Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
5, N
o
. 1
,
Febr
u
a
r
y
201
5,
pp
. 92
~101
I
S
SN
: 208
8-8
7
0
8
92
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Left and Right Hand Movements
EEG Signals Classification
Using Wavelet Transform and
Probabilistic Neural Network
A.B.
M.
Aowlad H
o
ss
ain,
Md. W
a
siur
Rahman
,
Manju
r
ul Ahs
a
n
Rih
een
Department o
f
Electronics
and C
o
mmunication Engineer
ing
Khulna University
of
Engin
eerin
g &
Technolog
y
,
Bangladesh
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Oct 7, 2014
Rev
i
sed
D
ec 11
, 20
14
Accepted Dec 26, 2014
Electroen
ceph
a
logram (EEG) signals ha
v
e
great importance in
the
area of
brain-com
puter
i
n
terfa
ce (BCI)
which
has div
e
r
s
e applications r
a
nging from
medicine
to entertainment. BCI
acquir
e
s brain
signals, ex
trac
ts inform
ativ
e
featur
es and gen
e
rates control signals
from the knowledge of these features
for functioning
of external devices. The
objective of this work
is twofold.
Firstl
y, to extr
ac
t suitable f
eatur
e
s
re
lated to hand
movements and
second
ly
,
to discriminate the left and right
hand movemen
t
s signals finding effectiv
e
classifier. Th
is work is a contin
uation of our previous stud
y
where beta b
a
nd
was found compatible for
hand
moveme
nt analy
s
is.
The discrete wav
e
let
transform (DWT) has b
een
used to sep
a
rate
beta
band
of the EEG signal in
order to extr
ac
t featur
es. Th
e pe
rform
ance of a
probabilist
i
c neu
r
al networ
k
(PNN) is investigat
ed to find
better
classifier of left and
right hand
movements EEG signals and
compared wi
th classical b
ack
propagation based
neural n
e
twork. The obtained
r
e
sults
shows that PNN (99.1%) has better
clas
s
i
fi
cat
ion ra
t
e
than
the
BP
(88.9%).
The r
e
s
u
l
t
s
of this
s
t
ud
y a
r
e exp
ect
ed
to be helpful in
brain computer in
terfacing for hand movements r
e
lated bio-
rehabi
lit
ation
ap
plic
ations.
Keyword:
Artificial n
e
u
r
al n
e
two
r
k
Back
pr
op
ag
atio
n algo
r
ith
m
Discrete wav
e
l
e
t
tran
sform
Electroe
n
cephalogram
Feature
extraction
Prob
ab
ilistic neu
r
al
n
e
two
r
k
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
A.B.M
.
A
o
wlad Hos
s
ain,
Depa
rt
m
e
nt
of
El
ect
roni
cs
an
d C
o
m
m
uni
cati
on
En
gi
nee
r
i
n
g
K
h
u
l
n
a
Un
iv
ersity o
f
Eng
i
n
e
er
ing
& Technolo
g
y
Kh
ul
na
-
9
2
0
3
, B
a
ngl
a
d
esh
.
Em
a
il: ao
wlad0
403
@ece.ku
e
t.ac.bd
1.
INTRODUCTION
Ov
er a m
ill
io
n
in
d
i
v
i
d
u
a
ls are su
fferi
n
g
fro
m
d
i
sab
ility an
nu
ally as a resu
lt o
f
strok
e
, t
r
aumatic b
r
ain
or s
p
i
n
al
co
rd
i
n
j
u
ri
es [
1
]
.
A
m
a
jor
po
rt
i
on
of di
sa
bl
e pe
o
p
l
e
have
rep
o
rt
ed t
r
o
u
b
l
e
s wi
t
h
ha
nd
fu
nct
i
o
n [2]
,
[3]. Failure of hand function cause
s seve
re problem
s in leading
of lif
e for the affe
cted persons.
Recent
research
es show th
at BCI is a n
e
w
h
o
p
e
i
n
treatm
e
n
tin
g th
e
d
i
sab
ilities. EEG si
g
n
a
l
is th
e m
o
st trend
y
resource
o
f
in
t
e
rpretin
g th
e brain
activ
ities i
n
th
e realm
o
f
n
on-inv
a
siv
e
BCI. Th
ey are
well stu
d
i
ed an
d th
ere
is ev
id
en
t th
at t
h
ey can b
e
used
in artificial han
d
m
o
v
e
m
e
n
t
s [4
].
EEG is
g
r
ap
h
i
cal rep
r
esen
tat
i
o
n
o
f
electrical activ
ities o
f
b
r
ai
n
wh
ich
i
s
record
ed
u
s
i
n
g
electrod
e
s
l
o
cat
i
ng
o
n
t
h
e
scal
p.
EE
G
ha
ve ce
rt
ai
n
ban
d
s
wi
t
h
se
pa
rat
e
fre
q
u
ency
ra
nge
s [
5
]
:
t
h
et
a
wa
ves
vari
es
i
n
t
h
e
ran
g
e o
f
4 Hz t
o
7 Hz an
d i
t
s
am
pl
i
t
ude gen
e
ral
l
y
arrou
n
d
20
μ
V, alpha
wave
varies wi
th in the range of 8 t
o
13
Hz
a
n
d
ab
out
3
0
-
5
0
μ
V
am
pl
i
t
ude. Fo
r
bet
a
wa
ve,
t
h
e
fre
q
u
e
n
ci
es vary
bet
w
ee
n 13
Hz
t
o
3
0
Hz
a
n
d
usu
a
l
l
y
have
a
l
o
w
vol
t
a
ge
b
e
t
w
een
5
-
3
0
μ
V. Differen
t ban
d
s
carry
inform
at
io
n
o
f
d
i
fferen
t brain
acti
v
ities.
EEG si
g
n
a
ls are ex
ten
s
iv
ely
stu
d
i
ed
b
y
n
u
mero
u
s
resear
ch
es to
classi
fy d
i
fferen
t
m
e
n
t
al o
r
brain
act
iv
ities
[6]
-
[
9]
.
Few
st
udi
es
ha
ve be
en pr
o
pose
d
o
n
han
d
m
ove
m
e
nt
cl
assi
fi
cat
i
ons usi
n
g su
pp
o
r
t
vect
or m
achi
n
e
(SVM
), lin
ear
d
i
scrim
i
n
a
tio
n
an
alysis, ad
ap
tiv
e Gau
ssi
an c
o
efficients, C-SVM and c
o
mbination of EE
G a
nd
MEG.
Howe
ver, m
o
st of the
m
are com
putationally com
p
lex and not s
o
effective
for re
al time applications
[10]-[14]. In t
h
is context, s
e
lec
tion of a
p
propriate and
com
p
atible f
eature is
very
crucial for effective
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Left a
n
d
Righ
t
H
and
Mo
vem
e
n
t
s EEG
S
i
gnals Cla
ssifica
tion
U
s
ing
Wa
vel
et Tran
sform
…
(
A
.B.M.A
. Hos
s
ain)
93
classif
i
catio
n
w
ith
sim
p
le classif
i
er
. In
ou
r pr
ev
i
o
us st
u
dy on
selection of
p
r
op
er fr
eq
u
e
n
c
y
b
a
n
d
fo
r
h
a
nd
m
ovem
e
nt
ana
l
y
s
i
s
[1
5]
, i
t
was
ob
ser
v
e
d
t
h
at
bet
a
ba
nd
(
9
7
.
