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.
10
,
No.
4
,
A
ugus
t
2020
,
pp. 352
8~35
36
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
4
.
pp3528
-
35
36
3528
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
ELM
and K
-
nn
m
ac
hin
e le
ar
nin
g in
classi
fication
of
B
reat
h sound
s s
ign
als
Z
.
Neil
i
,
M.
F
ez
ari
, A
. Red
j
at
i
Depa
rtment
o
f
E
le
c
troni
c
,
LAS
A L
abor
at
or
y
,
B
ad
ji
Mokht
ar
Univ
ers
ity
Annab
a, Alge
ri
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
7
, 2
019
Re
vised Jan
6
,
2020
Accepte
d
Ja
n
29
, 2
020
The
ac
qu
isit
ion
of
Brea
th
sounds
(BS)
signal
s
from
a
hu
m
an
respir
at
o
r
y
s
y
stem
with
a
n
el
e
ct
roni
c
st
et
hoscope
,
pro
vide
and
of
fer
prom
ine
nt
informati
on
which
hel
ps
the
doct
ors
to
dia
g
nosis
and
cl
assificat
ion
of
pulmonar
y
d
ise
ase
s.
Unfortuna
t
ely
,
thi
s
BS
si
gnal
s
with
o
ther
biol
ogi
ca
l
signal
s
hav
e
a
n
on
-
stat
ion
ar
y
na
ture
ac
cor
d
ing
t
o
the
v
ariati
on
of
the
lung
volume,
and
this
nat
ure
m
ake
s
it
diff
ic
ul
t
to
an
aly
z
e
and
class
if
y
b
et
we
en
seve
ral
d
isea
ses.
In
thi
s
stud
y
,
we
were
foc
use
d
on
compari
ng
the
ab
ili
t
y
of
the
ext
reme
le
arn
ing
m
ac
hin
e
(EL
M)
and
k
-
nea
rest
nei
ghb
our
(K
-
nn)
m
ac
hine
l
ea
rn
in
g
al
gori
thms
in
the
class
ifi
c
at
ion
of
adve
nt
it
i
ous
and
norm
al
bre
at
h
sounds
.
T
o
do
so,
the
empirical
m
ode
de
co
m
positi
on
(EMD)
was
used
in
thi
s
work
to
ana
l
y
z
e
BS
,
th
is
m
et
hod
is
rar
ely
used
in
th
e
br
ea
th
sounds
ana
l
y
sis.
Af
te
r
t
he
EMD
dec
om
positi
on
of
th
e
signal
s
int
o
In
tr
insic
Mode
Functi
ons
(IMF
s
),
t
he
Hjor
th
des
cri
ptors
(Act
ivi
t
y
)
and
Perm
uta
tion
Ent
r
o
p
y
(PE)
feature
s
we
re
ex
tra
c
te
d
fro
m
ea
ch
IMF
s
an
d
combined
for
cl
assifi
ca
t
ion
stage
.
The
stud
y
has
found
tha
t
t
he
combinati
on
of
fea
tur
es
(acti
vity
and
PE
)
y
i
el
d
ed
an
accu
racy
of
90.
71%
,
95%
using
EL
M
and
K
-
nn
res
pec
t
ive
l
y
i
n
bina
r
y
class
ifi
cation
(norm
al
and
abnor
m
al
bre
at
h
sounds
),
and
83.
57%
,
86.
42%
in
m
ult
i
cl
ass classificat
i
on
(five classes).
Ke
yw
or
d
s
:
Em
pirical
m
od
e d
ec
om
po
sit
ion
Extrem
e lea
rn
ing m
achine
Hjort
h desc
ript
or
s
K
-
near
e
st nei
ghbo
ur
Per
m
utati
on
ent
ropy
Copyright
©
202
0
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
:
N.
Zaka
ria
,
Dep
a
rtm
ent o
f E
le
ct
ro
nic
,
Faculty
of S
ci
e
nces,
Lab
or
at
oi
re
d’Au
t
om
atiqu
e e
t Sign
a
ux
de A
nn
a
ba
(L
ASA)
,
Ba
dj
i M
okh
ta
r
Un
i
ver
sit
y A
nnaba
,
P.O. Bo
x 1
2,
2300
0 Anna
ba,
Alge
ria
.
Em
a
il
:
z
aki.n
el
i@y
ahoo.fr
1.
INTROD
U
CTION
R
ecentl
y
m
an
y
research
e
s
in
the
Breat
h
Sounds
(B
S)
areas
is
cond
ucted
by
app
l
yi
ng
m
any
te
chn
iq
ues
f
or
analy
zi
ng
a
nd
cl
assify
a
re
sp
irat
ory
sig
na
l
in
orde
r
to
di
agnosis
the
pulm
on
ary
path
ology
,
su
c
h
as
(A
st
hm
a,
COPD,
P
ne
um
on
ia
…).
H
ow
e
ve
r,
the
ac
qu
isi
ti
on
of
l
ung
sou
nds
play
s
an
im
po
rta
nt
ro
le
to
detect
an
d
i
de
ntific
at
ion
of
pulm
on
ary
di
seases.
T
he
do
ct
or
s
li
ste
nin
g
to
t
he
lu
ng
sou
nds
th
rou
gh
the
ste
th
os
co
pe
placed
on
t
he
chest
or
poste
rior
to
t
he
patie
nt
w
hich
is
s
uffer
i
ng
from
a
ty
pe
of
lu
ng
disease,
bu
t t
he
pro
ble
m
is
so
m
et
i
m
e
s the d
eci
sion
of
the docto
rs
con
ce
r
ning ty
pes
of
p
at
holo
gy
is n
ot
accurate
, th
is i
s
du
e
to
m
any
reasons
s
uch
a
s
a
few
ex
peri
ences
c
oncer
ni
ng
au
sculta
ti
on
an
d
diag
nos
is.
To
ad
dr
es
s
this
pro
blem
a
research
st
ud
y
ha
s
been
wide
ly
con
duct
ed
in
this
area
f
ocus
on
th
ree
pr
inci
pal
do
m
ai
ns
(ti
m
e d
om
ai
n,
f
re
qu
e
ncy
do
m
ai
n
an
d
ti
m
e
-
frequ
e
ncy
do
m
ai
n).
Breat
h
s
ound
is
a
com
plex
sig
nal
[1
]
su
c
h
ot
h
er
bi
o
lo
gi
cal
sig
na
ls,
ha
ve
a
no
n
-
li
nea
r
a
nd
non
-
sta
ti
on
a
ry
natu
re
,
t
hese
pro
per
ti
es
of
t
he
sig
nal
le
ad
it
to
assesse
d
by
di
ff
e
ren
t
t
echn
i
qu
e
s
in
s
ign
al
processi
ng
ha
ve
the
sam
e
pro
per
ty
of
t
he
br
eat
h
si
gnal
s.
H
ow
e
ver,
these
te
ch
niq
ue
s
us
ed
f
or
thre
e
tran
s
form
ation
do
m
ai
ns
wh
ic
h
are
m
entione
d
previ
ously
.
I
n
[
2]
Islam
,
A
et
al
the
arti
fici
al
neu
ral
network
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
ELM
and
K
-
nn
mach
i
ne
le
arn
ing
i
n
cl
assi
fi
cation of
Breat
h so
unds si
gnals
(
Z.
