical
Models for EEG Data Selecti
on
The sam
p
ling
frequen
cy for lab re
cords
of EEG
is 100Hz. Acco
rdi
ng to the internation
a
l
10-2
0
sy
stem
, data of 21
e
l
ectro
d
e
s
a
r
e
recorded:
C3,
CZ, C4, FP1
,
FPZ, FP2, F7, F8, FZ, F3,
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2381 – 238
6
2382
F4, O1, OZ, O2, P3, PZ,
P4, T5, T6, T3, T4. One bl
ock (rep
re
se
nting a sm
all perio
d of time) of
EEG data is retrieve
d at a time, with 512 sam
p
ling
points for
ea
ch blo
c
k and
a record tim
e
of
5.12 seconds. The EEG of norma
l people will mainly show
α
rhythm, with
α
wa
ve appea
ring
in
the ba
ck of
the he
ad
an
d wea
k
eni
ng
gradually f
r
om b
a
ck to
front. Acco
rd
ing to
differe
nt
intensities of
α
wave in va
rious
part
s
of
head, the
21
con
d
u
c
tive el
ectro
d
e
s
can
be divide
d int
o
four
categ
o
ri
es, na
mely the fro
n
t hea
d ele
c
tr
o
d
e
s
, the si
de h
e
ad ele
c
trode
s, the
cent
ral
are
a
electrode
s, o
cci
pital ele
c
trode
s. The sp
ecific
cla
ssifi
cation is a
s
follows:
(1) T
he first category: the central a
r
ea el
ectro
d
e
s
(C3,
CZ, C4).
(2) T
he seco
nd cate
gory: the front hea
d
electro
de (F
Z, F3, F4, FP1, FPZ, FP2,
F7, F8).
(3) T
he third
categ
o
ry: the occipital ele
c
t
r
ode
(P3, PZ, P4, O1, OZ,
O2, T5, T6).
(4) T
he fourth
catego
ry: the side he
ad el
ectro
de (T3, T4).
2.3. Computer Proces
sing of the EEG
Data
In order to facilitate the analysis of t
he raw data of the EEG, MATLAB program
m
ing is
use
d
to de
si
gn a de
dicated EEG Tool
box. In EEG Toolbox
wh
en ra
w d
a
ta is a
c
cessed
and
saved in m
a
trix, the horizontal line
s
re
pre
s
ent
the ti
me point
s of lab re
co
rd
s (namely sa
mp
ling
points) a
n
d
the ve
rtical
co
lumns
stand
for
elect
r
od
es.
Before
an
alysis,
all d
a
ta o
f
every
subj
e
c
t
sho
u
ld be a
c
ce
ssed in EEG Toolbox an
d there sh
oul
d be visual EEGs, with on
e page in
clud
ing
a block of EEG data.
In acco
rd
an
ce with
the
e
l
ectro
d
e
cla
s
sifi
catio
n
m
e
thod
de
scribe
d ab
ove, 4
overall
Fishe
r
di
scri
minant a
naly
s
e
s
divide
sample
data i
n
to four cate
gorie
s. Fi
rst,
acco
rdin
g to
the
cla
ssifi
cation
we will put the
21 ele
c
trode
EEG
dat
a whi
c
h
ha
s mathem
atical mod
e
ls into 4
matrixes
(
1
,
2
...4)
i
Xi
. Then we put E
E
G data of the cu
rrent
block into the matrix X, whose
measurement
is
512 × 21. Now
we
will use Fish
er di
scrimi
nation analysi
s
to predi
ct the
electrode
cla
ssifi
cation results, and the result
s sh
ould
be displayed i
n
the form of the vector.
Fishe
r
di
scri
mination anal
ysis will
b
u
ild a
linea
r
cla
ssi
fier surfa
c
e i
n
the featu
r
e
space to
proje
c
t the sp
ace
sampl
e
p
o
ints alo
ng th
is dire
ct
ion, a
nd identify ca
tegorie
s a
c
co
rding to value
s
of these
proj
ection
s i
n
thi
s
di
re
ction.
