Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
12
,
No.
3
,
Decem
ber
201
8
, p
p.
984
~
9
9
4
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
2
.i
3
.pp
984
-
994
984
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
nd
ex.
ph
p/ij
eecs
Compari
son of Mu
ltiscale E
ntrop
y Techniq
ues for L
un
g Sound
Classific
atio
n
Achm
ad
Riz
al
1
,
Ris
an
uri
Hida
yat
2
,
Hanu
ng
Ad
i
Nugr
oho
3
1
,2,3
Depa
rt
m
ent o
f
Elec
tr
ical Engi
nee
ring
&
Infor
m
at
ion
T
ec
hnolo
g
y
,
Univer
sita
s Gad
j
ah
Mada
,
Jl
Graf
ika
no
2,
Mlati,
Sleman,
D
.
I
.
Yo
g
y
ak
arta,
Indone
sia
1
School
of El
ec
t
ric
a
l
Eng
ineeri
n
g,
T
el
kom
Unive
rsit
y
Jl T
e
le
kom
unikas
i
no
1,
T
ers.
Bu
ah
Ba
tu, Boj
ong
Soang,
Bandung
,
Indone
si
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
27
, 201
8
Re
vised
A
ug
24
, 2
018
Accepte
d
Oct
8
, 2
018
Lung
sound
is
a
biol
ogi
cal
sign
al
tha
t
c
an
b
e
u
sed
to
d
eterm
ine
th
e
h
ea
l
th
le
ve
l
of
th
e
res
pira
tor
y
tract.
Vari
ous
digital
signal
proc
essing
te
chn
iques
have
bee
n
dev
eloped
for
aut
om
at
ic
class
ifi
c
at
ion
of
lung
sounds
.
Ent
rop
y
i
s
one
of
the
par
a
m
et
ers
used
to
m
ea
sure
the
b
io
m
edi
ca
l
sign
al
complexit
y
.
Multi
sca
l
e
ent
ro
p
y
is
int
rodu
ced
to
m
ea
sure
th
e
en
trop
y
of
a
signal
a
t
a
par
ticula
r
sca
le
r
ange
.
Over
ti
m
e,
var
ious
m
ult
isca
le
ent
rop
y
t
ec
h
nique
s
have
bee
n
proposed
t
o
m
ea
sure
the
complexi
t
y
of
b
iol
ogical
s
ign
al
s
and
other
ph
y
sic
al
signa
ls.
In
thi
s
pape
r,
som
e
m
ult
isca
le
ent
rop
y
techniq
ues
for
lung
sound
cl
assifica
ti
on
ar
e
compar
ed.
Th
e
resul
t
o
f
the
compari
so
n
indi
c
at
es
tha
t
the
Mult
isca
l
e
Perm
uta
t
io
n
Ent
rop
y
(M
PE)
produc
es
t
he
highe
st
ac
cur
acy
of
97.
9
8%
fo
r
five
lung
sound
dat
ase
ts.
The
resul
t
was
a
chi
ev
ed
for
the
sca
l
e
1
-
10
produc
ing
te
n
fe
atures
for
ea
ch
lu
ng
sound
dat
a.
T
his
result
is
bet
t
er
tha
n
other
seve
n
ent
ropies
.
Multi
sca
l
e
en
trop
y
an
aly
sis
c
an
improve
the
a
cc
ur
acy
of
l
ung
sound
cl
assi
fic
a
ti
on
with
ou
t
req
uiri
ng
an
y
fe
at
ure
s
oth
er
tha
n
ent
rop
y
.
Ke
yw
or
ds:
Mult
isc
al
e entr
op
y
Lu
ng
sou
nd
Coarse
-
grai
ne
d p
ro
ce
dure
Mult
il
ay
er
perce
ptr
on
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
:
Ach
m
ad
Ri
zal
,
Dep
a
rtm
ent o
f El
ect
rical
En
gi
n
eeri
ng & In
f
orm
ation
Tec
hn
ology
,
Un
i
ver
sit
as
Ga
dj
a
h
Ma
da
,
Jl Gr
a
fika
no 2
, Mla
ti
, S
le
m
a
n,
D.I. Y
og
ya
ka
rta,
I
ndonesi
a
.
Em
a
il
:
rizal
.s3
te
14@m
ai
l.ug
m
.ac.id
1.
INTROD
U
CTION
Lu
ng
s
ound is o
ne of
t
he
bi
olo
gical
sig
nals that e
m
erg
e fr
om
the r
espirati
on
process
. An
y chan
ges i
n
it
are
ge
ner
at
e
d
f
r
om
fo
reig
n
bodies
or
phy
siolo
gical
cha
nges
in
the
res
pi
rator
y
tract
ca
us
e
d
by
diseas
es
[
1]
.
Diff
e
re
nces
in
patte
rn
s
of
l
ung
sou
nd
s
ca
n
be
heard
by
a
doct
or
us
i
ng
a
s
te
tho
sc
op
e
t
o
di
agnos
e
diseas
es
[2
]
.
Au
sc
ultat
ion
t
echn
i
qu
e
,
on
the
oth
e
r
ha
nd,
is
ver
y
su
bjec
ti
ve
fo
r
bein
g
dep
e
ndent
up
on
existi
ng
ex
pe
rience
and ex
pe
rtise
of the
docto
r.
Var
i
ou
s
te
c
hniqu
e
s
f
or
a
naly
zi
ng
l
ung
s
ounds
us
in
g
c
om
pu
te
rs
ha
ve
bee
n
de
vel
op
e
d.
S
om
e
of
these
te
chn
iq
u
es
i
nc
lud
e
ti
m
e
-
dom
ai
n
analy
sis
te
chn
i
qu
es
,
su
c
h
as
sta
ti
sti
cal
analy
sis
base
d
on
Hjort
h
descr
i
ptor
[
3]
,
em
pirical
m
od
e
dec
om
po
s
it
ion
(EM
D)
[4
]
,
or
f
ractal
analy
sis
[
5].
Seve
ral
re
se
arch
e
rs
pro
po
se
d
t
o
pe
rfor
m
lun
g
s
ound
analy
sis
in
the
f
reque
ncy
do
m
ai
n,
suc
h
as
quantil
e
ve
c
tor
fr
e
qu
e
ncy
[
6]
or
MFC
C [7
]
. Me
anwhil
e,
wav
el
et
an
al
ysi
s h
as
been u
sed
in [8
]
f
or cla
ssifyi
ng a
bnorm
al
lun
g
s
o
un
ds
.
On
e
of
the
popu
la
r
bio
l
og
ic
al
sign
al
analy
sis
m
et
ho
ds
is
m
ulti
scal
e
entropy
(MSE)
,
pro
posed
by
Costa
et
al
.
with
a
c
oa
rse
-
grai
ne
d
proce
dure
for
m
ulti
scal
e
process
an
d
sam
ple
entr
op
y
for
e
ntr
op
y
m
easur
em
ent
[9
]
.
Subse
quent
ly
,
sever
al
va
riants
of
MSE
e
m
erg
e
su
c
h
as
ref
ine
d
-
M
SE
,
com
po
sit
e
-
MSE
or
adap
ti
ve
MS
E
(A
ME
)
[
10
]
.
Othe
r
resea
rchers
m
od
ifie
d
their
ent
ropy
m
easur
em
ent
te
chn
i
qu
e
s,
w
hich
resu
lt
in
m
ulti
scal
e
per
m
utati
on
entr
op
y
(MPE
)
[11],
m
ult
isc
al
e
app
r
oxim
at
e
entropy
(MAp
E
N)
[12],
and
m
ul
ti
scal
e
fu
zzy
entropy
[13
]
.
In
the
case
of
lu
ng
sounds
,
m
ulti
scal
e
en
tro
py
has
bee
n
us
ed
to
an
al
yz
e
the
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
Compari
son
of
Mu
lt
isc
ale E
nt
ro
py
Tec
hniq
ue
s for Lu
ng
Sound
Cl
as
sif
ic
ation
(
Ac
hma
d
R
izal
)
985
lung
sou
nd
s
of
pu
lm
on
ary
al
veo
l
it
is
patie
nt
s
[14].
The
re
s
ults
ind
ic
at
ed
t
hat
MSE
pro
duces
m
or
e
consi
ste
nt
featur
e
s
com
par
ed
to
the
s
pe
ct
ral
m
et
ho
d.
