TELKOM
NIKA
, Vol.16, No
.4, August 20
18, pp. 1468
~14
8
0
ISSN: 1693-6
930,
accredited First Grade by
Kemenristekdikt
i
, Decree No: 21/
E/KPT/2018
DOI
:
10.12928/TELKOMNIKA.v16i4.8281
1468
Re
cei
v
ed
Jan
uary 4, 2018;
Re
vised Ap
ril
1, 2018; Accepted Ma
y 21
, 2018
Asthma Identification Using Gas Sensors and
Support Vector Machine
Hari Ag
us Sujono*
1
, Muhammad Riv
a
i
2
, Muhammad Amin
3
1,2
Department of Electrical En
gin
eeri
ng, Institut
T
e
knolo
g
i S
epu
luh N
o
p
e
m
ber, Surab
a
y
a,
Indones
ia
3
Department o
f
Pulmono
lo
g
y
and R
e
spir
ator
y Med
i
ci
ne, Airl
ang
ga U
n
ivers
i
t
y
, Surab
a
ya, Indo
nesi
a
1
Department o
f
Electrical Eng
i
ne
erin
g, Institut
T
e
knolog
i Ad
hi T
a
ma Surab
a
y
a, Surab
a
ya,
Indones
ia
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: haria
gus.itats
@
y
ah
oo.com
1
, muhamma
d_ri
v
ai@e
e.its.ac.i
d
2
A
b
st
r
a
ct
T
he exha
le
d
breath
an
alysi
s is a pr
oce
d
u
r
e of
meas
urin
g sever
a
l typ
e
s
of gases
tha
t
aim t
o
identify various
diseas
es in the
hum
a
n body. The purpos
e
of
this study is to analy
z
e the gases cont
ained
in the
exha
le
d bre
a
th in
or
der to rec
ogn
i
z
e
he
althy a
n
d
asth
ma s
u
b
j
ects w
i
th varying s
e
verity. An
electro
n
ic
nos
e
cons
isting
of
seven
g
a
s se
n
s
ors e
qui
pp
ed
w
i
th the S
u
p
p
o
r
t Vector Mac
h
ine
class
i
ficati
o
n
meth
od
is
use
d
to a
naly
z
e
th
e
gas
es to
deter
mi
ne t
he
pati
e
nt'
s
cond
ition.
Non-l
i
n
ear
bin
a
r
y classific
a
tio
n
is
used to id
entif
y healthy a
nd
asthma sub
j
ect
s
, w
hereas the multic
lass cl
a
ssificatio
n
is a
ppli
ed to rec
o
g
n
i
z
e
the subjects of asthma with different severit
y
. The re
sult of this study showed
that the system
pr
ovided
a
low accuracy t
o
distinguish
the subjects of asthm
a
with var
y
ing se
v
e
rity. This system
ca
n only differentiate
betw
een
partia
lly co
ntroll
ed
a
nd u
n
co
ntroll
e
d
asth
ma
su
bj
ects w
i
th goo
d
accuracy.
Ho
w
e
ver, this sys
tem
can prov
id
e hi
gh sens
itivity,
specificity, an
d
accuracy to di
stingu
ish b
e
tw
een h
e
a
l
thy an
d asth
ma su
bj
ects.
The use of fiv
e
gas sensors
in
the electronic nos
e system
has the
best accuracy in t
he classific
a
tion
results of 89.
5
%
. T
he gases
of carbo
n
mon
o
xid
e
, nitric
oxi
de, vol
a
tile
org
anic co
mpo
u
n
d
s, hydro
gen,
an
d
carbo
n
d
i
oxi
d
e
conta
i
ne
d i
n
t
he
exha
le
d br
e
a
th ar
e th
e
do
mi
na
nt in
dicati
ons
as b
i
o
m
ar
kers of asth
ma
.T
he
perfor
m
a
n
ce of
electron
ic nos
e w
a
s high
ly d
epe
nd
ent
on th
e abi
lity of sen
s
or array to an
aly
z
e
g
a
s type
in
the sa
mp
le. Th
erefore, i
n
furt
her study
w
e
w
ill e
m
pl
oy
the s
ensors
havi
ng
hig
her se
nsitiv
i
t
y to detect low
e
r
conce
n
tratio
n of the mark
er g
a
ses.
Ke
y
w
ords
: Asthma, Exhal
ed
breath, Gas se
nsors, Supp
ort vector mac
h
i
n
e
Copy
right
©
2018 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The exh
a
led
breath
contai
ns m
any g
a
ses, mo
stly ox
ygen (O
2
), ca
rbon
dioxid
e (CO
2
),
water vapo
r,
nitric
oxide
(NO
)
a
nd va
ri
ous vo
latile o
r
gani
c co
mpo
und
s
(V
OCs) [1].
These
g
a
s
e
s
can
be me
asured
accu
rat
e
ly usin
g Ga
s Chro
matog
r
aphy
(G
C) [
2
]. However,
this techniq
u
e is
expen
sive, b
e
ca
use the
t
e
sting
p
r
o
c
e
s
s ta
ke
s
a lo
n
g
time
and
requires inte
rpretation
fro
m
an
expert [3]. An
electro
n
ic no
se ca
n be used as an
alte
rnative to analize the exh
a
led breath with
low produ
ctio
n co
sts, non
-invasiv
e, fast
er time in sa
mple mea
s
u
r
ements, an
d portabl
es [4,
5
].
The
elect
r
oni
c n
o
se h
a
s
been
u
s
ed
to
diag
no
se
ki
dney di
se
ase
[6,7], diabet
es [8], a
nd l
ung
can
c
e
r
[
9
]
.
