Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
9
, No
.
2
,
Febr
ua
ry
201
8
,
pp.
403
~
409
IS
S
N:
25
02
-
4752
, DO
I: 10
.11
591/
ijeecs
.
v9.i
2
.
pp
403
-
409
403
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Perform
ance of S
up
port Vect
or M
achin
e in
Class
ifying E
EG
Signal
of Dysle
xic Child
ren usi
ng RBF
Kernel
AZ
A.
Z
ainuddi
n
1
, W.
Mans
or
2
,
Kh
uan Y
. Lee
3
, Z
. Ma
hmoodin
4
1
,2
Facul
t
y
of
El
e
ct
ri
ca
l
Eng
ine
er
i
ng,
Univer
si
ti T
e
knologi
MA
RA, 40450 Shah
Ala
m
,
Sela
ngor
,
Ma
lay
s
ia
1,2,3,4
Com
puta
ti
o
nal
In
te
l
li
gen
ce
Dete
c
ti
on
RIG
,
Pharm
ac
eut
i
ca
l
Li
fe
Scie
n
ce
s C
ORE,
Unive
rsiti
Te
knologi MA
RA,
40450
Shah
Ala
m
,
Sela
ngor
,
Ma
lay
s
ia
2,
3
Medic
a
l
Eng
in
ee
ring
T
ec
hnolo
g
y
Section,
Univ
ersit
i
Kuala L
um
pur,
53100
Gom
bak,
Se
la
ngor
,
Malay
s
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
1
9
, 201
7
Re
vised
Dec
2
2
, 2
01
7
Accepte
d
J
an
14
, 2
01
8
D
y
slexia
is
ref
err
ed
as
le
arn
i
ng
disabi
li
t
y
t
hat
ca
uses
lear
ner
havi
ng
diffi
cu
lt
i
es
in
dec
oding,
re
adi
ng
and
writi
ng
words
.
Thi
s
disabi
lit
y
associ
at
es
with
learni
ng
pr
oce
ss
ing
reg
ion
in
the
hum
an
br
ai
n.
Ac
ti
vi
ties
in
thi
s
reg
io
n
ca
n
be
exa
m
in
e
d
using
el
e
ct
ro
e
nce
pha
logra
m
(
EE
G)
which
r
ecord
el
e
ct
r
ic
a
l
ac
t
ivi
t
y
dur
ing
le
arn
ing
proc
es
s.
Thi
s
stud
y
l
ooks
int
o
per
fo
rm
anc
e
of
Support
Vec
tor
Mac
hine
(SV
M)
using
RBF
ker
nel
in
c
la
ss
if
y
ing
EE
G
signal
of
Norm
al
,
Poor
and
Cap
abl
e
D
y
slexic
ch
il
dr
e
n
during
writi
n
g
words
and
non
-
words
.
Discre
te
W
avele
t
Tr
ansform
(DW
T)
with
Daube
ch
i
es
orde
r
2
was
emplo
y
ed
t
o
ext
ra
ct
the
po
wer
of
beta
and
the
t
a
wav
es
of
EE
G
signal.
Bet
a
and
Th
eta/
Bet
a
r
atio
form
the
inpu
t
fe
at
ur
e
s
for
cl
assifi
er.
Multi
c
la
ss
one
ver
sus
one
SV
M
was
used
in
th
e
cl
assif
ic
a
ti
on
wher
e
RBF
ker
ne
l
par
amete
rs
and
box
constra
in
t
v
al
ues
were
v
arie
d
with
the
f
ac
to
r
of
10
to
ana
l
y
z
e
per
form
anc
e
of
the
class
ifi
er
.
It
was
found
tha
t
the
best
per
form
anc
e
of
SV
M
with
91%
over
all
a
cc
ur
a
c
y
was
obtained
when
both
ker
n
e
l
sca
l
e
and
box
constraint
ar
e
set
to
one
.
Ke
yw
or
d
s
:
Dysle
xia
Ele
ct
ro
e
nc
ep
ha
logram
RB
F k
e
rn
el
Suppor
t
V
ect
or Mac
hin
e
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
AZA. Zai
nudd
in
,
Faculty
of Elec
tric
al
Engineer
ing
,
Un
i
ver
sit
i Te
knol
og
i M
ARA
40450 S
hah A
l
a
m
, S
el
ango
r,
Ma
la
ysi
a
Em
a
il
:
zub
er@
un
i
kl.edu.m
y
1.
INTROD
U
CTION
Dysle
xia
is
ne
uro
bio
lo
gical
i
neffici
ency
of
so
m
e
par
t
in
th
e
brai
n
that
m
akes
t
he
pe
op
le
exp
e
rience
diff
ic
ulty
in
ac
qu
i
rin
g
fl
uen
t
sk
il
ls
in
rea
ding
al
th
ough
the
y
hav
e
recei
ve
d
a
pprop
riat
e
academ
ic
edu
ca
ti
on
at
the
sam
e
le
vel
as
norm
al
child
re
n
[
1]
.
Desp
i
te
this
le
arn
in
g
disabili
ty
,
dysle
xic
childre
n
po
s
sess
the
s
a
m
e
or
above
I
Q
le
vel
com
par
ed wit
h n
or
m
al
ch
il
dr
e
n
[2]
.
Seve
ral
stu
dies
ha
ve
bee
n
c
onduct
e
d
to
ide
ntify
co
gnit
ive
stre
ng
t
hs
a
nd
weaknesse
s
of
the
c
hildr
e
n
us
in
g
com
pu
te
r
m
od
el
analysis
fr
om
Gibs
on
te
st
[3]
.
Wh
il
e
Ma
la
ysi
a
Mi
nistry
of
Edu
cat
io
n
use
s
the
Dysle
xia
chec
k
li
st
as
the
instr
um
ent
to
i
den
ti
fy
the
prob
a
bili
ty
of
th
e
childre
n
ha
vi
ng
le
ar
ning
di
sabili
ty
sp
eci
fic to
d
ysl
exia b
y m
easuri
ng
t
heir
ca
p
a
bi
li
t
y i
n
sp
el
li
ng
,
rea
ding,
an
d wr
it
in
g.