5%
)
has
hi
g
h
er
m
a
ppi
n
g
preci
si
o
n
a
n
d
bet
t
e
r
con
v
e
r
ge
nce
r
a
t
e
t
h
an
t
h
e
ot
her
ba
n
d
s,
al
p
h
a
(9
3.
2%
) a
n
d t
h
et
a (
8
7.
8%
). T
h
e
r
ef
ore
,
b
e
t
a
ba
nd
can
b
e
m
o
st
su
itab
l
y u
s
ed
fo
r hand
m
o
v
e
men
t
an
alysis.
Th
is wo
rk
is an
ex
tension
in
o
r
d
e
r to
find
effectiv
e classifier of
left and
ri
g
h
t
han
d
m
o
v
e
m
e
n
t
EEG sign
al
wh
ich
is in
itiated
fro
m
th
e findin
g
s
of
o
u
r
p
r
ev
iou
s
work
.
In t
h
i
s
st
u
d
y
,
we pl
an t
o
rec
o
g
n
i
ze l
e
ft
and ri
g
h
t
han
d
m
ovem
e
nt
EEG si
gnal
s
fi
ndi
ng si
m
p
l
e
but
effective
class
i
fier
based
on wa
velet tra
n
s
f
orm
and
ne
u
r
al
net
w
o
r
k a
p
p
r
oaches
. Applying
the selected
feat
ure
s
,
we t
r
y
t
o
cl
assi
fy
left
an
d ri
g
h
t
h
a
nd m
ovem
e
nt
s usi
n
g
neu
r
al
net
w
or
k ba
se
d m
e
t
hods.
Di
ffe
rent
feat
ure
s
are pl
a
nne
d t
o
cal
cul
a
t
e
usi
ng wa
vel
e
t
anal
y
s
i
s
of EEG si
g
n
al
s.
Wavel
e
t
t
r
an
sf
orm
i
s
a powe
r
ful
t
o
ol
t
o
p
r
ocess
bi
o
m
edi
cal
si
gnal
s
f
o
r
va
ri
o
u
s a
ppl
i
cat
i
o
ns
[1
6
]
. It
i
s
a
r
o
b
u
st
t
echni
q
u
e t
o
r
e
prese
n
t
t
h
e si
gnal
s
i
n
tim
e
-frequency
dom
a
in. The
c
a
pability of tim
e-fre
que
ncy
dom
ain analysi
s
of
wa
velet transform
can be
use
f
ul
t
o
sepa
rat
e
di
f
f
ere
n
t
ba
nd
s a
nd c
o
nseq
ue
nt
l
y
t
o
ext
r
act
i
m
po
rt
ant
feat
ure
s
. N
u
m
e
rous t
echni
que
s ha
v
e
been
p
r
op
o
s
ed
for classificatio
n
of b
i
o-ev
en
ts
o
r
abn
o
rm
al
itie
s.
Artificial n
e
ural n
e
twork
s
hav
e
b
een
app
lied
i
n
several
stu
d
ies
f
o
r
EE
G a
n
al
y
s
is [1
7]
-[
1
9
]
.
A
N
N
s
is
e
x
e
c
ut
ed
t
h
r
o
ug
h
t
r
ai
ni
n
g
al
g
o
rith
m
s
with
sp
ecified
l
earni
n
g
cri
t
e
ri
a t
o
im
i
t
a
t
e
t
h
e l
earni
n
g
m
echani
s
m
s
of bi
ol
ogi
cal
ne
ur
o
n
s
[20]
. Di
ffe
rent
t
y
pes and st
ru
ct
ures
o
f
n
e
u
r
al
n
e
two
r
k
s
are
u
s
u
a
ll
y u
s
ed. In
t
h
is work, th
e probab
ilistic n
e
u
r
al
n
e
two
r
k
(PNN) and
classical b
ack
pr
o
p
agat
i
o
n
ba
sed
ne
ural
net
w
o
r
k
(B
P) a
r
e
appl
i
e
d
a
n
d
t
h
ei
r pe
rf
orm
a
nc
es are
com
p
are
d
t
o
i
d
e
n
t
i
f
y
t
h
e l
e
ft
and
ri
ght
ha
nd
EEG si
g
n
al
s.
The re
st
of t
h
e
pape
r i
s
o
r
ga
n
i
zed as f
o
l
l
o
w
s
:
t
h
e next
sect
i
on
desc
ri
bes t
h
e pr
op
ose
d
m
e
t
h
o
dol
ogy
whi
c
h desc
ri
be
s t
h
e wavel
e
t
b
a
sed dec
o
m
pos
i
t
i
on of EE
G b
a
nd
s, feat
u
r
e ext
r
act
i
o
ns an
d t
w
o ne
ural
net
w
o
r
k
s
base
d ap
p
r
oac
h
es
fo
r cl
assi
fi
cat
i
on. T
h
i
r
d s
ect
i
on
prese
n
t
s
t
h
e de
scri
pt
i
o
n
of
dat
a
u
s
ed i
n
t
h
i
s
st
udy
as
wel
l
as
anal
y
s
i
s
an
d
di
scussi
o
n
o
n
o
b
t
ai
ned
resul
t
s
whi
c
h a
r
e f
o
l
l
o
wed
by
c
o
ncl
u
si
on
i
n
fo
u
r
t
h
s
ect
i
on.
2.
METHO
D
OL
OGY
Th
e pur
po
se
of
th
is stud
y h
a
s tw
o
m
a
in
step
s; f
i
rst
l
y
, t
o
ext
r
act
feat
ure
s
fr
om
bet
a
band
of
EEG
si
gnal
s
.
T
h
en
,
we t
r
y
t
o
det
ect
t
h
e l
e
ft
a
n
d
ri
ght
ha
nd
m
ovem
e
nt
s EEG
si
g
n
al
fi
ndi
ng
ef
f
ect
i
v
e cl
assi
fi
er.
We
u
s
ed
d
i
screte
wav
e
let tran
sfo
r
m
to
select b
e
ta b
a
nd
in
o
r
d
e
r to c
o
m
pute diffe
re
nt
features. These feat
ures are
appl
i
e
d
t
o
B
P
and
P
N
N
i
n
o
r
der
t
o
c
o
m
p
are
t
h
ei
r e
ffect
i
v
e
n
ess i
n
ha
nd
m
ovem
e
nt
cl
assi
fi
cat
i
on.
The
bl
oc
k
di
ag
ram
of t
h
i
s
p
r
o
p
o
sed
m
e
tho
d
o
l
o
gy
i
s
s
h
ow
n i
n
Fi
gu
re
1.
Fi
gu
re
1.
B
l
oc
k
di
ag
ram
for
han
d
m
ovem
e
nt
s EEG
si
g
n
al
s
cl
assi
fi
cat
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 1, Feb
r
uar
y
20
1
5
:
9
2
– 10
1
94
2.
1. Wa
vel
e
t
B
a
sed Dec
o
m
p
osi
t
i
o
n an
d
F
e
at
ures
E
x
tr
a
c
ti
on
The features are extracted
from bio-signals can cha
r
acteriz
e th
e beha
vi
o
r
s
of t
h
e co
rres
p
on
di
n
g
bi
o-
events
. Feat
ures extra
c
tion
f
r
om
t
h
e p
h
y
s
i
o
l
o
gi
cal
si
g
n
al
s i
s
i
m
port
a
nt
fo
r
vari
ous
a
p
pl
i
cat
i
ons.