Si
m
il
ar
to
bio
log
ic
al
sig
nal
analy
sis,
wav
el
et
entropy
is
us
ed
f
or
trans
form
er
design
[
15]
,
wh
i
le
fau
lt
diagnosis
was
anal
yz
ed
us
in
g
sa
m
pl
e
entropy
[16]
and
a
ppr
ox
im
at
e
entr
op
y
[17].
The
perform
ance
com
par
iso
n
of
m
ulti
scal
e
e
ntr
op
y
f
or
va
ri
ou
s
entr
opy
m
easur
em
ent
te
chn
i
qu
e
s
has
nev
e
r
previ
ous
ly
been
done.
I
n
this
stu
dy,
we
com
par
ed
the
us
e
of
m
ult
isc
al
e
entro
py
for
var
i
ous
ent
ropy
m
easur
em
en
t
t
echn
i
qu
e
s
f
or
lung
s
ound
cl
as
sific
at
ion
.
Entr
op
y
m
easur
em
ent
m
et
ho
d
in
this
stu
dy
incl
ude
d
Sh
a
nnon
e
ntr
opy
[
18]
,
s
pectral
ent
ropy
[18],
Re
nyi
e
nt
ropy
[
19]
,
wa
velet
entr
opy
[
20
]
,
ap
pro
xim
at
e
entr
op
y
[21],
s
a
m
ple
entr
op
y
[22],
per
m
utatio
n
ent
ropy
[
23
]
,
a
nd
Tsal
li
s
e
ntr
op
y
[24].
T
ho
s
e
e
ig
hth
ent
ropies
m
easur
em
ents
are
a
re
prese
ntati
on
of
t
wo
ty
pes
entr
op
y
m
easur
em
ent
te
chn
iq
ues:
sp
ect
ral
e
ntropy
an
d
e
m
bed
di
ng
e
nt
ropy
.
Em
bed
di
ng
e
ntr
op
y
wa
s
cal
culat
ed
directl
y
fr
om
the
sign
al
in
the
tim
e
do
m
ai
n,
and
th
e
sp
e
ct
ral
e
ntr
opy
was
cal
culat
ed
f
r
om
the
sign
al
am
plit
ud
e
in
the
fr
e
que
ncy
dom
ai
n.
By
this
resear
ch,
we
ob
ta
ine
d
t
he
be
st
m
ulti
scal
e
entr
op
y
te
ch
ni
qu
e
s
f
or
a
ut
om
at
ic
lun
g
s
ound
cl
assifi
cat
ion
.
Ne
xt,
we
co
uld
reco
m
m
end
the
best
m
e
tho
d
fo
r
a
uto
m
at
ic
lun
g
s
ound
a
naly
sis.
From
our
ex
pe
rim
en
t,
MPE
produc
ed
th
e
highest ac
c
ur
a
cy
a
m
on
g t
hos
e eigh
t e
ntr
opie
s.
This
pa
per
is
orga
nized
as
f
ol
lows
.
Sect
ion
2
ex
plains
the
exp
e
rim
ental
m
et
ho
d,
t
he
lu
ng
s
ound
data
us
e
d
in
this
st
udy,
ent
ropy
m
e
asur
em
ent
m
e
t
hods
a
nd
m
ultiscale
entr
op
y
t
echn
i
qu
e
s.
Me
anwhil
e,
the
re
su
lt
s
and analy
sis
of the test
r
e
su
lt
s
d
esc
ribe
d
in
S
ect
ion
3.
Sect
ion 4
contai
ns t
he
c
on
cl
us
io
ns o
f
this
pa
per.
2.
MA
TE
RIA
L
AND
METH
OD
S
Figure
1
pr
ese
nts
the
m
et
ho
d
in
this
pa
per.
We
her
e
us
e
d
f
ive
cl
asses
of
l
ung
s
oun
d
data
as
the
i
nput
data.
Me
a
nwhi
le
,
in
t
he
cl
ass
ific
at
ion
sta
ge,
we
us
e
d
m
ult
il
ay
er
per
ce
ptr
on
an
d
thre
e
-
f
old
cr
os
s
-
valid
at
ion
.
The follo
wing
su
bse
ct
ions e
xpla
in the
dat
a,
entr
op
y m
easur
em
ent, an
d cl
assifi
cat
ion
m
eth
od i
n detai
l.
Figure
1. Bl
oc
k diag
ram
o
f
m
ulti
scal
e entr
opy i
n
this
pa
per
2
.
1.
Lun
g so
und da
ta
The
l
ung
s
ounds
wer
e
gat
hered
from
sever
a
l
sources
on
t
he
inter
net
[
25,
26
]
a
nd
CD
co
m
pan
ion
of
te
xtb
oo
k
[
27
]
.
All
the
data
w
ere
co
nv
e
rted
i
n
the
f
or
m
of
wav
e
file
s
with
a
sam
pling
fr
eq
ue
ncy
of
8000
Hz.
Fu
rt
her
m
or
e,
the
data we
re
c
ut into
one
res
pirato
ry cy
cl
e.
Table
1
pr
ese
nt
s
the
detai
l
of
lung
sou
nd
dat
a.
So
m
e
of
the
data
have
bee
n
us
e
d
in
a
pr
e
vious
stu
dy
[3
]
.
N
or
m
al
br
onc
hial
is
a
kind
of
norm
al
lung
s
ound
with
a
n
e
xp
i
ra
tory
durati
on
rel
at
ively
long
er
that
insp
irat
io
n
ph
a
se.
It
has
l
oud
and
high
-
pitch
sou
nd
with
a
pau
se
betwe
en
ins
pirati
on
and
e
xpirat
io
n
[28].
Crackle i
s a n
onm
us
ic
al
an
d
exp
l
os
ive s
ound
that has
a sh
or
t du
rati
on. Cra
ckle sound in
dicat
es secreti
on
as in
chro
nic
bro
nc
hiti
s
(co
a
rse
c
rack
le
)
or
no
t
correla
te
d
wi
th
secreti
on
a
s
in
c
ongestiv
e
hear
t
fail
ur
e
(f
i
ne
crackle)
[
28]
.
Asth
m
a
is
a
di
sease
that
pr
oduce
s
a
wh
eez
ing
s
ound
t
hat
has
a
co
ntin
uous
patte
r
n,
m
us
ic
al
so
un
d
a
nd
do
m
inant
fr
e
quen
cy
m
or
e
than
400
Hz
[29].
F
r
ic
ti
on
r
ub
or
pl
eur
al
r
ub
is
ass
ociat
ed
wit
h
pl
eur
al
inflam
m
at
ion
or
pleu
ral
tum
or
s
that
pr
oduce
nonm
us
ic
al
,
exp
l
os
ive
and
us
ua
ll
y
biphasic
sou
nd
[
28
]
.
So
m
et
i
m
es,
it
do
e
s
not
nee
d
a
ste
tho
sc
op
e
to
li
ste
n
to
the
stridor,
or
i
gin
a
ti
ng
f
ro
m
the
la
ryn
x
or
trac
he
a
an
d
hav
i
ng a
do
m
inant
fr
e
quency
> 10
00 H
z
[8]
.
Table
1.
L
un
g soun
d data
Data c
lass
Nu
m
b
e
r
o
f
data
No
r
m
al
bro
n
ch
ial
22
Crackl
e
21
Asth
m
a
18
Friction
r
u
b
18
Strido
r
20
On the
data, t
he
nor
m
al
iz
ation
process
is car
ried o
ut as in (
1) an
d (2).
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
25
02
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
984
–
994
986
(
)
=
(
)
−
1
∑
(
)
=
1
(1)
(
)
=
(
)
√
1
∑
(
−
)
2
1
wi
t
h
μ
=
1
∑
(
)
=
1
(2)
wh
e
re
x(
n)
is
t
he
in
pu
t
sig
nal
,
N
is
the
le
ngt
h
of
the
in
put
s
ign
al
,
a
nd
μ
re
fer
s
to
t
he
m
ea
n
of
the
in
put
s
ign
al
.
Be
cause in
(1)
m
ean (μ)
has
bee
n
m
ade as
ze
ro, s
o
(
2) ca
n b
e re
wr
it
te
n
as
(3).
(
)
=
(
)
√
1
∑
(
)
2
1
(3
)
The
res
ult
fr
om
(3
)
is
the
m
ean
of
si
gn
al
(
μ)
=
0
an
d
sta
nd
a
r
d
dev
ia
ti
on
(σ)
=
1.