Ne
il
i
)
3529
(ANN)
an
d
suppo
rt
v
ect
or
m
achine
(
SV
M
)
cl
assifi
ers
ha
ve
been
use
d
f
or
cl
assify
ing
norm
al
and
ast
hm
at
ic
su
bject
s
with
,
sp
ect
ral
s
ubba
nd
was
e
xtract
ed
f
r
om
the
lun
g
sou
nd
cy
cl
e
,
w
it
h
a
m
axi
m
u
m
cl
assifi
c
at
ion
accuracy
of
89.
2%
a
nd
93.
3%
by
th
e
A
N
N
a
nd
S
VM
c
la
ssifie
rs,
res
pe
ct
ively
,
Mondal
,
A
[
3]
et
a
l
app
ly
the
em
pirical
m
od
e
deco
m
po
sit
ion
t
o
lu
ng
sounds
fo
c
us
e
d
on
patte
r
n
re
cogniti
on
al
go
rithm
s
fo
r
cl
as
sify
ing
into
pulm
on
ary
d
ysf
un
ct
i
on
with
a
n
acc
ura
cy
of
94.
16
%
,
in
[4
]
A.
Ri
zal
et
al
cl
assify
the
lu
ng
s
ound
s
us
i
ng
Tsal
li
s
Entr
opy
and
us
in
g
M
LP
cl
assifi
er
w
it
h
an
accu
rac
y
95
.
35%,
Pa
nc
al
di,
F
et
al
[
5]
dia
gnos
is
t
he
lu
ng
diseases
(inter
sti
ti
al
lun
g
diseases
)
by
us
in
g
em
pirical
obser
vatio
n
as
pro
po
se
d
s
ol
ution
with
a
n
overall
accuracy
of
90.
0%,
A
.Chee
m
a,
M.Sing
h
[6
]
us
e
an
E
MD
m
et
ho
d
f
or
detect
Psyc
ho
l
og
ic
al
stres
s
fr
om
phonoca
r
diogr
aph
y
sig
nal
the
ave
rag
e
a
ccur
acy
of
93.
14%
to
cl
assify
ing
stre
ss
ed
an
d
no
n
-
st
resse
d,
in
[7
]
R.
Pala
ni
app
a
n
cl
assify
a
pu
lm
on
ary
s
ign
al
us
i
ng
A
ut
or
e
gr
essi
ve
Coef
fici
ents
an
d
k
-
Near
est
Neighb
or
as
a
cl
assifi
er
with
an
accu
ra
cy
of
95.18%
.
In
this
w
ork
w
e
analy
zed
a
breat
h
so
un
ds
sign
al
s
usi
ng
em
pirical
m
od
e
deco
m
po
sit
ion
with
H
j
ort
h
desc
ript
ors
(
Acti
vity
)
a
nd
Perm
utati
on
e
ntropy
as
fe
at
ur
es,
we
re
e
xtracte
d
from
each
IMFs
pro
duce
d
by
EMD,
fi
nally
,a
com
par
at
iv
e
stud
y
has
be
en
asses
sed
betwee
n
a
n
ex
trem
e
le
arn
in
g
m
achine
(ELM
)
and
K
-
near
est
nei
ghbor
(
K
-
N
N)
f
or
disti
nguish
i
ng
be
tween
no
rm
al,
a
nd
adv
e
ntit
iou
s
re
sp
irat
ory
s
ounds
.
2.
RESEA
R
CH MET
HO
D
Our
wor
k
was
div
i
ded
into
two
pri
ncipal
sta
ges
nam
el
y
(Multi
cl
ass
cl
assifi
cat
ion
sta
ge,
Bi
nar
y
cl
assifi
cat
ion
sta
ge)
.
T
he
fou
r
ste
ps
propose
d
for
both
sta
ge
stud
y
nam
ely
(d
at
abase,
pre
-
proce
ssin
g,
f
eat
ur
e
extracti
on
a
nd
cl
assifi
cat
ion
)
are
pr
ese
nted
in
Fi
gure
1
(
a
)
f
or
m
ulti
c
la
ss
cl
assifi
cat
ion
,
an
d
the
sec
ond
sta
ge
is
represe
nted
i
n
Figure
1
(b
)
f
or
b
ina
ry cla
ssifi
cat
ion
.
(a)
(b)
Figure
1
.
Tw
o pr
i
ncipal sta
ge of
the
br
eat
h
s
ounds si
gn
al
cl
assifi
cat
ion
(a)
m
ulti
cl
ass cla
ssific
at
ion
(b) binary
cl
assi
f
ic
at
ion
2.1. D
atabase
In
this
pap
e
r
t
he
data
base
of
br
eat
h
s
ounds
sign
al
s
use
d
f
or
a
naly
sis
are
the
R.A.
L
.E
(
Re
sp
irat
io
n
Acousti
c
Lab
or
at
or
y
En
vir
on
m
ent)
Lu
ng
So
un
ds
,
is
the
only
com
m
ercial
ly
avail
able
database
,
is
an
edu
cat
io
nal
program
to
hel
p
do
ct
or
s
a
nd
res
earche
rs
i
n
res
pirato
ry
sig
nal
s
proce
ssin
g
a
r
ea
offe
r
m
or
e
t
han
50
br
eat
h
s
ounds
we
re
recorde
d
us
in
g
a
co
ntact
acce
le
r
om
et
er
(S
ie
m
e
ns
-
EMT
25C)
cov
e
rin
g
nor
m
al
and
abno
rm
al
respi
rator
y
sou
nd
s
[
8
]
a
re
sam
pled
at
10,
240Hz.
As
this
da
ta
base
(R.
A.L.E)
has
a
fe
w
data,
therefo
re
to
e
ns
ure
t
he
c
re
di
bili
ty
of
this
com
par
at
ive
st
ud
y
we
us
e
d
ano
t
her
data
wer
e
colle
ct
ed
fro
m
the inter
net:
-
The A
us
culta
ti
on A
s
sist
ant,
2015
[
9,
10
]
-
Arnall
, 201
5
[
11
]
-
The
C
D of t
he boo
k [
12
]
In
al
l
a
75
brea
th
sounds
div
i
ded
i
nto
fi
ve
cl
asses
(No
rm
al
bro
nch
ia
l,
Wheez
e,
Crackle,
Pleural
r
ub,
Strid
or)
wer
e
us
e
d
in
our
stud
y,
eac
h
sou
nd
is
an
eff
ect
of
par
ti
cula
r
disease
su
c
h
as
Wh
eeze
in
dicat
e
that
the
patie
nt
suf
fer
in
g
f
r
om
as
thm
a
and
COPD
(C
hro
nic
Ob
st
ru
ct
ive
Pulm
on
ary
disea
ses),
c
rack
le
i
nd
ic
at
e
pn
e
um
on
ia
or l
ung
ca
nce
r.
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.
10
, No
.
4
,
A
ugus
t
2020
:
3528
-
3536
3530
2.2.
Br
ea
th
sounds
p
re
-
proc
essing
Breat
h
sou
nds
sign
al
s
are
sub
j
ect
to
seve
ral
artefact
s
su
c
h
as
hear
t
s
ound
s
and
noise
w
hi
ch
si
m
ulate
real
-
li
fe
co
nd
i
t
ion
s.