Be
side
visu
al
,
aud
it
ory
,
proce
ssing
a
nd
w
ord
te
st
to
e
xam
ine
the
et
i
ology
of
dysle
xia,
fur
ther
stu
dies
wer
e
car
ried
ou
t
us
i
ng
i
m
a
ging
te
chn
i
ques
su
ch
as
f
unct
ion
al
Ma
gn
et
ic
Re
so
nan
c
e
Im
aging
fMRI
[4]
,
Po
sit
r
on
Em
issi
on
T
om
og
ra
phy
PET
[
5]
,
Ma
gn
et
oen
c
ep
ha
logram
MEG
[
6]
wh
ic
h
exam
ine
cogniti
ve
proces
s
associat
ed
wit
h
le
ar
ning
disa
bili
ti
es.
Howe
ver,
EE
G
a
nal
ysi
s
is
the
s
ubje
ct
of
interest
in
this
stu
dy
due
to
it
s
pr
act
ic
al
it
y and
c
os
t
-
e
ff
ect
iv
e
with
high te
m
po
ral res
olu
ti
on.
Ele
ct
rical
a
ct
ivit
ie
s
of
t
he
br
ai
n
ca
n
be
rec
orded
an
d
m
on
it
or
e
d
noni
nv
a
sively
us
i
ng
EE
G
el
ect
r
od
e
s
at
ta
ched
to
th
e
scal
p.
T
his
sign
al
s
hows
act
ivit
ie
s
of
t
he
br
ai
n
re
gion
du
rin
g
e
xec
uting
a
ta
sk
s
uch
as
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
9
,
No.
2
,
Fe
br
uary
201
8
:
40
3
–
40
9
404
decodin
g,
rea
di
ng
,
a
nd
wr
it
in
g.
EE
G
si
gn
al
consi
sts
of
se
ver
al
f
re
qu
e
nc
ie
s
bands
.
Del
ta
wav
es
δ
(
1
-
4H
z
),
Theta
wa
ves
θ
(4
-
7Hz)
,
Al
pha
wav
e
s
α
(8
-
12Hz
),
Be
ta
wa
ves
β
(13
-
30hz
)
an
d
G
am
m
a
wav
e
s
γ
(
31Hz
a
nd
above
)
that i
ndic
at
e d
iffe
ren
t
act
ivit
ie
s an
d
l
evel of a
war
e
ne
ss in
t
he brai
n.
Hen
ce
,
se
ver
al
stud
ie
s
we
re
cond
ucted
t
o
e
xtract
a
nd
cl
assify
EEG
sig
na
l
in
ide
ntifyi
ng
I
ntell
igent
Quotie
nt
(IQ)
[7]
as
well
EE
G
relat
ed
pro
bl
e
m
su
ch
a
s
sl
eep
st
ud
ie
s
[
8]
,
epile
ptic
[9,10]
m
ental
ta
sk
[
11
]
,
m
ental
i
m
aginar
y
[12]
,
m
oto
r
i
m
aginar
y
[
13]
,
br
ai
n
-
c
om
pu
te
r
interface
[
14,
15
]
a
nd
le
a
rni
ng
disabili
ty
[16]
t
o
nam
e a f
ew. L
at
er,
the
ex
t
rac
te
d
EE
G
si
gn
al
w
as
sub
j
ect
ed t
o
cl
assifi
cat
io
n for ide
ntific
a
ti
on
.
Var
i
ou
s
cl
assifi
cat
ion
te
c
hn
i
qu
e
s
ha
ve
bee
n
in
vestigat
e
d
to
ide
ntify
dy
sle
xia
accura
te
ly
.
On
e
of
them
is
SV
M,
wh
ic
h
is
kn
own
as
good
pe
rfor
m
ance
cl
assifi
e
r
com
par
ed
to
oth
e
r
cl
assifi
ers.
S
V
M
is
a
su
pe
r
vised
bina
ry
cl
assifi
cat
i
on
al
gorithm
that
finds
the
op
ti
m
al
separ
at
ing
bounda
r
y
in
hyperpla
ne
by
m
axi
m
isi
ng
the
m
arg
in
of
t
wo
cl
asses/
trai
ning
data.
SVM
has
great
abili
ty
in
so
lvin
g
hi
gh
dim
ension
a
nd
nonline
a
r
feat
ur
es
.
H
ow
e
ve
r
,
the
perf
or
m
a
nce
of
SV
M
in
cl
assify
ing
dysle
xia
us
in
g
the
op
ti
m
u
m
value
ob
ta
ine
d by
va
ryi
ng the scale
of k
e
rn
el
pa
ra
m
et
er h
as
no
t
bee
n rep
or
te
d.
It
is
antic
ipate
d
that
by
tu
ni
ng
the
kernel
par
am
et
er
of
the
S
VM,
t
he
cl
assifi
er
can
pro
du
ce
hi
gh
accuracy
in
c
la
ssifyi
ng
dysle
xia
and
pe
r
form
bette
r
t
han
oth
e
r
cl
a
ssifie
rs.
T
his
pap
e
r
descr
i
be
s
the
cl
assifi
cat
ion
of
EE
G
sig
nals
of
norm
al
,
poor
dysle
xic
a
nd
capab
le
dysle
xi
c
childre
n
usi
ng
m
ulti
cl
ass
SV
M
bin
a
ry
le
arn
er
thr
ough
one
ve
rsu
s
on
e
c
od
i
ng
desig
n.
Va
r
yi
ng
scal
e
of
SV
M
an
d
RB
F
kernel
par
a
m
et
er
is
carried
out t
o
f
ind
t
he op
ti
m
um
p
ara
m
et
ers.
2.
RESEA
R
CH MET
HO
D
In
this
wor
k,
the
ex
am
inati
o
n
of
the
SV
M
pe
rfor
m
ance
in
cl
assify
in
g
dysle
xia
was
carried
out
thr
ough
seve
ra
l
sta
ges
w
hich
include
sub
j
e
ct
identific
at
ion
,
EE
G
si
gn
al
acqu
isi
ti
on,
notc
h
a
nd
high
pass
filt
ering
,
po
we
r
featu
re
e
xtrac
ti
on
,
kernel
pa
r
a
m
et
er
scal
e
tun
in
g,
c
ro
s
s
vali
dation
an
d
cl
as
sific
at
ion
as
s
how
n
in Figu
re
1.