T
h
e
r
e are
num
bers of
t
i
m
e
and fre
q
u
e
n
cy
o
r
bot
h d
o
m
a
i
n
feat
ures
are gene
ral
l
y
use
d
. A great
adva
nt
age
o
f
wavel
e
t
tran
sform
b
a
sed
an
alysis is th
at it
can
rep
r
esen
t th
e sig
n
a
l in
bo
th
ti
m
e
an
d
frequ
en
cy domain
wh
ich
is b
e
tter
com
p
ared t
o
the Fourier
t
r
ansf
o
r
m
and
sho
r
t
t
i
m
e Fouri
e
r
t
r
a
n
sf
or
m
.
Theref
ore,
we c
h
oos
e
wavel
e
t
tran
sform
e
d
sig
n
a
ls fo
r selectin
g
d
i
ff
eren
t EEG
b
a
n
d
s
signals wh
ich
will b
e
in
ten
d
e
d
fo
r features ex
tractio
n
.
Wav
e
let is a small wav
e
fo
rm o
f
effectiv
ely b
r
ief
d
u
rat
i
on t
h
at
has
a zer
o
net
area a
n
d
he
nce
of
feri
n
g
p
o
t
en
tiality
to
cap
ture ev
en
ts
o
ccurs in
a sh
ort p
e
ri
o
d
o
f
ti
me. Th
e wav
e
let tran
sfo
r
m
d
eco
m
p
o
s
es th
e sig
n
a
l
in
to
d
i
fferen
t scales with
d
i
fferen
t lev
e
ls o
f
reso
l
u
tio
n
b
y
d
ilatin
g
a
m
o
th
er wav
e
let. Fo
r
d
i
screte ti
m
e
s
i
g
n
a
ls,
d
i
screte
wav
e
l
e
t tran
sform
(DWT) is equ
i
v
a
len
t
to
an
octav
e
filter b
a
n
k
[2
1
]
, [22
]
.
Th
is
m
u
ltireso
l
u
tion
analysis (MRA)
decom
pose
s
a signal int
o
scales with
diffe
re
nt tim
e
and
fre
qu
enc
y
resolutio
n.
We ca
n
di
scri
m
i
nat
e
the di
f
f
ere
n
t
b
a
nd
s of EE
G si
gnal
t
h
r
o
ug
h
t
h
e decom
pos
i
t
i
on of M
R
A
i
n
t
o
di
ffe
rent
l
e
vel
s
.
Sel
ect
i
on o
f
m
a
xi
m
u
m
decom
posi
t
i
on l
e
ve
l
s
depe
nd
s o
n
t
h
e fre
q
u
ency
ban
d
s
.
The
r
ef
o
r
e, t
h
ro
u
gh
wa
vel
e
t
anal
y
s
i
s
, we c
a
n co
nse
r
ve t
h
e t
i
m
e
-freq
u
en
cy
com
ponent
s
of E
E
G si
gna
l
at
di
ffere
nt
r
e
sol
u
t
i
o
n a
nd
scal
es
.
Th
e reso
lu
tion
o
f
th
e si
g
n
a
l is d
e
term
in
ed
b
y
th
e filterin
g
o
p
e
ratio
ns and
th
e scale is d
e
term
in
ed
b
y
u
p
sam
p
l
i
ng and
do
w
n
sam
p
l
i
ng o
p
erat
i
o
ns
. As a res
u
l
t
,
t
h
e D
W
T ca
n be
achi
e
ve
d by
s
u
ccessi
ve l
o
w
pass an
d
h
i
gh
p
a
ss filter at d
i
screte time d
o
m
ain
is
sh
own
in
Fi
g
.
2
.
Wh
ere
x
[
n
]
i
s
t
h
e i
nput
si
gnal
,
w
h
i
c
h
passe
s
th
ro
ugh
a h
i
gh
p
a
ss filter o
f
i
m
p
u
l
se
response
h
[
n
] and
si
m
u
l
t
an
eo
usly p
a
sses th
rou
gh th
e l
o
w p
a
ss fi
lter wit
h
an im
pulse response
g
[
n
]. The o
u
t
p
u
t
of th
e h
i
gh
p
a
ss
filter p
r
ov
id
es t
h
e d
e
tail co
efficien
ts
D
an
d t
h
e
out
pu
t
o
f
t
h
e low
p
a
ss filter p
r
ov
id
es th
e app
r
ox
im
a
t
e co
efficien
ts
A
. The filter
o
u
tp
u
t
is g
i
v
e
n
in (1) an
d
(2
)
where
k
varies fr
om
-
∞
to
∞
[
2
2
]
.
k
n
h
k
x
Y
high
2
(1
)
k
n
g
k
x
Y
low
2
(2
)
Fi
gu
re
2.
W
a
v
e
l
e
t
decom
posi
t
i
on t
r
ee
u
p
t
o
l
e
vel
3
We ha
ve dec
o
m
posed t
h
e E
E
G si
g
n
al
i
n
b
o
t
h
t
i
m
e and fr
eque
ncy
d
o
m
a
i
n
usi
n
g D
W
T
t
o
sel
ect
t
h
e
EEG
ban
d
s
.
T
h
e dec
o
m
pose
d
si
g
n
al
s are t
h
en u
s
ed t
o
com
put
e
di
ffe
re
nt
f
eat
ures.
We
ha
ve u
s
ed
d
b4
w
a
vel
e
t
wh
ich
was
foun
d su
itab
l
e aft
e
r so
m
e
trial a
n
d error atte
m
p
t
s
u
p
t
o
dec
o
m
posi
t
i
on l
e
ve
l
fo
ur
t
o
si
x.
F
o
u
r
t
h
,
fifth
and
si
x
t
h lev
e
l co
effici
en
ts
o
f
wav
e
let d
eco
m
p
o
s
ition
co
rresp
ond
s th
e
b
e
ta, alpha and
th
eta
b
a
n
d
of
fre
que
nci
e
s res
p
ect
i
v
el
y
.
We have
u
s
ed
t
h
e beta
ba
nd for features
e
x
tracti
o
n.
Th
ere are nu
mb
ers
o
f
tim
e a
n
d
frequ
en
cy
d
o
m
ain
feat
u
r
es u
s
ed
i
n
literatu
res. Mo
st
o
f
th
e cases
m
a
gni
t
ude
, fr
eque
ncy
,
ene
r
gy
, po
we
r,
a
n
d
di
ffe
rent
st
at
i
s
t
i
cal
m
e
asures
are
c
o
n
s
i
d
ere
d
fo
r f
eat
ure
s
cal
cul
a
t
i
on.
A
f
t
e
r a ri
go
r
ous
st
udy
,
we
ha
v
e
co
nsi
d
e
r
ed
e
i
ght
e
ffect
i
v
e
f
eat
ures
w
h
i
c
h
are
descri
bed
bel
o
w.
Let,
N
is th
e len
g
t
h
o
f
th
e EEG sign
al
x
[
n
].
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Left a
n
d
Righ
t
H
and
Mo
vem
e
n
t
s EEG
S
i
gnals Cla
ssifica
tion
U
s
ing
Wa
vel
et Tran
sform
…
(
A
.B.M.A
. Hos
s
ain)
95
i.
Ener
gy
:
Ener
gy
o
f
a si
gnal
ca
n be
de
fi
ne
d as a si
m
p
l
e
sq
uare i
n
t
e
gral
.
It
m
i
ght
be an i
m
port
a
nt
feat
u
r
e o
f
EEG signal.
F
o
r discrete
EE
G sequence
the e
n
ergy ca
n
be c
a
lculated as:
N
n
n
x
Energy
1
2
]
[
(3
)
ii.