T
he
bre
at
h
so
un
ds
sign
al
s
(R.
A.L.
E)
that
hav
e
been
filt
ered
by
a
hig
h
-
pa
s
s
filt
er
with
7.5
Hz
by
1s
t
order
B
utterw
or
t
h
to
r
e
m
ov
e
DC
off
set
,
and
a
lo
w
-
pass
filt
er
at
2.5
Hz
by
8
th
order
Butt
erwor
t
h
[
13,
14
]
,
an
d
c
oncer
ning
the
da
ta
wer
e
colle
c
te
d
we
ap
ply
a
m
ean
and
am
plit
ud
e
norm
alizat
ion
to
re
du
ce
the
e
ff
ect
of
hea
rt
s
ounds.
Finall
y,
al
l
sa
m
ples
ar
e
dow
ns
am
pled
to
8000
Hz
sam
pling
fr
e
qu
ency
accor
ding
to
C
ORSA
(co
m
pute
rized
res
pirat
or
y
sou
nd
a
naly
sis)
[1
5
]
,
in
this
stud
y,
the
16
-
bit
reso
luti
on
a
nd
o
ne
r
es
pirat
or
y
cyc
le
are use
d.
2.3. Empi
ri
ca
l mode
dec
om
po
siti
on
The
Em
pirical
Mo
de
Decom
po
sit
ion
(E
MD)
m
et
ho
d
is
a
ne
w
a
da
ptive
si
gn
al
ti
m
e
-
fr
e
qu
e
ncy
processi
ng
m
eth
od
pr
opos
e
d
by
NE
H
ua
ng
in
1998
by
N
ASA
an
d
ot
he
rs
[
16
]
.
It
is
e
sp
eci
al
ly
su
it
able
for
n
onli
nea
rity
,
analy
sis
an
d
pr
ocessin
g
of
no
n
-
sta
ti
onary
si
gn
al
s
.
T
he
Hil
ber
t
tra
nsfo
rm
trans
form
s
the
well
-
known
H
il
be
rt
-
Hu
a
ng T
ransf
orm
(
HH
T
).
EMD
is
act
ually
a
m
et
ho
d
of
de
com
po
sin
g
sig
nals.
It
is
con
sist
ent
wit
h
the
co
re
i
de
a
of
F
ourier
trans
form
and
wa
velet
tran
sform
.
Ever
y
one
wan
ts
t
o
deco
m
po
se
th
e
sig
nal
into
a
supe
rpositi
on
of
ind
e
pende
nt
c
om
po
nen
ts,
onl
y
the
F
ourier
trans
f
or
m
and
the
wavel
et
tra
ns
f
or
m
it
is
re
qu
i
red
to
ha
ve
a
basic
functi
on,
but
EMD
c
om
pletely
abandons
the
co
ns
trai
nt
of
t
he
basis
f
u
nctio
n,
an
d
only
perf
or
m
s
sign
al
deco
m
po
sit
io
n
based
on
the
tim
e
scal
e
featur
e
of
the
data
it
sel
f,
and
has
a
dap
ta
bili
ty
.
Since
no
basis
f
un
ct
ion
is
require
d,
E
MD
ca
n
be
use
d
for
al
m
os
t
any
ty
pe
of
sig
na
l
deco
m
po
sit
ion,
es
pecial
ly
for
the
dec
om
po
sit
ion
of
n
onli
nea
r,
non
-
sta
ti
on
a
ry
sign
al
s
[
17,
18
]
.
T
he
pu
rpose
of
EMD
is
to
deco
m
po
se
the
sig
nal
into
a
super
posit
ion
of
m
ulti
ple
intrinsic
m
od
e
f
unct
ions
(
IMFs).
I
n
ad
diti
on,
t
he
IMF
m
us
t
s
at
isfy
the
fo
ll
owin
g
two
c
onditi
ons
(
t
he
f
unct
ion
m
us
t
hav
e
the
sam
e
nu
m
b
er
of
local
e
xtre
m
e
po
ints
an
d
zero
cr
os
sin
gs
within
the
entire
ti
m
e
range
,
a
nd
a
t
any
point
in
tim
e,
the
env
el
op
e
of
th
e
local
m
axi
m
u
m
the
envel
op
e
of
the (u
pp
e
r
e
nvel
op
e
)
a
nd the
local
m
ini
m
u
m
(lo
wer en
velo
pe) m
us
t be zero
on av
e
ra
ge.
The
EM
D
m
eth
od is
base
d o
n
the:
The
si
gn
al
has
at
least
two
e
xt
rem
e p
oin
ts,
one m
axi
m
u
m
an
d o
ne
m
ini
m
um
.
The
c
har
a
ct
eris
ti
c tim
e scal
e i
s d
e
fine
d by th
e tim
e b
et
ween t
he
tw
o
e
xtre
m
e p
oin
ts
.
If
the
data
la
cks
ext
rem
e
po
ints
but
has
defor
m
at
ion
po
i
nt
s,
the
extr
em
e
po
i
nts
ca
n
be
ob
ta
ine
d
by
da
ta
diff
e
re
ntiat
ion
on
ce
or se
ve
ral tim
es, an
d t
he
n
the
d
ec
om
posit
ion
r
e
su
lt
s a
r
e obtai
ned b
y
integ
rati
on.
The
al
gorithm
f
lo
w
is as
foll
ows:
-
I
de
ntify al
l ext
rem
a o
f
x(t
)
-
I
nter
pola
te
b
et
ween m
ini
m
a (
resp. m
axi
m
a),
en
ding
up w
it
h
s
om
e env
e
lo
pe
em
in(t)
(
res
p.
em
ax(
t)
)
-
Com
pu
te
the
m
ean m
(t)
= (e
m
in(t)+em
ax(
t))/2
-
Extract t
he deta
il
d
(t)
=
x(t)
−
m
(t)
-
Iterate
on t
he r
esi
du
al
m
(t)
2.4.
Fe
at
ure
s
e
xt
r
act
i
on
A
hel
pful
featu
re
f
or
e
xpress
a
bio
m
edical
si
gn
al
nam
ely
Hj
ort
h
descr
i
ptors
(
HD)
di
vid
e
d
into
th
ree
m
ai
n
par
am
et
e
rs
as
f
ollows:
-
Act
iv
it
y
:
is
the
m
os
t
us
efu
l
pa
ram
et
ers
in
bi
ologica
l
sign
al
s,
si
m
ply
i
ts
var
ia
nce
of
the
sign
al
re
pr
ese
nt
s
the en
e
r
gy:
=
0
2
(1)
-
Mo
bili
ty
: M
ob
il
it
y i
s g
iven by
:
=
1
2
/
0
2
(2)
-
C
omplexi
ty
:
giv
es a
co
m
pu
ta
ti
on
al
value
for
the s
hap
e
of th
e sig
nal:
=
√
(
(
+
1
)
2
(
)
2
−
(
)
2
(
−
1
)
2
)
2
(3)
-
Perm
uta
ti
on
E
ntropy
:
Ba
ndt
and
Po
m
pe
are
in
vestigat
e
d
the
(
PE)
Pe
rm
utati
on
entr
op
y
to
m
easu
re
t
he
com
plexity
of
the
non
-
li
near
it
y
an
d
non
-
sta
ti
onary
na
ture
in
tim
e
series
sig
nals
[
1
9
]
.
t
he
S
ha
nnon
entr
op
y i
s calc
ulate
d
in
PE
fo
r
the
d
if
fer
e
nt s
ym
bo
l i
n
th
e si
gn
al
a
nd ca
n b
e cal
culat
ed
as
fo
ll
ows:
=
∑
=
1
∗
(
)
/
(
)
(4
)
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
ELM
and
K
-
nn
mach
i
ne
le
arn
ing
i
n
cl
assi
fi
cation of
Breat
h so
unds si
gnals
(
Z.