Fi
she
r
di
scri
mi
nation analy
s
is pro
c
e
dures in this rese
a
r
ch
prog
ram
are
base
d
on t
he multi-cha
nnel EEG d
a
ta. First we
must e
s
tabl
ish math
ema
t
ical
model
s, nam
ely the discri
mination fun
c
tions, then to
predi
ct EEG
data catego
ries a
c
cording
to
the discrimin
ation rul
e
s.
Fishe
r
di
scri
minati
on a
n
a
l
ysis
can
be
explaine
d b
y
the followi
ng
mathemati
c
al
formula:
yu
X
(1)
In the formula,
u
is the feature vector
corre
s
po
ndin
g
to the maximum
c
h
arac
teris
t
ic root
of each
1
EA
.Multi-dime
n
si
onal
discrimin
a
tio
n
need to
u
s
e
many feature
vectors to form
multiple discri
m
inant functi
ons.
We put uncl
a
ssified
EEG data X into four Fishe
r
discrimi
nant
function
s, an
d get the minimum dista
n
ce and ma
tch it with the corresp
ondi
ng whole. We
can
use Fi
she
r
di
scrimin
ant function
s to pre
d
ict cl
a
s
sifica
tions for EEG
data of each
block, and
sh
ow
the predi
cted
cla
ssifi
cation
results visu
all
y
combin
e
d
with the actual
cla
ssifi
cation
in Fishe
r
discrimi
nant
analysi
s
predi
ction map.
2.4. The algo
rithm of Fish
er Discrimin
ant Analy
s
is
The algo
rithm
[
4]:
Input: EEG d
a
ta of the current blo
ck.
Output: Pred
icted cl
assification re
sults
and
a
c
cura
cy
of Fisher di
scrimi
nant an
a
l
yses.
Step 1 : Calculate mean v
e
ctors.
()
()
1
1
,(
1
,
2
)
j
n
jj
ji
i
j
X
Xj
l
n
(2)
Step 2 : Calculate overall
mean vecto
r
.
()
1
1
,(
1
,
2
,
,
)
l
j
k
k
X
nX
j
l
n
(3)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study on Fi
sh
er Anal
ysi
s of
Electroe
ncep
halog
rap
h
Da
ta (Yuan Shi)
2383
Step 3 : Use
estimate valu
e of sample t
o
cal
c
ulate th
e deviation of
group
s, as m
ean
vectoers o
r
o
v
erall mea
n
vector a
r
e u
n
known.
(1
)
(
1
)
(
)
(
)
1
()
()
()
(
)
ll
ll
E
n
X
X
X
X
n
XX
XX
(4)
Step 4 : Calculate deviatio
n
of group
s.
1
(
1
)
(
1
)
(
1
)
(
1
)
()
()
(
)
(
)
11
()
()
()
()
l
n
n
ll
ll
ii
i
i
ii
A
X
XX
X
X
XX
X
(5)
Step 5 : Calculate
u
is the feature ve
ctor
corre
s
p
ondin
g
to the maximum ch
ara
c
t
e
risti
c
root of ea
ch
1
EA
.
Step 6 : Build Fishe
r
discri
minantfun
c
tio
n
s
ll
y
uX
.
Step 7 : Put u
n
cla
s
sified EEG data X into four Fisher
discrimi
nantf
unctio
n
s, an
d
get the
minimum di
stance and mat
c
h it with the corre
s
p
ondin
g
whol
e.
()
(
)
1
()
m
i
n
(
)
ij
jk
yX
y
y
X
y
(6)
S
t
ep 8 :
P
r
edict
cla
ssif
i
cat
i
on re
sult
s a
n
d
cal
c
ulat
e t
h
e cla
ssif
i
cat
i
o
n
ac
cur
a
cy
.