The
value
of
σ
=
1
would be
used
in the calc
ulati
on of sa
m
ple en
tr
op
y a
nd a
pproxim
at
e entropy.
Figure
2
sho
w
s
an
exam
ple
of
norm
al
bro
nch
ia
l
a
nd
it
s
lung
sou
nd
s
pectr
um
fr
eq
ue
ncy,
wh
il
e
Figure
3
disp
la
ys
the
crackle
so
un
d
exam
ple.
Fr
om
Figu
re
2
an
d
Fig
ur
e
3,
we
can
see
the
diff
e
ren
ce
between
the
two
ty
pes
of
si
g
nals
in
t
he
tim
e
do
m
ain
an
d
the
fr
e
quency
do
m
ai
n.
This
is
to
be
disti
nguish
e
d
us
in
g
m
ul
ti
scal
e entro
py
both
sp
ect
ral entr
opy an
d em
bed
ded ent
ropy.
Figure
2. N
orm
al
lun
g so
un
d
a
nd the
sp
ect
ru
m
Figure
3. Crac
k
le
lu
ng s
ound
and the
sp
ect
r
um
2.2
.
M
ultisc
al
e pro
cess
Mult
isc
al
e
entro
py
wa
s
us
e
d
to
view
t
he
physi
ology
sign
al
’
s
entropy
at
va
rio
us
scal
es
[9]
.
W
e
use
d
the co
a
rse
-
grai
ned pr
ocedur
e
for
the
m
ulti
sc
al
e p
r
ocess
as i
n (4).
(
)
=
1
∑
=
(
−
1
)
+
1
,
1
≤
≤
(4
)
wh
e
re
x(i
)
is
t
he
i
nput
sig
nal,
τ
is
scal
e
a
nd
y
j
(
τ
)
is
a
sig
nal
in
s
cal
e
τ.
I
n
ge
ne
ral,
t
he
si
gnal
y
j
(
τ
)
is
the
outp
ut
sign
al
on
a
sc
al
e
τ
as
t
he
a
ve
rag
e
of
a
num
ber
τ
of
t
he
input
sig
nal
x(i
).
In
it
ia
ll
y,
this
m
et
ho
d
was
us
e
d
i
n
conj
un
ct
io
n
w
it
h
a
sam
ple
of
e
ntr
opy
to
it
s
entr
op
y
c
al
culat
ion
s
[9]
.
I
n
this
pa
pe
r,
the
coa
rse
-
gr
ai
ned
proce
dure
was use
d
to
gethe
r wit
h
the
foll
ow
ing
e
ntr
opy m
e
asur
em
ent to e
xtract the
f
eat
ures
of lu
ng sou
n
ds.
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
Compari
son
of
Mu
lt
isc
ale E
nt
ro
py
Tec
hniq
ue
s for Lu
ng
Sound
Cl
as
sif
ic
ation
(
Ac
hma
d
R
izal
)
987
2.3
.
En
tropy
measure
ment
In
t
his
pa
pe
r,
we
us
ed
ei
ght
entr
op
y
m
easur
em
ent
m
et
ho
ds
to
be
com
bine
d
with
the
c
oa
rse
-
gr
ai
ne
d
proce
dure. Thi
s secti
on expla
ins e
ac
h
e
ntrop
y
m
easur
em
ent m
e
tho
d.
2.3.1
.
Sh
an
n
on Entr
opy
Sh
a
nnon
ent
ropy
(ShE
N)
is
a
sign
al
c
om
p
le
xity
m
et
ric.
Sh
E
N
s
hows
i
nfor
m
at
ion
co
ntain
in
t
he
sign
al
as
exp
re
ssed
i
n (5).
ℎ
=
−
∑
=
1
2
(5
)
wh
e
re
pi
is
distrib
ution p
roba
bili
ty
o
f
sig
nal,
and
i
is a
level
of the
sig
nal.
2.3.2
.
Spec
tra
l
Entr
opy
Sp
ect
ral e
ntr
op
y (SE
N)
is
S
hE
N
m
easur
em
ent in th
e
fr
e
que
nc
y do
m
ai
n
as e
xpresse
d
as
(6).
=
−
∑
ℎ
=
0
2
(
1
)
(6
)
wh
e
re
P
f
is
the
power
de
ns
it
y
of
fr
e
qu
e
ncy
band,
an
d
f
i
and
f
h
are
the
fr
e
qu
e
ncy
lim
i
t
of
the
sign
al
.
The
powe
r of
t
he
si
gn
al
was
nor
m
al
iz
ed
so t
hat
∑p
n
=
1.
2.3.3
.
Ren
yi E
nt
r
op
y
Re
nyi entr
op
y
(REN
)
is t
he
ge
ner
al
form
o
f ShE
N [19]
. Ma
them
a
ti
cal
l
y REN
is e
xpres
se
d
as
(7).
=
1
1
−
2
(
∑
=
1
)
,
q ≠ 1
(7
)
wh
e
re
q
=
ord
e
r of
Re
nyi entr
op
y.
Prac
ti
cal
l
y, REN is
d
e
fined f
or or
der
q = 2
.
2.3.4
.
Wavel
et
Entr
opy
Wav
el
et
Ent
ropy
(
W
E
)
is
cal
culat
ed
f
ro
m
the
energy
of
eac
h
sub
-
ba
nd
i
n
the
wa
velet
tran
sform
at
ion
resu
lt
s
[20].
W
E is ex
presse
d as (
8).
=
−
∑
<
0
(8
)
wh
e
re
pi
is t
he
relat
ive w
a
vele
t ener
gy
ob
ta
in
ed fr
om
(
9)
.
=
(9
)
wh
e
re
Ei
is
the
energy
for
i
-
th
reso
l
utio
n,
and
Et
is
the
total
energy.
T
he
ad
va
ntages
of
WE
that
it
is
no
t
influ
e
nce
d
by
no
ise
.
W
E
det
ect
s
so
m
e
s
m
a
ll
chan
ges
in
non
-
sta
ti
on
a
ry
sign
al
s
an
d
doe
s
no
t
de
pe
nd
on
any
par
am
et
ers
[20
]
. I
n t
his
pa
per
,
w
e
us
e
d D
b2
as m
ot
her
w
a
ve
le
t and
dec
ompo
sit
io
n
le
vel
= 7
,
as i
n
[
30]
.
2.3.5
.
Appro
xi
mat
e
Entr
opy
Appro
xim
at
e
entr
op
y
(ApE
n)
is
a
si
gn
al
com
plexity
pa
ram
et
er
by
m
easur
in
g
t
he
num
ber
of
occurre
nces
of
a
sig
nal
patte
r
n
al
ong
t
he
si
gn
al
[
21
]
.
I
f
the
s
equ
e
nce
of
the
sign
al
al
o
ng
N
{
(
)
:
,
1
≤
≤
}
is give
n
to
for
m
a
vector
unti
l
−
+
1
as in
(10
).
X
i
m
=
{
u
(
i
)
,
u
(
i
+
1
)
,
…
,
u
(
i
+
m
−
1
)
}
,
for
i
=
1
,
…
,
N
−
m
+
1
(1
0
)
wh
e
re
is
the
l
eng
t
h
of
wind
ow
wh
ic
h
woul
d
be
c
om
par
ed
.
F
or
eac
h
≤
−
+
1
,
the
def
i
ne
d
(
)
is
(
−
+
1
)
−
1
m
ult
ipli
ed
by
a num
ber
of th
e the
in
from
.
By
d
efi
ning
(11)
∅
m
(
r
)
=
(
N
−
m
+
1
)
−
1
∑
ln
C
i
m
(
r
)
N
−
m
+
1
i
=
1
(1
1
)
wh
e
re
is nat
ural
log
a
rithm
.
Pincus
def
ine
d
A
pE
n
as i
n (12)
[
21
]
.
ApEn
(
m
,
r
)
=
li
m
N
→
∞
[
Φ
m
(
r
)
−
Φ
m
+
1
(
r
)
]
(1
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
25
02
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
984
–
994
988
ApEn
was
esti
m
at
ed
us
in
g
st
at
ist
ic
as
(13).
ApEn
(
m
,
r
,
N
)
=
Φ
m
(
r
)
−
Φ
m
+
1
(
r
)
(1
3
)
Com
m
on
ly
A
pE
n was calc
ula
te
d
usi
ng m
= 2
a
nd r =
0.2
×
st
and
a
r
d dev
ia
ti
on.