B
i
omar
ke
rs
a
r
e
phy
si
cal
sy
mpt
o
ms
of
la
borato
r
y me
a
s
ureme
n
ts th
at ca
n serve
as
indicators of
biologi
cal
or p
a
thop
hysiolo
g
ical
p
r
oce
s
se
s o
r
in respon
se to the
r
ap
eutic
interventions
[10].
Several
studi
es
have
anal
yzed
exhale
d
bre
a
th fo
r a
s
thma
by u
s
i
ng
Cyran
o
se
320 th
at
contai
ns 3
2
different poly
m
er na
no
co
mposite
se
nsors. T
he an
al
ytical method
used to a
nal
yze
the sen
s
o
r
re
spo
n
se is Pri
n
cip
a
l Com
p
onent Analys
i
s
(PCA
). The
result
s sh
ow that electron
ic
nose can
dist
ingui
sh exh
a
l
ed b
r
eath
bet
wee
n
he
althy and
asth
ma
subj
ect
s
, ho
wever, it i
s
n
o
t
good
eno
ugh
to disting
u
ish the subje
c
t
s
of a
s
thma
with differe
nt severity [11].
The pe
rforma
nce
of the electro
n
ic no
se
dep
end
s on the f
eature
s
of
th
e cla
ssifi
catio
n
algo
rithm fo
r exhale
d
bre
a
th.
One of the
classificatio
n
method
s that
gets a lo
t o
f
attention as state of th
e art in patt
e
rn
cla
ssifi
cation
is the S
upp
ort Vector Ma
chine
(
SVM) [
12-1
8
]. The
SVM is a
n
ef
fective techni
que
for qu
antitatively analyzi
n
g
gas mixture
s
as thi
s
can
solve the
cro
s
s se
nsitivity probl
em of
g
a
s
s
e
ns
or array [19].
The pu
rpo
s
e
of this study
is to analyze
the gases
contai
ned in
the exhaled
breath in
orde
r to recogni
ze he
althy and a
s
th
ma su
bje
c
ts
with varying
severity. The
exhaled b
r
e
a
th
sampl
e
s
were taken from
patients dia
g
nosed with
a
s
thma an
d then analyze it
in the laborat
ory.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Asthm
a
Identification Usin
g
Gas Sen
s
o
r
s
and Suppo
rt Vector….(Hari Agus Sujon
o
)
1469
An elect
r
oni
c nose con
s
isting of seve
n ga
s sen
s
o
r
s
equip
ped
with the SV
M cla
s
sificati
on
method i
s
used to analyze
the gase
s
containe
d in
the exhale
d
b
r
eath
sampl
e
to determine
the
patient's con
d
ition. The
n
on-lin
ea
r bin
a
ry
classifi
cation i
s
u
s
e
d
t
o
ide
n
tify he
althy and
a
s
thma
subj
ect
s
, wh
erea
s th
e mu
lticlass
cla
ssi
fication i
s
u
s
ed to id
entify the subje
c
ts of asthm
a
with
different s
e
verity.
2. Researc
h
Method
2.1. Subjects
and resea
rc
h design
The subje
c
ts are 30 p
a
tients dia
gno
sed
with asthma, and
30 healthy
subj
ect
s
volunteered i
n
the study. All subje
c
ts a
r
e adult
s
, not smokers, ag
ed bet
we
en
30-6
0
years, and
do not suffer
from acute or chroni
c illness. In
current
clini
c
al
pr
acti
ce, asthm
a
i
s
diagnosed and
monitored from symptom
s
and
physi
ologi
cal mea
s
ureme
n
ts u
s
ing the
Glo
be Initiative for
Asthma (GINA) and Asth
ma Cont
rol T
e
st (ACT). O
ne of the sta
ndard ch
arac
teristics of GI
NA is
lung fun
c
tion examination
by measu
r
ing
Forced Ex
piratory Volume
in the first se
con
d
(FEV
1
) or
Peak Expirat
o
ry Flow (P
EF) perfo
rm
ed by
force
d
expiratory
maneuvers through sta
ndard
pro
c
ed
ures.
From thi
s
me
asu
r
em
ent, the deg
ree
of
a
s
thma i
s
divid
ed into three
cla
s
ses, n
a
m
e
ly
controlled, pa
rtly controll
ed
, and uncontrolled a
s
thma
[20,21].
The
cla
ssifi
cation of a
s
th
ma subje
c
ts
is b
a
se
d
o
n
stand
ard
s
wit
h
diag
no
stic results
usin
g ACT
sh
own
in T
able
1. Each
subje
c
t is a
s
ked to
exhale
the ai
r colle
cted i
n
1L Te
dla
r
b
a
g
after breathin
g
in and
out for 5 min
u
tes
with cle
an ro
om air. Th
e e
x
haled b
r
eath
is ca
rri
ed o
u
t at
the a
s
thma
cl
inic
at Dr. So
etomo Su
ra
b
a
ya afte
r gett
i
ng a
pproval
from
Ho
spital
Re
se
arch
an
d
Develo
pment
. Figure 1 sho
w
s the
re
sea
r
ch de
sig
n
performed fo
r the cla
ssifi
catio
n
of asthma.
Table 1. The
ACT value
s
ACT Score
25
20 -
2
4
<
20
Controlled
P
artl
y
-C
ontrolled
U
ncontrolled
Figure 1. The
rese
arch d
e
sign for asth
m
a
identificatio
n usin
g ga
s sensor
2.2. Electronic nose s
y
stem
The mai
n
pa
rt of the electronic
nose sy
stem
is
a 24
0 ml ch
ambe
r co
nsi
s
ting
o
f
a gas
sen
s
o
r
a
r
ray. Each
sen
s
o
r
ha
s
sen
s
itivity and
se
le
cti
v
ity to certai
n ga
se
s flo
w
i
ng throug
h t
h
e
cham
be
r. Th
e se
nsor resi
stan
ce
will ch
ange from fre
s
h ai
r to ga
s
sampl
e
,
. T
h
e s
e
ns
or
o
u
t
put
voltage is det
ermin
ed by e
quation (1).