2.5. The Ana
l
y
s
is Results
of Fisher Di
scriminant
Analy
s
is
Fis
h
er
dis
c
r
i
mination analys
is
procedure c
an pr
edic
t
c
l
as
s
i
fi
c
a
tions
of all blocks of EE
G
data bel
ongi
n
g
to differe
nt subj
ect
s
. Because spa
c
e
i
s
limited, we will
give cla
s
sificatio
n
results
of a block of 21 ele
c
trod
es for one su
bj
ect in det
ail, and fore
ca
st cla
ssifi
cation
results in oth
e
r
blocks are si
milar h
e
re. F
i
she
r
di
scrimi
nant
an
alysis will be
con
d
u
cted i
n
the
twenty-secon
d
block of EEG data for Subj
ect 1.
Note 1): f
o
recast
results in the upper left corner
are
for firs
t-catego
r
y
electrodes, results in the upper left
corner a
r
e for
second-categor
y
electrodes, results in the low
e
r left corner
a
r
e for t
h
ird-catego
r
y
electrodes, and r
e
sults in the low
e
r
right corner
are f
o
r fourt
h
-catego
r
y
el
ectrodes. 2)
:
x-coor
dinate is
for electrode name
s
of all ty
pes, an
d
y
-
coo
r
dinate is
for categories. 3)
: red * indicates forecast classif
i
ca
tion,
and blue O
r
epresents the act
ual
classificat
i
on. When forecast
class
i
fication and
actual class
i
ficat
i
on
ar
e consistent, * w
ill coincide
w
i
th
O
.
Figur
e 1.
Fish
er f
o
re
ca
st
re
sult
s f
o
r the twenty-se
con
d
block Subje
c
t 1
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2381 – 238
6
2384
Figure 1 sh
o
w
s Fi
she
r
predicte
d
re
sult
s fo
r the twe
n
ty-se
c
on
d bl
ock of EEG data of
Subject
1. In
the current
bl
ock, we h
a
ve
accu
rate
an
alysis for ele
c
trod
es in
the
first
cate
gory
in
the ce
ntral a
r
ea and th
e third
categ
o
ry
in occipital e
l
ectro
d
e
s
, but
se
con
d
-cate
gory ele
c
trod
es
FZ, F3, F
4
in
the front h
e
ad i
s
cla
ssifie
d
into th
e first categ
o
ry, a
nd fou
r
th-cat
egory
ele
c
tro
d
e
s
T3 and T4 in
the side h
ead
is cla
ssifie
d
into the se
con
d
categ
o
ry. The EEG of normal pe
ople
will
sho
w
α
rhythm,
with wave
α
wea
k
e
n
ing
in a
ro
w fro
m
ba
ck of
th
e he
ad, the
central
area, h
ead
side, fro
n
t head. In pre
d
i
c
ted results,
FZ, F3,
and F4 are cla
ssifie
d
into the ce
ntral a
r
ea,
indicating wa
ve
α
of FZ, F
3
and
F4 h
a
ve strong
er el
ectro
d
e
s
than
norm
a
l fro
n
t head; T
3
an
d
T4
are
cla
s
sified
into the front
head, i
ndi
cat
i
ng wave
α
of
T3 a
nd T
4
h
a
ve we
aker
e
l
ectro
de
wave
than no
rmal
temporal ele
c
trod
e wave. The esti
mat
ed sp
ect
r
al
den
sity of EEG data for
the
curre
n
t blo
ck
(sh
o
wn in Fig
u
re 2
)
is
retri
e
ved by
usi
n
g psd functio
n
. By doing this, we found
that
wav
e
α
of Subje
c
t 1 for t
he current
bl
ock conforms with that of
electrode
s in
rea
r
hea
d, a
nd
tends to
wea
k
en g
r
a
dually
along the di
rection of fro
n
t head. Wave
α
of front hea
d elect
r
ode
s
are
stron
g
e
r
, and
are not
wea
k
ened
com
pared with el
ectr
ode
s in the central a
r
ea. A
s
a re
sult, the
s
e
electrode
s a
r
e wrongly
cla
ssifie
d
into th
e first catego
ry. Wave
α
of temporal ele
c
trod
es
with t
he
same intensit
y of wave
α
i
n
fro
n
t he
ad
electrode
s, a
nd thi
s
i
s
wh
y T3 a
nd
T4
are
wron
gly
put
into the
se
cond
cate
gory
.