2.3.6
.
Sampl
e
Entropy
Sam
ple
entro
py
(S
a
m
pEn
)
w
as
propose
d
by
R
ic
h
m
an
and
Moorm
an
to
ov
erc
om
e
the
weakness
of
ApEn
[
22
]
.
I
n
ApEn
,
bias
is
pr
ese
nt
by
sel
f
-
m
at
ches
.
Sam
pE
n
is
a
m
easur
e
of
the
pro
ba
bili
ty
of
a
row
of
m
data
that
w
ould
be
the
sam
e
as
oth
e
r
in
a
s
eries
of
the
si
gnal
with
a
t
ole
ran
ce
of
r,
w
hi
ch
w
ould
rem
a
i
n
the
sam
e if the r
ow
of
m
d
at
a is
increase
d
to
m
+1.
Mat
hem
ati
cal
ly
Sa
m
pEn
i
s expr
e
ssed
as
(14).
Sa
mp
En
(
m
,
r
)
=
li
m
N
→
∞
−
ln
A
m
(
r
)
B
m
(
r
)
(1
4
)
wh
e
re
(
)
is
the
pro
ba
bili
ty
of
t
wo
se
qu
e
nces
will
m
a
tc
h
f
or
a
num
ber
of
m
+1
sam
ples
wit
hin
t
olera
nce
r
.
Me
anwhil
e,
(
)
is
the
pro
ba
bili
t
y
of
tw
o
se
qu
e
nces
that
woul
d
m
at
ch
fo
r
m
nu
m
ber
of
sam
ples
withi
n
the
tolerance
of
r
.
I
n
bo
t
h
par
a
m
et
ers,
sel
f
-
m
at
ches
was
a
voide
d.
Furthe
r
m
or
e,
by
m
aking
=
{
[
(
−
−
1
)
(
−
)
]
/
2
}
(
)
an
d
=
{
[
(
−
−
1
)
(
−
)
]
/
2
}
(
)
so S
am
pEn
ca
n be e
xpr
essed
a
s (1
5).
Sa
m
pEn
(
m
,
r
,
N
)
=
−
ln
A
B
(1
5
)
2.3.7
.
Permu
t
at
i
on
En
tr
opy
Perm
utati
on
en
tro
py
(PE)
is
the
m
easur
em
e
nt
of
si
gn
al
co
m
plexit
y
by
id
entify
ing
the
pro
ba
bili
ty
of
cod
e
se
quence
in
the
sig
nal
[23].
PE
obse
rv
e
s
the
per
m
utatio
n
patte
rns
of
the
dif
fer
e
nt
el
em
ents
in
the
sign
a
l
.
It is ex
presse
d as (
16).
PE
=
−
∑
p
j
log
2
p
j
n
j
=
1
(1
6
)
wh
e
re
is
the
r
el
at
ive
fr
e
quen
cy
of
possi
ble
patte
rn
an
d
n
i
s
is
pe
rm
utati
on
order,
n
≥
2.
I
n
this
pa
per,
we
us
e
d n = 6 as
in [1
1].
2.3.8
.
Ts
alli
s
Entropy
Tsal
li
s
entrop
y
(TE)
is
com
m
on
ly
us
ed
to
descr
ibe
th
e
ph
ysi
cal
be
ha
vior
of
a
syst
e
m
[2
4].
TE
descr
i
bes
the
s
yst
e
m
with
the
eff
ect
of
lo
ng
-
te
rm
m
e
m
or
y,
long
-
range
inte
racti
on
a
nd
m
ulti
fr
act
al
sp
ace
-
tim
e
const
raint [28
]
.
I
t i
s n
on
-
e
xten
sive in
w
hich
t
her
e a
re two
id
entic
al
syst
e
m
s
, th
en
the
nu
m
ber
of
e
ntropy
is no
t
the
sam
e as b
ot
h
a
dd
e
d
t
og
et
he
r.
T
E m
at
he
m
at
ic
al
ly
is exp
r
essed
a
s (1
7).
TE
=
1
−
∑
p
i
q
W
i
=
1
q
−
1
(
17)
wh
e
re
q
is
no
n
-
e
xtensi
vity
orde
r,
pi
is
disc
rete
pro
bab
il
it
y,
and
W
is
th
e
m
ic
ro
sco
pic
config
ur
at
io
n
of
t
he
syst
e
m
.
In
t
hi
s
pa
per
we
us
e
t
he
order
of
non
-
e
xte
nsi
vity
q
=
2
pro
ve
d
to
pro
duce
the
hi
gh
e
st
accuracy i
n [31].
3.
RESU
LT
S
A
ND
DI
SCUS
S
ION
3.1.
Measure
ment Resul
t a
nd
S
tatistics
Analysis
Figure
4
s
how
s
the
resu
lt
s
of
the
m
ult
isc
ale
process
us
i
ng
the
coa
rse
-
grai
ne
d
proce
dure
f
or
the
norm
al
br
on
c
hi
al
and
lu
ng
s
ound
crac
kle
s
ound.
In
ge
neral
,
the
sig
nal
sh
a
pe
di
d
not
change
bu
t
t
he
data
le
ng
th
bec
om
e
s
N/τ,
with
N
is
the
or
igi
nal
data
le
ng
t
h,
a
nd
τ
is
the
scal
e.
The
value
of
each
sam
ple
on
the
scal
e
τ
was
the
ave
rag
e
val
ue
of
τ
sam
ple
da
ta
from
the
ori
gin
al
da
ta
.
T
he
coa
rse
-
gr
ai
ne
d
proce
dure
re
du
ce
d
the
var
ia
nce
of
the
sign
al
,
a
s
the
scal
e
τ
increase
th
e
n
the
var
ia
nce
will
de
crease.
T
he
de
crease
in
var
i
ance
will
change
th
e
entropy
m
ea
su
rem
ent
valu
e.
The
c
hange
of
ent
ropy
va
lue
was
util
iz
ed
as
a
featu
re
for
pu
lm
on
ary s
ou
nd classi
ficat
io
n.
Fig
ur
e
5 dis
play
s the
var
i
a
nce as
a fu
ncti
on of scale τ
.
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
Compari
son
of
Mu
lt
isc
ale E
nt
ro
py
Tec
hniq
ue
s for Lu
ng
Sound
Cl
as
sif
ic
ation
(
Ac
hma
d
R
izal
)
989
Figure
4.
N
orm
al
b
ronc
hial and crac
kle lu
ng s
ound
for
sca
le
1
–
scal
e
5
Figure
5. The
var
ia
nce
of
eac
h
cl
ass
of
data
Figure
6
-
Fi
gure
13
sho
w
th
e
aver
a
ge
valu
es
of
t
he
m
ultiscale
entr
op
y
m
easur
em
ent
resu
lt
s
in
fi
ve
cl
asses
of
l
ung
s
ounds.
T
he
entr
opy
m
eas
ur
em
ent
res
ults
te
nded
to
be
low
e
xcep
t
Tsal
li
s
entr
opy
that
gen
e
rated
a
c
on
si
der
a
ble
va
lue
but
ne
gati
ve
value.
A
s
m
al
l
value
wa
s
infl
uen
ce
d
by
the
norm
al
i
zat
ion
process
s
o
that
the
aver
a
ge
va
lue
=
0
w
hile
the
var
ia
nce
va
lue
bec
am
e
=
1.
W
e
c
ou
l
d
identify
that
spe
ct
ral
entr
op
y
an
d
w
avelet
entropy
gen
e
rated
r
el
at
ively
disti
nct
patte
rn
s
betwe
e
n
cl
asses.
The
resu
lt
sug
gested
that
the
entr
op
y
m
easur
em
ent
in
the
fr
e
quency
do
m
ai
n
produ
ced
m
or
e
disti
nct
char
act
e
risti
cs
than
in
th
e
tim
e
d
om
ai
n.
Fi
gure
10
a
nd
Fi
gure
11
disp
la
y
t
ha
t
wh
e
eze
s
ound
on
the
Sam
pEn
a
nd
ApEn
hav
e
se
par
at
e
d
val
ues
to
oth
e
r
cl
ass
es.
Sam
pEn
is
an
i
m
pr
ovem
ent
of
ApEn
t
hat
has
s
om
e
sel
f
-
m
at
ching
pro
blem
s
[4
5].