Ne
il
i
)
3531
2.4.1.
Statis
tical
anal
ys
is
In
this
st
ud
y,
a
sta
ti
sti
cal
analy
sis
of
m
ean
and
sta
nda
rd
de
viati
on
was
use
d
t
o
te
st
the
s
ign
ific
a
nce
of the acti
vity
and PE
f
eat
ure
s.
S
D
a
nd Mea
n
a
re e
xpresse
d resp
ect
ively
as foll
ows:
=
√
1
∑
(
−
)
2
=
1
(5)
̅
=
(
∑
)
/
(6)
w
he
re:
each
valu
e
of
t
he data
set
.
M
the total
num
ber
o
f data
points.
2.5.
Cl
as
sific
ati
on
In
this
stu
dy,
two
cl
assifi
er
s
wer
e
us
ed
for
two
cl
assifi
cat
ion
ty
pes
(m
ulti
cl
ass
cl
assifi
c
at
ion
,
bi
nary
cl
assifi
cat
ion
),
o
ne
is
the
e
xtrem
e
le
arn
ing
m
achine
(EL
M)
an
d
the
ot
her
is
a
k
-
nea
r
est
neig
hbour
(K
-
N
N).
detai
le
d of
t
hes
e cla
ssifie
rs
a
r
e presente
d
in
t
his secti
on:
2.
5.1.
E
xt
re
m
e learnin
g
m
ac
hine
Hu
a
ng
et
al
.
[
20
]
propose
a
n
al
gorithm
fo
r
so
lvin
g
a
sin
gle
hid
de
n
la
ye
r
n
eur
al
net
wor
k
wh
ic
h
is
a
n
extrem
e
le
arn
ing
m
achine
(E
LM).
T
he
bigg
est
featu
re
of
ELM
is
that
tr
aditi
on
al
neura
l
netw
orks,
esp
eci
al
ly
con
ce
r
ning
a
s
ing
le
hidden
l
ay
er
fee
dforwa
rd
ne
ur
al
netw
orks
(S
LF
Ns
),
are
faste
r
tha
n
tra
diti
on
al
le
arn
i
ng
al
gorithm
s w
hile
gu
a
ra
nteei
ng learni
ng accu
r
acy
[
21
].
Fo
r
a
si
ng
le
hi
dd
e
n
la
ye
r
ne
ur
al
netw
ork
s
how
n
in
Fig
ur
e
2,
ass
um
e
that
there
is
an
arb
it
ra
ry
sam
ple
(
,
)
of
wh
i
ch [
22
]:
=
[
,
,
…
,
]
ϵ
,
=
[
,
,
…
,
]
ϵ
(7)
Fo
r
a
sing
le
hi
dd
e
n
la
ye
r
n
e
ural
n
et
work wit
h
a
hidde
n
la
ye
r node,
it
can b
e ex
pr
es
sed
as
:
∑
(
=
1
.
+
)
=
,
=
1
,
…
,
(8)
Am
on
g
t
hem
(
)
, t
he
act
ivati
on fun
ct
io
n, w
hich
:
=
[
,
1
,
,
2
,
…
,
,
]
is
the
in
put
w
ei
gh
t
an
d
the
ou
t
pu
t
w
ei
ght,
is
the
offset
of
the
first
hidden
la
ye
r
unit
.
.
Re
pr
ese
ntati
on
an
d
inn
e
r pro
du
c
t.
.
The
goal
of
a
sing
le
hi
dden
la
y
er
neural
ne
twork
le
ar
ning
is
to
m
ini
m
i
ze
the
error
of
the
outp
ut,
wh
ic
h
ca
n be e
xpresse
d
as
:
∑
‖
−
‖
=
0
=
1
(9)
That e
xists
,
and
so t
hat
∑
(
=
1
.
+
)
=
,
=
1
,
…
,
(10)
Ca
n be
e
xpress
ed
as
a m
at
rix
=
(11)
Am
on
g
them
it
is
t
he
outp
ut
of
the
hidden
la
ye
r
node,
w
hich
is
the
ou
t
pu
t
weig
ht
a
nd
is
the expect
e
d o
utput.
=
(
1
,
…
,
,
1
,
…
,
,
1
,
…
,
)
(12)
=
[
(
1
.
1
+
1
)
…
(
.
1
+
)
⋮
…
⋮
(
1
.
+
1
)
…
(
.
+
)
]
×
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.
10
, No
.
4
,
A
ugus
t
2020
:
3528
-
3536
3532
=
[
1
⋮
]
×
,
=
[
1
⋮
]
×
,
(13)
In ord
e
r
t
o be a
ble to t
rain
a
si
ng
le
hidde
n
la
ye
r
ne
ur
al
net
work,
we h
op
e
to
get
1
,
an
d
to
m
ake
:
‖
(
̂
,
̂
)
̂
−
‖
=
min
,
,
‖
(
,
)
−
‖
(
14)
w
he
re
=
,
…
,
this is e
qu
i
valent t
o
m
i
nim
iz
ing
the
lo
ss fun
ct
i
on
=
∑
(
∑
(
=
1
.
+
)
−
)
=
1
2
(15)
So
m
e
tradit
io
nal
al
gorit
hm
s
base
d
on
t
he
gradie
nt
desc
ent
m
et
ho
d
c
a
n
be
us
e
d
t
o
so
lve
s
uch
pro
blem
s,
bu
t
the
ba
sic
gradi
ent
-
based
le
ar
ning
al
go
rithm
nee
ds
t
o
a
dju
s
t
al
l
par
am
et
ers
duri
ng
the
it
erati
ve
process
.
In
the
ELM
al
gori
th
m
,
on
ce
th
e
in
pu
t
wei
gh
t
an
d
the
bias
of
th
e
hi
dd
e
n
la
ye
r
are
ra
ndom
ly
determ
ined,
t
he
outp
ut
m
at
rix
of
the
hidden
la
ye
r
is
un
i
quel
y
determ
ined.
T
he
t
rainin
g
sing
le
hi
dd
e
n
l
ay
er
neural
netw
ork
can be t
ran
s
f
orm
ed
into a lin
ear syst
em
=
an
d t
he ou
t
pu
t
wei
gh
t
can
be dete
rm
ined
̂
=
ϯ
(
16)
Am
on
g
them
ϯ
is
the
Mo
ore
-
Penrose
ge
neral
iz
ed
in
ve
rse
of
the
m
at
rix.
A
nd
it
ca
n
be
prov
e
d
that
̂
the no
rm
o
f
the
so
l
ution o
btained
is m
ini
m
a
l
and
un
i
qu
e
[
22
].
