Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4265
~
4271
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp
4265
-
42
71
4265
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Voice A
s
sess
m
ent
s for
D
et
ec
tin
g
P
atie
nt
s with
Parkin
son’s
D
iseas
es in
D
iffer
ent
S
tag
es
El
mehdi B
en
malek
,
Jam
al
E
lm
ha
mdi
,
A
bdeli
lah
J
il
bab
Depa
rtment
o
f
E
le
c
tri
c
al
Engi
n
eering,
ENSET
R
a
bat
,
Moham
ed
V
Univer
sit
y
,
Mor
occ
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Feb
20
, 201
8
Re
vised
Ju
l
6
,
201
8
Accepte
d
Aug
7
, 2
01
8
Rec
en
tly
,
a
wide
ran
ge
of
spee
ch
signal
proc
essing
al
gorit
hm
s
(d
y
sphoni
a
m
ea
sures)
ai
m
in
g
to
de
tect
p
at
i
ent
s
with
Park
i
nson’s
disea
se
(
PD
).
So
we
have
computed
19
d
y
sphoni
a
m
ea
sures
from
sus
ta
in
ed
vowels
co
ll
e
ct
ed
from
375
voic
e
sam
ple
s
from
hea
l
th
y
and
pe
opl
e
suffe
r
from
PD
.
All
t
he
fe
at
ure
s
are
an
aly
s
ed
an
d
the
m
ore
releva
nt
ones
ar
e
sele
cted
b
y
th
e
Princi
pa
l
component
ana
l
y
sis
(PCA
)
to
cl
assif
y
th
e
subjects
in
4
cl
asses
ac
cor
ding
t
o
the
UP
DRS
(un
ifi
ed
Parkinson’
s
disea
se
Rat
ing
Scal
e)
score
.
W
e
used
k
-
fo
lds
cro
ss
valid
at
ion
m
e
thod
wit
h
(k=
4)
v
al
id
at
io
n
sche
m
e;
75%
f
or
training
and
25%
for
te
st
ing,
al
ong
wi
th
t
he
Support
Ve
ctor
Mac
hin
es
(SV
M)
with
it
s
diffe
ren
t
t
y
p
es
of
ker
ne
ls.
Th
e
be
st
result
obt
ai
n
e
d
was
92.
5%
usi
ng
the
PC
A
and
th
e li
ne
ar
S
VM
.
Ke
yw
or
d:
Cl
assifi
cat
ion
Par
kin
s
on’s dis
ease
PCA
SV
M
Vo
ic
e
featu
res
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
:
El
m
ehd
i B
enm
al
ek
,
Dep
a
rtm
ent o
f El
ect
rical
En
gi
neer
i
ng,
ENS
E
T Rabat,
Moham
ed
V Un
i
ver
sit
y,
Ra
bat, Mo
r
occ
o
.
Em
a
il
: elm
ehd
i.be
nm
al
ek@
um
5s
.n
et
.m
a
1.
INTROD
U
CTION
Par
kin
s
on’s dis
ease (P
D)
is a
neur
od
e
ge
ner
a
ti
ve
disor
der
th
at
r
esults f
ro
m
the d
eat
h
of
dopam
inerg
ic
cel
ls
in
the
substanti
a
nigra
wh
ic
h
is
a
bas
al
ganglia
struc
ture
locat
e
d
in
the
m
idb
rain
.
Su
c
h
ne
urol
ogic
al
diseases
pr
ofo
undly
aff
ect
th
e
patie
nts’
qu
al
it
y
of
li
fe
and
t
heir
fam
il
ie
s
[1
]
.
Ag
e
is
on
e
of
the
m
os
t
i
m
p
or
ta
nt
risk
facto
r
w
hi
ch
ex
plain
t
hat
PD
is
gen
e
ral
ly
seen
in
pe
ople
o
ver
t
he
a
ge
of
50.
Dia
gnos
is
of
P
D
is
ver
y
diff
ic
ult
we
use
neu
r
ologica
l
te
sts
and
br
ai
n
scans
to
diag
nose
it
.
These
m
et
hods
are
ve
r
y
exp
ensi
ve
an
d
nee
d
high level
of e
xp
e
rtise
.
Since
m
os
t
of
the
people
with
PD
suffe
r
fro
m
sp
eech
dis
order
s
[
2]
,
[
3],
it
co
ul
d
be
c
onsidere
d
as
th
e
m
os
t
reaso
nab
l
e
way
for
dete
ct
ion
of
PD
[4]
.
The
range
of
sy
m
pto
m
s
pr
esent
in
sp
eec
h
dis
order
s
i
nc
lud
es
reduce
d
lo
udne
ss,
inc
reased
vo
cal
trem
or
,
and
breat
hin
es
s.
V
ocal
dis
order
s
do
not
ap
pear
a
bru
ptly
,
they
ar
e
the
res
ult
of
a
slo
w
e
voluti
on
w
ho
se
early
s
ta
ges
m
ay
be
unnoti
ced.
V
oice
assessm
ents
has
pro
ven
to
be
a
n
eff
ect
ive
t
oo
l f
or
PD
d
et
ect
io
n,
f
or
this p
urp
os
e,
t
he
proces
sing
of
the qua
li
ty
of
sp
eec
h,
an
d
the
ide
ntifi
cat
ion
of
the
ca
us
es
of
it
s
degrada
ti
on
in
the
c
onte
xt
of
P
D
bas
ed
on
phon
ologica
l
and
ac
ousti
c
cues
ha
ve
bec
om
e
on
e
of t
he
m
ain
inte
rest of cli
nicia
ns
a
nd s
pe
ech
path
ologist
s.