The
cal
culat
ion
of
bo
t
h
e
ntr
op
ie
s
is ab
ou
t t
he
sa
m
e;
h
ence t
he
r
esulti
ng en
t
rop
y al
so
has
a
sim
il
ar ch
aracte
r
ist
ic
.
Figure
6. S
hannon
e
ntr
opy m
easur
em
ent for
f
ive
lung s
ound cla
sses
Figure
7. S
pectral entr
opy m
e
asur
em
ent for f
ive lu
ng
so
un
d
cl
asses
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
25
02
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
984
–
994
990
Figure
8.
Re
ny
i entr
op
y m
easur
em
ent f
or
fiv
e lung
so
un
d
cl
asses
Figure
9.
W
a
ve
le
t entropy m
easur
em
ent for
f
ive
lu
ng
so
un
d
cl
asses
Figure
10. A
pproxim
at
e entropy m
easur
em
e
nt for
five
l
ung
s
ound classe
s
Figure
11. S
am
ple en
t
ropy m
e
asur
em
ent for f
ive lu
ng
so
un
d
cl
asses
Figure
12. P
e
r
m
uta
ti
on
e
ntrop
y m
easur
em
e
nt for
five
l
ung
s
ound classe
s
Figure
13. Tsall
is entr
op
y m
easur
em
ent for fi
ve
lu
ng
so
un
d
cl
asses
We
us
e
d
t
he
a
naly
sis
of
var
i
ance
(AN
OVA
)
sta
ti
sti
cal
te
st
to
te
st
t
he
se
pa
rati
on
bet
wee
n
t
he
cl
asses
of
featu
re
res
ul
ti
ng
f
ro
m
the
m
ult
isc
al
e
ent
ropy.
T
he
perform
ance
of
m
ul
ti
scal
e
entro
py
was
deter
m
ined
us
in
g
T
he
A
N
OVA
F
-
value
[32].
A
la
r
ger
F
-
va
lue
i
nd
ic
at
es
a
bette
r
pe
r
form
ance
for
s
epar
at
in
g
t
he
c
la
sses
te
ste
d
com
par
e
d
to
the l
ow
e
r on
e
. In
t
his stu
dy, F
-
val
ue wa
s m
easur
ed on
the scal
e
of
1
-
20,
1
-
15 scale
,
a
nd so
on
to
see
t
he
changes
in
t
he
discrim
inati
on
pe
rfo
rm
ance
at
diff
ere
nt
scal
es.
Table
2
shows
t
he
resu
lts
ANO
VA test
.
Table
2
sho
ws
that
the
m
ulti
s
cal
e
sa
m
ple
entropy
pro
du
ce
d
the
la
rg
e
st
F
-
value
on
the
sc
al
e
of
1
-
20
and
t
he
scal
e
of
1
-
15.
F
or
t
he
scal
e
of
1
-
10
a
nd
scal
e
of
1
-
5,
the
high
est
F
-
val
ue
w
as
achieve
d
by
Re
nyi
entr
op
y.
F
or
t
he
scal
e
of
1
to
the
scal
e
of
1
-
4,
the
hig
he
st
F
-
val
ue
was
achieve
d
by
Tsal
li
s
entropy.
Me
anwhil
e, m
ulti
scal
e Sh
a
nnon en
t
ropy
pro
du
ce
d
t
he
lo
we
st F
-
value
i
n
a
l
l condit
ions.
It
was
widely
per
cei
ved
t
ha
t
the
F
-
value
decr
ease
d
as
t
he
re
duct
ion
s
cal
e
excep
t
on
m
ulti
scal
e
Tsal
li
s
entropy
an
d
m
ulti
scale
Re
nyi
entr
op
y.
In
m
ulti
scaleTsall
is
entr
opy,
F
-
value
incr
eased
on
the
s
cal
e
of
1
-
15
an
d
then
decr
ease
d
in
t
he
ne
xt
scal
e.
Me
anwhil
e,
in
m
ulti
scal
e
Ren
yi
entropy,
F
-
val
ue
of
1
-
20
scal
e
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
Compari
son
of
Mu
lt
isc
ale E
nt
ro
py
Tec
hniq
ue
s for Lu
ng
Sound
Cl
as
sif
ic
ation
(
Ac
hma
d
R
izal
)
991
increase
d
to
a
scal
e
of
1
-
10
t
hen
dec
rease
d.
This
i
nd
ic
at
ed
that
re
duci
ng
the
scal
e
or
re
du
ci
ng
the
nu
m
ber
of
featur
e
s
did
no
t
al
ways
red
uc
e
the
rate
of
se
par
at
io
n
of
dat
a
bu
t
m
ay
increas
e
it
.
This
oc
curred
beca
us
e
of
the
high scale
(sca
le
1
5 t
o sca
le
20) t
he varia
nce
of the
s
ig
nal t
end
e
d
t
o decre
ase [
10
]
.
Th
us
the
entr
opy
ge
ne
rated
w
as
not
sig
nifica
ntly
diff
e
re
nt
f
ro
m
oth
er
data.
The
sm
allest
F
-
value
f
or
al
l
the
m
ulti
sc
al
e
entr
op
y
w
as
ob
ta
ine
d
in
scal
e
1
or
m
e
asur
em
ent
of
entr
op
y
i
n
the
or
i
gin
al
si
gn
a
l.
Thi
s
su
ggest
e
d
that
the
m
ulti
scal
e
schem
e
pr
od
uc
es
a
bette
r
pe
rfor
m
ance
com
par
ed
t
o
e
ntr
op
y
m
easur
e
m
ent
of
on
a sin
gle scal
e s
ign
al
on
ly
.
Table
2.
F
-
Values
from
A
NOVA Test
MS
-
Entro
p
y
F v
alu
e
scale
1
-
20
scale
1
-
15
scale
1
-
10
scale
1
-
5
scale
1
-
4
scale
1
-
3
scale
1
-
2
scale
1
Multis
cale Sa
m
p
le
entro
p
y
8
0
9
.60
7
0
9
.18
5
1
3
.69
2
0
1
.71
1
4
5
.39
4
9
.66
5
7
.19
2
4
.29
Multis
cale
Tsalli's
en
trop
y
6
3
7
.19
6
3
8
.20
6
1
5
.42
4
8
1
.38
4
2
5
.38
3
5
4
.12
2
6
3
.91
1
4
8
.62
Mu
ltiscale Ap
p
rox
i
m
ate
Entro
p
y
5
7
9
.57
4
1
5
.65
2
5
2
.37
9
2
.46
6
9
.20
4
9
.66
3
3
.38
1
9
.51
Multis
cale Pe
r
m
u
t
atio
n
entro
p
y
5
1
2
.33
4
0
8
.36
3
3
4
.95
3
6
0
.62
3
4
4
.47
2
7
5
.38
1
5
4
.25
4
0
.08
Multis
cale
W
av
ele
t entro
p
y
4
0
0
.53
3
2
2
.37
1
5
0
.48
4
1
.75
3
2
.13
2
4
.57
1
5
.22
1
0
.63
Multis
c
ale Spect
ra
l entro
p
y
3
7
1
.63
3
2
2
.34
2
8
7
.77
2
4
8
.40
2
1
0
.23
1
5
1
.73
8
1
.69
3
1
.33
Multis
cale Ren
y
i e
n
trop
y
3
6
0
.94
4
4
2
.09
6
2
0
.06
5
1
2
.14
4
3
4
.54
3
4
0
.49
2
3
3
.71
1
1
8
.13
Mutlis
cale Shan
n
o
n
entro
p
y
8
6
.68
5
7
.74
3
3
.48
1
5
.84
1
2
.18
9
.75
8
.35
3
.64
3.2.
Accur
ac
y T
esting u
sin
g Multil
ayer
Pe
rceptr
o
To
determ
ine
the
perform
ance
of
m
ulti
sca
le
entropy
in
the
lung
sou
nd
cl
assifi
cat
ion
,
we
us
ed
m
ul
ti
la
ye
r
perce
ptr
on
(ML
P
)
as
a
cl
assifi
er.
ML
P
is
one
var
ia
nt
of
arti
fici
al
neura
l
netw
orks
tha
t
have
su
pe
r
vised
le
ar
ning
pro
pe
rtie
s.
B
ecause
the
MLP
was
s
upe
rv
ise
d,
the
n
w
e
us
ed
t
he
N
-
f
old
c
ro
ss
-
valid
at
ion
(N
-
f
old
C
V)
f
or
t
he
trai
ni
ng
and
te
sti
ng
process
[33].