Figure
1.
Flo
w
Char
t
of EE
G Si
gn
al
A
naly
sis
2.1
Subj
ec
t
I
den
tificat
i
on
and T
as
k
Proce
dure
W
i
reless
bio
si
gn
al
ac
qu
isi
ti
on
syst
em
g.
naut
il
us
was
us
e
d
to
captu
re
EE
G
sig
nal
from
the
scal
p
of
the
childre
n.
Head
c
ov
e
r
con
sist
s
of
8
c
hannel
el
ect
rodes
that
are
co
m
plied
with
internati
onal
10
to
20
el
ect
ro
de
placem
ent
syst
e
m
was
us
e
d
duri
ng
the
recordi
ng.
These
el
ec
tro
des
we
re
po
sit
ion
ed
at
C3,
P3
,
T
7
and
FC
5
in
the
le
ft
side
of
the
br
ai
n
a
nd
C
4,
P4
,
T
8
an
d
FC
6
at
the
rig
ht
side
of
the
brai
n
as
sh
ow
n
in
Fi
gure.
2.
T
he
syst
em
acq
uire
d
EE
G
si
gn
al
,
am
plifie
d
an
d
sam
pled
it
us
in
g
a
sam
pling
f
requen
cy
256Hz
befor
e
transm
itti
ng
th
e sig
nal w
irel
e
ssly
to
a
per
s
onal
com
pu
te
r
for rec
ordin
g
a
nd a
naly
zi
ng
.
Subj
ec
t Identi
fi
c
a
ti
on (7 to 12
y
ea
r
s
ol
d) & Task
Proc
edure
(Word
& Non
-
w
o
rd)
EE
G
Si
gnal
Ac
qui
s
i
ti
on (8 c
hannel
el
ec
trodes)
SV
M Cl
as
s
i
fi
c
ati
on
Fil
ter: Notch
(50
Hz
) & Hi
gh P
as
s
(0.5Hz)
End
Feature
ex
t
racti
on: DW
T db2
Power F
eatures
:
Beta
and Th
eta
Band
RBF Kernel
Hi
gh A
c
c
urac
y
?
Y
es
No
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Perf
orma
nce
of
S
up
port Vect
or
M
ac
hin
e i
n C
lassif
yi
ng
E
E
G Sig
na
l
of
Dy
sle
xi
c
…
(
AZA.
Za
i
nuddin
)
405
Figure
2. Ele
ct
rode
Plac
em
en
t i
n
Left
an
d
Ri
gh
t
Hem
isph
er
e of
Brai
n
.
In
t
his
st
ud
y,
the
EE
G
data
w
ere
rec
orde
d
from
33
s
ubj
ect
s
with
t
he
a
ge
r
ang
i
ng
f
ro
m
7
to
12
ye
a
rs
old
.
From
the
total
sub
j
ect
s,
t
he
distrib
ution
is
8
norm
al
,
17
poor
dysle
xics
an
d
8
ca
pab
le
dysle
xics.
This
data
was
ac
quired
with the
assist
a
nt fro
m
D
ysl
e
xi
a A
ss
ociat
ion
of Mal
ay
sia
and Ra
kan D
ysl
e
xia Mal
ay
sia
grou
p.
Tw
o
cat
egories
of
w
ord
were
pr
epa
re
d
for
the
subj
ect
;
known
w
ord
or
w
ord
that
was
fa
m
iliar
to
the
su
bject
with
w
hich
ca
n
be
vis
ualiz
ed
in
t
heir
m
ind
or
ha
ve
a
sp
eci
fic
m
ea
ning.
Anothe
r
cat
egor
y
is
non
-
wor
d
wh
ic
h
has
no
t
seen
befo
re
by
the
su
bject
or
w
ord
that
ha
ve
no
sp
eci
fi
c
m
eaning
in
par
ti
cula
r
an
d
is
no
t
ref
e
rr
in
g
t
o
a
ny
thing
.
T
hr
ee
s
et
s
of
w
ord
a
nd
non
-
w
ord
we
re
pr
e
par
e
d
ba
sed
on
their
ag
e
ap
pro
pr
ia
te
t
o
thei
r
academ
ic
le
vel
.
Set
A
was
f
or
sub
j
ect
of
a
ge
7
to
8,
set
B
was
for
s
ubj
e
ct
of
a
ge
9
to
10
an
d
set
C
was
f
or
su
bject
of
age
11
to
12.
Ta
ble
1
shows
five
ta
sk
s
perf
or
m
ed
by
the
sub
je
ct
wh
il
e
their
br
ai
n
act
ivit
ie
s
are
recorde
d.
Table
1.
T
asks
That
Wer
e
Perform
ed
Durin
g EEG
Sig
nal
R
ecordin
g
Task
Descripti
o
n
Task
1: Bas
elin
e
Su
b
ject
was
ask
ed
to
relax
an
d
tr
y
t
o
th
in
k
o
f
n
o
th
in
g
in
p
articular
f
o
r
4
0
seco
n
d
s.
Task
2: Si
m
p
le
Wo
rd
Three
si
m
p
le
wo
r
d
s
were
sh
o
wn
o
n
e
b
y
o
n
e
an
d
th
e
su
b
ject
was
ask
ed
to write
th
e
wo
rd o
n
a piece of
pap
er
th
en
r
elax
.
Task
3: Co
m
p
lex
W
o
rd
Then
an
o
th
er
th
ree
co
m
p
lex
wo
rds
were
sh
o
wn
o
n
e
b
y
o
n
e
an
d
th
e
su
b
ject
was
ask
ed
to
write
th
e
wo
rd
th
ey
saw
o
n
a
p
i
ece
o
f
p
ap
er
th
en
r
elax
.
Task
4: Si
m
p
l
e N
o
n
-
W
o
rd
Three
si
m
p
le
n
o
n
-
wo
rds
were
sh
o
wn
o
n
e
b
y
o
n
e
an
d
th
e
su
b
ject
was ask
ed
to writ
e
th
e word th
ey
s
aw
o
n
a piece of
pap
er
th
en
r
elax
.