Am
on
g
the
m
os
t
interest
in
g
recent
wor
ks
are
tho
se
co
ncerne
d
with
cl
ass
of
ne
uro
deg
e
ne
rati
ve
diseases
su
c
h
as
PD
,
m
ulti
pl
e
scl
ero
sis
am
ong
o
t
her
,
that
aff
ect
m
oto
r,
cogniti
ve
capa
bili
ti
es,
and
pa
ti
ent's
sp
eec
h
[5
]
,
[
6]
.
Ther
e
are
r
ecent
stud
ie
s
us
in
g
m
achine
le
arn
in
g
too
ls
su
ch
as
Sup
port
Vect
or
Ma
chine
(S
VM
)
cl
assifi
er,
Ga
us
sia
n
ra
dial
basis
ke
rnel
fu
nc
ti
on
s
,
re
gr
essi
on,
ne
ura
l
networks,
D
Mne
ural
an
d
de
ci
sion
tree
[
7]
,
[
8],
a
nd
aco
us
ti
c
m
easur
em
ents
(
f
eat
ur
es
)
of
dys
phonia
f
or
the
detect
ion
of
voic
e
dis
orde
rs,
these
include
f
unda
m
ental
fr
eq
ue
nc
y
or
pitc
h
of
vo
cal
os
ci
ll
at
i
on
(F0);
Jit
te
r
wh
ic
h
is
the
c
yc
le
-
to
-
cy
cl
e
va
riat
ion
of
fun
dam
ental
fr
e
qu
e
ncy;
Shim
m
er
that
represents
t
he
ext
ent
of
va
riat
ion
in
s
peec
h
am
plit
ud
e
from
c
yc
le
to
cy
cl
e;
m
easur
es
of
noise
-
to
-
ha
rm
on
ic
s
rati
o
com
ponen
ts
in
t
he
vo
ic
e;
t
he
N
on
li
near
dy
nam
ic
al
com
plexity
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4265
-
4271
4266
and
f
undam
ental
fr
eq
ue
ncy
va
riat
ion
a
nd
Sign
al
fr
act
al
sc
al
in
g
e
xponent
[1
]
,
[4]
,
[9
]
.
Stud
ie
s
ha
ve
s
how
n
var
ia
ti
ons
in
al
l
these
m
easurem
ents
in
peop
le
with
PD
[
10
]
.
All
these
stud
ie
s
has
been
perform
ed
fo
r
bin
a
ry
cl
assifi
cat
ion
,
so
f
or
a
n
ea
rly
diag
no
sis
of
PD
,
m
ulti
cl
ass
cl
assifi
cat
ion
base
d
on
se
veri
ty
of
sy
m
pto
m
s
has
been
ac
hieve
d
with
di
ff
e
ren
t
cl
assifi
ers
us
i
ng
the
L
ocal
Le
arn
i
ng
-
Ba
sed
Feat
ur
e
Sele
ct
ion
featu
re
sel
e
ct
ion
al
gorithm
and
t
he
ce
ps
tral
a
na
ly
sis [11]
,
[
12]
,
In
t
his
stu
dy,
we
wan
t
t
o
disti
nguish
PD
patie
nts
on
dif
fer
e
nt
sta
ges
of
sym
pto
m
s’
sever
it
y
f
ro
m
healt
hy
c
on
tr
ol
us
i
ng
these
ac
ou
sti
c
m
easur
e
m
ents.
S
o
we
a
i
m
ed
to d
isc
ri
m
inate
375
s
ub
j
ect
s o
n
4 groups; 55
healt
hy
co
ntr
ol,
178
in
ea
rly
118
in
interm
ediat
e
an
d
24
in
adv
a
nce
d
sta
ge
accor
ding
t
o
the
UPDRS
s
c
or
es
.
Each
pa
rtic
ipant
wa
s
in
vited
to
pr
onou
nce
the
s
us
ta
ine
d
vowel
/a
/
a
nd
hold
it
at
c
om
fortable
le
vel,
f
r
om
each
vo
ic
e
sam
ple
we
ha
ve
extra
c
te
d
19
ac
ousti
c
featur
es
,
to
re
du
ce
t
he
num
ber
of
t
hese
aco
us
ti
c
featur
e
s
and
get
on
ly
the
m
os
t
releva
nt
ones,
we
ap
plied
t
he
pr
i
ncipal
c
ompone
nt
analy
sis,
a
nd
for
cl
as
sific
at
ion
we
use
d
k
-
fo
l
ds
c
ro
ss
validat
ion m
et
ho
d al
ong wit
h
th
e
SV
M cl
assifi
e
r
s w
it
h i
ts dif
fere
nt k
e
r
nels.
2.
RESEA
R
CH MET
HO
D
2.1.
Dataset
The
dataset
c
ol
le
ct
ed
in
this
stud
y
belo
ng
t
o
T
he
Pati
ent
Vo
ic
e
A
naly
sis
(PV
A)
datas
et
[
8
],
[
13
]
,
it
con
ta
in
s
voic
e
recordi
ngs
of
voic
e
ph
on
at
io
ns
sel
f
-
repo
rted
sy
m
pto
m
asses
sm
ent
PD
RS
(
Par
kin
s
on’s
Di
sease
Ra
ti
ng
Scal
e)
and
dem
og
ra
phic
inf
or
m
at
ion
ab
ou
t
the
cal
le
rs.
Each
row
in
the
dataset
cor
r
esp
onde
d
to
one
repor
t
from
a
Par
kin
s
on’s
pa
ti
ent
and
the
dy
sp
honia
m
easur
em
ents
are
r
epr
ese
nted
in
t
he
c
olu
m
ns
.
T
her
e
are
375
us
ers
t
ota
l
(r
epeate
d
a
nd
use
le
ss
rec
ords
a
re
rem
oved).
All
pa
rtic
ipants
wer
e
a
sk
e
d
to
rec
or
d
the
su
sta
ine
d
vow
el
“a”
ho
ld
as
lon
g
as
po
s
sible
at
a
co
m
fo
rta
ble
le
vel.