Be
cause
the
num
ber
of
data
in
each
cl
ass
is
18
to
22
data,
the
n,
the
N
-
fo
l
d
C
V
us
e
d
N
=
3
s
o
tha
t
each
dataset
would
ha
ve
6
–
8
dat
a.
Ta
ble
3
pr
ese
nts
it
s
r
esult.
Table
3
s
hows
that
the
m
ultiscale
pe
rm
utatio
n
e
ntr
opy
pr
oduce
d
th
e
highest
acc
ur
acy
i
n
m
os
t
of
the
scal
es.
Mult
isc
al
e
Perm
uta
ti
on
Entr
opy
(MPE)
pro
duced
t
he
lowe
r
accuracy
com
par
e
d
to
m
ulti
s
cal
e
sa
m
ple
entropy
(MSE)
on
the
s
cal
e
of
1
–
5
and
scal
e
of
1
–
4.
PE
ha
d
sever
al
ad
va
ntages
c
om
par
ed
with
Sam
pEn
:
PE
cal
culat
ed
the
per
m
utati
on
of
the
sequ
e
nce
patte
rn
of
the
s
ign
al
,
w
hile
the
Sa
m
pEn
cal
culat
ed
the
num
ber
of
the
sam
e
patte
rn
s o
f
t
he
sig
nal
with
a
ce
rtai
n
tolerance
.
T
his
i
m
pl
ie
s
that
the
Sam
pEn
cal
culat
ion
ta
kes
lo
ng
e
r
com
pu
ta
ti
on
ti
m
e
rather
than
the
PE.
Sim
i
l
arit
ie
s
betwee
n
those
two
pa
tt
ern
s
we
re
de
te
rm
ined
by
sever
al
par
am
et
ers,
suc
h
as
the
n
-
or
der
per
m
utati
on
on
PE,
m
ea
nwhile
,
in
Sa
m
pEn
,
they
wer
e
determ
ined
by
the
si
m
il
a
rit
ie
s toleran
ce
r
a
nd le
ng
t
h of patt
er
n
m
.
Table
3.
Acc
uracy
(
%)
for va
r
iou
s
m
ulti
scal
e
en
tr
opie
s usi
ng th
ree
-
f
old
c
r
os
s
-
validat
io
n
MS
Ent
rop
y
Scale us
e f
o
r
en
tro
p
y
m
e
asu
re
m
en
t
scale
1
-
20
scale
1
-
15
scale
1
-
10
scale
1
-
5
scale
1
-
4
scale
1
-
3
scale
1
-
2
s
cale
1
Multis
cale Sa
m
p
le
entro
p
y
8
5
.86
8
6
.87
8
6
.87
8
6
.87
8
2
.83
7
2
.73
6
9
.7
5
0
.51
Multis
cale
Tsalli's
en
trop
y
8
8
.89
8
9
.9
9
1
.92
7
9
.8
7
3
.74
7
2
.73
7
0
.71
6
9
.7
Multis
cale A
p
p
rox
i
m
ate
Entro
p
y
8
8
.89
8
8
.89
8
3
.84
7
9
.8
7
1
.72
6
5
.66
5
9
.6
5
1
.52
Multis
cale Pe
r
m
u
t
ati
o
n
entro
p
y
9
6
.97
9
6
.97
9
7
.98
9
3
.94
9
2
.93
8
9
.9
8
6
.87
5
9
.6
Multis
cale
W
av
ele
t entro
p
y
7
4
.75
6
8
.69
6
7
.68
7
0
.71
7
0
.71
7
1
.72
6
6
.67
4
3
.43
Multis
cale Spect
ra
l entro
p
y
9
3
.94
9
5
.96
9
5
.96
9
0
.91
8
9
.9
8
0
.81
7
5
.76
4
9
.49
Multis
cale Ren
y
i e
n
trop
y
8
8
.89
9
0
.91
9
3
.94
78
.79
7
0
.71
7
4
.75
6
9
.7
6
8
.69
Mutlis
cale Shan
n
o
n
entro
p
y
8
3
.84
8
4
.85
8
1
.82
7
6
.77
7
6
.77
6
6
.67
5
8
.59
3
7
.37
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
25
02
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
984
–
994
992
Table
4
s
ummari
zes
the
best
res
ults
f
or
eac
h
of
m
ulti
scal
e
entr
op
y
i
nclu
ding
a
scal
e
t
ha
t
produce
s
the
highest
acc
ur
acy
,
the
acc
ur
acy
of
the
si
ng
le
s
cal
e
a
nd
aver
a
ge
c
om
pu
ta
ti
on
tim
e
for
a
sin
gle
data.
Test
s
wer
e
pe
rfor
m
ed
on
a
PC
with
the
sp
eci
ficat
ion
of
In
te
l
(
R)
Co
re
(TM
)
i3
-
3
22
0
CP
U
@
3.3
0GHz
,
4
GB
of
RAM.
Table
4
disp
la
ys
that
the
co
m
pu
ta
ti
on
al
tim
e
fo
r
the
ov
e
rall
te
chn
iq
ue
was
acce
pta
ble
excep
t
f
or
t
he
m
ul
ti
scal
e
approxim
at
e
entro
py.
It
t
ook
a
ve
ry
long
ti
m
e,
so
it
was
le
ss
r
el
ia
ble
us
e
d
in
real
ap
plica
ti
ons.
A
sel
f
-
m
at
ching
process
cause
d l
ong
c
om
pu
ta
ti
on
ti
m
e o
ccurr
ed
in
ApE
n
c
om
pu
ta
ti
on
.
Table
4.
Scal
e
for
the
h
i
gh
e
st
accuracy a
nd c
om
pu
ta
ti
on
ti
m
e
Multis
cale
Entro
p
y
Bes
t scale
Hig
h
est
Accurac
y
(
%)
Accurac
y
f
o
r
sin
g
le scale (
%)
Av
erage co
m
p
u
tati
o
n
ti
m
e
(s)
Multis
cale Sa
m
p
le
entro
p
y
1
-
5
8
6
.87
5
0
.51
2
7
.69
Multis
cale
Tsalli's
en
trop
y
1
-
10
9
1
.92
6
9
.7
0
.69
Multis
cale A
p
p
rox
i
m
ate
Entro
p
y
1
-
15
88
.89
5
1
.52
2
0
8
7
.4
7
Multis
cale Pe
r
m
u
t
atio
n
entro
p
y
1
-
10
9
7
.98
5
9
.6
1
.61
Multis
cale
W
av
ele
t entro
p
y
1
-
20
7
4
.75
4
3
.43
1
.00
Multis
cale Spect
ra
l entro
p
y
1
-
10
9
5
.96
4
9
.49
0
.93
Multis
cale Ren
y
i e
n
trop
y
1
-
10
9
3
.94
6
8
.69
0
.70
Mutlis
cale Shan
n
o
n
entro
p
y
1
-
15
84
.85
3
7
.37
0
.84
Table
4
s
hows
that
the
MPE
pro
du
ce
d
the
hi
gh
est
accu
rac
y
on
the
scal
e
1
-
10
with
an
a
ccur
acy
of
97.98%
.
So
m
e
m
ulti
scal
e
ent
ropy
only
us
e
d
scal
e
1
-
5
to
pro
du
ce
th
e
hi
gh
e
st
accuracy
,
but
sti
ll
below
the
MPE.
He
nce,
it
was
no
t
the
best
ch
oic
e
.
Re
gardin
g
com
pu
ti
ng
ti
m
e,
the
MPE
r
equ
i
red
a
reas
on
a
ble
com
pu
ti
ng
ti
m
e,
i.e.,
1.6
2
s
f
or
t
he
si
ng
le
da
ta
.
Me
an
wh
il
e,
m
ulti
scal
e
Ap
E
n
to
ok
a
lo
ng
c
om
pu
ta
ti
onal
tim
e
.
This
was
cause
d
by
the
cal
culat
ion
of
the n
um
ber
of
si
m
i
lar
patte
rn
s o
f
da
ta
carried
al
on
g
the
seq
uen
ce
of
t
he
sign
al
s
an
d
th
e
e
m
erg
ence
of
sel
f
-
m
at
ching
.
Mult
isc
al
e
Sam
pEn
(MSE
)
ov
e
rcam
e
th
e
weakness
m
ulti
scal
e
ApEn
, so MSE
r
e
qu
ire
d sh
or
t
er c
om
pu
ta
ti
on tim
e than
the
m
ul
ti
scal
e A
pE
n.