The int
ensi
t
ies of
wave
α
i
n
el
ectro
des sho
w
ed
spe
c
tral
de
n
s
ity
estimate
s is
basi
c
ally in a
g
ree
m
ent wit
h
those
in predicte
d
re
sult
s. Fish
er di
scrimina
n
t anal
ysis
will build a linear cl
assification su
rface in the feature
space to project the space
sampl
e
point
s
along thi
s
di
rection, a
nd
combinin
g dat
a to be
anal
y
z
ed i
n
the
wh
ole with l
o
we
st proje
c
tion
value
along thi
s
direction.
We
wi
ll take F
3
a
s
an exam
pl
e t
o
illust
rate th
e process of t
he di
scrimin
a
t
ion
analysi
s
. We
put F3 into fo
ur Fishe
r
discrimi
nati
on fu
nction
s. Thu
s
the distan
ce
from proj
ect
ed
F3 to fou
r
whole
s
is
2.50
61e
+01
7
, 4.
4660
e
+018
, 5.7533 th
e
e +018, 3 .
8302
e
+ 01
8
respe
c
tively.
In these valu
es, the dista
n
ce fro
m
proj
ected F3 to the first wh
ole
is the smalle
st,
thus it’s
put into the first
category. Th
ro
ugh p
r
oj
e
c
tio
n
Fish
er di
scriminatio
n an
alysis
ca
n m
o
re
accurately di
stingui
sh
su
p
e
rio
r
ity levels of wave
α
with
better p
r
edi
cted re
su
lts.
The
ri
g
h
t
predi
cted
nu
mber for th
e
curre
n
t blo
c
k is 1
6
, an
d t
he
wro
ng
nu
mber is
5, wi
th 76.1% a
s
the
acc
u
rac
y
rate.
Note 1): psd f
unction is used to obtai
n the po
w
e
r spectral densit
y
estimates
in the 22nd block for Subject 1. The x-
coordinate is for t
he freque
nc
y
,
an
d the ordinate
-
co
ordinat
e is for th
e po
w
e
r
estimate. 2): the f
r
equenc
y of
w
a
ve
α
is 8
~
13Hz; the large
r
t
he ratio of the sq
uare of
wave
α
t
o
the overall are
a
is, the stronger
α
wave is. We can see intensities of
α
wave in electrodes.
Figure 2 Power sp
ect
r
al de
nsity of the 22nd blo
c
k for Subject 1
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Study on Fi
sh
er Anal
ysi
s of
Electroe
ncep
halog
rap
h
Da
ta (Yuan Shi)
2385
Fishe
r
discri
m
ination
anal
ysis i
s
cond
u
c
ted fo
r
63
subje
c
ts. Pred
icted
cla
s
sifications of
10 bl
ocks of
EEG data
are
ran
domly
sel
e
cted
from
ea
c
h
su
b
j
ec
t,
w
i
th
an
a
v
er
ag
e
acc
u
ra
c
y
ra
te
of 82.3% (sh
o
wn i
n
Ta
ble
1). Compa
r
e
d
with th
e Ma
halan
obis di
stance
discrim
ination a
naly
s
is,
Fishe
r
discri
mination
ana
lysis
has hig
her
pre
d
ictio
n
a
c
cura
cy rate, and
more a
c
curate
EE
G
feature (mai
n
l
y
wave
α
) e
x
traction. P
r
e
d
icted
re
sult
s ca
n reflect i
n
tensitie
s
of
wave
α
in
e
a
c
h
electrode,
bu
t these
can
not re
move
amplitude
m
odulatio
n, ap
proximatio
n
and
parti
cipa
nts'
individual
differen
c
e
s
as
well a
s
the
em
erge
nce of
ot
her ban
d
wa
ves’ im
pact
o
n
cl
assificatio
n
res
u
lts.