They
al
so
pro
vid
e
d
the
f
ollow
i
ng
inf
or
m
at
ion
;
age,
ge
nder,
a
ge
of
diag
nosis,
ye
ars
si
nce
fi
rst
sym
pto
m
,
if
they
are
on
t
reatm
ent
or
not,
with
(m
ean
62
.
17
y
ears
old,
m
axi
m
u
m
84
and
m
ini
m
u
m
34
,
sta
nd
a
rd
dev
ia
ti
on
:
8.370
254,
var
ia
nce:
69.
88011,
popula
r
sta
nd
a
rd d
e
viati
on
:
8.359
432,
va
rian
ce p
opular:
67.
9286)
.
Am
on
g
375
pe
rsons
for
wh
i
ch
the
data
we
re
rec
orde
d,
w
e
cl
assify
55
s
ubj
ect
s
as
heal
thy,
178
in
early
sta
ge,
11
8
in
inte
rm
edi
at
e
sta
ge,
an
d
24
as
a
dva
nce
d
sta
ge
base
d
on
UPDRS
sc
or
es
.
V
oice
re
cordin
gs
and
the
pr
e
-
processin
g
a
re
n
ot
suffici
e
nt
in
the
asses
sm
e
nt
of
vo
ic
e
dis
orders.
T
he
refor
e
,
it
is
esse
nt
ia
l
to
dev
ise
an
d
des
cribe
voic
e
sa
m
ples
us
ing
a
set
of
ac
ousti
c
featur
e
s,
w
hich
a
re
represe
nted
as
a
featu
re
vec
t
or
us
e
d
f
or s
peec
h
a
naly
sis.
2.2.
Feature e
xt
r
ac
tion
In
this
dataset
,
19
li
nea
r
an
d
non
-
li
near
featu
res
we
re
extrac
te
d.
Table
1
co
ntains
al
l
the
f
eat
ur
es
a
nd
a
br
ie
f
desc
ript
ion
s
[
14
]
.
16
f
eat
ur
es
a
re
bas
ed
on
f
our
factor
s:
F
0
(
f
unda
m
ental
fr
eq
ue
nc
y
or
pitc
h)
,
se
ver
al
m
easur
es
of
va
riat
ion
in
f
undam
ental
fr
equ
ency
an
d
am
plit
ud
e
an
d
m
easur
e
s
of
rati
o
of
no
ise
to
ton
al
com
po
ne
nts in
the voice
, t
hes
e m
easur
e
m
ents are
t
he
m
os
t i
m
po
rtant f
act
ors
of
t
he v
oice
sign
al
.
Table
1.
Feat
ur
es
Ext
racted
Featu
re
n
u
m
b
er
Featu
res
Descripti
o
n
1
MDVP:
Fo (H
z)
Av
erage vo
cal fu
n
d
a
m
en
tal
f
requ
en
cy
2
MDVP:
Fhi (
Hz)
Maxi
m
u
m
vo
cal fu
n
d
a
m
en
tal
f
requ
en
cy
3
MDVP:
Flo (
Hz)
Mini
m
u
m
vo
cal
f
u
n
d
a
m
en
tal
f
requ
en
cy
4
Jitter (
%
)
Sev
eral
m
easu
res
o
f
variatio
n
in f
u
n
d
a
m
en
tal f
requ
en
cy
5
Jitter (
Ab
s)
6
MDVP:
RAP
7
MDVP:
PPQ
8
Jitter: D
DP
9
Sh
i
m
m
e
r
Sev
eral
m
easu
res
o
f
variatio
n
in a
m
p
litu
d
e
10
Sh
i
m
m
e
r
(dB
)
11
Sh
i
m
m
e
r:
APQ3
12
Sh
i
m
m
e
r:
APQ5
13
MDVP:
APQ
14
Sh
i
m
m
e
r:
DDA
15
NHR
Two
m
e
asu
res of
r
atio
of
no
ise to
tonal
co
m
p
o
n
en
ts in
th
e vo
ice
16
HNR
17
RPDE
No
n
lin
ear
d
y
n
a
m
i
cal
co
m
p
lex
it
y
m
e
asu
res
18
DFA
Sig
n
al f
racta
l scali
n
g
exp
o
n
en
t
19
PPE
No
n
lin
ear m
e
asu
re
of
f
u
n
d
a
m
en
tal f
r
eq
u
en
cy
variatio
n
Jit
te
r
(%)
:
Expresse
d
as
a
per
centa
ge,
this
is
the
aver
ag
e
abs
olu
te
differ
ence
betwee
n
consecuti
ve
per
i
od
s
of
fun
dam
ental
f
re
qu
e
ncy,
div
ide
d b
y t
he
ave
ra
ge p
erio
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Voice Asse
ssme
nts for
D
et
ect
ing
Patie
nts
wi
th P
ar
ki
nso
n’
s
….
(
Elmeh
di Be
nmalek)
4267
(
%
)
=
1
∑
|
−
−
1
|
−
1
=
1
1
∑
=
1
(1)
Wh
e
re
is
the
per
i
od
of
f
undam
ental
fr
e
qu
encies
of
window
num
ber
“
i”
an
d
N
is
t
he
total
nu
m
be
r
of
windows.
Jit
te
r
(
ABS)
:
Jit
te
r
abs
ol
ute
is
the
cy
cl
e
-
to
-
cy
cl
e
var
ia
ti
on
of
fun
dam
ental
fr
e
qu
e
ncy,
i.
e.
the
aver
a
ge
a
bsolu
te
d
iffe
re
nce
be
tween c
onsec
utive
per
i
od
s
, e
x
presse
d
as:
(
)
=
1
−
1
∑
|
−
−
1
|
−
1
=
1
(2)
Wh
e
re
is
the
e
xtracted
F
0
pe
r
iod
le
ngths
,
a
nd
N
are
is
the
n
um
ber
of
extr
act
ed
F
0
per
io
ds
.