Scal
e vari
an
ce
(
SV
):
Scale va
riance
is m
easure of log-va
riance
that can
be e
x
pre
ssed as:
2
log
/
]
[
var
log
n
x
SV
(4
)
iii
. RMS v
a
lue
:
The root m
ean squa
re (RMS) is a statistica
l
measur
e o
f
va
r
y
i
ng si
g
n
al
s w
h
i
c
h can s
h
ow
t
h
e st
ren
g
t
h
of signal. T
h
e
avera
g
e
powe
r
of a signal is s
qua
re
of its RMS val
u
e. T
h
e
RMS val
u
e ca
n
be calculate
d as:
N
n
x
value
RMS
N
n
1
2
]
[
(5
)
iv
.
Ro
ll o
f
f
:
The r
o
l
l
o
f
f
i
s
a
m
easure of t
h
e fre
q
u
e
n
cy
bel
o
w w
h
i
c
h
85
% of t
h
e m
a
gni
t
u
de
di
st
ri
but
i
o
n o
f
t
h
e
sp
ectru
m
is co
n
cen
t
r
ated
. It i
s
also
a m
easu
r
e of sp
ectral shap
e an
d can be written as:
2
/
1
]
[
85
.
0
N
n
n
x
R
(6
)
v.
Vari
ance
:
V
a
r
i
a
n
c
e
i
s
t
h
e
m
e
a
n
v
a
l
u
e
o
f
t
h
e
s
q
u
a
r
e
o
f
t
h
e
d
e
v
i
a
t
i
o
n
o
f
t
h
e
s
i
g
n
a
l
.
H
o
w
e
v
e
r
,
m
e
a
n
o
f
EEG signal
is close
to
zero. Hence
,
we
calc
u
l
a
t
e
t
h
e
vari
a
n
ce
of
EE
G f
o
l
l
owi
n
g t
h
e e
q
u
a
t
i
on:
N
n
n
x
N
VAR
1
2
]
[
1
1
(7
)
vi
.
A
p
proxi
m
at
e ent
r
opy
(
A
pE
n)
:
Ap
pr
o
x
i
m
at
e ent
r
opy
i
s
a m
easure
o
f
ra
n
d
o
m
ness or re
g
u
l
a
rity [23
]
. A l
o
w valu
e
o
f
t
h
e
entropy indicat
es that
t
h
e t
i
m
e
seri
es
i
s
det
e
rm
i
n
i
s
ti
c;
on t
h
e ot
her
h
a
nd
, a hi
g
h
val
u
e i
ndi
cat
es t
h
e ran
dom
ness.
ApE
n
has
been
used
to characte
r
ize diffe
rent
biomedical events
. For
ApE
n
(
m
,
r
,
N
), t
w
o i
n
pu
t
param
e
t
e
rs, a ru
n l
e
ngt
h
m
,
and a
to
leran
ce windo
w
r
, m
u
st be considere
d
for
N
l
e
ngt
h t
i
m
e seri
es. I
f
N
dat
a
poi
nt
s f
o
rm
a tim
e seri
es
X
=
[
x
(1
),
x
(2
),
x
(
3
)
,
.
.
.
,
x
(
N
)]
.,
we ca
n c
o
m
put
e
ApEn
as fo
llows:
a) Fo
rm
a subseq
uence
of l
e
ngt
h
N-m
+1
co
n
s
isting
x
[1
],
x
[2]
,
…
..,
x
[
N-m
+
1
]
.
Sim
i
larly
,
differe
nt
subseque
nces a
r
e:
X
[
i
] =
[
x
(
i
),
x
(
i
+
1
), …,
x
(
i+
m
-1
)]
, whe
r
e
i
=1,2,….
,
n-
m
+1
b) Calculate the distance
d
[
X
(
i
),
X
(
j
)] bet
w
een
X
(
i
) an
d
X
(
j
), as t
h
e maxim
u
m
absolute differe
n
ce
bet
w
ee
n t
h
ei
r
r
e
spect
i
v
e sc
al
ar c
o
m
pone
nt
s:
)
1
(
)
1
(
max
=
)]
(
)
(
[
1,2,..m
=
k
k
j
x
k
i
x
j
,X
i
X
d
(8
)
c) F
o
r a
gi
ve
n
X
(
i
), c
o
unt
t
h
e
num
ber
of
j
, (
j
= 1,
2,
…,
N-m
+1) so t
h
at
d
[
X
(
i
),
X
(
j
)]
≤
r*S
T
D
. Whe
r
e,
ST
D
i
s
t
h
e st
a
nda
r
d
devi
at
i
o
n
of
se
que
nce
and
r
can b
e
var
y
in
g b
e
t
w
een
0
and
1
.
Th
e to
tal coun
tin
g nu
m
b
er
can be
de
not
e
d
as
θ
(
i
), whe
r
e
i
= 1, 2,..
,
N-m
+1. T
h
en
, de
fi
ne a pa
ram
e
t
e
r
C
r
m
(
i
) as bel
o
w an
d c
o
m
put
e i
t
for
each
i
.
C
r
m
(
i
) =
θ
(
i
)/(
N-m
+1
) (
9
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 1, Feb
r
uar
y
20
1
5
:
9
2
– 10
1
96
d) Com
pute the natural loga
rithm
of each
C
r
m
(
i
) and ave
r
age it ove
r
i
,
1
m
))
(
(
log
1
1
)
(
m
N
i
m
r
e
i
C
m
N
r
(1
0)
e) Inc
r
ease the
dim
e
nsion
m
to
m
+1. Re
peat
steps a
)
–d) to
get
C
r
m+
1
(
i
) an
d
φ
m
+1
(
r
).
f)
ApE
n
ca
n be
calculated by
t
h
e following form
ula:
ApE
n
(
m
,
r
,
N
) =
φ
m
(
r
) -
φ
m
+1
(
r
) (
1
1
)
We have
com
p
ut
e
Ap
E
n
t
a
ki
n
g
t
h
e
val
u
e
of
r
as 0.
15
.
v
ii.
Zer
o
C
r
oss
i
ng
(
ZC
):
Zero crossi
ng indicates
the
num
b
er of ti
mes th
e EE
G signal crosse
s the z
e
ro line. It
can be
calculated as:
Threshold
)
*
sgn(
1
1
1
1
n
n
N
n
n
n
x
x
x
x
ZC
(1
2)
whe
r
e,
Otherwise
0,
Threshold
,
1
)
sgn(
x
x
v
iii.
Mod
ified
Mean
Ab
so
lu
te Va
l
u
e
(
MMAV
):
Mean Abs
o
lut
e
Value (
MAV
) is the
m
oving
avera
g
e
of fu
ll
-wav
e rectified EEG.
It is calcu
lated
tak
i
ng
th
e
avera
g
e
of a
b
solute
value
of t
h
e EE
G s
i
gnal.
MM
A
V
i
s
th
e ex
ten
s
ion o
f
MA
V
with
ad
d
ition
a
l
wei
g
h
ting
fu
nct
i
o
n
w
[
n
] as show
n in
equatio
n
(
1
3
)
.
N
n
n
x
n
w
N
MMAV
1
]
[
]
[
1
(1
3)
whe
r
e,
w
[
n
] are taken as
1
for the sam
p
les
N
n
N
75
.
0
25
.
0
an
d 0.5 fo
r r
e
st
o
f
th
e sam
p
les.
2.