(1)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 146
8-1480
1470
whe
r
e
is the
load resi
stan
ce,
is sen
s
in
g re
sist
a
n
c
e
,
is the con
s
ta
nt voltage ap
plied to
the sen
s
o
r
a
nd
is the tra
n
sie
n
t output voltage. Each
sen
s
o
r
re
sp
onse is an a
nalog
sign
al,
whi
c
h i
s
the
n
filtered, am
pl
ified and
con
v
erted to
digital form
and
then
sen
d
s it to the
com
put
er
every se
con
d
.
Pumps are
activated to flow fres
h ai
r from sili
ca
gel or ga
s from the exhal
ed
breath
bag in
to the sen
s
o
r
cham
ber.T
h
e
flow rate
i
s
maintaine
d
a
t
100 ml/min. Before carrying
the ga
s
sa
m
p
le, the
se
nsor
ch
ambe
r i
s
d
r
ie
d
with
f
r
esh
air to
cl
ean th
e
sen
s
or from
re
sid
ual
gas. T
he h
u
m
idity rate in
the fre
s
h ai
r is redu
ced
b
y
flowing the
air throug
h
a tube
contai
ning
silica gel mat
e
rial.
The
sampli
ng
pro
c
e
s
s of the exhal
ed b
r
eath i
s
carri
ed out in
the
certai
n time
orde
r. In
the
pe
riod of 1
st
to15
th
se
cond
s, the
sen
s
or re
sp
on
se
is in
a ba
se
d
line time,
t
b
. During this
time,
the valve is OFF so that fresh air
can
fl
ow from B to C. In the peri
od of 16
th
to 55
th
se
con
d
s,
t
h
e
exhaled b
r
eat
h is inje
cted to the sen
s
o
r
cham
be
r.
This process is
calle
d rea
c
tio
n
stage time,
t
r
.
At this moment, the valve
is ON
so that the gas in tedl
ar bag is
su
cked from A to C. In the period
of 56
th
to150
th
se
co
nd
s, the
inje
ction i
s
stoppe
d an
d
se
nso
r
a
r
ray is
clea
ned,
and
this p
r
o
c
e
s
s i
s
calle
d a
s
pu
rge sta
ge time
,
t
p
. At this stage, the valv
e is
OFF a
gai
n. After the 1
5
0
th
second
s,
the
valve remai
n
s OF
F until f
u
rthe
r sampli
ng too
k
pl
ace
.
The blo
c
k d
i
agra
m
of the
elect
r
oni
c n
o
s
e
system i
s
sh
own i
n
Fig
u
re
2, whil
e the
photo
se
tup
of the ap
para
t
us a
nd its co
mputer interf
ace
is sh
own in F
i
gure
3. The
sen
s
o
r
re
sp
o
n
se d
u
ri
ng sampling
pro
c
ess is
sho
w
n
in Figure 4. The
portion
of th
e
se
nsor re
sp
onse at t
he
reactio
n
stage
over a
pe
rio
d
of
30
th
to
49
th
sec
ond
s i
s
a
feature extra
c
tion
which can be
con
s
id
ered to
re
pr
es
e
n
t
th
e
o
v
era
ll s
e
ns
or
res
p
on
se
fo
r
the
cla
ssif
i
cat
i
on pro
c
e
ss.
Figure 2. The
electro
n
ic n
o
s
e sy
stem
Figure 3. The
exhaled brea
th analys
i
s
using the ele
c
tronic n
o
se
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Asthm
a
Identification Usin
g
Gas Sen
s
o
r
s
and Suppo
rt Vector….(Hari Agus Sujon
o
)
1471
Figure 4. The
time resp
on
se of a gas se
nso
r
Table 2. The
gas
sen
s
o
r
s
use
d
in the el
ectro
n
ic n
o
se
system
The
sen
s
o
r
a
rray u
s
e
d
on
the ele
c
troni
c nose con
s
ist
s
of seven g
a
s
sen
s
o
r
s
which it
s
sele
ctivity and se
nsitivity are
sh
own in
Table
2.
Ea
ch
sen
s
o
r
i
s
intende
d to
detect
gases in
exhaled b
r
ea
th that indicate the pre
s
ence of
asth
ma in the subje
c
t. In some studi
es
have
sho
w
e
d
that
asthma
p
a
tie
n
ts h
a
d
high
e
r
con
c
e
n
tratio
n value
s
fo
r
NO [2
2], an
d
carbon
mo
no
xide
(CO) [23]. Th
ere i
s
al
so a
relation
shi
p
betwe
en the
VOC in exh
a
l
ed breath
su
bject
with lun
g
dise
ase [24]. There
are
several fa
ctors asso
ci
ate
d
with exhale
d
bre
a
th by asthma
patie
nts
inclu
d
ing
an i
n
crea
se in
hy
drog
en
(H
2
) d
ue to dig
e
stiv
e syste
m
, CO
2
affected by
exposure
to a
i
r
pollution, am
monia
(NH
3
)
due to
a
c
id
-b
ase
statu
s
in
the ai
rway
of
asthmati
cs
[2
5], and
hydro
gen
sulfide (H
2
S)
as ha
rmful hy
dro
c
a
r
bo
n ele
m
ents in the
human b
ody [26].