Jit
te
r
(RA
P
):
it
is
def
ine
d
as
the
Re
la
ti
ve
A
ver
a
ge
Pe
rturb
at
ion
,
the
a
verage
abs
olu
te
di
ff
ere
nc
e
bet
w
een
a
pe
rio
d
and
t
he
aver
a
ge of
it
a
nd it
s tw
o neig
hbours
, divide
d by the a
ve
rage pe
rio
d.
Jit
te
r
(P
P
Q)
re
pr
ese
nts
t
he
Pe
rio
d
Pe
rtu
rb
at
i
on
Q
uoti
ent,
de
fine
d
as
the
a
ver
a
ge
a
bsolut
e
dif
fer
e
nc
e
betwee
n
a
pe
ri
od
a
nd
the
a
ve
rag
e
of
it
an
d
i
ts
four
cl
ose
s
t
neig
hbors,
div
i
ded
by
the
a
ve
rag
e
per
i
od
[
15
],
[
16
].
Sh
im
m
er:
This
is
the
aver
age
abso
lute
diff
e
ren
ce
betwee
n
the
a
m
plit
ud
es
of
co
ns
ec
utiv
e
per
io
ds
,
div
i
ded
by
the av
e
ra
ge
am
plit
ud
e
ℎ
=
1
−
1
∑
|
−
−
1
|
−
1
=
1
1
∑
=
1
(3)
Sh
im
m
er
(
APQ5):
It
is
def
i
ned
as
t
he
fiv
e
-
point
Am
plitu
de
Pe
rtu
rb
at
i
on
Q
uoti
ent,
t
he
a
ver
a
ge
abs
olu
te
dif
fere
nce
bet
ween
the
am
pli
tud
e
of
a
pe
rio
d
an
d
the
ave
rag
e
of
the
am
plit
ud
es
of
it
and
it
s
fou
r
cl
os
est
neig
hb
ours,
div
i
ded
by
the
ave
rage
a
m
plit
ud
e.
HN
R:
Ha
rm
o
ni
cs
to
No
ise
Ra
t
io
,
NH
R:
No
ise
to
Har
m
on
ic
s Rat
io.
Re
currence
Pe
rio
dicit
y
Den
sit
y
Entro
py
(R
PD
E
)
is
based
on
the
noti
on
of
rec
urren
ce
[
17
]
,
w
hich
can
be
seen
as
a
gen
erali
zat
ion
of
pe
rio
dicit
y
[
18
]
.
This
m
easur
e
ad
dr
e
sses
the
abili
ty
of
the
vocal
fo
lds
t
o
su
sta
in
sta
ble
vo
cal
f
old
osc
il
la
ti
on
,
qu
a
ntif
yi
ng
the
de
viati
on
s
f
ro
m
exact
per
iod
ic
it
y.
Pit
ch
Period
E
nt
ropy
(P
PE
)
m
easur
e
s
the
im
paired
con
t
ro
l
of
sta
bl
e
pitch
durin
g
su
sta
ine
d
phonat
ions
[
1],
a
s
ym
pto
m
com
mo
n
to
people
with
P
D
[
19
].
Detre
nded
Fluctuat
i
on
A
naly
sis
(
D
FA
)
is
a
scal
in
g
analy
sis
m
eth
od
us
e
d
to
quantify
long
ra
nge
power
-
la
w
a
uto
c
orrelat
ion
s
i
n
sign
al
s
w
hich
are
no
n
-
sta
ti
on
ary,
th
us
ov
e
r
com
ing
so
m
e
of
t
he
pro
blem
s o
f
sc
al
ing
a
naly
sis t
echn
i
qu
e
s
wh
i
ch
a
re
on
ly
s
uitable
for st
at
iona
ry sig
nals [
18
],
[
20
].
2.3.
Feature sel
ec
t
ion
an
d
vali
dation
In
m
os
t
sit
uatio
ns,
we
fin
d
oursel
ves
with
a
nu
m
ber
of
va
r
ia
bles
wh
ic
h
te
nd
s
t
o
excee
d
the
num
ber
of
ob
se
r
vations.
Dim
ension
a
li
ty
red
uctio
n
process
procee
ds
by
ap
plyi
ng
a
feature
sel
ect
ion
al
gorith
m
.
In
order
t
o
ha
ve
a
bette
r
representat
io
n
of
the
data,
re
dundant
a
nd
us
el
ess
in
for
m
a
ti
on
will
be
thu
s
ci
rcu
m
ven
te
d.
The pri
ncipal
obj
ect
iv
es
of
t
he
r
ed
uctio
n of di
m
ension
can
be
d
esc
ribe
d by
[
21
]
. So t
o
im
p
rove
the
ta
sk
of
cl
assifi
cat
ion
an
d
to
ai
d
the
visua
li
zat
ion
an
d
t
he
com
pr
ehe
nsi
on
of
the
da
ta
,
we
ha
ve
to
id
entify
the
m
or
e
relev
ant
feat
ur
es
in
order
to
re
du
ce
the
sto
ra
ge
of
sp
ace
necessa
r
y,
m
ini
m
iz
e
tim
e
con
s
um
pti
on
a
nd
CPU
-
e
xpen
ditur
e
.
Howe
ver,
the
el
i
m
inati
on
of
certai
n
in
for
m
at
ion
can
in
crease
the
cl
as
sific
at
i
on
er
r
or,
co
ns
ide
rin
g
this
inform
at
io
n
can
pro
ve
to
be
inf
orm
ati
ve
if
they
are
us
ed
[
22
]
.