Task
5: Co
m
p
lex
No
n
-
W
o
rd
Three
co
m
p
l
ex
n
o
n
-
wo
rds
were
sh
o
wn
o
n
e
b
y
o
n
e
an
d
th
e
su
b
ject
was ask
ed
to writ
e
th
e word th
ey
s
aw
o
n
a piece of
pap
er
th
en
r
elax
.
Altog
et
her
17
0
dataset
s
we
r
e
colle
ct
ed
where
each
datas
et
con
ta
in
s
8
-
e
le
ct
ro
de
rec
ordin
g.
He
nce,
the
total
nu
m
ber
of
data
recorde
d
was
1360
.
Ou
t
of
this,
sixty
-
five
per
ce
nt
(6
5%
)
of
the
dataset
was
use
d
f
or
trai
ning
data a
nd the
rem
ai
ni
ng
thirty
-
five
pe
rcen
t
(35 %)
of the
dataset
was use
d
f
or te
sti
ng
data.
2.2
EE
G Sig
n
al P
re
-
pr
ocessin
g and Fe
at
ure
s
Extr
act
i
on
The
r
eco
rd
e
d
EEG
si
gn
al
s
w
ere
filt
ered
us
i
ng
a
notc
h
filt
er
t
o
el
im
inate
powe
r
li
ne
noise
at
50Hz
and
a
high
pas
s
filt
er
with
a
cuto
ff
f
reque
nc
y
of
0.5H
z
to
re
m
ov
e
dc
off
set
.
The
data
wer
e
analy
ze
d
us
in
g
a
pro
gr
am
wr
it
te
n
in
Ma
tl
ab
.
Since
E
EG
s
ign
al
is
non
-
sta
ti
on
ary,
ti
m
e
-
scal
e
analy
sis
is
m
or
e
s
uitab
le
for
extract
in
g
the
unde
rly
ing
i
nfor
m
at
ion
tha
n
oth
e
r
m
et
ho
ds
.
The
ra
w
EE
G
signa
ls
we
re
e
xtracted
us
i
ng
D
WT
to
dec
om
po
se
t
he
sig
nal
int
o
f
reque
ncy
sub
-
bands
a
s
s
how
n
in
Fig
ur
e
3.
I
n
this
w
ork,
i
nput
featu
res
w
ere
not
norm
al
iz
ed
becau
se t
he ou
t
pu
t
v
a
riat
ion
was sm
a
ll
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
9
,
No.
2
,
Fe
br
uary
201
8
:
40
3
–
40
9
406
Figure
3. D
WT
D
ec
om
po
sit
ion
of EE
G
Si
gnal
Ou
t
of
seve
ral
wav
el
et
fam
ily
,
Daubiechie
s
of
orde
r
2
(
db2)
wa
s
e
m
plo
ye
d
to
pro
vid
e
EEG
sig
nal
tim
e
-
fr
eq
uen
cy
scal
e
re
pr
e
sen
ta
ti
on
as
it
s
a
bi
li
t
y
to
local
iz
e
featu
res
a
nd
sm
oo
thin
g
over
EE
G
si
gnal
[
17
]
.
The
detai
l
coe
ff
ic
ie
nt
D5
is
t
heta
ba
nd
that
ind
ic
at
es
dr
owsiness
a
nd
t
he
detai
l
coeffic
i
ent
D
3
is
beta
band,
wh
ic
h
ref
e
rs
t
o
act
ive
at
te
ntio
n
a
nd
w
as
the
su
bject
of
inte
r
est
in
this
stu
dy
.
Wh
en
a
ta
s
k
is
pe
rfo
rm
ed
by
the
su
bject
,
t
he
br
ai
n
wa
ves
wil
l
sh
ift
towa
rds
increasin
g
be
ta
band
f
requen
cy
wh
il
e
the
rest
of
the
band
fr
e
qu
e
ncy
will
b
e
red
uce
d.
Theta
-
Be
ta
rati
o
is
an
i
nd
ic
at
ion
of
the
r
el
at
ion
s
hi
p
bet
wee
n
inter
nal
,
(slow
act
ivit
y)
an
d
seq
uen
ti
al
,
(f
ast
act
ivit
y)
[
18,
19]
.
T
heta
band
re
pr
ese
nt
s
the
s
ubco
ns
c
iou
s
m
ind
an
d
beta
ba
nd
repr
esents
the
co
nsc
iou
s
m
ind
.
Br
ai
n
a
ct
ivati
on
t
hrough
theta
-
beta
r
at
io
was
exam
ined
t
o
a
naly
ze
the
br
ai
n
sta
te
at
a
pa
rtic
ul
ar
sit
e
betwee
n
lo
gical
and
s
ponta
neous
processi
ng.
Hi
gh
e
r
rat
io
ind
ic
at
es
th
et
a
is
do
m
inant
w
hile
lowe
r
rati
o
ind
ic
at
es
beta is d
om
inant.
2.3
Clas
sific
at
i
on
In
t
his
sta
ge,
m
ul
ti
cl
ass
cl
as
sific
at
ion
with
on
e
versus
on
e
was
em
plo
ye
d
to
cl
assify
norm
al
,
po
or
dysle
xic
an
d
c
apab
le
dysle
xi
c.
SV
M
with
RB
F
kernel
w
as
then
a
ppli
ed
to
the
e
xtracted
ba
nd
po
wer
featur
e
s
of
Be
ta
an
d
T
he
ta
-
Be
ta
rati
o.
SV
M
cl
assifi
c
at
ion
is
base
d
on
fi
nd
i
ng
m
a
xim
u
m
m
arg
in
sepa
rati
on
bo
unda
ry
betwee
n
t
wo
c
la
sses.
I
n
li
nea
r
f
orm
,
the
sep
arati
on
can
be
done
strai
ght
f
orward
but
f
or
nonlin
ear
co
ndit
ion
,
the
data
has
t
o
be
placed
i
n
f
e
at
ur
es
s
pace
w
her
e
t
he
se
pa
ra
ti
on
is
perform
ed
in
hy
per
s
pa
ce.