2. Cl
as
si
fi
ca
ti
on
usi
n
g Ne
ural
Netw
orks
Art
i
f
i
c
i
a
l
neu
r
al
net
w
o
r
ks
(
A
N
N
s
)
are si
m
i
l
a
r t
o
bi
ol
ogi
cal
neu
r
o
n
al
net
w
o
r
ks t
h
at
are usef
ul
fo
r
pattern rec
o
gni
tion, classi
fication et
c [20]. ANN
learni
ng
is accom
p
lished by
traini
ng al
gorithm
s
base
d
on the
l
earni
n
g
m
echani
s
m
s
of
bi
ol
ogi
cal
neu
r
o
n
s
.
There
are vari
ous
t
y
pes of ne
ural
net
w
or
k v
a
ry
i
ng fu
n
d
am
ent
a
l
l
y
in
th
e way th
ey learn
.
In
th
i
s
stu
d
y
, a feed
fo
rwar
d
b
a
ck
prop
ag
atio
n
n
e
ural n
e
twork
and
a prob
ab
ilistic
neu
r
al
net
w
or
k
are use
d
.
A fee
d
fo
rwa
r
d bac
k
pr
o
p
ag
at
i
on
neu
r
al
n
e
t
w
o
r
k c
o
nsi
s
t
s
of
a n
u
m
b
er
of si
m
p
l
e
ne
ur
o
n
s l
i
k
e
p
r
o
cessi
n
g
un
i
t
. Ev
ery un
it i
n
a layer is con
n
ected
with
all th
e un
its in
t
h
e
p
r
ev
iou
s
layer.
Alon
g wit
h
inpu
t
and
o
u
t
p
ut
l
a
y
e
rs o
f
ne
ur
o
n
s,
hi
d
d
en
pr
oces
si
ng l
a
y
e
r i
s
al
so use
d
t
o
h
o
l
d
t
h
e no
nl
i
n
ea
ri
t
y
and com
p
l
e
x
i
t
y
of
t
h
e pr
o
b
l
e
m
.
Sel
ect
i
on o
f
p
r
o
p
er
num
ber
of
hi
d
d
en l
a
y
e
rs i
s
im
port
a
nt
. The p
r
o
p
o
se
d fee
d
fo
r
w
ar
d bac
k
pr
o
p
agat
i
o
n
ne
ural
net
w
or
k s
t
ruct
u
r
e i
s
s
h
o
w
n i
n
Fi
g.
3.
We h
a
ve c
o
ns
i
d
ere
d
8 a
b
ov
em
ent
i
one
d fe
at
ures
.
There
f
ore,
ne
u
r
al
net
w
o
r
k i
s
desi
g
n
e
d
wi
t
h
ei
ght
i
n
p
u
t
n
o
d
es a
nd
o
n
e
o
u
t
p
ut
n
o
d
e. T
h
e num
ber
of
n
ode
s i
n
h
i
dd
en
layer
has b
een
set on tr
ial-
an
d-
err
o
r b
a
sis. Sig
m
oi
d t
r
ans
f
er
fu
nc
t
i
on has
b
een chosen
to deal
with
no
nl
i
n
ea
ri
t
y
. The
net
w
or
k i
s
t
r
ai
ne
d wi
t
h
g
r
a
d
i
e
nt
desce
n
t
a
l
go
ri
t
h
m
.
Perf
o
r
m
a
nce of t
h
e
net
w
or
k i
s
s
p
e
c
i
f
i
e
d
as M
ean
Sq
ua
r
e
Er
ro
r
(M
SE)
.
The
t
r
ai
ni
ng
was st
op
pe
d
w
h
en
t
h
e
M
S
E
b
e
t
w
een t
h
e
net
w
o
r
k
out
put
s a
n
d
t
h
e
t
a
rget
s
was l
e
s
s
er t
h
a
n
or e
q
u
a
l
t
o
0.
0
0
0
0
1
.
The l
ear
ni
n
g
r
a
t
e
i
s
fi
xed
at
0.
05
. T
h
e
num
ber
o
f
t
r
ai
ni
ng
epoc
hs
was fi
xe
d uni
f
o
rm
l
y
at
350
0.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Left a
n
d
Righ
t
H
and
Mo
vem
e
n
t
s EEG
S
i
gnals Cla
ssifica
tion
U
s
ing
Wa
vel
et Tran
sform
…
(
A
.B.M.A
. Hos
s
ain)
97
Fi
gu
re
3.
Pr
o
p
o
se
d
back
p
r
o
p
a
gat
i
o
n
ne
u
r
al
net
w
or
k st
ruct
ure
We h
a
v
e
u
s
ed
a p
r
o
b
a
b
ilistic n
e
ural n
e
t
w
ork to
find
a b
e
tter left and
righ
t
h
a
nd
m
o
v
e
m
e
n
t
s sign
als
classifier compari
ng to clas
sical
B
P
based N
N
. P
N
N h
a
s spee
dy
t
r
ai
ni
n
g
p
r
oce
ss
wi
t
h
i
n
nat
e
l
y
paral
l
e
l
structure a
n
d
one-pass
training e
f
fort
[24
]
. It
do
es no
t
n
e
ed
an
iterativ
e
t
r
aining
process
a
n
d the t
r
aining tim
e
is j
u
st th
e lo
ad
ing
tim
e o
f
th
e train
i
ng
m
a
trix
. Th
erefo
r
e
,
l
earni
ng
rat
e
of
PN
N i
s
fast
er t
h
a
n
m
a
ny
neu
r
al
n
e
two
r
k
s
m
o
d
e
ls, wh
ich in
crease its efficacy for
real tim
e
app
licatio
n
s
.
Th
e i
n
pu
t layer
p
a
rt
d
i
stribu
t
e
s th
e
in
pu
t to
th
e n
e
u
r
on
s in
t
h
e pattern
layer. Receiv
i
n
g
th
e
p
a
t
t
ern
fro
m
th
e i
n
pu
t layer, t
h
e
n
e
uro
n
of th
e
p
a
ttern
layer co
m
p
u
t
es its ou
tpu
t
in
acco
rd
an
ce to the prob
ab
ility d
e
n
s
ity fun
c
tion (p
df) fo
r a sing
le sam
p
le.
Th
e
PNN
arch
itectu
r
e
h
a
s t
h
ree layers: th
e in
pu
t layer, th
e
rad
i
al
b
a
sis lay
e
r, and
t
h
e
o
u
t
p
u
t
layer as
sho
w
n i
n
Fi
g.
4. T
h
e i
n
put
vect
o
r
p
re
presents the fe
ature
vector
with size 8×
1,
Q
is th
e num
b
e
r o
f
in
pu
t/targ
et
p
a
ir d
a
tasets an
d
K
is th
e nu
m
b
er
of classes. In th
is algorith
m
,
at th
e first layer in
pu
t vecto
r
s
calculate the distance from
the traini
ng i
n
p
u
t
an
d p
r
od
uc
es a ne
w vect
or
whi
c
h i
s
cl
ose
d
wi
t
h
t
h
e
t
r
ai
ni
n
g
i
n
p
u
t
.
T
h
e
n
, t
h
e radi
al
basi
s l
a
y
e
r sum
s
al
l
the c
ont
ri
b
u
t
i
o
n
s
o
f
t
h
e i
n
p
u
t
v
ect
or a
n
d p
r
od
uces a
n
out
put
vect
o
r
o
f
prob
ab
ilitie
s. In
rad
i
al b
a
sis layer, v
ect
o
r
d
i
stan
ces, ||
W
-
p
|
|
of di
m
e
nsi
o
n
Q
×1 are calculated
us
ing
dot
pr
o
duct
bet
w
e
e
n i
n
p
u
t
vect
or
p
and t
h
e eac
h r
o
w
of w
e
i
g
ht
m
a
t
r
i
x
W
wi
t
h
di
m
e
nsi
on of
Q
×8. The
n
,
the bias
vector
b
is com
b
in
ed
with ||
W
-
p
|
|
by
a
n
el
em
ent
-
by
-el
e
m
e
nt
m
u
l
t
i
p
l
i
c
at
i
on.