In
this
stud
y
we
us
ed
the
Pr
i
nc
ipal
Com
po
ne
nt
A
naly
sis
(P
C
A)
,
w
hich
c
onsid
ered
t
he
m
or
e
recog
nized
li
near
te
c
hniq
ue
for
dim
ension
al
it
y
reducti
on, th
e
PCA p
e
rfor
m
s a l
inear m
app
ing
of
the
data to a lo
we
r
-
dim
e
ns
io
nal sp
ace i
n
su
c
h
a w
ay
that the
var
ia
nce
of
the
data
in
the
lo
w
-
dim
ension
al
representat
io
n
is
m
axi
m
iz
ed.
Pr
evi
ous
sp
ee
ch
analy
sis
has
sho
w
n
sat
isfact
or
y
res
ults usi
ng this
r
edu
ce
d
im
ension
al
it
y m
et
ho
d [
23
]
.
Af
te
r
extracti
ng
al
l
featu
res
a
nd
sel
ect
ing
t
he
m
or
e
releva
nt
on
es
,
we
cl
as
sify
voic
e
sam
ples
ba
sed
on
t
hese
feature
s
into
four
gr
oups
;
Healt
hy
cases,
pe
ople
with
P
D
in
ea
r
ly
,
interm
ediate
and
ad
va
nce
d
sta
ges
.
Subseque
ntly
,
we
buil
t
a
m
atr
ix
ba
sed
on
t
he
se
p
aram
et
ers.
The
c
olu
m
ns
of
the
m
at
rix
r
epr
ese
nt
the
fe
at
ur
es
and
t
he
r
ow
s
r
epr
ese
nt
the
voic
e
sam
ples.
In
this
stu
dy,
w
e
us
ed
k
-
f
old
s
cro
ss
validat
io
n
m
et
ho
d
with
(k
=
4)
al
ong
with
di
f
fer
e
nt
kernel
of
the
S
VM
cl
assifi
er;
Train
ing
an
d
te
sti
ng
proce
dures
are
a
ppli
ed:
75%
f
or
trai
ning
an
d
25%
f
or
te
sti
ng.
The
dataset
is
div
ide
d
i
nto
4
subsets,
eac
h
t
i
m
e,
on
e
of
th
e
4
subsets
is
use
d
as
the
te
st
set
and
the
ot
her
3
s
ub
s
et
s
are
put
tog
et
he
r
to
f
orm
a
trai
nin
g
set
.
The
n
the
aver
a
ge
er
ror
acro
s
s
al
l
4
tria
ls
is
com
p
uted.
T
he
adv
a
ntage
of
this
m
et
ho
d
is
that
it
m
at
te
rs
l
ess
how
the
da
ta
gets
div
ided
,
ever
y
data point
gets
to b
e
in
a
test
s
et
ex
act
ly
once
, and gets t
o be
in
a t
rainin
g s
et
3
ti
m
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4265
-
4271
4268
3.
RESU
LT
S
A
ND AN
ALYSIS
3.1.
Obtained
resu
lts usin
g
li
ne
ar
kernel
The
T
a
ble
2
re
pr
ese
nt
the
obt
ai
ned
res
ults
of
cl
assifi
cat
ion
us
in
g
t
he
li
ne
ar
S
VM,
a
nd
s
el
ect
ing
th
e
m
or
e
relevan
t
featur
e
s
by
the
PCA
m
et
ho
d,
with
92.5
%
overall
accu
rac
y.
Fo
r
each
cl
a
ss
we
ha
ve
the
ROC
curve.
In this
m
od
el
w
e h
a
ve
for
:
1.
The
healt
hy
co
ntr
ol:
we
ha
ve
49
wer
e
c
orre
ct
ly
cl
assifi
ed,
6
we
re
m
isc
lassified
a
nd
c
on
sidere
d
as
earl
y
sta
ge,
with a
percenta
ge of
89% tr
ue p
os
it
iv
e rate;
2.
The
ea
rly
sta
ge
cl
ass:
we
ha
ve
171
we
re
c
orrectl
y
cl
assifi
ed,
7
were
m
isclassifie
d
(
2
a
s
healt
hy,
a
nd
5
as
interm
ediat
e st
age),
with a
pe
rcen
ta
ge of
t
rue p
os
it
ive r
at
e
96%;
3.
The
interm
ediat
e
sta
ge
cl
ass:
we
hav
e
11
3
wer
e
co
rrec
tl
y
cl
assifi
ed,
5
w
ere
m
isc
la
ssifi
ed
(4
as
in
ear
ly
sta
ge,
a
nd
1
as
adv
a
nce
d
sta
ge
)
wit
h
a
per
ce
nt
age of 9
6%
tr
ue posi
ti
ve
rate
;
4.
The
a
dv
a
nce
d
sta
ge
cl
ass:
we
ha
ve
14 w
e
re co
r
rectl
y
cl
assifi
ed,
10 w
ere m
isc
la
ssifie
d
(
as
in
inte
rm
edi
at
e
sta
ge)
,
w
it
h a
pe
rcen
ta
ge of
58% tr
ue p
os
it
iv
e rate.
Table
2
.
Re
s
ults
Usi
ng Linea
r
SV
M
Confus
i
on m
atr
ix
ROC cu
r
ve
Cl
ass 1
Cl
ass 2
Cl
ass 3
Cl
ass 4
3.2.
Obtained
resu
lts usin
g
q
uad
rat
ic
kernel
The
T
able
3
re
pr
ese
nt
t
he
ob
t
ai
ned
res
ults
of
cl
assifi
cat
io
n
us
i
ng
the
qua
dr
at
ic
S
VM
a
nd
the
PC
A,
with acc
ur
acy
of
87.
5%. F
or
each class
we have
the R
OC
curve.
In this
m
od
el
w
e h
a
ve
for
:
1.