Figure
2
.
SLF
N: add
it
ive
hidden
no
des
2.5.2.
K
-
neare
st
nei
gh
b
ou
r
The
K
-
Near
est
Neig
hbors
(
K
-
N
N)
al
gorith
m
is
a
cl
assificati
on
al
gorith
m
and
one
of
the
easi
est
to
unde
rstan
d
m
a
chine
le
a
rn
i
ng
al
gorithm
s.
In
1968,
it
wa
s
pr
opos
e
d
by
Co
ve
r
an
d
Har
t
[
23
]
.
T
he
sim
ple
st
an
d
m
un
dan
e
cl
ass
ifie
r
m
ay
be
th
e
ki
nd
of
m
e
m
or
a
ble
cl
assifi
e
r,
rem
e
m
ber
al
l
the
trai
ning
da
ta
,
f
or
the
ne
w
data,
it
m
a
tc
hes
the
trai
ning
data
di
rectl
y,
if
there
is
trai
nin
g
dat
a
of
the
sam
e
at
tribu
te
,
use
it
directl
y
,
com
e
as
a
cl
assifi
cat
ion
of
new
dat
a.T
he
k
-
N
N
al
gori
thm
find
s
the
k
records
cl
os
es
t
to
the
new
da
ta
fr
om
the
trai
ning
set
and the
n de
te
rm
ines the ca
te
gory of the
new
data b
a
sed
on their
pri
m
ar
y cl
assifi
cat
ion.
The
al
gorithm
involves
thr
e
e
m
ai
n
facto
rs
:
-
The
trai
ning se
t
-
The dist
ance
or sim
i
la
r
m
eas
ur
e
. In
t
his stu
dy a E
uclidia
n
distance
has be
en use
d:
(
,
)
=
∑
√
2
−
2
=
1
(17)
-
The
siz
e
of
k
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
ELM
and
K
-
nn
mach
i
ne
le
arn
ing
i
n
cl
assi
fi
cation of
Breat
h so
unds si
gnals
(
Z.
Ne
il
i
)
3533
In
the
vali
dation
sta
ge
,
the
da
ta
set
X
is
div
ided
into
a
trai
ning
set
Y
(tra
ining
set
)
a
nd
a
te
st
set
Z
(test
set
),
for
the
case
that
the
sa
m
ple
siz
e
i
s
insu
f
fici
ent
s
uch
as
in
ou
r
stud
y,
an
d
i
n
or
der
to
f
ull
us
e
of
al
l
data
set
to
te
st
the
al
gorithm
s
eff
ect
,
data
bas
e
X
is
ra
ndom
l
y
div
id
ed
into
k
packet
s,
on
e
of
w
hich
is
use
d
as
a
te
st
set
each
tim
e,
and
the
rem
a
ining
k
-
1
pac
kets
are
trai
ned
as
a
tr
ai
nin
g
set
,
by
us
in
g
k
-
f
old
cro
ss
-
validat
io
n
m
eth
od
[
2
4
].
3.
RESU
LT
S
A
ND
DI
SCUS
S
ION
In
t
his
w
ork
t
he
al
l
ex
per
im
ents
we
re
perf
or
m
ed
by
usi
ng
MAT
LAB
R201
3b
a
nd
on
a
pc
with
a
co
nfi
gu
rati
on
of
I
ntel
CP
U
Co
re
i5
,
4
GB
RAM,
a
nd
W
in
dows
10
op
e
rati
ng
syst
em
.
In
[2
5
]
the
auth
or
s
app
ly
Hjor
t
h
descr
i
ptors
as
featur
e
s
a
nd
fi
nd
that
the
ac
ti
vity
featur
e
i
s
the
best
feat
ur
e
c
om
par
ed
with
m
ob
il
i
ty
and
com
plexity
as
sh
ow
n
in
(
1)
,(2)
a
nd
(3)
.
Ther
e
f
or
e,
i
n
our
w
ork
,
the
act
ivit
y
featur
e
was
exp
l
oited
f
or
enh
a
nce
d
this
stud
y
with
c
om
bin
ed
it
wi
th
the
per
m
utati
on
e
ntropy
featur
e
sho
wn
in
(
4
),
and
f
or
m
ed
a
featu
res
vector
s
to
fed
i
nt
o
tw
o
m
achin
e
le
arn
i
ng
al
gorithm
s
nam
ely
ELM
an
d
K
-
NN
,
to
c
om
par
e
the
m
in
the
cl
assif
ic
at
ion
of
breat
h
s
ounds
sig
na
ls.
T
he
EM
D
de
com
po
ses
BS
sign
al
s
into
a
s
et
of
IMFs,
Fig
ur
e
3
s
hows
s
om
e
IMFs
of
t
he
norm
al
su
bject
.
T
he
featu
res
(A
ct
ivit
y
an
d
Perm
utati
on
E
ntr
op
y
)
wer
e
e
xtracte
d
from
each
IMF
an
d
te
ste
d
us
in
g
a
sta
ti
s
ti
cal
m
easur
e
of
(
m
ean
an
d
st
and
a
r
d
de
viati
on
S
D
descr
i
bed
in
(
1) a
nd (2)
)
as
ta
bu
la
te
d
in
Tabl
e 1
.
(a)
IMF1
⋮
IMF10
(b)
Figure
3
.
(a
)
N
or
m
al
b
reath
s
ounds
a
nd
(b)
t
heir
IMF
(1
-
10)
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.
10
, No
.
4
,
A
ugus
t
2020
:
3528
-
3536
3534
Table
1
Stat
ist
ic
al
an
al
ysi
s of
featur
e
s e
xtract
ed fr
om
b
reath
so
un
ds
Breath
so
u
n
d
s
Activ
ity
Mean ±
st
an
d
ard d
ev
iatio
n
Per
m
u
tatio
n
entro
p
y
Me
an
±
stan
d
ar
d
dev
iatio
n
No
r
m
al
bro
n
ch
ial
0
.00
1
3
8
7
±
0
.00
4
3
8
7
0
.56
4
9
7
±
0
.20
2
8
1
9
W
h
eeze
Crackl
e
0
.00
5
8
2
9
±
0
.0
0
0
4
6
8
0
.00
0
6
4
5
±
0
.00
0
7
6
5
0
.56
7
9
2
±
0
.20
9
8
8
8
0
.58
6
6
7
9
±
0
.21
6
7
1
4
Pleu
ral
rub
0
.00
2
6
0
8
±
0
.00
0
5
1
3
0
.56
6
7
5
±
0
.22
3
9
2
3
Strido
r
0
.00
1
5
9
4
±
0
.00
1
4
7
2
0
.56
3
3
5
2
±0
.2
1
5
6
4
1
Fr
om
T
able
1
we
infe
rr
e
d
t
hat
there
is
si
gn
i
ficant
disc
r
i
m
inati
on
in
the
act
ivit
y
and
PE
f
eat
ur
e
s
of
di
ff
e
ren
t
cl
a
sses.
Ca
n
be
obser
ve
d
a
m
ea
n
an
d
SD
are
di
ff
ere
nt
from
e
ach
cl
ass
in
act
ivit
y
featur
es,
bu
t
in
PE
featu
res
a
li
ttle
diff
ere
nt
between
cl
a
sses.