Table 1. Average a
c
cura
cy
rate of using
Fi
she
r
di
scri
minant analy
s
is to p
r
edi
ct EEG
cla
ssif
i
cat
i
on
s f
o
r 63
subje
c
t
s
Accuracy
rate
70%-7
5%
75%-8
0%
80%-8
5%
85%-9
0%
Average
accuracy
rate
Subject(s) 4
25
33
1
82.3
％
3. Results a
nd Analy
s
is
Acco
rdi
ng to
different int
ensitie
s of
α
wave in
variou
s p
a
rts, the 21
co
ndu
ctive
electrode
s
can be
divide
d into fou
r
categori
e
s.
M
a
thematical
model
s for
F
i
she
r
di
scrimi
nant
analysi
s
ca
n
be esta
blish
ed
by usi
ng data
of 21
e
l
ectro
d
e
s
. Fi
she
r
di
scrimi
nant a
nalysi
s
is
con
d
u
c
ted fo
r 6
3
n
o
rm
al
su
bje
c
ts. F
o
re
ca
st
cla
s
sification
s of
10 bl
ocks
of EEG d
a
ta
are
rand
omly sel
e
cted f
r
om
e
a
ch
su
bje
c
t, with an
average a
c
cu
ra
cy rate
of 82.3
%
. It has hig
her
predi
ction a
c
curacy rate, and mo
re a
c
curate EEG feature (mai
nly wave
α
) extraction,
a
nd
make
s bette
r distinctio
ns i
n
intensitie
s
of wave
α
in
electrode
s. T
he tests
sho
w
that the Fishe
r
discrimi
nant
analysi
s
m
e
thod
can
extract EEG fe
at
ure
s
of
norm
a
l peo
ple i
n
a mo
re favo
rable
manne
r, and
can b
e
appli
e
d in EEG data cla
ssifi
catio
n
s[5],
[6].
Fishe
r
di
scri
minant an
alysis
ca
n not remove amplit
ude mo
dulati
on, app
roxim
a
tion and
partici
pant
s' i
ndividual diff
eren
ce
s a
s
well a
s
the e
m
erg
e
n
c
e of other b
and
waves’ impa
ct
o
n
cla
ssif
i
cat
i
on
res
u
lt
s,
w
h
ich lea
d
s t
o
wro
ng
cla
ssif
i
cat
i
on
s.
B
y
usin
g t
heo
rie
s
of
mult
iv
ari
a
t
e
statistics, we
find out rea
s
o
n
s for ele
c
tro
de misju
dgm
ent as follo
ws:
(1) Fisher
di
scriminant
analysis
will bui
l
d a linear cl
assi
fication surface in the feature
spa
c
e to p
r
oj
ect the spa
c
e
sampl
e
point
s alon
g this d
i
rectio
n, and i
dentify categ
o
rie
s
a
c
cordi
n
g
to values of these p
r
oje
c
tions i
n
this direction. Si
n
c
e
it is impo
ssibl
e
to get a
n
in
finite numb
e
r
of
sampl
e
data,
the
overall
p
r
oje
c
tion
are
a
s
ca
n o
n
ly
estimated
by
usin
g limited
EEG data
wit
h
in
the cent
ralize
d
training
set. This may affect the prediction accuracy
.
(2)
Misj
udgm
ent exists in
Fishe
r
di
scri
minati
on
anal
ysis
whe
n
it’s use
d
to a
nal
yze EEG
data fo
r n
o
rm
al pe
ople. In
fact, Fisher d
i
scrimin
ant re
duces dime
n
s
ion
s
, n
a
mel
y
it proj
ect
s
t
h
e
data of the 512-dime
nsio
nal EEG vector onto a
st
raight line, a
nd then cl
assifies p
a
ttern
s
according to t
he obtain
ed o
ne-di
men
s
ion
a
l cha
r
a
c
te
ri
stic (al
s
o calle
d scala
r). It views the Fish
er
discrimi
nant
analysi
s
as a
feature extraction
alg
o
rit
h
m, we rese
arch th
e
way
of obtai
ning
the
best p
r
oje
c
ti
on directio
n
from the p
e
rspe
ctive
of feature
extra
c
tion so th
at
one-dime
nsio
nal
proje
c
tion
ch
ara
c
teri
stic can dist
ing
u
ish four types of electro
d
e
s
in most favorably manne
r[7-8].