The
healt
hy
co
ntr
ol:
we
hav
e
44
we
re
co
rr
e
ct
ly
cl
assifi
ed,
11
we
re
m
isc
l
assifi
ed
(all
as
in
early
sta
ge),
with a
pe
rcen
ta
ge of
80% t
ru
e
posit
ive r
at
e;
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Voice Asse
ssme
nts for
D
et
ect
ing
Patie
nts
wi
th P
ar
ki
nso
n’
s
….
(
Elmeh
di Be
nmalek)
4269
2.
The
ea
rly
sta
ge
cl
ass:
we
ha
ve
16
4
we
re
c
orrectl
y
cl
assifi
ed,
12
wer
e
m
isc
la
ssifie
d
(
3
a
s
healt
hy,
a
nd
11
as interm
ediat
e stage)
, wit
h
a
per
ce
ntage
of t
ru
e
posit
ive
rate 9
2%;
3.
The
interm
ediat
e
sta
ge
cl
ass:
we
ha
ve
106
w
ere
co
rr
ect
ly
cl
assifi
ed,
12
we
re
m
isclassifie
d
(
11
as
in
ea
rly
sta
ge,
a
nd
1
as
adv
a
nce
d
sta
ge
)
wit
h
a
per
ce
nt
age of 9
0%
tr
ue posi
ti
ve
rate
;
4.
The
ad
va
nce
d
sta
ge
cl
ass: we
h
ave
14
we
re co
r
rectl
y cl
assif
ie
d,
10
we
re mi
scl
assifi
ed
(1
as in
early
st
ag
e
and 9 as i
n
inte
rm
ediat
e stage,)
, wit
h
a
p
e
rce
ntage o
f 58
%
tru
e
posit
ive
rate.
Table
3
.
Re
s
ults usi
ng qua
dr
at
ic
SV
M
Confus
i
on m
atr
ix
ROC cu
r
ve
Cl
ass 1
Cl
ass 2
Cl
ass 3
Cl
ass 4
3.3.
Obtained
resu
lts usin
g
cu
bic
kernel
The
T
a
ble
4
re
pr
ese
nt
the
obta
ined
resu
lt
s
of
cl
assifi
cat
ion
us
in
g
the
c
ubic
SV
M,
a
nd
se
le
ct
ing
the
m
or
e
releva
nt
by
the
PC
A,
w
it
h
accu
racy
of
85.
1%.
F
or
ea
ch
cl
ass
we
ha
ve
t
he
R
OC
c
urve
.
In
this
m
od
el
we
hav
e
for
:
1.
The
healt
hy
co
ntr
ol:
we
hav
e
41
we
re
co
rr
e
ct
ly
cl
assifi
ed,
1
4
we
re
m
isc
l
assifi
ed
(all
as
in
early
sta
ge),
with a
pe
rcen
ta
ge of
75% t
ru
e
posit
ive r
at
e;
2.
The
ea
rly
sta
ge
cl
ass:
we
ha
ve
166
we
re
c
orrectl
y
cl
assifi
ed,
1
4
we
re
m
isc
la
ssifie
d
(
4
as
healt
hy,
an
d
8
as interm
ediat
e stage)
, wit
h
a
per
ce
ntage
of t
ru
e
posit
ive
rate 9
3%;
3.
The
interm
ediat
e
sta
ge
cl
ass:
we
ha
ve
104
w
ere
co
rr
ect
ly
cl
assifi
ed,
14
we
re
m
isclassifie
d
(all
as
in
ea
rly
sta
ge)
with a
percenta
ge of
88% tr
ue p
os
it
iv
e rate;
4.
The
a
dv
a
nce
d
sta
ge
cl
ass:
we
hav
e
8
wer
e
c
orrectl
y
cl
assifi
ed,
16
wer
e
m
isc
la
ssifie
d
as
in
interm
ediat
e
sta
ge,
with a
percenta
ge of
33% tr
ue p
os
it
iv
e rate.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4265
-
4271
4270
Table
4.
Res
ults usi
ng cubic S
VM
Confus
i
on m
atr
ix
ROC cu
r
ve
Cl
ass 1
Cl
ass 2
Cl
ass 3
Cl
ass 4
Fr
om
al
l
pr
ev
iou
s
res
ults,
it
is
seen
t
hat
the
m
axi
m
u
m
cl
assifi
cat
ion
accuracy
of
92.5%
wa
s
achieve
d
us
i
ng
the
li
near
S
VM.
Com
par
e
d
with
pr
e
vious
stu
dies
do
ne
,
the
pro
pose
d
m
et
ho
d
gi
ve
bette
r
resu
lt
s
t
han
the
cepstral
analy
sis
ap
proac
h
(
86.
7%)
[
12]
,
bu
t
this
fi
nd
i
ng
s
cou
l
d
be
im
pr
oved
by
us
i
ng
f
eat
ur
e
sel
ect
ion
al
gor
it
h
m
ded
ic
at
ed
for
m
ulti
cl
ass
cl
assifi
cat
ion
and
c
om
bin
ig
th
e
voic
e
featu
res
with
t
he
ce
ps
tral
analy
sis
wh
ere
a
scor
e
of
96
%
has
been
ac
hieve
d
in
[11],
bu
t
the
appr
oa
ch
was
m
or
e
com
plex
than
the
one
pro
po
se
d
i
n
th
is
stud
y.
T
he
resu
lt
s
s
how
a
lso
that
t
he
fe
at
ur
e
sel
ect
io
n
play
crit
ic
al
r
ole
in
cl
assifi
cat
ion
op
ti
m
iz
ation
.
And
the
m
isc
lassificat
ion
is
e
xp
la
ine
d
by
th
e
relat
ive
m
erits
of
t
he
UPDR
S
scal
e
for
acc
ur
at
el
y
determ
ining
th
e
degree
of
d
is
ease
progressi
on.