2.3. Data
An
aly
s
is
In the
sam
p
li
ng p
r
o
c
e
ss,
the
sen
s
o
r
array mea
s
u
r
e
s
the level
of
gases in th
e
exhaled
breath
an
d t
hen
se
nd
s it
to the
com
puter fo
r
an
alysis. T
he
d
a
ta pa
ckag
e
of the
sa
m
p
ling
pro
c
e
ss
on a
subje
c
t is a
data set a
s
matrix [150x
7
]
consi
s
ting o
f
a resp
on
se
of seven sen
s
ors
for 1
50
se
co
nds.
Only th
e si
gnal
resp
onse at
the
rea
c
tion
sta
g
e
is u
s
e
d
fo
r analy
s
is.
Da
ta
analysi
s
co
nsists of three
stage
s, name
l
y pre-p
r
o
c
e
s
sing, feature extraction a
n
d
cla
ssifi
catio
n
.
In the
pre
-
p
r
o
c
e
ssi
ng i
n
cl
u
des ba
seli
ne
corre
c
tion
an
d no
rmali
z
ati
on. Th
e b
a
sel
i
ne
co
rre
ction
is
con
d
u
c
ted by
subt
ra
cting e
a
ch
se
nsor
si
gnal
re
sp
o
n
se with th
e av
erag
e
sign
al
respon
se
at the
baseline,
,
.
Th
e result is the
pre-pro
c
e
s
se
d sign
al re
sp
onse define
d
at equation (2
).
,
,
∑
,
(2)
where i = 1,2, .....
.,N
i
(N
i
is the number
of s
a
mple
s
u
bjec
t
s
),
s
=
1,2,...
.,N
s
(N
s
i
s
the
num
be
r of
s
e
ns
or
s)
, t
b
=
1,2, ..
.,N
b
(N
b
is the maximum time pe
riod on the b
a
se line
stag
e). No
rmali
z
a
t
ion
of the d
a
ta
is
de
sire
d
to redu
ce
a p
a
ttern
variation
du
e to vari
ations in th
e
vapor
con
c
e
n
tration
[27]. The normalize
d
data
is expre
s
sed
by equation (3).
Sensors
G
ases
Sensitivities
(pp
m
)
MQ7
CO
10
–
10000
MQ8
H
2
100
–
1000
MQ131
NOx
0
.01
–
2
MQ136
H
2
S1
–
100
MQ137
NH
3
5
–
200
MQ138
VOCs
10
-
1000
TGS41
6
1
C
O
2
350
–
10000
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 146
8-1480
1472
,
,
,
(3)
Output si
gnal
respon
se fo
r each
se
nsor at
the rea
c
tion stag
e afte
r bein
g
proce
s
sed by
equatio
n (2)
and (3
),
,
is a feature extraction sig
nal fo
r use on cla
ssifi
cation proce
s
s. In
this pro
c
e
s
s, a dataset [
20x7] is derived
from the
,
on the sign
al re
spon
se at the sampli
ng
perio
de of 30
th
- 49
th
sec
o
n
d
s.
SVM is a m
a
chi
ne lea
r
ni
ng metho
d
that wo
rks o
n
the pri
n
cip
l
e of Structu
r
al Ri
sk
Minimization
(SRM
)
with th
e aim
of findi
ng the
be
st
h
y
perpla
ne th
a
t
sep
a
rates th
e two
cl
asse
s in
the input
sp
a
c
e. Th
e m
e
th
od p
r
op
osed
in 199
2 by
Vl
adimir N. Va
pnik, Be
rn
hard E. Boser
a
nd
Isabell
e
M. Guyon has b
e
en widely ap
plied espe
cia
lly in the field
of bioinform
a
tics [28]. Th
ere
are
two
type
s of
cla
s
sification
con
d
u
c
ted by th
e S
V
M, namely
non-li
nea
r
bi
nary
cla
s
sification
and n
o
n
-
line
a
r m
u
lti-cl
ass cla
s
sification
. For m
u
lti-cl
ass cla
ssifica
tion
empl
oys one-agai
nst
-
o
n
e
approa
ch an
d
Gaussia
n
ra
dial ba
sis fun
c
tion kern
el o
r
RBF kern
el.
The SVM
co
nce
p
t ca
n b
e
explaine
d si
mply as
a
se
arch for the
best hyp
e
rpl
ane that
serve
s
a
s
a
sep
a
rato
r of two cla
s
ses
i
n
the
input
space. Fig
u
re
5
sh
ows
so
me p
a
ttern
s
of
membe
r
s
of two cl
asse
s: +1 and -1. The
pattern in
co
rporate
d
in cl
a
ss
-1 is
symb
olize
d
by gre
en
(ci
r
cl
e), whil
e
the pattern in cla
ss
+1 is
symbol
i
z
ed b
y
red (box). T
he cla
s
sificati
on pro
b
lem
can
be tran
slated
by finding
the
line
or hype
rplane
t
hat
se
parate
s
the t
w
o
group
s. V
a
riou
s alterna
t
ive
discrimi
nation
boun
dari
e
s a
r
e
sho
w
n i
n
F
i
gure
5(a)
. Th
e be
st se
pa
ra
tor hype
rpla
n
e
between th
e
two cl
asse
s
can
be fo
un
d by mea
s
u
r
ing the
hype
rplan
e
'
s
ma
rgins
and
se
arching
for t
he
maximum poi
nt.
Margi
n
is the
distan
ce bet
wee
n
the hyperpl
ane a
n
d
the closest
pattern of ea
ch cl
ass.
The clo
s
e
s
t pattern is
cal
l
ed a sup
p
o
r
t vector. The
solid line in
Figure 5
(
b)
sho
w
s the best
hyperpl
ane,
whi
c
h is l
o
cated at the cen
t
er of t
he two
cla
sse
s, whil
e the green a
nd re
d dots t
hat
are in the
bla
ck
circle a
r
e t
he su
ppo
rt vectors. The
eff
o
rt to locate the hype
rplan
e
is at the he
art
of the learnin
g
pro
c
e
ss in
SVM.