W
e
ha
v
e
use
d
ra
di
al
basi
s
t
r
ans
f
er fu
nct
i
on
a
s
,
2
=
rad
b
as
(n)
n
e
. The
sprea
d
val
u
e
of t
h
e ra
di
al
basi
s f
u
nct
i
o
n
was
used a
s
a
sm
oot
hi
ng
fac
t
or
whi
c
h
was
co
nsi
d
e
r
ed
as
0.
1.
The
n
,
t
h
e o
u
t
p
ut
o
f
ra
di
al
basi
s l
a
y
e
r i
s
f
o
un
d a
s
:
a
i
=
rad
b
as
(||
W
i
−
p
||.×
b
i
), w
h
ere
W
i
is th
e it
h
row
o
f
W
, and
b
i
is th
e
i
th
elem
e
n
t of
bias
vector
b
.
Fin
a
lly, in
th
e o
u
t
p
u
t
co
m
p
eti
tiv
e layer, th
e v
ector
a
is in
itially
m
u
ltip
lie
d
with
layer weig
h
t
m
a
trix
M
t
o
c
o
m
put
e an o
u
t
p
ut
vect
or
d
. There is
n
o
b
i
as in
th
is
layer. Th
e co
mp
etitiv
e fun
c
tion
is u
s
ed
to
p
i
ck
th
e
max
i
m
u
m
p
r
obab
ilities an
d
pro
d
u
ces ‘1
’ fo
r th
e in
ten
d
e
d
class and
‘0’
fo
r t
h
e
o
t
h
e
r class.
Fig
u
re
4
.
Probab
ilistic n
e
ural n
e
twork stru
ctu
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 1, Feb
r
uar
y
20
1
5
:
9
2
– 10
1
98
3.
RESULT ANALYSIS
We ha
ve use
d
expe
ri
m
e
nt
al
EEG dat
a
o
f
l
e
ft
and ri
g
h
t
h
a
nd m
ovem
e
nt. The dat
a
set
was col
l
ect
ed
fr
om
t
h
e web
s
i
t
e
[2
5]
. T
h
e s
u
bject
o
f
t
h
e
dat
a
set
i
s
a 2
1
y
e
ar
ol
d,
ri
g
h
t
ha
nde
d m
a
l
e
wi
t
h
no
k
n
o
w
n
m
e
di
cal
co
nd
itio
ns. The EEG con
s
ists o
f
actu
a
l
rando
m
m
o
v
e
m
e
n
t
s of left and
ri
g
h
t
h
a
n
d
reco
rd
ed with eyes
clo
s
ed.
There
are
nine
teen electrodes
.
T
h
e
or
der
of
t
h
e el
ect
ro
des
i
s
FP
1 F
P
2
F
3
F4
C
3
C
4
P
3
P4
O
1
O2
F
7
F8
T
3
T4 T
5
T6
FZ
C
Z
PZ. T
h
e E
E
G si
g
n
al
s wa
s acui
s
i
t
e
d usi
ng
Ne
ur
ofa
x
E
E
G Sy
st
em
and p
o
w
er l
i
n
e f
r
e
que
ncy
was 5
0
Hz
. Th
e l
e
ngt
h
of t
h
e
si
gnal
f
o
r eac
h el
ect
ro
de i
s
32
0
0
sam
p
l
e
s
and sam
p
l
e
rat
e
i
s
500
Hz. Fi
gu
re
5
sho
w
s t
h
e ra
w
EEG
dat
a
fo
r l
e
ft
and
ri
g
h
t
h
a
nd m
ovem
e
nt
. The
raw si
gn
al
s are t
h
en
de
com
posed i
n
d
i
ffere
nt
b
a
nd
s
u
s
ing
d
i
screte wavelet tran
sform
wh
ich
are sh
own
i
n
Fi
g. 6.
Sel
ect
i
on
of
decom
p
o
s
i
t
i
on l
e
vel
s
de
pen
d
s
o
n
alph
a,
b
e
ta and
th
eta
frequ
e
n
c
y b
a
nd
wh
ich
was foun
d in
five, fo
urth
and six
t
h
d
e
co
mp
o
s
ition
respect
i
v
el
y
.
M
o
st
o
f
t
h
e c
o
m
put
at
i
on an
d
pr
ocessi
ng
i
n
t
h
i
s
st
u
d
y
a
r
e
p
e
rf
orm
e
d i
n
M
A
TL
AB
[
2
6]
.
The
perform
a
nce of t
h
e classification
of the
EEG
data
is evalu
a
ted
in
term
s o
f
t
h
e three
p
a
ram
e
ters
i.e., sen
s
itiv
ity (SE), sp
ecificity (SP) an
d accu
r
acy.
Sen
s
itiv
ity (SE) ind
i
cates th
e
cap
acity o
f
corr
ectly id
en
tified
p
o
s
itiv
e cases and
d
e
fi
n
e
d as:
FN
TP
TP
SE
; TP=Tru
e Po
sitiv
e an
d FN =
False Neg
a
tiv
e.
Specifity (SP)
indicates the
c
a
pacity of correctly identified ne
gative case
s
and e
x
presse
d
as:
TN
FP
TN
SP
;
TN
= Tr
ue
Ne
gat
i
v
e a
n
d
FP
= False Po
sitive
Accuracy indic
a
tes that the
pro
portion of c
o
rrect classified
ev
ents
. It ca
n
be calculated as:
F
N
T
N
F
P
T
P
TN
TP
Accuracy
Fi
gu
re 5.
R
a
w EEG dat
a
fo
r
l
e
ft
an
d ri
g
h
t
ha
nd
m
ovem
e
nt
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Left a
n
d
Righ
t
H
and
Mo
vem
e
n
t
s EEG
S
i
gnals Cla
ssifica
tion
U
s
ing
Wa
vel
et Tran
sform
…
(
A
.B.M.A
. Hos
s
ain)
99
Fi
gu
re
6.
Se
par
a
t
e
d al
p
h
a
bet
a
, an
d t
h
et
a b
a
n
d
fr
om
EEG u
s
i
ng
D
W
T
We ha
ve use
d
8×5
7
l
e
ft
han
d
dat
a
set
s
and 8
×
57 ri
ght
ha
n
d
dat
a
set
s
whi
c
h were e
x
t
r
act
ed fr
om
bet
a
ban
d
f
o
r t
r
ai
ni
ng a
nd t
e
st
i
ng
dat
a
set
s
.
W
e
h
a
ve com
p
ared
t
h
e per
f
o
r
m
a
nce of B
P
-
NN a
nd P
NN cl
assi
fi
er t
o
fi
n
d
t
h
e bet
t
e
r
cl
assi
fi
er. Tabl
e 1 an
d 2 s
h
ow t
h
e
resul
t
of t
w
o cl
assi
fi
ers pe
rf
orm
a
nce used i
n
t
h
i
s
st
udy
.
Table
1 indicat
es the
num
b
er
of correct a
n
d
false cla
ssi
fi
cat
i
ons
o
f
l
e
ft
an
d
ri
g
h
t
han
d
m
ovem
e
nt
s EEG
si
gnal
usi
n
g B
P
ba
se
d
NN
an
d
PN
N.