Ke
r
nel
is
a
string
that
sp
eci
fies
the
kernel
f
un
c
ti
on
an
d
is
us
e
d
to
m
ap
the
data
fr
om
inp
ut
sp
ace
into
a
ne
w
sp
ace.
T
he
re
are
three
ty
pe
s
of
kernel
f
unct
io
n
that
ca
n
be
u
s
ed.
They
a
re
know
n
as
Linea
r,
Po
ly
nom
ia
l
and
RB
F.
Po
ly
no
m
ia
l
and
RB
F
kerne
l
are
us
e
d
f
or
m
app
in
g
no
n
-
l
inear
data
int
o
hype
rsp
ace
.
T
he
S
VM
cl
assifi
er
can
be
w
ri
tt
en
as
in E
qu
at
io
n (
1) an
d
the
RB
F
ke
rn
el
functi
on i
s sho
wn in E
qu
at
ion
(2).
N
i
i
i
i
b
x
x
k
y
x
f
,
(1)
2
2
2
||
'
||
e
xp
'
,
x
x
x
x
k
(2)
2
2
1
(3)
Eq
uation
(
3)
sh
ows
or
kern
el
width
that
is
a
po
sit
ive
nu
m
ber
sp
eci
fyi
ng
the
ke
rn
el
sc
al
e
factor
wh
ic
h
is
us
e
d
to
sp
eci
fy
the
sh
a
pe
of
“
pea
k”
ei
ther
br
oa
de
r
or
pointed
bum
p.
The
SV
M
cl
assifi
er
with
RB
F
kernel i
s
giv
e
n by E
qu
at
io
n (4).
b
x
x
y
x
f
N
i
i
i
i
2
2
2
||
||
e
x
p
(4)
The
SV
M
cl
as
sifie
r
with
RB
F
kernel
has
two
par
am
et
ers;
ker
nel
scal
e
(
)
and
box
co
nst
raint
(C).
Box
co
ns
tr
ai
nt
is
a
reg
ulati
on
par
am
et
er
wh
ic
h
co
ntr
ols
tradeoff
betwee
n
m
arg
in
m
axi
m
iz
at
ion
and
e
rror
s
of
trai
ning
data. S
VM w
it
h (C
)
is
shown i
n
E
qu
at
ion
(5)
b
x
x
y
C
x
f
N
i
i
i
i
2
2
2
||
||
e
x
p
(5)
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Perf
orma
nce
of
S
up
port Vect
or
M
ac
hin
e i
n C
lassif
yi
ng
E
E
G Sig
na
l
of
Dy
sle
xi
c
…
(
AZA.
Za
i
nuddin
)
407
To
obta
in
the
op
ti
m
al
par
am
et
ers,
var
yi
n
g
scal
e
on
SV
M
with
RB
F
kernel
was
carrie
d
out.
In
the
first
analy
sis
the
box
c
onstra
int
was
va
ried
from
0.
001
to
1000
by
incre
a
sing
fa
ct
or
of
10
wh
il
e
ke
rne
l
scale
was
set
to
1.
I
n
the
sec
ond
a
na
ly
sis
the
kern
el
scal
e,
was
va
ried
fr
om
0.
001
t
o
1000
by
increasi
ng
fact
or
of
10
wh
il
e
the
box
c
on
st
raint
was
fixe
d
to
1.
Cros
s
-
validat
ion
with
K
-
fo
l
d
eq
ual
to
te
n
fo
l
ds
was
ap
pl
ie
d
to
pr
e
dicts cl
assif
ic
at
ion
acc
ur
ac
y wit
h
the
lo
w
est
error is
perf
or
m
ed
with
tra
ining data.
Confus
i
on
m
a
trix
f
o
r
m
ulticlass
wer
e
th
en
em
plo
ye
d
in
orde
r
to
ve
rify
the
pe
rfor
m
ance
of
cl
assifi
cat
ion
m
od
el
.
The
se
ns
it
ivit
y,
sp
eci
fici
ty
and
acc
uracy
we
re
dete
rm
ined
us
in
g
Eq
uation
(
6)
,
(
7)
a
nd
(8) respecti
vel
y.
N
P
P
PR
e
F
T
T
T
S
y
S
e
n
s
i
ti
v
it
,
(6)
P
N
N
NR
p
F
T
T
T
S
y
S
p
e
c
if
ic
it
,
(7)
N
P
N
P
N
P
c
F
F
T
T
T
T
A
A
c
c
u
r
a
c
y
,
(8)
3.
RESU
LT
S
AND A
N
ALYSIS
Table
2
sho
w
s
the
re
su
lt
of
k
-
fo
l
d
c
ro
ss
-
validat
io
n
er
ror
f
or
va
rio
us
C
and
ke
rn
el
scal
es.
It
is
obvious
that sc
al
e 1
for bo
t
h C
and
giv
es t
he
low
e
st er
ror,
wh
ic
h
is
23%.
Table
2.
K
-
Fo
l
d
Cr
os
s
-
Vali
da
ti
on
E
rror
Scale
Cro
ss
Validatio
n
0
.00
1
0
.01
0
.1
1
10
100
1000
Bo
x
Co
n
strain
t,
C
0
.45
0
.43
0
.40
0
.23
0
.20
0
.21
0
.20
Kernel Sc
ale,
0
.71
0
.71
0
.71
0
.23
0
.38
0
.55
0
.52
The
sensiti
vity
ver
s
us
C
plo
t
of
the
m
ulti
cl
a
ss
SV
M
cl
assifi
er
wh
e
n
C
is
var
ie
d
from
0.
001
to
10
00
is
sh
own
i
n
Figure
4(a)
.
As
c
an
be
see
n,
inc
reasin
g
C
m
or
e
than
0.1
decre
ases
the
cl
assifi
er
sensiti
vity
fr
om
100%
to
92%
for
po
or
dysle
xic,
w
hile
for
capab
l
e
dysle
xi
c
the
sensiti
vity
rap
idly
incr
eases
from
25
%
to
75%.
I
n
co
ntr
ast
,
the
sensiti
vity
fo
r
norm
al
su
bject
do
e
s
no
t
cha
nge
and
sta
ys
at
100%
.