The available
data is denoted as
̅
, where
a
s ea
ch label
is denoted
∈
1,
1
for
i = 1,2, ....,
l
.
It
is
a
s
sum
ed
that
both
cla
s
se
s -1
and +1
can
be
p
e
rfe
c
tly sep
a
rate
d by
d
dimen
s
ion
e
d
hyperplan
e, defined as:
.
̅
0
. The
̅
pattern that belo
ngs to cla
s
s -1
(neg
ative sa
mple) can be
formulated as
a pattern that satisfies in
equality
.
̅
1
. While
the
̅
pattern t
hat belo
n
g
s
t
o
the
cla
ss
+
1 (p
ositive
sa
mple)
ca
n b
e
formulate
d
a
s
a patte
rn tha
t
satisfies inequality
.
̅
1
. Th
e greate
s
t margi
n
can
be found by maximizing
the
distan
ce val
u
e between th
e hype
rplan
e
and its ne
arest poi
nt,
‖
‖
. This
can
be fo
rmulate
d
a
s
a
Quad
ratic P
r
ogra
mming
(QP) problem,
which is to
fi
nd the minim
a
l point of eq
uation sho
w
n
in
equatio
n (4
), and (5
).
min
‖
‖
(4)
.
̅
1
0
,
∀
(5)
This p
r
obl
em can b
e
solve
d
by various
comp
ut
ationa
l techniq
u
e
s
, inclu
d
ing La
grange Multipli
er.
,
,
‖
‖
∑
.
̅
1,
1,2,
…
…
.
(6)
is a Lag
ra
ng
e multiplier th
at is either
0 or po
sitive
0
. The optimal
value of equ
ation (6
)
can be calcul
ated by minimizing
L
to
and
b
and ma
ximizing
L
to
. Consid
erin
g
the nature a
t
the optimum
point of slop
e
L
= 0, equ
ation (6
) ca
n
be modified
as a maxim
i
zation p
r
obl
em
contai
ning
only, as indicated by (7).
Maximize:
̅
∑
∑
̅
.
̅
,
(7)
Subject to
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Asthm
a
Identification Usin
g
Gas Sen
s
o
r
s
and Suppo
rt Vector….(Hari Agus Sujon
o
)
1473
0
f
o
r
1
,
2
,
….,
;
∑
0
(a)
(
b
)
Figure 5. The
SVM finds the best hype
rp
lane that se
p
a
rate
s the two cla
s
ses
From the results of this ca
lcul
ation, it
will be obtained
which m
o
stly positive valu
e
.
Data correlat
ed with po
sitive
is called a supp
ort ve
ctor. In gene
ral, probl
em
s in real wo
rld
domain
s
are rarely lin
ea
rly
sep
a
rate
but
mostly n
on-li
near.
To
solv
e the
non
-line
a
r
pro
b
lem
s
, t
he
SVM is modifi
ed by ente
r
in
g the Kern
el
functio
n
. In no
n-line
a
r SVM,
the data
x
i
s
mappe
d by th
e
function
Φ
̅
to
a higher-dime
n
sio
nal vecto
r
spa
c
e. In
th
is new vecto
r
space, the hyperpla
ne
that sep
a
rate
s the t
w
o
cla
s
se
s can b
e
constructe
d. T
o
cla
s
sify non
-linea
r d
a
ta, the SVM form
ula
must be m
o
d
i
fied. Therefo
r
e, the two li
miting fiel
ds
of (5) h
a
ve to
be ch
ang
ed
so that they
are
more flexible
for certain condition
s wit
h
the addition of the
variable
(
0
shown in the
equatio
n (8
).
̅
.
1
,∀
(8)
Thus th
e equ
ation (4
) is ch
ange
d to:
min
‖
‖
∑
(9)
C is
cho
s
en t
o
cont
rol the t
r
ade
off betwe
en margin an
d miscla
ssifi
cation. A large
C value mea
n
s
it will give a larger penalty for the misclassifi
cation.
Another meth
od of
solving
the no
n-lin
ea
r data
proble
m
s i
n
SVM i
s
by ma
ppin
g
data to
a
highe
r
dime
n
s
ion spa
c
e (f
eature sp
ace
)
[29],
with
d
a
ta in that sp
ace
can
be
separated lin
e
a
rly
by usin
g the
tran
sform
a
tion of
Φ:
→
.
Thus
the traini
ng
algorith
m
de
pend
s o
n
th
e data
throug
h the
dot pro
d
u
c
t in
H
(e.g.
Φ
.Φ
). If there
is a
kernel
K
function,
su
ch
as
,
Φ
.Φ
, thus in the trainin
g
algorithm requ
ires o
n
ly the kern
el
K
fu
nc
tion
without havin
g to know the
exact
Φ
transformatio
n. By
transfo
rmin
g
→
,
then the value
w
be
come
s
∑
and the lea
r
ni
ng functio
n
b
e
com
e
s:
∑
.
(10)
The featu
r
e
space u
s
u
a
lly has a
highe
r
dimen
s
ion
re
sulting
a fe
ature
sp
ace tha
t
may
b
e
very large. To solve this probl
em
, then it uses the
"kernel tri
c
k"
,
.
, for
whi
c
h the eq
uation (1
0) b
e
com
e
s:
∑
,
(11)
with
is the suppo
rt vecto
r
. There a
r
e
some kern
el functio
n
s that
are
comm
on
ly used in th
e
SVM, one of whi
c
h i
s
used
in this stu
d
y is the
RBF kernel.
In pa
rticula
r
, it is co
mmonly u
s
ed
in
.