From
thi
s,
we
ca
n
co
m
bin
e
the
m
e
to
te
st
and
c
om
par
e
the cla
ssific
at
ion acc
uracy
of
bo
t
h K
-
nn
a
nd
ELM
classi
fier
s.
T
hese
featu
r
es h
a
ve bee
n
f
or
m
ed
as
fo
ll
ows:
Feat
ur
es
=
[A
c
ti
vity
, P
E]
.
In
or
der
to
ve
r
ify
the
reli
ability
of
the
ou
tc
om
e
of
the
cl
a
ssifie
rs,
the
k
-
fo
l
d
cro
s
s
-
validat
ion
was
us
e
d.
Fig
ur
e
4
sh
ows
how
the
te
n
-
f
old
w
ork
s.
Af
te
r
se
ve
ra
l
te
sts
to
choos
e
the
k
value,
we
fou
nd
that
k=10
is prom
ise
d
val
ue,
t
her
e
fore
it
has bee
n used i
n
this
stu
dy.
Figure
4
.
K
-
f
ol
d
c
ro
ss
-
validat
ion m
et
ho
ds
In
the
li
te
rat
ur
e
rev
ie
w
m
any
researc
he
rs
base
d
on
act
ivit
y
or
en
tro
py
featur
e
s
extracti
on,
nev
e
rtheless
,
this
stud
y
has
c
om
bin
ed
both
act
ivit
ie
s
and
PE
featur
e
s
for
obser
ved
the
abili
ty
of
bo
t
h
ELM
and
K
-
nn
to
cl
assify
diff
e
r
ent
BS
sig
nals.
I
n
Table
2
the
cl
assifi
cat
ion
sta
ge
is
descr
i
bed,
an
d
giv
e
the
cl
assifi
cat
ion
pe
rfo
rm
ance
of
feat
ur
es
(
Acti
vity
,
PE)
e
xtracted
f
ro
m
I
MFs
us
i
ng
EL
M
with
RB
F
Kernel
,
Po
ly
nom
ial
Ker
nel
a
nd
K
-
nn
with
d
ist
a
nce
e
uclidia
n
wh
ic
h
is
desc
r
ibed
in
e
qu
at
i
on
(
17
)
,
a
nd
1
to
10
nu
m
ber
of n
ei
g
hbours
.
Table
2
.
Cl
assi
ficat
ion
pe
rform
ance of (Acti
vity
, P
E)
fro
m
I
MFs
of BS si
gnal
s
i
n
m
ulti
cl
ass cl
assifi
cat
ion
sta
ge
Clas
sif
ier
K
-
Fo
ld
K neig
h
b
o
u
rs
Kernel
Av
erage accur
acy
(
%)
EL
M
(
A
ctiv
ity
,
PE
)
EL
M
(
A
ctiv
ity
,
PE
)
10
10
/
/
RBF
Po
ly
n
o
m
ial
8
3
.57
7
7
.86
K
-
n
n
(
Activ
it
y
,
PE
)
10
1
/
8
6
.42
K
-
n
n
(
Activ
it
y
,
PE
)
10
2
/
8
0
.71
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
10
10
10
10
10
10
10
10
3
4
5
6
7
8
9
10
/
/
/
/
/
/
/
/
8
2
.14
8
0
.00
8
2
.14
8
0
.00
8
1
.42
8
2
.14
8
1
.00
7
7
.14
As
sho
wn
in
Tables
2,
T
he
ELM
with
RB
F
Kernel
an
d
K
-
nn
with
1
neighb
our
ga
ve
the
highe
r
cl
assifi
cat
ion
a
ccur
acy
of
83.
57%
an
d
86.42%
re
sp
ect
ivel
y
.
The
ELM
by
RB
F
kernel
is
bette
r
tha
n
ELM
with
Po
ly
nom
ia
l
ker
nel
in
m
ul
ti
cl
ass
cl
a
ssific
at
ion
cas
e,
a
nd
k
-
nn
by
1
neig
hbour
is
bette
r
than
rest
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
ELM
and
K
-
nn
mach
i
ne
le
arn
ing
i
n
cl
assi
fi
cation of
Breat
h so
unds si
gnals
(
Z.
Ne
il
i
)
3535
neig
hbours
.
W
e
can
say
t
hat
the
abili
ty
of
t
he
K
-
nn
is
hi
gher
t
han
ELM
in
the
cl
assi
ficat
ion
of
t
he
Breat
h
so
un
ds
sig
nals
into
seve
ral
cl
asses
(Nor
m
al
bron
c
hial,
Wh
eeze
,
Crac
kle
,
Pleural
r
ub,
Strid
or).
W
e
c
an
be
s
een
in
Ta
ble
3,
the
acc
uracy
found
from
k
-
nn
is
95%
by
6
-
8
-
10
neig
hbours
a
nd
f
ro
m
ELM
with
P
olyno
m
ia
l
Kernel
is
90.
71%
be
tt
er
than
RB
F
kernel
in
bi
nar
y
cl
assi
ficat
ion
case
.
Howe
ver,
acco
rd
i
ng
t
o
these
resu
lt
s
,
we
can
say
th
at
,
this
com
par
at
ive
stud
y
s
hows
t
hat
the
c
apab
il
it
y
of
th
e
k
-
nn
cl
assi
fier
is
higher
c
om
par
ed
with
that
of
t
he
ELM
cl
ass
ifie
r
in
the
cl
assifi
cat
io
n
of
breat
h
sou
nds
si
gnal
s
f
r
om
our
te
st
co
nd
it
ion
s.
The
abili
ty
of
the
k
-
nn
is
higher
tha
n
ELM
in
the
cl
assifi
cat
ion
of
the
br
eat
h
sounds
si
gn
al
s
int
o
bi
na
ry
and
m
ul
ti
cl
ass cla
s
sific
at
ion
case
s
.
Table
3
.
Cl
assi
ficat
ion
pe
rform
ance of (Acti
vity
, P
E)
fro
m
I
MFs
of BS si
gnal
s in
b
i
nar
y c
la
ssific
at
ion
st
age
Clas
sif
ier
K
-
Fo
ld
K neig
h
b
o
u
rs
Kernel
Av
erage accur
acy
(%)
EL
M
(
A
ctiv
ity
,
PE
)
EL
M
(
A
ctiv
ity
,
PE
)
10
10
/
/
RBF
Po
ly
n
o
m
ial
8
9
.29
9
0
.71
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
K
-
nn
(Activit
y
,
PE
)
10
10
10
10
10
10
10
10
10
10
1
2
3
4
5
6
7
8
9
10
/
/
/
/
/
/
/
/
/
/
93.
57
9
2
.14
9
4
.29
9
2
.86
9
4
.28
9
5
.00
9
4
.29
9
5
.00
9
3
.00
9
5
.00
4.
CONCL
US
I
O
N
In
t
his
stu
dy,
t
he
perform
ance
of
the
ELM
an
d
K
-
nn
cl
assifi
ers
wer
e
c
om
par
ed
us
in
g
the
Hjort
h
descr
i
ptors
(
A
ct
ivit
y)
and
Pe
rm
utati
on
Entr
op
y
(P
E
)
feat
ures
in
disti
ngui
sh
in
g
betwee
n
br
eat
h
s
ounds
sign
al
s
with
c
om
bin
at
ion
the
se
fe
at
ur
es
(
Acti
vity
,
PE).