If four types o
f
electro
d
e
s
a
r
e ove
r
lap
p
in
g wh
en p
r
oje
c
ted, an
d dat
a to be a
nalyzed i
s
ju
st in t
he
overlap
p
ing
a
r
ea, it i
s
p
r
o
b
ably to mi
scl
a
ssify the
sam
p
le. We ta
ke t
he first-categ
o
ry an
d
se
co
nd
category electrodes
(Figure 3)
as examples to illustrate that
one-di
mensional characteri
sti
c
obtaine
d fro
m
variou
s
projecte
d di
re
ctions, m
a
y a
ffe
c
t
gr
e
a
t
ly in c
l
as
s
i
fied
per
fo
r
m
an
ce
s
.
F
o
r
instan
ce, first
-
cate
go
ry and
second
cate
gory ele
c
tro
d
e
s have
som
e
overlap
p
ing
area in spa
c
e. If
it is p
r
oj
ecte
d to the
x-a
x
is directio
n, Fis
her di
scrimina
nt an
al
ysis
ca
n e
a
s
ily di
stingui
sh
electrode
s p
r
ojecte
d in the regio
n
of R4 an
d R6
. Ho
wever, it can
not accu
rately distingu
ish
electrode
s p
r
oje
c
ted
i
n
the
re
gion
of
R5
where many
ele
c
trod
es ove
r
lap, an
d th
us
miscl
assification occu
rred.
We hop
e to find a lin
e
L, so that proje
c
ted el
e
c
trod
es
can
be
sep
a
rate
d as
far as p
o
ssibl
e
. As it can b
e
see
n
in
Fig
u
re 4.9, the e
ffect is better
to proje
c
t alo
ng
the L dire
ctio
n than the x dire
ction. Although th
e
r
e
wi
ll still be som
e
overla
pped
electrode
s, the
electrode
s i
n
the ove
r
lapp
e
d
a
r
ea
R2
h
a
v
e bee
n
signi
ficantly redu
ced. In fa
ct, fo
ur
cato
geri
e
s
o
f
electrode
s
p
a
rtially overl
ap in the
512-dime
nsio
nal spa
c
e,
and
can
not
be a
c
curately
disting
u
ished
after they are
proje
c
ted, an
d then it may lead to miscla
ssifi
cation.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2381 – 238
6
2386
Figure 3. Fish
er proje
c
tion
maps
of the
collectivity of both
(3) T
here are individual
differen
c
e
s
in
EEG
[9]. Different
subj
ect
s
have different EEG
amplitude
s, frequ
en
cie
s
, and waveforms; ea
ch bl
o
ck of EEG signal
s differs as well. Fisher
discrimi
nant
analysi
s
met
hod can not
eliminate i
ndi
vidual differe
nce
s
of the n
o
rmal EEG d
a
ta
whe
n
it is used to analyze each blo
c
k of EEG data,
all of which result in miscl
assificatio
n
in
electrode
s ca
tegorie
s [10].
4. Conclusio
n
Fishe
r
di
scri
minant an
al
ysis i
s
con
ducte
d for
63 no
rmal
subj
ect
s
. Fo
recast
cla
ssifi
cation
s of 10 bl
ocks
of EEG data are
ra
n
d
o
mly sele
cte
d
from ea
ch
subje
c
t, wit
h
an
averag
e accu
racy rate of 82.3%. It has highe
r pr
edict
ion accu
ra
cy rate, and mo
re accurate EEG
feature (main
l
y
wave
α
) e
x
traction, a
n
d
ma
ke
s bett
e
r di
stinctio
n
s
in inte
nsiti
e
s of
wave
α
in
electrode
s. T
he test
s sho
w
that the F
i
she
r
di
scrimi
nant an
alysi
s
method
ca
n
extract EE
G
feature
s
of
n
o
rmal
peo
ple
in a m
o
re f
a
vorabl
e ma
nner,
and
ca
n be
appli
e
d
in EEG d
a
ta
cla
ssif
i
cat
i
on
s.