̅
0
.
̅
1
.
̅
1
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 146
8-1480
1474
sup
port vect
or ma
chin
e cla
ssifi
cation
[30]. The
RBF ke
rnel
on two sam
p
les
and
rep
r
e
s
ente
d
as featu
r
e vectors in some
input spa
c
e i
s
define
d
in e
quation (12
)
,
2
⁄
(12)
The
cla
ssification reli
abilit
y is indi
cate
d by
the sen
s
itivity and specifi
c
ity resulted by
either p
o
sitiv
e
(di
s
ea
se
d) or neg
ative (healthy)
. In
the cla
ssifi
cation, the nu
mber of di
se
ase
subj
ect
s
id
ent
ified a
s
the
di
sea
s
e
subje
c
t
s
a
r
e
de
noted
by
t
p
(t
rue
po
sitive), the
di
sease
subj
ect
s
identified a
s
the healthy
subj
ect
s
a
r
e den
oted b
y
f
p
(false p
o
sitive), the
healthy subj
ects
identified a
s
the disea
s
e
subj
ect a
r
e d
enoted
by
f
n
(false
neg
ative) an
d the h
ealthy su
bje
c
ts
identified a
s
t
he he
althy subje
c
ts a
r
e
d
enoted
by
t
n
(true
neg
ative). Th
e sen
s
i
t
ivity, specificity,
and a
c
cura
cy
for two and t
h
ree
cla
s
ses
are defin
ed a
s
in Table 3 a
nd 4, re
spe
c
tively.
Table 3. Defi
nition of Sensitivity, S
pecivicity and Accu
racy for two classe
s
Predicted Result
Sensitiv
ity
Specivi
c
ity
Accur
a
cy
Positive
Negative
Actual
condition
Positive
p
t
n
f
n
p
p
f
t
t
p
n
n
f
t
t
n
n
p
p
n
p
f
t
f
t
t
t
Negative
p
f
n
t
Table 4. Defi
nition of Accu
racy for th
ree
classe
s
Figure 6. The
SVM Algorithm for cla
ssifi
cation
3. Results a
nd Analy
s
is
The el
ectroni
c no
se
ha
s
been te
sted
to detect
an
d identify ex
halled
bre
a
th
sam
p
le
s
from subje
c
t
s
with
health
y and a
s
thm
a
su
bje
c
ts a
nd amo
ng a
s
thma
subj
ects with different
Predicted Result
Accur
a
cy
A B
C
Actual
Condition
A
A
t
AB
f
AC
f
C
B
A
t
t
t
T
CB
CA
BC
BA
AC
AB
f
f
f
f
f
f
T
T
Acc
B
BA
f
B
t
BC
f
C
CA
f
CB
f
C
t
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Asthm
a
Identification Usin
g
Gas Sen
s
o
r
s
and Suppo
rt Vector….(Hari Agus Sujon
o
)
1475
severity usi
n
g the pro
c
ed
ure de
scri
be
d in the method. Figu
re
7 sho
w
s the
resp
on
se of
the
seven
sen
s
ors to th
e fou
r
exhaled
breat
h sample
s
du
ring
a p
e
rio
d
e
of 15
0
se
con
d
s. Fig
u
re 8
(a)
sho
w
s the av
erag
e respon
se for th
e two cate
gori
e
s.
It is possibl
e to obtain
a
combi
nation
of
sen
s
o
r
s th
at clea
rly indi
cat
e
each of the
tw
o catego
rie
s
. Sensor
re
spon
se
s of CO
, NO, H
2
S, and
CO
2
a
r
e
smal
ler for
health
y subje
c
ts, b
u
t not for
H
2
, NH
3
, an
d V
O
Cs. Figu
re
8 (b
) sho
w
s t
h
e
sen
s
o
r
re
sp
o
n
se
s for the f
our catego
rie
s
. We
have
condu
cted five
types of classificatio
n
s.
(a)
(b)
(c
)
(d)
Figure 7. The
sen
s
or
re
spo
n
se
s to (a
) he
althy (b
) controlled a
s
thma
(c) partly cont
rolled a
s
thm
a
,
and (d
) un
co
ntrolled a
s
th
ma
(a)
(b)
Figure 8. The
sen
s
or
re
spo
n
se
s to (a
) tw
o categ
o
rie
s
and, (b
) four
categ
o
rie
s
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1
9
17
25
33
41
49
57
65
73
81
89
97
10
5
11
3
12
1
12
9
13
7
14
5
Re
s
p
o
n
se
(V)
Ti
m
e
s
(s
)
CO
H2
NO
H2S
NH3
VOC
CO
2
0
0.2
0.4
0.6
0.8
1
1.2
1
9
17
25
33
41
49
57
65
73
81
89
97
10
5
11
3
12
1
12
9
13
7
14
5
Re
s
p
o
n
se
(V)
Ti
m
e
s
(s
)
CO
H2
NO
H2S
NH
3
VO
C
CO
2
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1
9
17
25
33
41
49
57
65
73
81
89
97
10
5
11
3
12
1
12
9
13
7
14
5
Re
spons
e
(V
)
Ti
m
e
s
(s
)
CO
H2
NO
H2
S
NH3
VOC
CO
2
0
0.2
0.4
0.6
0.8
1
1.2
1
9
17
25
33
41
49
57
65
73
81
89
97
10
5
11
3
12
1
12
9
13
7
14
5
Respons
e
(V)
Ti
me
s
(s)
CO
H2
NO
H2
S
NH
3
VO
C
CO
2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
C
O
H2
NO
H2S
N
H
3
VOC
C
O
2
R
e
spo
n
se
(V
)
Sen
s
or
s
HE
A
L
TH
Y
AST
H
MA
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
0.