Furthe
r
m
or
e,
increasin
g
C
a
bove 1
giv
e
no
ch
a
ng
e
s to
classi
fier se
ns
it
ivit
y fo
r
all
classe
s.
(a)
(b)
Figure
4. Mult
ic
la
ss SV
M C
la
ssific
at
ion
Per
f
or
m
ance
Wh
e
n C
is v
a
ried
for N
or
m
al
, P
oor Dysl
exic a
nd
Ca
pab
le
Dysle
xic (
a
)
Se
ns
it
iv
it
y (b
) Speci
fici
ty
Figure
4(b)
s
hows
t
he
spe
ci
fici
ty
of
m
ulti
cl
ass
SV
M
cl
assifi
cat
ion
perform
ance
wh
ic
h
was
m
easur
ed
for
var
i
o
us
range
of
C
(
0.001
to
1000).
It
can
be
seen
t
hat
the
sp
eci
fici
ty
for
cl
assify
ing
c
apab
le
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vol
.
9
,
No.
2
,
Fe
br
uary
201
8
:
40
3
–
40
9
408
dysle
xic
an
d
norm
al
su
bj
ect
decr
ease
s
from
100%
to
98%
and
95
%
resp
e
ct
ively
,
wh
il
e
fo
r
po
or
dysle
xic
th
e
sp
eci
fici
ty
incre
ases f
ro
m
6
3%
to 8
8% wh
e
n
C i
s s
et
at 1
. T
h
e
res
ult rem
ai
ns
un
c
ha
ng
e
d wh
e
n
C
is ab
ov
e
1.
Althou
gh
C
in
the
ran
ge
of
0.001
to
0.1
pe
rfor
m
s
bette
r
in
sp
eci
fici
ty
fo
r
norm
al
a
nd
ca
pab
le
dysle
xic,
it
do
es
not
pe
rfor
m
well
for
poor
dysle
xic.
T
hu
s,
it
can
be
co
nclu
ded
t
hat
C
equ
al
s
t
o
1
is
the
op
ti
m
al
s
et
ti
ng
that
giv
es
the
best
ov
e
rall
sensiti
vity
and
sp
eci
fici
ty
fo
r
c
la
ssifyi
ng
nor
m
al
,
po
or
dysle
xic
an
d
capab
le
dysle
xi
c.
Figure
5(a)
a
nd
(
b)
s
hows
the
sensiti
vity
and
s
pecifici
ty
fo
r
norm
al
,
po
or
dysle
xic
a
nd
ca
pab
le
dysle
xic
res
ulted
from
SV
M
cl
assifi
ca
ti
on
wh
e
n
value
is
var
ie
d
from
0.
001
to
1000.
Wh
e
n
is
set
fr
om
0.001
to
0.1
,
the
SV
M
se
ns
it
ivit
y
fo
r
poor
and
ca
pab
le
dysle
xic
is
fluctu
at
ed,
betwe
en
0%
an
d
10
0%.
Wh
il
e
for
norm
al
su
bject
it
is
no
t
sensiti
ve
at
al
l.
Howe
ver,
w
he
n
is
set
to
1,
the
sensiti
vity
increase
s
to
10
0%
f
or
norm
al
,
92
%
for
po
or
dysle
xic
an
d
75%
for
capa
ble
dy
sle
xic.
A
bove
scal
e
of
10,
the
sensiti
vity
drops
trem
end
ou
sly
wh
e
n
cl
assi
fyi
ng no
rm
al
an
d poo
r dysl
exic.
The
sam
e
tren
d
is
ob
se
r
ved
in
the
sp
eci
fici
ty
for
in
the
ra
ng
e
of
0.001
t
o
0.1.
At
scal
e
equ
al
t
o
1,
sp
eci
fici
ty
fo
r
cl
assify
ing
nor
m
al
su
bj
ect
is
95%,
w
hile
for
poor
dysle
xic
and
ca
pab
le
dy
sle
xic,
it
is
88
%
and
98% r
e
sp
ect
ive
ly
. T
he
best se
ns
it
ivit
y and sp
eci
fici
ty
are
ob
ta
ined for all
gro
ups
wh
e
n
is
set
to 1.
(a)
(b)
Figure
5. Mult
ic
la
ss SV
M C
la
ssific
at
ion
Per
f
or
m
ance
Wh
e
n Ke
rn
el
Scal
e is
Va
ried
f
or
N
or
m
al
, P
oor
Dysle
xic a
nd
Ca
pab
le
Dysle
xic (
a
)
Se
n
sit
iv
it
y (b
) Speci
fici
ty
It is
obser
ve
d
t
hat in F
i
gure
6, cl
assifi
er acc
uracy
for
C is
hi
gh, whic
h
is i
n t
he
ra
nge
of
94% to
89%.
Howe
ver,
cl
as
sifie
r
acc
ur
acy
is
not
sta
ble
f
or
,
wh
ic
h
inc
r
eases
an
d
decre
ases
bet
ween
91%
to
9%.
Wh
en
bo
t
h
an
d
C
ar
e
1,
the
S
VM
accuracy
is
91%
.
The
acc
ur
acy
decr
eases
whe
n
both
par
am
eter
s
is
set
above
1.
Th
us
, t
he op
ti
m
al
v
al
ue
f
or
C an
d
is
1
si
nc
e these
values
giv
e
good acc
uracy
.
Figure
6. Mult
ic
la
ss SV
M C
la
ssific
at
ion
O
ve
rall
A
cc
ur
acy
f
or RB
F Ker
nel
4.
CONCL
US
I
O
N
This
w
ork
wa
s
carried
ou
t
to
exam
ine
the
cl
assifi
cat
ion
pe
rfor
m
ance
of
m
ulti
cl
ass
SV
M
in
disti
nguish
i
ng
EEG
sig
nal
of
norm
al
,
poor
and
capa
ble
dy
sle
xic
child
re
n.
The
ext
ra
ct
ion
of
featu
res
wh
ic
h
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Perf
orma
nce
of
S
up
port Vect
or
M
ac
hin
e i
n C
lassif
yi
ng
E
E
G Sig
na
l
of
Dy
sle
xi
c
…
(
AZA.