The
purpos
e
of
this
stu
dy
is
to
sh
ow
the
eff
ect
ive
ness
of
us
in
g
vo
ic
e
rec
ordi
ng
to
cl
assify
people
with
P
a
rk
i
ns
on
’s
dis
ease
by
the
sever
it
y
of
sym
pto
m
s
us
ing
only
19
featur
e
s.
4.
CONCL
US
I
O
N
Cl
inici
ans
and
vo
ic
e
path
ologist
s
ha
ve
bec
om
e
pr
ogressi
vely
watch
fu
l
to
any
te
ch
niques,
wh
i
c
h
m
igh
t
pro
vid
e
sup
plem
entary
inf
or
m
at
ion
to
he
lp
t
hem
i
n
th
e
e
valuati
on
an
d
the
dia
gnos
is
of
P
D.
In
this
pap
e
r,
we
pres
ented
ne
w
te
c
hn
i
qu
e
that
ca
n
se
par
at
e
bet
wee
n
healt
hy
people
a
nd
P
D
patie
nts
at
diff
e
re
nt
sever
it
y
sta
ges
base
d
on
vo
ic
e
featu
res.
As
a
resu
lt
,
we
ac
hieve
d
92.5
%
of
acc
uracy
usi
ng
li
nea
r
S
V
M
an
d
the
PCA.
T
he
r
esults
sh
ow
al
s
o
that
the
featu
re
sel
ect
ion
pla
y
crit
ic
al
ro
le
in
cl
assifi
cat
ion
op
ti
m
iz
at
i
on
.
And
the
m
isc
la
ssifi
ed
sam
ples
ar
e
us
ually
m
in
gled
with
the
near
est
cl
ass,
wh
ic
h
cl
inica
lly
exp
la
ined
by
the
relat
ive
m
eri
ts
of
the
UPDRS
scal
e
for
accuratel
y
deter
m
ining
the
de
gr
ee
of
diseas
e
progressi
on.
These
resu
lt
s
are
very
encour
a
ging
,
in
f
uture
w
orks
we
co
ns
i
der
to
determ
i
ne
co
rr
el
at
io
n
betwee
n
the
vo
ic
e
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Voice Asse
ssme
nts for
D
et
ect
ing
Patie
nts
wi
th P
ar
ki
nso
n’
s
….
(
Elmeh
di Be
nmalek)
4271
disorde
rs
an
d
the
sym
pto
m
s,
wh
ic
h
will
be
of
great
help
t
o
the
m
edici
ne
and
co
uld
al
s
o
exten
de
d
f
or
oth
er
vo
ic
e
path
ol
ogie
s.
ACKN
OWLE
DGE
MENTS
These
Dataset
s
wer
e
generat
ed
thr
ough
c
ollaborat
ion
bet
ween
Sage
Bi
on
et
wor
ks
,
Pat
ie
ntsLikeMe
a
nd
Dr.
Ma
x
Lit
tl
e
as
par
t
of
t
he
Pa
ti
ent
Vo
ic
e
A
naly
sis
stud
y
(
PVA).
They
wer
e
obta
ined
thr
ough
Syna
ps
e
I
D
[
syn
232174
5
].
REFERE
NCE
S
[1]
M.
A.
Li
t
tl
e
,
et
al.
,
“
Suita
bi
li
t
y
of
d
y
sphon
ia
m
e
asure
m
ent
s
for
t
el
emon
it
or
ing
of
Parkinson'
s
dise
ase
,”
B
iomedi
ca
l
Engi
ne
ering, IEEE
Tr
ansacti
ons
on
,
vol/
issue
:
56
(
4
),
pp
.
1015
-
1
022
,
2009
.
[2]
A.
Ho,
et
a
l.
,
“
Speec
h
impairme
nt
in
a
la
rg
e
sa
m
ple
of
pa
ti
en
ts
with
Pa
rk
inson’s
disea
se
,”
Be
ha
vi
oral
Neurolog
y
,
vol.
11
,
pp
.
131
–
137
,
1998
.
[3]
J.
A.
Loge
m
ann
,
et
a
l.
,
“
Freque
nc
y
and
co
-
o
cc
u
rre
nce
of
vocal
-
t
rac
t
d
y
sfunc
ti
on
s
in
spee
ch
of
a
la
r
ge
sam
ple
o
f
Parkinson
patien
ts
,”
Journal
o
f
S
pee
ch
Hearing
Disor
der
,
vol.
4
3,
pp
.
47
–
57
,
19
78.
[4]
D.
A.
Rahn
,
et
a
l.
,
“
Phonator
y
i
m
pai
rm
ent
in
Pa
rkinson’s
disea
se:
Ev
ide
n
ce
fro
m
nonli
nea
r
d
y
n
amic
an
aly
sis
a
n
d
per
turbation anal
y
sis
,
”
J
.
Voice
,
v
ol.
21
,
pp
.
64
-
71
,
2007
.
[5]
V.
Parsa
and
D.
G.
Jam
ie
son,
“
In
te
ra
ct
ions b
et
we
en
spee
ch
code
rs
and
d
isorder
ed
spee
ch
,
”
Spe
ec
h
Comm
unic
ati
on
,
vol
/i
ss
ue:
40
(
7
)
,
pp.
365
–
385
,
20
03
.
[6]
S.
B.
Davis,
“
Ac
oustic
ch
aract
er
i
stic
s
of
norm
al
a
nd
pat
hologica
l
voic
es,
”
Spe
ec
h
and
Language:
Adv
anc
es
inB
asi
c
Re
search
and
Pr
act
i
ce
,
vol
.
1
,
pp
.
271
–
335
,
1979
.
[7]
R.
Das,
“
A
compari
son
of
m
ulti
ple
cl
assifi
ca
t
ion
m
et
hods
for
di
a
gnosis
of
Parkin
son
disea
se,
”
Expert
Syst
ems
wit
h
Appl
ic
a
ti
ons
,
vo
l/
issue:
37
(2)
,
pp
.