7
0.
8
0.
9
C
O
H2
N
O
H
2
S
N
H3
VOC
C
O
2
R
e
sp
o
n
se
(V)
Sen
s
o
r
s
Hea
l
th
y
As
thma
C
o
ntr
o
lled
As
thma
Pa
rty
‐
C
o
ntr
o
lled
As
thma
U
n
co
ntr
o
lled
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 16, No. 4, August 2018: 146
8-1480
1476
3.1. Classific
a
tion of
Heal
th
y
and Asth
ma using se
v
e
n sensors
In this study
, we used the ele
c
tro
n
ic
nose to di
stinguish between exhal
ed
breat
h
sampl
e
s of healthy and
asthma su
bject
s
usin
g
non-lin
ear
binary SVM classificatio
n
.The
comp
ositio
n
of the
subj
e
c
t data
b
a
s
e
is
sho
w
n
in
Table
5. Th
e
exhale
d
b
r
e
a
th sampl
e
s of
subj
ect
s
we
re colle
cted from patients
who hav
e b
e
en diagn
osed
with early diagno
si
s stan
dard
usin
g ACT.
We rand
omly colle
cted 2
0
sampl
e
s fo
r
the trainin
g
sets from e
a
ch cate
gory. T
he
remai
n
ing
10
sam
p
le
s fro
m
ea
ch
cate
gory
were u
s
ed for the test s
e
t. In this
s
t
udy, we aim to
test the abili
ty of the electroni
c n
o
se
to
distingui
sh each cate
gory of healt
h
y and a
s
th
ma
subj
ect
s
usi
n
g binary cl
assification.
Table 5. The
comp
ositio
n of the subje
c
t
databa
se for
two cate
gori
e
s
Cate
g
ories of su
b
j
ects
Numbe
r
A
g
e
Health
y
30
32
–
60
Asthma
30
30
–
58
Table
6 sho
w
s t
he resul
t
s of cl
assifi
cation fo
r two cate
go
rie
s
. For th
e 1
0
asthm
a
subj
ect
s
a
s
the test
set,
the syste
m
coul
d
corre
c
tly identify 10.0 sam
p
le
s
as a
s
thm
a
a
nd
inco
rretly 0.0 sample a
s
h
ealthy. For the 10 heal
thy subj
ect
s
as the test set, the system
co
uld
corre
c
tly ide
n
tify 7.8 samples
as h
e
a
lthy and
in
corretly
2.2 sampl
e
s as asthma.
In
this
identificatio
n
re
sults, the
se
n
s
itivity,
spe
c
ificity, a
nd a
c
cura
cy
are
100, 7
8
.0, and 8
9
%
,
r
e
spec
tively.
Table 6. The
non-li
nea
r bin
a
ry SVM classificatio
n
re
su
lt of two categorie
s u
s
ing
seven
sen
s
o
r
s
Predicted Result
Sensitiv
ity
Specific
ity
Accur
a
cy
Positive
Negative
Actual
Condition
Asthma Positive
10.0
0.0
100 78
89
Health
y Negative
2.2
7.8
3.2. Classific
a
tion of
Heal
th
y
and Asth
ma
w
i
th Different Sev
e
rity
We
have
coll
ected
three
categ
o
rie
s
of
asth
ma
su
bj
ects sepa
rat
ed by
severit
y
, each
con
s
i
s
ting
of ten
sampl
e
s. Ten
sam
p
l
e
s fo
r h
ealth
y subj
ect
s
were ta
ke
n
ra
ndomly from
30
sampl
e
s. T
h
e
com
p
o
s
ition
of the subje
c
t
databa
se
is
sho
w
n i
n
Ta
b
l
e 7. We rand
omly coll
ecte
d
six sampl
e
s f
o
r the trainin
g
sets from
each ca
teg
o
ry. The remai
n
ing four
sa
mples from e
a
ch
cat
e
g
o
ry
we
r
e
u
s
ed f
o
r t
h
e
t
e
st
set
.
Tabl
e 8
sho
w
s t
h
e re
sult
of
cl
a
ssif
i
cat
i
on
f
o
r
f
our
cat
e
g
o
ri
e
s
.
CA subje
c
t
could b
e
di
stin
guished fo
rm
healthy
subj
ect
with 82%
accuracy, 9
0
%
sen
s
itivity and
75% spe
c
ifici
t
y.
PA subject could be differentia
ted fo
rm healthy subje
c
t with 87.5% accura
cy,
75%
sen
s
itivity and
100%
sp
ecifi
c
ity. UA
subj
ec
t could be disti
ngui
she
d
form
he
althy
su
bject
with 66% accura
cy, 65% sensit
ivity and 67.5% spe
c
ifi
c
ity.
Table 7. The
comp
ositio
n of the subje
c
t
databa
se for
four cat
ego
rie
s
Cate
g
ories of su
b
j
ects
Numbe
r
Health
y
10
Controlled asthm
a
(
CA
)
10
Partl
y
-cont
rolled asthma
(
PA
)
10
Uncontrolled asthma
(
UA
)
10
Table 8. The
non-li
nea
r bin
a
ry SVM classificatio
n
re
su
lt of four cate
gorie
s
Predicted Result
Sensitiv
ity
Specific
ity
Accur
a
cy
Positive
Negative
Actual
Condition
CA Positive
3.6
0.4
90 75
82.5
Health
y Negative
1.0
3.0
PA Positive
3.0
1.0
75 100
87.5
Health
y Negative
0.0
4.0
UA Positive
2.6
1.4
65 67.5
66.25
Health
y Negative
1.3
2.7
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Asthm
a
Identification Usin
g
Gas Sen
s
o
r
s
and Suppo
rt Vector….(Hari Agus Sujon
o
)
1477
3.3. Classific
a
tion of Th
re
e Categorie
s
of As
thma
w
i
th Differ
en
t Sev
e
rit
y
In this study,
we
have te
st
ed the
ele
c
tro
n
ic
no
se to
di
stingui
sh
thre
e different
cat
egori
e
s
of asthma
su
bject
s
with di
fferent seve
ri
ty
of CA, PA and UA u
s
i
ng the non
-li
near m
u
lti cl
ass
SVM classification. Table 9
sho
w
s
the
re
sults of no
n-li
near m
u
lti cla
ss SVM cl
assification for th
is
study. For th
e four
CA su
bject
s
a
s
the
test se
t, the
system
coul
d co
rrectly id
entify 0.9 as
CA,
inco
rrectly 1.