Za
i
nuddin
)
409
are
Be
ta
an
d
T
heta
-
Be
ta
rati
o
was
ca
rr
ie
d
out
us
i
ng
wa
vel
et
db2
a
nd
thes
e
featu
res
wer
e
us
e
d
as
the
in
pu
t
t
o
the
cl
assifi
er.
The
bo
x
co
ns
t
raint
of
SV
M
and
th
e
RB
F
ke
rn
el
pa
ram
et
e
r
wer
e
var
ie
d
to
fin
d
the
op
ti
m
u
m
resu
lt
s.
Cr
os
s
-
valid
at
io
n
al
so
was
carried
ou
t.
T
he
res
ults
ob
ta
ine
d
in
this
stud
y
sh
ow
s
that
RB
F
ker
nel
par
am
et
er
aff
e
ct
s
the
cl
assifi
cat
ion
perform
ance.
Sett
ing
to
1
in
the
RB
F
kernel
a
nd
C
t
o
the
sam
e
value
in
the
SV
M
yi
el
ded
the
high
est
accuracy,
wh
ic
h
is
at
91%.
The
S
VM
with
RB
F
kernel
co
uld
cl
ass
ify
the
norm
al
,
po
or
dy
sle
xic
and
ca
pab
le
dysle
xic
childre
n
accu
r
at
el
y
with
hig
h
sensiti
vity
and
sp
eci
fici
ty
us
ing
the
op
ti
m
u
m
p
ara
m
et
ers.
ACKN
OWLE
DGE
MENT
This
w
ork
w
as
su
pp
or
te
d
by
Fund
am
ental
Re
search
Gr
a
nt
Schem
e
(F
RGS
),
Ma
la
ysi
a
(6
00
-
RM
I/FRGS
5/3
(
137/
2015)
).
The
a
uthors
would
li
ke
t
o
than
k
Mi
nistry
of
Higher
Ed
ucati
on,
Ma
la
ysi
a,
Re
search
Ma
na
gem
ent
In
sti
tute
an
d
Facult
y
of
Ele
ct
rical
Eng
i
neer
i
ng,
Un
i
ver
s
it
i
Tek
no
l
og
i
MAR
A
,
Sh
a
h
Alam
,
fo
r
fina
ncial
su
pp
or
t,
facil
it
ie
s
and
var
io
us
co
ntributi
ons,
an
d
to
Dysle
xia
Asso
ci
at
ion
Ma
la
ys
ia
for
their assist
an
ce
.
REFERE
NCE
S
[1]
B.
a
.
Sha
y
wi
tz
et al
.
,
“
Th
e
neu
rob
iol
og
y
of
d
y
s
le
xi
a,
”
Cli
n
.
N
eurosci. Res.
,
vol
.
1
,
p
p.
291
–
299
,
200
1.
[2]
B.
Sklar
,
J.
Han
ley
,
and
W
.
W
.
Sim
m
ons,
“
A
c
om
pute
r
anal
y
s
i
s
of
EE
G
Spe
ctral
Sign
at
ure
s
f
r
om
Norm
al
and
D
y
slexic
Ch
il
dre
n,
”
I
EEE
Tr
ans. Biomed. E
ng
.
,
v
ol.
49
,
no
.
1
,
pp
.
20
–
26,
1973
.
[3]
H.
M.
Al
-
Barham
toshy
and
D.
M.
Motaweh,
“
Diagnosis
of
D
y
slexi
a
using
computat
ion
anal
y
s
is,”
2017
Int
.
Con
f.
Informatic
s,
H
ea
l.
Te
chnol.
ICIH
T 2017
,
vol
.
6
,
n
o.
2
,
pp
.
462
–
48
2,
2017
.
[4]
M.
Vande
rm
oste
n,
F.
Hoeft
,
an
d
E.
S.
Norton
,
“
In
te
gra
ti
ng
MRI
br
ai
n
im
agi
ng
studie
s
of
pre
-
r
ea
d
in
g
chi
ldr
en
with
cur
ren
t
th
eor
i
es
of
deve
lopment
a
l
d
y
sl
exi
a
:
a
rev
i
ew
and
quant
i
ta
t
ive
m
et
a
-
ana
l
y
si
s,”
Curr
.
Opin.
Be
hav
.
Sci.
,
vo
l
.
10,
no
.
June
,
pp
.
155
–
161,
Aug. 2016.
[5]
Y.
Sun,
J.
Le
e,
a
nd
R.
Kirb
y
,
“
Brai
n
Im
agi
ng
Findings
i
n
Dy
sl
exia,
”
P
edi
atr
.
Neo
natol
.
,
vol.
51,
n
o.
2,
pp.
89
–
96,
Apr.
2010.
[6]
S.
I.
Dim
it
ri
adi
s
et
al
.
,
“
Altered
te
m
pora
l
co
rre
l
at
ions
in
r
esti
ng
-
stat
e
connect
iv
i
t
y
f
luctuations
i
n
chi
ldr
en
with
rea
ding
difficul
ties de
t
ec
t
ed
v
ia
MEG,”
Neuroim
age
,
vo
l. 83, pp.
307
–
317,
De
c. 2
013.
[7]
A.
H.
Jahid
in,
M.
N.
T
ai
b
,
N.
M.
Ta
h
ir,
and
M
.
S.
A.
M.
Al
i,
“
IQ
Cla
ss
ifi
c
at
io
n
via
Br
ai
nwav
e
Feat
ur
es:
Rev
iew
on
Artificial
In
telli
gen
ce Tec
hniq
ues,
”
Int. J. Elec
tr.
Comput.
Eng
.
,
vol
.
5
,
no
.
1
,
pp
.
84
–
91
,
2015
.
[8]
S.
Motamedi
-
Fa
khr,
M.
Mos
hre
fi
-
Torbati,
M.
Hill
,
C
.
M.
Hill
,
and
P.
R.
W
hit
e,
“
Signal
proc
essing
te
chni
qu
es
appl
i
ed
to
hum
a
n
slee
p
E
EG
sig
nal
s
—
A re
vi
ew,”
B
iomed. Si
gnal
Proce
ss
.