Tabl
e
2
sh
ows
t
h
e
com
p
ari
s
o
n
of
o
v
e
r
al
l
per
f
o
r
m
a
nces by
c
o
nsi
d
e
r
i
n
g
sensitivity, spe
c
ificity and overall accu
racy. The sensitivity of BP and P
NN
has bee
n
found as 86.35% and
9
1
.74
%
resp
ectiv
ely. Th
e sp
ecificity o
f
BP
an
d PNN
has
been
f
o
un
d
as
98
.3
1%
an
d 1
0
0
% res
p
ect
i
v
el
y
.
The
ove
rall accura
cy of B
P
a
n
d PNN is calc
u
lated as
88
.9% and
99.1%. From
thes
e data and pe
rformance
param
e
t
e
r, i
t
can
be c
oncl
ude
d t
h
at
PN
N
ha
s bet
t
e
r cl
assi
fi
cat
i
on
rat
e
t
h
a
n
B
P
.
M
o
reo
v
e
r, t
h
e q
u
i
c
k t
r
ai
ni
n
g
p
r
o
cess with in
h
e
ren
tly p
a
ral
l
el stru
cture
o
f
PNN
will
in
ev
itab
l
y sp
eed
u
p
classifier
decisio
n
ex
ecu
tio
n ti
m
e
and im
prove
its effectivenes
s
fo real tim
e applications.
Tabl
e
1. C
l
assi
fi
cat
i
on R
e
s
u
l
t
s
o
f
Han
d
M
o
v
e
m
e
nt
EEG
Si
gnal
s
Han
d
BP
PN
N
No. of Classification
No. of Classification
Co
rrect
False
Co
rrect
False
Lef
t
4
8
9
5
6
1
Rig
h
t
5
3
4
5
7
0
Tab
l
e 2
.
Sen
s
it
iv
ity,
Sp
ecifici
ty,
and Acc
u
ra
cy of the Class
i
fiers
Sensitivity Specificity
Accurac
y
BP 86.
35%
91.
74%
88.
9%
PNN 98.
31%
100%
99.
1%
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 1, Feb
r
uar
y
20
1
5
:
9
2
– 10
1
10
0
4.
CO
NCL
USI
O
N
Th
e m
a
in
fo
cus of t
h
is stud
y
is to
d
e
tect left
an
d
right
hand E
E
G signal finding effective classifier.
The EEG
ban
d
s
ha
ve bee
n
t
a
ken
out
t
h
r
o
ug
h di
scret
e
wa
v
e
l
e
t
t
r
ansf
o
r
m
.
Ei
g
h
t
e
x
t
r
act
e
d
feat
u
r
es fr
o
m
bet
a
b
a
nd
h
a
v
e
b
e
en
used
t
o
train a classical b
a
ck
pro
p
a
g
a
tion b
a
sed
n
e
ural n
e
two
r
k
an
d
a p
r
ob
ab
ilistic n
e
ural
net
w
or
k i
n
or
d
e
r t
o
fi
nd ef
fe
ct
i
v
e cl
assi
fi
er for l
e
ft
and ri
ght
ha
n
d
m
o
v
e
m
e
nt
si
gnal
s
di
scri
m
i
nat
i
on.
Fi
ft
y
sev
e
n
lef
t
an
d r
i
g
h
t
h
a
nd
mo
v
e
m
e
n
t
EEG f
eatu
r
es
d
a
tasets w
e
r
e
used
in
th
is study. Th
e classif
i
catio
n
per
f
o
r
m
a
nce p
a
ram
e
t
e
rs sho
w
t
h
at
P
NN
(9
9.
1%)
has
bet
t
e
r classification
rate th
an
th
e
BP (88
.
9
%
).
Also
, th
e
fast train
i
n
g
p
r
o
cess
with innately p
a
rallel stru
ctur
e
o
f
PNN
will in
ev
itably sp
eed up classificatio
n
ti
me an
d
enha
nce i
t
s
us
eful
ness i
n
re
al
t
i
m
e
appl
i
cat
i
ons.
The
fi
ndi
ng
s o
f
t
h
i
s
st
udy
a
r
e e
x
pect
ed t
o
be
u
s
eful
i
n
artificial h
a
nd
m
o
v
e
m
e
n
t
s th
ro
ugh
b
r
ai
n
com
p
u
t
er in
terfacin
g
fo
r
b
i
o-rehab
ilitatio
n
app
l
icatio
n
s
.
ACKNOWLE
DGE
M
ENTS
The aut
h
o
r
s w
oul
d t
h
an
k t
h
e
aut
h
ori
t
y
of K
hul
na U
n
i
v
e
r
si
t
y
of Engi
neeri
ng & Tec
h
nol
ogy
,
KUE
T
,
Bangla
d
esh and the
Brain Com
puter Inte
rfa
ce rese
a
r
ch La
b at
NUST, Pa
kistan
for E
E
G data.
REFERE
NC
ES
[1]
M
.
A. Lebed
e
v
and M
.
A.L.
Nicole
lis
, “
B
rai
n
-M
achine Int
e
rfaces
: P
a
s
t
, P
r
es
ent and F
u
t
u
re”,
Trends in
Neurosciences
,
vol. 29
, pp
.536
—546, 2006
.
[2]
L.R. Hochb
e
rg,
et al.
, “Neuronal Ensemble Control of Prosthetic
Devices
b
y
a H
u
m
a
n with Tetra
p
legi
a”,
Nature
,
vol.
442, pp. 16
4–171, 2006
.
[3]
J.
R.
Wolpa
w
,
et al.
, “Brain-Co
m
puter Interface Technolog
y
:
A
Re
view of the First In
ternational Meeting”,
IE
EE
Transactions on
Rehabil
itation Engineering
, vol.
8, pp
.164—173,
2000.
[4]
T. Ebrah
i
m
i
, J.M. Vesin, and G.Garci
a
, “
B
rai
n
-com
puter inter
f
ace in Multim
e
d
ia Com
m
unication”,
IEEE Sign
al
Processing Mag
a
zine
, vo
l. 20, p
p
. 14-24
, 2003
.
[5]
S.
Sa
ne
i a
n
d J.
Cha
m
be
rs,
EEG S
i
gnal
Processing
, John Wiley
&
Sons, England
,
2
007.
[6]
G. Pfurtschel
ler
,
C. N
e
uper
,
A.
Sc
hlögl, and
K.
Lugger, “Separabilit
y
of
EEG Si
gnals Recorded
During Right
an
d
Left M
o
tor Im
ager
y Us
ing Adaptive Autor
e
gres
s
i
ve P
a
ra
m
e
ters
”,
IEEE Transactions
on
Rehabilitatio
n
Engineering
, vol. 6
,
pp
. 316-325
, 1998.
[7]
E. Has
e
ls
t
e
in
er
and G. P
f
urts
ch
el
ler, “Using Time-Depende
nt
Neural Network
s
for EEG Clas
sification”,
IEEE
Transactions on
Rehabil
itation Engineering
, vol.
8, pp
. 457-463
,
2000.
[8]
D. Najumnissa and T.R
.
Rangaswamy
,
“Detection and Classificati
on of Epilep
tic Seizures using
Wavelet featur
e
extra
c
tion
and Adaptive Neuro
-
F
u
zz
y
Inf
e
ren
c
e S
y
s
t
em
”,
In
ternational Journal of
Computational Engineerin
g
Research
, vol. 2
,
pp. 755-761, 20
12.
[9]
Y. Song, “A rev
i
ew of developm
ents of EEG-bas
ed au
tomatic medical support s
y
stem
s for epilep
s
y
diagnosis and
seizure
de
tec
tion
”
,
Journal o
f
Bio
m
edical Science
and Engin
eering
, vol. 4
,
pp
. 788-
796, 2011
.