1 sa
mple
s a
s
PA, incorre
c
tly 2.0 sampl
e
s a
s
UA. Fo
r the fou
r
PA
subj
ect
s
a
s
t
h
e
test set, the system
could
corre
c
tly identify 2.2
samples a
s
PA, inco
rrectly 1.8
sample
s a
s
CA,
and in
co
rrectl
y 0.0 sample
s a
s
UA. Fo
r the fou
r
UA
subj
ect
s
a
s
t
he te
st set, the
system
co
uld
corre
c
tly iden
tify 1.6 sampl
e
s a
s
UA, incorrectl
y 1.0
sample a
s
CA, and in
co
rrectly 1.4 sampl
e
s
as PA. In this identification
,
the accura
cy is 39%
. The re
sult of identific
ation
shows that by only
providin
g accura
cy value, the sy
stem
ca
nnot distin
gui
sh amo
ng ca
tegori
e
s. Th
erefore, we h
a
ve
employed th
e binary cl
assificatio
n
for these th
ree categori
e
s. Ta
ble 10 sh
ows the non
-lin
ea
r
binary
SVM
classificatio
n
f
o
r th
ree
cate
gorie
s of a
s
t
h
ma
su
bje
c
ts. For the fo
ur CA
sa
mple
s as
the test
set, the
system
co
uld
co
rre
ctly i
dentif
y 0.4
sa
mples a
s
CA, and
in
co
rretl
y 3.6
sampl
e
s
as PA. Thi
s
i
ndicates th
at
the se
nsitivity is b
ad (i.e.
1
0
%). For the
four PA sam
p
les
as the t
e
st
set, the syste
m
could
co
rre
ctly
identify 3.3 sampl
e
s a
s
PA, and inco
rre
ctly 0.7 sa
mples a
s
CA.
Table 9. No
n-linear m
u
lti cl
ass SVM cla
s
sificatio
n
re
su
lt
Predicted Result
Accur
a
cy
CA PA
UA
Actual
Condition
CA 0.9
1.1
2.0
39%
PA 1.8
2.2
0
UA 1.0
1.4
1.6
Table 10. No
n-linie
r Binary
SVM classification result
Predicted Result
Sensitiv
ity
Specific
ity
Accur
a
cy
Positive
Negative
Actual
Condition
CA Positive
0.4
3.6
10.0%
82.5%
46,25%
PA Negative
0.7
3.3
CA Positive
1.6
2.4
40%
75%
57,5%
UA Negative
1.0
3.0
PA Positive
4.0
0.0
100%
57.5
78,75%
UA Negative
1.7
2.3
This in
dicate
s that the
sp
ecificity is
go
od (i
.e. 82.5%). Similarly, to dis
t
inguish between
CA an
d
UA
subj
ect
s
, the
syste
m
's ab
ility sh
o
w
s
p
oor se
nsitivity
(i.e.
40%) and quite
g
o
od
spe
c
ificity (i.e. 75%). Finally, to differentia
te between PA and
UA subje
c
ts, the syste
m
’s
cap
ability sho
w
s g
ood a
c
cu
racy (i.e. 78.7
5
%).
3.4. Classific
a
tion of
Heal
th
y
a
nd Asth
ma using fiv
e
sensor
s
In this
study, we u
s
e
d
the
respon
se
fr
om five
s
e
ns
ors
,
i.e
.
C
O
,
NO
, VO
C
,
H
2
and CO
2
sen
s
o
r
s, to
differentiate
betwe
en he
a
l
thy and asth
ma su
bje
c
ts
usin
g non
-lin
ear bi
nary S
V
M
cla
ssif
i
cat
i
on.
Table 1
1
s
h
ow
s t
he
cla
s
sif
i
cat
i
o
n
re
sults obtai
ning
the value
s
o
f
the sen
s
itivity,
spe
c
ificity, an
d accuracy a
r
e
100, 79, an
d 89.5%, respectively.
Table 11. Th
e non-li
nier bi
nary SVM cla
ssifi
cation
result of two categori
e
s u
s
in
g
five senso
r
s
Predicted Result
Sensitiv
ity
Specific
ity
Accur
a
cy
Positive
Negative
Actual
Condition
Asthma Positive
10.0
0.0
100 79
89.5
Health
y Negative
2.1
7.9
3.5. Classific
a
tion of
Heal
th
y
and Asth
ma using thr
ee sens
o
rs
In this study, we used th
e re
spon
se f
r
om
three se
nso
r
s o
n
ly, i.e. CO, NO a
nd VOC
sen
s
o
r
s, to
differentiate
betwe
en h
ealthy and
asthma
subj
ects. T
able
12
sho
w
s th
e
cla
ssifi
cation
results obtai
n
i
ng the value
s
of the se
n
s
i
t
ivity, specificity,
and accu
racy are 9
0
, 79
,
and 84.5%, resp
ectively.
Evaluation Warning : The document was created with Spire.PDF for Python.