Mult
isc
al
e
SEN
pr
oduce
d
t
he
highest
ac
cur
acy
a
fter
m
ul
ti
scal
e
PE
as
so
m
e
signa
l
diff
e
rence
s
betwee
n
the
cl
asses
cou
l
d
be
seen
cl
early
in
the
sp
ect
ra
l
fr
eq
uen
cy
.
H
ow
e
ve
r,
m
ulti
s
cal
e
W
E
c
ou
l
d
not
pro
du
ce
a
quit
e
high
acc
ur
ac
y
for
bei
ng
inf
luence
d
by
the
m
oth
er
wa
velet
sel
ect
ion
,
th
e
sam
pling
fr
e
qu
e
nc
y
and
t
he
le
vel
o
f
dec
om
po
sit
io
n.
T
hese
facto
r
s
w
ou
l
d
af
fect
the
f
or
m
ed
sub
-
ba
nd.
Th
e
se
le
ct
ion
of
DB
2
and
deco
m
po
sit
io
n
le
vel
7
did
no
t
pro
du
ce
a
high
acc
ur
ac
y.
A
pprop
riat
e
W
a
velet
filt
er
an
d
t
he
le
ve
l
of
deco
m
po
sit
io
n sel
ect
ion
will
be t
he next
rese
arch
.
Sh
E
N,
RE
N,
a
nd
T
E
ha
ve
a
s
i
m
i
la
r
form
.
Sh
E
N
is
a
s
peci
al
case
of
RE
N
a
nd
TE.
S
hE
N
is
RE
N
or
TE
with
orde
r
q
=
1.
T
he
sel
e
ct
ion
of
orde
r
q
in
RE
N
a
nd
TE
will
pro
duc
e
the
di
ff
e
ren
t
accuracy.
The
us
e
of
TE
with
a
dif
fer
e
nt
orde
r
f
or
l
ung
s
ound
analy
sis
was
pr
es
ente
d
in
[31].
T
he
re
sul
t
sh
owed
that
non
-
extensi
vity
o
r
de
r q =
2 pro
du
c
es the
best
featu
re
for t
he
l
ung
s
ound a
naly
sis.
The
e
ntr
opy
m
easur
em
ent
i
n
the
m
ulti
sca
le
sche
m
e
increased
lung
sound
cl
assifi
cat
ion
accu
racy
com
par
ed
with
the
sing
le
scal
e
schem
es.
Seve
ral
pr
e
viou
s
s
tud
ie
s
us
ed
only
on
e
e
ntropy
values
c
ouple
d
with
the
oth
e
r
featu
res
of
the
lu
ng
so
un
d
cl
assifi
cat
ion
.
Sam
ple
entropy
was
c
om
bin
ed
with
sk
ew
ness
,
kurt
os
is
,
and
la
cu
nar
it
y
repor
te
d
by
M
ondal
et
al
[
34]
.
The
res
ulti
ng
acc
ur
acy
wa
s
92.
86%
for
norm
al
and
ab
norm
al
lung
s
ound.
M
eanwhil
e,
kurt
os
is,
m
ean
cr
ossi
ng
i
rr
e
gula
r
it
y
and
Re
nyi
entr
op
y
us
e
d
in
[
35]
.
T
he
re
su
lt
s
ob
ta
ine
d
t
he
ac
cur
acy
of
95.
1%
f
or
the
trai
ni
ng
data
a
nd
93.
5%
for
te
sti
ng
data.
Tsal
li
s
entr
op
y
a
nd
25
othe
r
char
act
e
risti
cs
us
e
d
by
Morill
o
et
al
.
f
or
l
ung
s
ound
from
congesti
ve
obs
tructi
ve
pulm
o
nar
y
disease
(
COPD)
patie
nts
[
36
]
.
Me
anwhil
e,
Jin
et
al
.
us
e
d
s
a
m
ple
entropy
as
the
featu
re
s
of
l
ung
s
ound
extr
act
ed
f
r
om
the
sh
ort
-
ti
m
e
Fo
uri
er tra
nsfo
rm
(
STFT
)
[
37]
.
The
sam
e
entr
op
ie
s
a
s
in
t
his
pa
per
we
re
us
e
d
in
[
38
]
.
Sing
le
e
ntr
opy
only
pro
du
ce
d
69.7
%
of
accuracy
us
i
ng
TE.
Com
po
sit
e
of
seve
n
e
ntropies
increa
se
d
the
accu
racy
becam
e
94
.9
%.
Com
pu
ta
ti
on
of
seve
n
entr
opie
s
was
m
or
e
co
m
plex
than
on
e
entr
op
y
with
the
m
ult
isc
al
e
process.
C
ompare
d
wit
h
this
pap
e
r
,
94.9 %
was
ac
hieve
d usin
g
s
even feat
ur
es
wh
il
e
97.9 %
w
as ac
hieve
d u
sing t
en feat
ur
es.
Ov
e
rall
the
stu
dies
we
re
co
nducte
d
on
si
ngle
scal
e
sign
al
.
W
e
co
uld
see
that
entr
op
y
s
ti
ll
need
ed
oth
e
r
feat
ur
es
f
or
l
ung
sou
nd
cl
assifi
cat
ion
.
Wh
il
e
on
the
m
ul
ti
scal
e
entro
py,
we
only
use
d
only
entr
opy
tha
t
cal
culat
ed
on
a
sign
al
with
a
differe
nt
sca
le
.
Alth
ough
it
was
no
t
direc
tl
y
co
m
par
able
,
m
ulti
sca
le
entropy
pro
vid
e
d
m
or
e
prom
isi
ng
res
ults.
Di
rect
co
m
par
ison
s
with
the
sam
e
dataset
m
a
y
be
m
ade
on
the
rese
arch
i
n
the futu
re.
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
Compari
son
of
Mu
lt
isc
ale E
nt
ro
py
Tec
hniq
ue
s for Lu
ng
Sound
Cl
as
sif
ic
ation
(
Ac
hma
d
R
izal
)
993
So
m
e
m
ult
isc
a
le
entropies
we
re
pro
posed
by
so
m
e
researchers
f
or
s
olv
in
g
var
i
ou
s
case
s
in
previ
ous
stud
ie
s.
Mult
is
cal
e
sam
ple
entropy
w
hich
w
as
m
or
e
kn
own
as
m
ulti
scale
entr
opy
(MS
E)
was
pro
pos
ed
by
Costa
et
al
[9
]
.
Mult
isc
al
e
perm
uta
ti
on
ent
r
opy
was
us
e
d
f
or
quantify
in
g
s
ign
al
com
plexi
ty
pr
op
os
ed
by
Azis
and
Ar
if
[11].
In
this
pa
pe
r,
we
intr
oduce
d
sever
al
m
ulti
sc
al
e
entropies
f
or
sig
nal
com
plexity
m
easur
em
ent.
Fo
r
exam
ple
m
ulti
scal
e
Tsal
lis
entr
op
y,
m
ult
isc
al
e
Re
nyi
entr
op
y,
m
ulti
scal
e
sp
ect
r
al
ent
ropy
an
d
m
ultiscale
wav
el
et
ent
ropy.
Mult
isc
al
e
entropy
pro
vi
des
bette
r
pe
rfor
m
ance
for
lung
sou
nd
analy
sis
com
par
ed
to
entr
op
y i
n
sin
gl
e scal
e [38].
4.
CONCL
US
I
O
N
Lu
ng
sou
nd
i
s
one
of
t
he
bio
lo
gical
sig
nals
w
hich
ha
ve
in
form
at
io
n
a
bout
the
he
al
th
of
th
e
resp
i
rator
y
sys
tem
.
To
reduc
e
su
bject
ivit
y
in
the
eval
uation
of
l
ung
s
ounds
t
a
va
riet
y
of
dig
it
al
sign
al
processi
ng
te
c
hn
i
qu
e
s
for
au
tom
a
ti
c
lun
g
s
ound
cl
assifi
cat
ion
ha
ve
bee
n
dev
el
opin
g.
On
e
of
the
m
et
hods
us
e
d
in
l
ung
s
ound
sig
nal
processin
g
is
t
he
entr
opy
m
eas
ur
em
ent
m
e
tho
d.