T
he
featu
res
e
xt
racted
wer
e
analy
zed
sta
ti
sti
cal
l
y
by
cal
culat
ing
a
m
ean
and
sta
nd
ar
d
de
vi
at
ion
to
ob
s
e
rv
e
the
di
ff
e
r
ence
betwee
n
the
m
fo
r
each
cl
ass
(Nor
m
al
bronchial
,
Wh
eeze
,
Crac
kle,
Pleural
rub,
Stri
dor).
T
he
cl
a
ssific
at
ion
acc
ur
acy
i
n
m
ulti
cl
ass
cl
assifi
cat
ion
case
of
the
E
LM
an
d
k
-
nn
cl
assifi
ers
is
83.57%
a
nd
86.
42%
res
pect
ively
,
an
d
in
bin
a
ry
cl
assifi
cat
ion
c
ase,
the
acc
ur
a
cy
is
90
.
71
%
,
95%
res
pecti
ve
ly
.
These
s
how
that
the
a
bili
ty
of
k
-
nn
i
n
our
te
s
t
conditi
ons
(d
a
ta
base,
m
et
ho
ds
of
analy
ses
the
br
eat
h
sign
al
s
,
and
fea
tures
us
e
d)
is
higher
tha
n
th
e
ELM
cl
assifi
er
in
m
ulti
cl
ass
and
bi
nar
y
cl
assifi
ca
ti
on
.
In
f
utu
re
work,
the
EMD
m
et
ho
ds
will
be
com
par
ed
wit
h
ano
t
her m
et
ho
d for
furthe
r
a
na
ly
sis of
br
eat
h so
unds si
gnal
s
u
si
ng a lar
ge d
at
abase.
REFERE
NCE
S
[1]
H.
Kita
oka
,
and
R.
T
aka
ki
,
B.
Suki
,
“
A
thr
ee
-
d
imensional
m
odel
of
th
e
hum
an
ai
rwa
y
t
ree,”
J
ournal
of
Applie
d
Phy
siolog
y
,
Dec
.
1999.
[2]
Md.A.
Islam,
et
al
.
,
“
Multi
cha
nn
el
lung
sound
ana
l
y
s
is
for
asthma
det
e
ct
ion
,
”
Co
mputer
Me
thods
and
Programs
i
n
Bi
omedi
ci
ne
,
vo
l
.
159
,
pp
.
111
–
1
23,
Jun.
2018.
[3]
A.
Mond
al
,
e
t
al
.
,
“
A
Novel
Feat
ure
Ext
r
acti
on
Techni
que
f
or
Pulm
onar
y
Sound
Anal
y
s
is
Based
on
EMD,”
Computer
Me
tho
ds and
Program
s in
B
iomedicine
,
vol
.
159
,
pp
.
19
9
-
209,
Jun.
2018
.
[4]
A.
Riz
a
l,
R
.
Hid
a
y
at
and
H.
A.
N
ugroho,
“
Pulm
o
nar
y
crackle
fe
ature
ext
r
action
us
ing
tsal
l
is
ent
rop
y
for
aut
om
at
i
c
lung
sound c
la
ss
ifi
c
at
ion
,
”
2016
1st I
nte
rnationa
l
Confe
renc
e
on
Bi
omedi
cal
Eng
i
nee
ring (
IBI
OMED)
,
Yogy
aka
r
t
a,
pp.
1
-
4
,
2016
.
[5]
F.
Panca
ld
i,
et
al
.
,
“
Anal
y
sis
o
f
pulmonar
y
sounds
for
the
di
agnosis
of
int
e
r
stit
ial
lung
d
ise
ase
s
sec
ondar
y
to
rhe
um
at
oid art
hr
it
is,
”
Computers
in
B
iol
ogy
and
Me
dicine
,
vol
.
9
6,
pp
.
91
-
97
,
Ma
y
2018.
[6]
A.
Chee
m
a,
et
al
.
,
“
An a
ppli
catio
n
of
phonoca
rdi
ogra
ph
y
sign
al
s for
psy
chol
ogi
cal
stress de
te
ction
using non
-
li
nea
r
ent
rop
y
-
base
d
f
ea
tur
es
in
empir
ic
a
l
m
od
e
dec
o
m
positi
on
dom
ai
n,
”
In
Applied
Soft
Computing
Journal
,
vol
.
77
,
pp.
24
-
33
,
Apr.
2019.
[7]
R.
Pala
n
ia
ppan
,
et
al
.
,
“
Pulm
onar
y
Acousti
c
Sign
al
Cla
ss
ifica
ti
on
using
Autore
gre
ss
ive
Coeff
icients
and
k
-
Nea
rest
Neighbor
,
”
Ad
va
nce
d
Manu
factu
ring R
ese
arch
a
nd
Intelli
g
ent A
p
pli
cations
,
vo
l
.
5
91,
pp
.
211
-
214
,
Jul.
2014
.
[8]
H.
Pasterka
m
p,
et
al
.
,
“
State
of
the
ar
t
adv
anc
es
be
y
ond
the
st
ethos
cope
,
”
Ame
ri
can
J
ournal
R
espiratory
Criti
ca
l
Care
Me
di
ci
n
e
,
vol
.
156
,
pp
.
974
-
987,
1997
.
[9]
The
A
uscul
ta
t
io
n
As
sistant
,
“
Bre
at
h
Sounds
,
”
20
15.
[10]
A.
Riz
a
l,
R
.
Hi
da
y
at
and
H.
A.
Nugroho,
“
Ent
r
op
y
m
ea
surem
e
nt
as
feature
s
e
xtra
c
ti
on
in
aut
o
m
at
ic
lung
soun
d
cl
assifi
ca
t
ion,”
Inte
rnational
C
onfe
renc
e
on
Control,
Elec
tronic
s,
R
ene
wab
le
En
ergy
and
Comm
unic
ati
o
ns
(
ICCRE
C)
,
Yog
y
ak
arta,
pp.
93
-
9
7,
2017
.
[11]
D.
Arnal
l
,
“
Pul
m
onar
y
br
eath
s
ounds,
”
2015
.
[12]
R.
W
il
kins
,
e
t
a
l
.
,
“
Lung
Sounds
:
A Prac
t
ical
Gu
i
de
with
Audio
C
D .
2nd
Ed,
”
Mos
by
,
Feb
.
1996
.
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.
10
, No
.
4
,
A
ugus
t
2020
:
3528
-
3536
3536
[13]
D.
Bardou,
et
a
l
.
,
“
Lung
sounds
cl
assific
a
ti
on
using
convol
utional
neur
a
l
ne
t
works
,
”
Arti
fici
al
int
ellige
n
ce
i
n
medic
in
e
,
v
ol
.
8
8,
p
p
.
58
-
69
,
20
18.
[14]
R.
Pala
ni
appa
n
,
et
al
.
,
“
A
compara
ti
v
e
stud
y
of
t
he
svm
and
k
-
nn
m
ac
hine
learni
n
g
al
gorit
hm
s
for
the
dia
gnosis
of
respir
at
or
y
pa
tho
logi
es
using
pul
m
onar
y
ac
ousti
c
signal
s,
”
BMC bioi
nfo
rm
ati
cs
,
vol
.