1568
–
1572
,
20
10.
[8]
A.
Tsana
s,
et
a
l.
,
“
Novel
spee
ch
signal
p
roc
e
ss
ing
al
gorit
hm
s
for
high
-
accur
acy
c
la
ss
ifica
t
io
n
of
P
ark
inson's
disea
se
,”
I
EE
E
T
rans
act
ions o
n
B
iomedi
ca
l
Eng
in
ee
ring
,
vol/is
sue:
59(5)
,
pp.
1264
-
1271
,
2012
.
[9]
M.
A.
L
it
t
le
,
et
al.
,
“
Expl
oiting
nonli
ne
ar
re
cur
r
enc
e
and
fr
acta
l
sca
li
ng
prope
r
ti
es
for
voice
d
isorder
de
te
c
ti
on
,
”
Bi
omed. E
ng
.
O
nli
ne
,
2007
.
[10]
P.
Zwirne
r
,
et
a
l
.
,
“
Phonator
y
fu
nct
ion
of
n
eur
ol
ogic
a
lly
impair
e
d
pat
i
ent
s
,”
Jour
nal
of
communic
ati
on
disorders
,
vol/
issue:
24
(
4
),
pp.
287300
,
199
1
.
[11]
E.
Benmale
k
,
et
al.
,
“
Multi
cl
ass
cl
assifi
ca
t
ion
of
Parkinson’s
disea
se
using
diffe
re
nt
cl
assifi
ers
and
LL
BF
S
fea
tur
e
sele
c
ti
on
al
gor
ithm
,”
Int
ernati
on
al
Journal
of
Sp
ee
ch
Techno
logy
,
vol
/i
ss
ue:
20(1)
,
pp
.
179
-
184
,
2
017.
[12]
E.
Benmal
ek,
et
al.
,
“
Multi
class
cl
assifi
ca
t
ion
of
Parkinson’s
disea
se
using
ce
pstr
a
l
ana
l
y
sis
,”
Int
ernati
onal
Journal
of
Spe
ec
h
Techn
ology
,
pp.
1
-
11
,
2017.
[13]
Pati
ent Voice
A
naly
s
is (
PV
A) Sy
napse
ID:
s
y
n2
321745
htt
ps:/
/www
.
sy
n
apse
.
o
rg,
[14]
V.
J
.
Phi
li
ppe
,
et al
.
,
“
Kerne
l
Me
t
hods i
n
Com
put
at
ion
al
B
iol
og
y
,”
MIT
Pr
ess,
Ca
m
bridge
,
2004
.
[15]
M.
Farrús,
et al
.
,
“
Jitt
er
and
shim
m
er
m
ea
surem
ent
s for
spe
ake
r
r
ec
ogni
ti
on
,”
INT
ER
SP
EE
CH
,
20
07.
[16]
R.
A.
Shirvan
and
E.
Ta
hami
,
“
Voice
anal
y
s
is
for
det
ecting
Par
kinson'
s
disea
se
using
gene
ti
c
a
l
gorit
hm
and
KNN
cl
assifi
ca
t
ion
m
e
thod
,”
Bi
om
edi
c
al
Eng
ine
ering
(
ICBME
)
,
2011
1
8th
Iranian
Con
f
ere
nce of.
I
EE
E
,
2011
.
[17]
H.
Kant
z
a
nd
T
.
Schre
ibe
r
,
“
Nonl
ine
ar
ti
m
e
serie
s
ana
l
y
sis,
”
C
ambridge
Univ
er
sit
y
Press
,
2nd
ed
it
io
n,
2004
.
[18]
M.
A.
Li
t
tl
e
,
e
t
a
l.
,
“
Expl
o
it
ing
N
onli
ne
ar
Rec
urr
e
nce
and
Fra
ct
a
l S
ca
li
ng
Propert
i
es
for
Voice
Dis
orde
r
Detect
ion
,
”
Bi
omedi
cal
Enginee
ring Onl
ine
,
vol
/i
ss
ue:
6(23),
2007.
[19]
L.
Cnockaert
,
et
al.
,
“
Low
fre
quency
vo
cal
m
o
dula
ti
ons
in
vowels
produc
ed
b
y
Parkinsonia
n
subjec
ts,”
Spe
e
ch
Comm
unic
ati
on
,
v
ol. 50, pp. 288
-
300,
2008
.
[20]
Z.
Ch
en,
et
a
l.
,
“
Eff
ect
of
nonst
ationar
ities
on
d
etrende
d
f
luc
tu
at
i
on
ana
l
y
sis
,
”
Ph
ysic
al
Revie
w
E
,
v
ol
/i
ss
ue:
65(4)
,
pp.
041107
,
200
2
.
[21]
Guéri
f
S.
,
“
Réduc
ti
on
de
dimen
sion
en
appr
ent
i
ss
age
num
éri
que
non
supervisé
e
,”
PhD
the
sis,
U
nive
rsit
é
Paris
13
,
p
p.
420
148
,
200
6
.
[22]
Ferc
hic
h
i
S
.
E
.
,
et
al
.
,
“
Feat
ur
e
select
ion
usin
g
an
SV
M
le
ar
ning
m
ac
hine
s
,”
Proce
ed
ings
of
the
422
3rd
Inte
rnational
Co
nfe
renc
e
on
Sign
als,
Circu
it
s and
Syste
ms
(
SCS
2
009)
,
pp.
1
-
6
,
20
09
.
[23]
A.
Benba,
e
t
al
.
,
“
Voice
assess
m
ent
s
for
det
e
cting
patient
s
wi
t
h
Parkinso
n’s
disea
ses
using
PC
A
and
NP
CA
,”
Inte
rnational
Jo
urnal
of
Sp
eech T
ec
hn
ology
,
vo
l/is
sue:
19(4),
pp.
743
–
75412
,
201
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.