Referen
ces
[1]
Gabor AJ, L
e
a
c
h RR, D
o
w
l
a
F
U
. Automated
seizur
e det
ecti
on us
ing
a se
lf-orga
n
izi
ng
ne
ural
net
w
o
rk.
Electroe
nce
p
h
a
lo
grap
hy an
d Clin
ical N
eur
op
hysiol
ogy
. 1
9
9
5
; 99(3): 25
7-2
66.
[2]
S Blanco, et al. Appl
yi
ng ti
me-fr
equ
enc
y
ana
l
y
sis to seizure EEG activit
y
.
IEEE Engineering in
me
dici
ne a
nd b
i
olo
g
y
mag
a
z
i
n
e
. 1997; 1
6
: 65
-71.
[3]
Williams WJ. Time-Frequ
enc
y Anal
ys
is of Bi
olo
g
ica
l
Sig
n
a
l
s.
IEEE Electrical Computer
Scienc
e
. 19
93;
12(1): 83-
86.
[4]
Sebasti
an M, Gunnar R, Jas
on M, et al.
F
i
sher discrim
in
ant ana
l
y
sis
w
i
th kernels.
Pr
ocee
din
g
s of
IEEE International Work s
h
op on
Neur
al
Net
w
orks for Signal Proc
essing
.
Madis
on, W
i
sc
onsi
n
. 19
99;
41-4
8
.
[5]
John T
r
inder,
John A va
n B
e
vere
n, Phil
ip
Smith,
et al. Correl
a
tio
n
bet
w
e
e
n
venti
l
ati
on a
nd EEG
arous
al dur
in
g slee
p ons
et in yo
un
g sub
j
ects
.
Journal of Ap
plie
d Physi
ol
og
y
. 2001; 83: 20
05-2
011.
[6]
P Comon. Ind
e
pen
de
nt comp
one
nt ana
l
y
sis
—
a ne
w
co
nce
p
t.
Signal Proc
essin
g
. 200
0; 36: 287-3
14.
[7]
W
Z
hao, R
Ch
ella
pp
a, a
nd A
Krish
nas
w
a
m
y
. D
i
scirmi
nant
an
al
ysis
of pr
i
n
cipl
e c
o
mpo
n
ents for fac
e
recog
n
itio
n.
Pr
oc. of Inter. C
onf. on A
u
to
matic F
a
ce a
nd
Gesture Rec
o
gniti
on
. Nar
a
,
Japa
n. 19
98;
336-
341.
[8]
Hen
d
ra Kusu
ma, NFN Wiraw
a
n, Adi So
e
p
rija
nto.
Gabo
rbase
d
F
a
ce Reco
gniti
on W
i
th Illumin
a
tio
n
Variati
on Usi
n
g Subs
pace
Lin
ear Discr
im
ina
n
t.
T
E
LKOMNIKA Indon
esia
n Jour
nal
of Electrica
l
Engi
neer
in
g
. 2012; 10(
1): 119
-128.
[9]
Z
hang S
h
a
o
-b
ai, Hu
ang
Da
n-da
n. Electro
ence
p
h
a
lo
grap
h
y
F
e
ature E
x
tractio
n
Usi
n
g Hig
h T
i
me
F
r
eque
nc
y
Re
soluti
on
Ana
l
ysis.
T
E
LKOMNIKA Indonesi
a
n Journ
a
l of El
ectrical En
gin
e
e
rin
g
. 2012;
10(6): 14
15-
14
21.
[10]
Kasper K, Sch
u
ster HG. Easil
y
calc
ul
abl
e m
easur
e for comple
xit
y
of spati
a
l tempor
al p
a
ttern.
Physic
a
l
Review
Onli
ne
Archive
. 20
10;
36 (2): 842-
84
8.
Evaluation Warning : The document was created with Spire.PDF for Python.