15
,
p
p
.
223
,
2014.
[15]
A.
S
ovij
ärv
i
,
et
al
.
,
“
Standardi
za
t
ion
of
computer
i
ze
d
respir
a
t
or
y
sound
anal
y
sis,
”
Crit
Care
Me
d
.
,
v
ol
.
15
6,
pp.
974
-
987
,
19
97.
[16]
I.
Branc
h
,
et
al
.
,
“
The
empirical
m
ode
dec
om
po
siti
on
and
the
H
il
ber
t
spe
ct
rum
for
nonli
nea
r
a
n
d
non
-
sta
ti
onar
y
ti
m
e
seri
es
an
alys
is,”
Proc
ee
d
ing
of the
Roy
a
l
So
c
ie
t
y
Math
ematic
al,
Phy
si
cal
and
Engi
ne
ering
Sc
i
enc
es
,
1998.
[17]
G.
Ril
li
ng,
e
t
a
l
.
,
“
On E
m
piri
cal M
ode
Dec
om
positi
on
an
d
i
ts Al
gorit
hm
s,”
2003
.
[18]
A,
Mondal,
et
al
.
,
“
A
Enha
nc
ement
of
lung
soun
ds
base
d
on
empirical
m
ode
de
co
m
positi
on
and
Fourier
tr
ansform
al
gorit
hm
,
”
Com
pute
r met
hods a
nd
progr
ams
in
biomedi
c
ine
,
vol
.
139
,
pp
.
119
-
1
36,
2017
.
[19]
C.
Bandt
,
and
B.
Po
m
pe,
“
Perm
u
ta
ti
on
En
trop
y
:
A
Natur
al
Com
p
le
xity
Mea
sur
e
for
Ti
m
e
Serie
s,”
Phy
sical
Revi
e
w
Lett
ers
,
vo
l
.
88,
pp.
17
-
29
,
Ap
r.
2002.
[20]
G.
Bin
-
Huang,
e
t
al
.
,
“
Ext
r
em L
ea
rning
Mac
hin
e: T
heor
y
an
d
applications,
”
Neuro
computi
ng,
vol
.
70,
pp.
489
-
501
,
Dec
.
2006.
[21]
G.
Bin
-
Huang,
et
al
.
,
“
Ext
r
eme
le
arn
ing
m
a
chines :
a
surve
y
,
”
Inte
rnational
J
ournal
of
Mac
h
ine
Learning
an
d
Cybe
rnet
ic
s
,
vol
.
2,
pp.
107
–
122,
Jun.
2011.
[22]
S.
Liao
and
C.
Feng,
“
Met
a
-
E
LM:
E
LM
with
EL
M
h
idde
n
nodes,
”
N
euroc
omputing
,
vo
l
.
128,
pp.
81
-
87,
Mar.
2014
.
[23]
C
over
,
T.
M,
et
a
l
.
,
“
Nea
rest
nei
ghbor
pat
t
ern
cl
assifi
ca
t
ion,”
IEE
E
transacti
on
s
on
inf
orm
ati
o
n
the
ory
,
v
ol
.
13
,
no.
1
,
pp
.
21
-
27
,
1967.
[24]
T.
Andersen
and
T.
Marti
ne
z,
“
Cross
val
ida
ti
on
a
nd
MLP
arc
hit
ecture
select
ion
,
”
I
n
IJCNN'99.
Inte
rnational
Joi
n
t
Confe
renc
e
on
N
eural
Ne
tworks.
Proce
ed
ings
,
vo
l
.
3
,
pp
.
1614
-
16
19,
1999
,
[25]
A.
Riz
al,
et
al
.
,
“
Lung
Sound
Cla
ss
ifi
c
at
ion
Us
ing
Empiric
al
Mode
Dec
om
po
siti
on
and
the
Hjorth
Descri
ptor
,
”
Ame
rican Journal
of
Applied
Scienc
es
,
vol
.
14
,
p
p.
166
-
173
,
Jan
.
2017.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Neil
i
Z
akaria
rec
ei
v
ed
his
m
aster
degr
e
e
in
m
ulti
m
edi
a
and
d
igi
t
a
l
comm
unic
at
ion
s
from
Badj
i
Mokhtar
Annaba
Univer
sit
y
in
2017.
Since
2
018
he
has
joi
n
ed
the
Autom
atic
and
Signa
ls
La
bora
tor
y
of
A
nnaba
Unive
rsit
y
,
His
rese
ar
ch
i
nte
rests
ar
e
in
bi
om
edi
ca
l
sign
al
proc
essing
and
biomedic
a
l appl
i
ca
t
ions.
Mohame
d
Fe
z
a
ri
is
a
profe
ss
or
in
el
e
ct
roni
cs
a
nd
computer
archit
e
ct
ure
at
th
e
Univer
sit
y
of
Badj
i
Mokht
ar
Annaba
,
Alg
eria
.
He
go
t
a
Bac
h
el
or
d
egr
ee
in
e
le
c
tri
c
al
engi
n
eering
from
the
Univer
sit
y
of
Oran,
1983
.
H
e
g
ot
an
MS
c
degr
ee
in
computer
scie
nc
e
from
th
e
Univer
si
t
y
of
Cal
ifornia
River
side,
1987.
He
h
olds
a
PhD
degr
ee
in
e
lectr
oni
cs
from
the
Univer
sit
y
of
B
adj
i
Mokhtar
Annab
a,
2006
.
His
res
ea
rch
intere
sts
i
ncl
ude
sp
ee
ch
p
roc
essing,
DS
P,
m
ic
roc
ont
rol
le
r
,
m
ic
roproc
essor, robotic
s
and
hu
m
an
-
m
ac
hine
intera
c
ti
on,
and
reh
abi
litation.
Ab
delghan
i
Re
djati
was
born
in
Annaba
,
Al
ger
ia,
in
1966
.
He
rec
e
ive
d
M.
Sc.
degr
e
e
in
El
e
ct
roni
c
(State
Engi
ne
er
in
El
e
ct
roni
cs
(June
1995
–
Badj
i
Mokhtar
Annaba
Unive
rsit
y
),
Alger
ia,
Magist
er
degr
e
e
in
Ind
ustria
l
Contro
l
i
n
2005
and
PhD
degr
ee
s
in
Si
gnal
Proce
ss
ing
from
Badj
i
-
Mokhtar
Univ
ersity
,
Annaba
Alger
ia,
in
2009.
Curre
ntly
,
h
e
is
As
sis
ta
nt
Prof
essor
and
Hea
d
of
El
e
ct
roni
c
Depa
rtment
at
Badji
-
Mokhtar
Univer
sit
y
,
Annab
a
Alger
ia
.
His
cur
ren
t
rese
arc
h
intere
s
ts
inc
lud
e
Spe
ec
h
pro
ce
ss
ing
and
embedd
e
d
s
y
stems
,
Inf
orm
at
ion
and
Com
m
unic
at
ion Te
chno
log
y
,
ele
ct
roni
c
design
c
i
rcu
it
s.
Evaluation Warning : The document was created with Spire.PDF for Python.