[10]
M.
H.
Alomari,
A.
Samaha, and K.
Al K
a
m
h
a,
“
A
utom
ated Cla
s
s
i
ficatio
n
of
L/R Hand Movement
EEG Signals
using Advanced
Feature Extr
ac
ti
on and M
ach
ine
Learn
i
ng”
,
International Journ
a
l of Ad
vanced
Computer Scien
c
e
and Applications
, vol. 4
,
pp
. 207-
212, 2013
.
[11]
A. Schloeg
l
, C. Neuper
,
an
d
G. P
f
urts
chel
ler
,
“
S
ubject
S
p
ec
ific
EEG
P
a
tt
er
ns
During M
o
t
o
r Im
aginar
y”
,
in
Conference of the IEEE Engin
e
eri
ng in Medicine and Biolog
y
Society, 1997
.
IE
M
B
S
1997
.
N
i
neteenth Annua
l
International
, 19
97,
pp. 1530-15
32.
[12]
E.J
.
Cos
t
a and
E.F
.
Cabr
al J
r
., “
EEG-Bas
ed
Dis
c
ri
m
i
nation Between
Im
agi
n
ati
on of Left
and Right Han
d
Movements Using Adaptiv
e Gaussian Represen
tation
”
,
Med
i
cal
Engin
eer
ing
&
Phys
i
c
s
, vol. 2
2
, pp. 345-348
,
2000.
[13]
F
.
W
a
n
g
,
K
.
K
i
m
,
S
.
W
e
n
,
Y
.
Z
h
a
n
g
,
a
n
d
C
.
W
u
, “EEG
Based Automatic Left-Righ
t
Hand Movemen
t
Cla
ssifica
tion”
, in
Control and
Decision
Conferen
ce, 2012.
CCDC 2012
.I
, 2012
,
pp. 1469
– 1472
.
[14]
F. Quandt,
et al.
, “Single Trial
Discrimination
o
f
Individual Fin
g
er Movement
s on One Hand: A Combined MEG
and EEG Stud
y
”
,
NeuroImage
, v
o
l. 59
, pp
. 3316–
3324, 2012
.
[15]
M.
A.
Rihe
e
n
,
M.
W.
Ra
hma
n
, a
nd A.
B.
M.
A. Hossa
in,
“S
election of Proper
Frequenc
y
Ba
nd a
nd Compa
t
ible
Features for Left and Right H
a
nd M
ovement from EEG Si
gnal Analy
s
is”,
Conference on
Computer and
Information Technology, 2013. I
CCIT 2013,
Si
xt
eenth
In
ternatio
nal,
2013
.
[16]
H. Olkkonen
,
Disc
re
te
Wav
e
le
t Transforms
- Biomedical Applications
, InTech
, R
i
jeka, Croatia, 20
11.
[17]
A.
Nait-Ali,
Ad
vanced Biosignal Processing
, Springer, 2009
.
[18]
A. Subasi and
E.
E
r
ce
l
e
bi
,
“Cla
ssi
fi
cation of EEG Signals Using Neural
Network and Logis
tic Regr
ession”,
Computer Metho
d
s and Programs in Biomedicine
, pp
.87—99, 200
5.
[19]
S.J. Husain and K.S. Rao,
“
E
pilept
ic Sei
z
ures
Cla
ssification from EEG Signa
ls Us
ing
Neural Networks”, in
Conference on
I
n
formation and
Network Techno
logy
, 2012,
ICIN
T 2012
.
International,
2012
pp.2
69—273.
[20]
S
.
H
a
ykin
,
Neur
al Networks-A C
o
mprehensive Foundation
, Prentice Hall, NJ, 199
9.
[21]
S.G. Mallat
,
“
A
theor
y
for m
u
lt
iresolution si
gna
l decom
position:
the wavele
t rep
r
esenta
tion”
,
IEEE Transactions
on Pattern Ana
l
ysis and Ma
chin
e Intelligen
ce
, v
o
l. 11
, pp
. 674-6
92, 1989
.
[22]
A. Aldroubi and
M. Unser,
Wa
velets in
Med
i
cin
e
and Bio
l
ogy
. CR
C Press, USA, 1
996.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Left a
n
d
Righ
t
H
and
Mo
vem
e
n
t
s EEG
S
i
gnals Cla
ssifica
tion
U
s
ing
Wa
vel
et Tran
sform
…
(
A
.B.M.A
. Hos
s
ain)
10
1
[23]
H. Ocak,
“
A
utom
atic de
te
ction
of epil
epti
c sei
z
ures in
EEG using discrete wav
e
le
t tr
ansform and approximate
entrop
y”
,
Expert Systems with
Ap
plications
, vo
l. 3
6
, pp
. 2027–203
6, 2009
.
[24]
L. Rutkowski,
New Soft Computing Techn
i
q
u
es for System
Modelling, Pa
ttern Classification and Image
Proc
e
ssing
, Springer, 2004
[25]
Availab
l
e in:
htt
p
://sit
es.googl
e.com
/
site/proj
ectb
c
i/
[26]
The Ma
thW
o
rks Inc.
, M
a
ssachusetts,
USA. Avail
a
ble
in:
http
://w
ww.m
a
thworks.com
/
BIOGRAP
HI
ES OF
AUTH
ORS
Dr. A.B.M
.
Aowlad Hos
s
a
in re
ceiv
e
d his
B
.
S
c
.
in El
ec
tric
al
and
El
ectron
i
c
Engi
neering
(EE
E
) fr
om
Khulna University
of Eng
i
neer
ing & Technol
o
g
y
(KUET) and
M.Sc. in EEE from Bangladesh
University
of Engineer
ing & Technolog
y
(BUET) in 2002 and 2005 respectively
.
He join
ed
in
KUET as lecturer in 2005
. Dr. H
o
ssain completed
his Ph.D. in B
i
omedical Engin
e
ering from K
y
ung
Hee University
,
Korea in 2012. Currently
he is an
Associate Professor in the Department of EC
E,
KUET. His research in
ter
e
sts ar
e biomedical sig
n
al and
image p
r
ocessing, ultras
ound imaging,
and
computer aided
diagnisis etc. He
is member
of different pro
f
e
ssional societies and reviewer
s of
differen
t
confer
ences and
journ
a
ls.
Mr.
Md.
Wa
siur Ra
hma
n
re
c
e
ive
d
his B.
Sc
. in
Electronics
and
Communication Engineering
(
E
C
E
)
from Khulna Un
iversity
of Eng
i
neering &
Tech
nolog
y
(KUET)
in 2013. Currently
h
e
is prepar
in
g
for higher
studies and r
e
search
.
His resear
ch in
te
rests are biomed
ical signa
l
and image processing.
Mr. Manjurul
Ahsan Riheen r
eceived h
i
s B.S
c
.
in Electronics and Co
mmunication Eng
i
neer
in
g
(ECE) from Kh
ulna Univ
ersity
of Engin
eering
&
Technolog
y
(
KUET) in 2013
. Now he is wor
k
ing
in cell phone r
e
search
and
deve
lopment d
i
vis
i
on of Walton
Hi-T
ech Industr
ies Limited as an
as
s
i
s
t
ant engin
e
er. His
res
ear
ch
interes
t
s
ar
e bi
om
edical s
i
gn
al
proces
s
i
ng, s
i
g
n
al proc
es
s
i
ng for
communications
and sensor
netw
orks.
Evaluation Warning : The document was created with Spire.PDF for Python.