I
n
t
his
re
s
earch
,
we
c
ompare
d
m
ul
ti
scal
e
entr
op
y
with
s
om
e
diff
ere
nt
entr
op
y
m
easur
em
ent
te
chn
iq
ues
fo
r
lu
ng
s
ound
featu
re
extra
ct
ion.
The
res
ults
showe
d
t
hat
the
m
ul
ti
scal
e
perm
uta
ti
on
e
ntropy
pr
oduce
d
t
he
highest
acc
ur
acy
of
97
.98
%
with
the
scal
e
of
1
-
10.
T
hese
res
ul
ts
are
m
uch
hig
he
r
t
han
usi
ng
Pe
rm
utati
on
entr
op
y
i
nd
i
vid
ually
on
the
ov
e
rall
sign
al
.
S
om
e
entr
op
y
m
easur
em
ent
par
am
e
te
rs
that
co
uld
be
al
te
red
t
o
change
s
uc
h
as
Tsal
li
s
entropy
or
Re
nyi
entr
op
y
for
a
diff
e
re
nt
ord
er
of
q.
Pe
r
form
ance
of
m
ulti
scal
e
entropy
with
va
rio
us
entr
opy
pa
ra
m
et
ers
can
be
e
xam
ined
in
s
ub
se
que
nt r
esea
rc
h.
ACKN
OWLE
DGE
MENTS
This
w
ork
ha
s
been
fina
nc
ia
ll
y
su
ppor
te
d
by
Mi
nistry
of
Re
sea
rch,
Tech
no
l
og
y,
and
Highe
r
Ed
ucati
on of R
epublic
of
Ind
onesi
a
under Pe
nelit
ia
n
Dise
rtasi
D
okt
or
Sche
m
e n
o:
014/P
N
LT3/PPM/
2018.
REFERE
NCE
S
[1]
Pasterka
m
p
H, K
raman
SS
,
W
odic
ka
GR.
R
espi
rat
or
y
Sounds
.
A
m J
R
espir
Crit
Care
Me
d
.
1997;
156(3):974
–
987.
[2]
Spiet
h
PM
,
Zha
ng
H.
Anal
y
z
in
g
Lung
Crac
kle
Sounds
:
Stet
ho
scope
s
And
Be
y
ond
.
Inte
nsiv
e
Care
Me
dic
in
e
.
2011;37(8):
1238
–
9.
[3]
Riz
al
A,
Hida
yat
R,
Nugroho
HA
.
Mult
iscal
e
Hjorth
Descrip
tor
for
Lung
So
und
Classifi
cat
i
on
.
Inte
rna
ti
on
a
l
Confer
ence
on
Scie
n
ce a
nd
Tech
nolog
y
20
15.
Yo
g
y
ak
arta. 2015;
1:160008.
[4]
Chen
X,
Shao
J,
Long
Y,
Qu
e
C,
Zh
ang
J,
Fang
J.
Ide
ntifi
ca
t
ion
of
Vel
cr
o
Ral
es
Based
on
Hilbe
rt
–
Hua
ng
Tra
nsform
.
Ph
ys
A
S
tat Me
ch
i
ts
Appl
.
2014;
401(
8):34
–
44.
[5]
Gnite
ck
i
J,
Mous
savi
Z,
Pasterk
a
m
p
H.
Classifi
ca
ti
on
of
lung
sounds
during
bronc
hial
provo
cat
ion
using
wav
ef
orm
fract
al
dimens
io
ns
.
Annual
Inte
r
nat
ion
al
Confer
e
nce
of
the
IEEE
Engi
nee
r
ing
in
Medic
in
e
and
Bi
olog
y
Soc
iet
y
.
2004;
3844
–
384
7.
[6]
Ma
y
org
a
P,
Dru
zga
lski
C
,
Gonz
al
e
z
OH
,
Lop
ez
HS
.
Modif
ie
d
Classifi
cation
of
Norm
al
Lung
Sounds
Appl
yi
n
g
Quantil
e
Ve
c
tors
.
Proce
edi
ngs
o
f
the
Annu
al
In
t
ern
ational
Conf
e
ren
ce
of
the
I
E
EE
Eng
ine
er
ing
in
Medicine
and
Biol
og
y
So
ciet
y
,
EMBS
.
2012;
4
262
-
4264.
[7]
Chen
C,
Huang
W
,
Ta
n
T,
Chan
g
C
-
C,
Ch
ang
Y
.
Us
ing
K
-
Nea
r
e
st
Neighbor
Cl
assific
a
ti
on
to
Dia
gnose
Abnorm
al
Lung
Sounds
.
S
e
nsor
s
.
2015
Jun
4;15(6):
13132
–
5
8.
[8]
Kanda
sw
am
y
A,
Kum
ar
CS
,
Ra
m
ana
tha
n
RP
,
J
a
y
ara
m
an
S,
Ma
lmurugan
N.
Ne
ura
l
cl
assifi
catio
n
of
lung
sound
s
using wa
velet
co
eff
icient
s.
Comp
ut
B
iol
M
ed
.
200
4;34(6):
523
–
537
.
[9]
Costa
M,
Gold
ber
ger
AL
,
Pen
g
CK.
Multi
sc
a
le
En
trop
y
Ana
l
y
s
is
of
Biol
ogi
ca
l
Sign
al
s.
Ph
ys
Rev
E
-
Sta
t
Nonli
near,
Soft
Matt
er
Phy
s
.
20
05;71:
1
–
18.
[10]
Hum
ea
u
-
heur
ti
e
r
A. The
Multi
sc
al
e
Ent
rop
y
Alg
orit
hm
and Its V
ari
an
ts:
A R
evi
e
w.
Ent
ropy
.
201
5;17:
3110
–
3123.
[11]
Aziz
W
,
Arif
M.
Mult
isca
le
P
ermutati
on
Ent
rop
y
of
Phy
siolog
ic
al
Time
Serie
s
.
In:
9th
Inte
rn
at
i
onal
Multi
top
ic
Confer
ence, IEE
E
INM
IC
2005.
Kara
chi 2005;
1
-
6.
[12]
W
ang
G
-
J,
Xie
C,
Han
F.
Multi
-
Scal
e
Approxi
m
at
e
Ent
rop
y
A
naly
s
is
of
Foreign
Exc
hange
M
ark
et
s
Eff
i
cienc
y
.
Syst
Eng
Proc
ed
ia
.
2012
;3(2011)
:201
–
208.
[13]
Xie
H,
Chen
W
,
He
W
,
Li
u
H.
Com
ple
xity
anal
y
sis
of
the
biomedical
signal
usi
ng
fuz
z
y
ent
rop
y
m
ea
surem
ent.
Appl
So
ft
Comp
ut
.
2011
;11(2):
2
871
–
2879.
[14]
Charl
eston
-
Vi
ll
a
lobos
S,
Albuer
ne
-
Sanchez
L,
Gonza
le
z
-
C
amare
na
R,
Mej
ia
-
Avila
M,
Carr
il
lo
-
Rodrigu
ez
G,
Alja
m
a
-
Corra
l
es
T.
Li
ne
ar
and
Nonlinear
Anal
y
si
s
of
Base
Lung
Sound
in
Ext
rins
ic
All
erg
i
c
Alve
oli
ti
s
Pat
ie
n
ts
i
n
Com
par
ison t
o
Hea
lt
h
y
Sub
je
c
ts.
Me
thods I
nf
M
ed
.
2013
;52(3):
2
66
–
76.
[15]
W
u
N,
Guo
Y,
W
ei
Y,
Fan
S,
Li
X.
Design
of
New
Tra
nsform
er
Protec
ti
on
Devic
e
Based
o
n
W
ave
le
t
En
er
g
y
Ent
rop
y
-
Neura
l N
et
work T
heo
r
y
and.
TEL
KOMNIKA.
2013
;11(11
):6746
–
6754.
[16]
Xu
X,
W
ang
S,
Li
u
J,
W
u
Z.
Fa
n
Fault
Di
agnosi
s
Based
on
W
av
el
e
t
Pack
et
and
Sam
ple
Ent
rop
y
.
TEL
KOMNIKA
.
2013;11(6):
3451
–
3462.
[17]
Dong
Z,
Ti
an
X,
Ze
ng
J.
Mec
hanica
l
Faul
t
Diagnosis
Based
on
LMD
-
Ap
proximate
Ent
ro
p
y
and
LSS
VM
.
TEL
KOMNIKA
.
2013;11(2):
803
–
808.
Evaluation Warning : The document was created with Spire.PDF for Python.