Contro
l
,
vo
l. 10, pp. 21
–
33,
Mar
.
2014
.
[9]
A.
T.
Tzallas,
M.
G.
Tsipouras
,
and
D.
I.
Fotia
d
is,
“
Epi
l
eptic
seiz
ure
de
te
c
tion
in
EE
Gs
using
ti
m
e
-
fre
quen
c
y
ana
l
y
sis.
,
”
IE
EE Tr
ans.
Inf. Tec
h
nol.
Bi
omed
.
,
vo
l.
13
,
no
.
5
,
pp
.
7
03
–
710,
2009
.
[10]
H.
Ulla
h
,
S.
Ma
hm
ud,
and
R.
H
.
Chowdhur
y
,
“
I
dent
ifica
ti
on
of
Brai
n
disorder
s
b
y
Sub
-
band
De
compos
it
ion
of
E
EG
signa
ls
an
d
Mea
surem
ent
of
Signal
to
Noi
se
Ratio,
”
Indon
es.
J
.
Elec
tr.
En
g.
Comput.
Sc
i.
,
vol.
4,
no.
3,
p.
568,
De
c. 2016.
[11]
I.
Güler
and
E.
D.
Ube
y
li,
“
Multi
class
support
v
ec
tor
m
a
chi
nes
f
or
EE
G
-
signal
s
cl
assifi
ca
t
ion.,”
IEE
E
Tr
ans.
In
f
.
Technol
.
B
iom
ed
.
,
vol. 11, no. 2,
pp.
117
–
126
,
20
07.
[12]
M.
S.
Basc
il
,
A.
Y.
T
esne
li,
F.
Te
m
urta
s,
Á
.
Pc
a,
and
Á.
L.
V.
Q.
Á.
Mlnn,
“
Multi
-
ch
annel
E
E
G
signal
fe
at
ur
e
ext
ra
ct
ion
and
pat
t
ern
re
cognit
ion
on
hori
zontal
m
ental
imagi
nat
ion
ta
sk
of
1
-
D
cur
sor
m
ovement
for
br
ain
computer
in
te
rf
a
ce
,
”
Australas. P
hys.
Eng
.
S
ci.
M
ed.
,
vol
.
38
,
no
.
2,
pp
.
229
–
239
,
2015.
[13]
S.
K.
Bashar
,
A.
B.
Das,
and
M.
I.
H.
Bhuiy
an
,
“
Motor
imager
y
m
ovements
det
ection
of
EE
G
signal
s
using
stat
isti
ca
l
f
ea
tu
r
es
in
the
Dual
Tre
e
Com
ple
x
W
ave
le
t
Tra
nsf
orm
do
m
ai
n,
”
2
015
Int.
Conf.
El
e
ct
r.
Eng
.
Inf
.
Comm
un.
Techn
ol.
,
pp.
1
–
6,
201
5.
[14]
X.
Li,
X.
Chen
,
Y.
Yan,
W
.
W
ei,
and
Z
.
J.
W
ang
,
“
Cla
ss
ifi
c
at
ion
of
EE
G
sign
al
s
using
a
m
ult
ip
le
ker
ne
l
l
ea
rn
ing
support
vector
m
ac
hin
e,”
Sensors
(
Basel
)
.
,
vol
.
14
,
no
.
7
,
pp
.
1278
4
–
12802,
2014
.
[15]
A.
B.
M.
A.
Hos
sain,
W
.
Rahman,
and
M.
A.
Rihe
en,
“
Le
ft
and
Rig
ht
Hand
Moveme
nts
EE
G
Signal
s
Cla
ss
ifi
c
at
ion
Us
ing
W
ave
let
Tr
ansform
and
Probabil
ist
ic
N
eur
al
Network,
”
Int
.
J.
El
e
ct
r.
Comput.
Eng.
,
vol
.
5
,
no.
1
,
pp
.
92
–
1
0
1,
2015
.
[16]
A.
Z.
A
.
Z
ai
nud
din,
K.
Y
.
L
ee,
W
.
Mansor,
and
Z.
Mahm
oodin
,
“
Optimize
d
KN
N
cl
assif
y
rule
for
EE
G
b
ase
d
diffe
ren
ti
a
ti
on
b
et
wee
n
c
apa
bl
e
d
y
slex
ic
and
no
rm
al
chi
ldre
n
,
”
in
2016
IEE
E
E
MBS
Confe
renc
e
on
Bi
omedic
al
Engi
ne
ering
and
Scienc
es
(
IECB
ES
)
,
2016,
pp
.
6
85
–
688.
[17]
T.
Gandh
i,
B.
K.
Panigr
ahi,
and
S.
Anan
d,
“
A
compara
ti
ve
stud
y
of
wave
let
f
amili
e
s
for
EE
G
signa
l
cl
assifi
ca
t
ion,”
Neurocomputi
ng
,
vol
.
74
,
no
.
17
,
pp.
3051
–
3057
,
2011.
[18]
J.
E
.
W
al
k
er
an
d
C.
A
.
Norm
an
,
“
The
Neuroph
y
siolog
y
of
D
y
s
le
xi
a:
A
Selective
Rev
ie
w
wi
th
Im
pli
cations
fo
r
Neurofe
edback
Remedia
t
ion
an
d
Result
s
of
Tre
at
m
ent
in
Twe
lv
e
Consecut
iv
e
Patients,
”
J
.
Neur
othe
r.
,
vol
.
10,
n
o.
1,
pp
.
45
–
55
,
De
c.
2006
.
[19]
C.
Spirone
ll
i
,
B
.
Penolazz
i
,
and
A.
Angrilli,
“
D
y
sfuncti
ona
l
he
m
i
spheric
as
y
m
m
etr
y
of
th
eta
and
b
et
a
EE
G
activit
y
during
l
ingui
sti
c
ta
sks
in
dev
el
op
m
ent
al
d
y
sl
exi
a
,
”
B
iol
.
Psy
chol.
,
vol.
77
,
no
.
2
,
pp
.
123
–
131
,
2008
.
Evaluation Warning : The document was created with Spire.PDF for Python.