Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 4
,
A
ugu
st
2016
, pp
. 18
00
~
1
810
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
4.9
902
1
800
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Hybrid Approach for Prediction
of Cardiovascul
ar Dis
e
as
e
Using Cl
ass Associati
o
n Rules an
d MLP
K. Srini
v
as
1
,
B. R
a
m
a
sub
b
a
Redd
y
2
,
B
.
Ka
vi
th
a R
a
ni
1
, R
a
vi
nd
ar
M
ogi
l
i
3
1
Professor, J
y
othishmathi Institu
t
e of
T
echnolog
y
& Scien
ce, Karimnagar, TS, In
dia
2
Professor, SVEC, Tirupati, AP,
India
3
Associate
Prof
essor, J
y
o
t
hishm
a
thi
Institu
te
of
Techno
log
y
&
Scien
c
e
,
Kar
i
m
n
agar,
TS, Ind
i
a
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Dec 28, 2015
Rev
i
sed
Feb
26
, 20
16
Accepted
Mar 10, 2016
In data m
i
nin
g
clas
s
i
fi
cat
ion
tec
hniqu
es ar
e used to predict group
m
e
m
b
ers
h
ip for data
ins
t
an
ces
.
T
h
es
e t
echniqu
es
are
cap
able
of pr
oces
s
i
ng
a
wider variety
of
data
and the output can
be
eas
i
l
y
interpr
e
ted
.
Th
e
aim
of an
y
classification algorithm is the
design
and conception of a standard model
with refer
e
nce to
the given
input.
The
model thus
generated may
b
e
deplo
y
ed
to clas
s
i
f
y
n
e
w exam
ples
or enable a bett
er com
p
rehens
ion of ava
ilabl
e dat
a
.
Medical data classification is the pro
cess of transforming descriptions of
medical diagnoses and procedur
es us
ed to find hidden information
.
Two
experim
e
nts
ar
e perform
ed t
o
identif
y th
e
predict
i
on ac
curac
y
of
Cardiovas
c
u
l
ar
Dis
eas
e (CVD).
A
h
y
brid
appr
oach for
cl
as
s
i
fica
tion
is
proposed in this paper b
y
combining the
results of the associate classifier and
artif
ici
a
l n
e
ura
l
networks (MLP).
Th
e first exp
e
riment
is perfo
rmed using
as
s
o
ciat
ive c
l
as
s
i
fier to
ident
i
f
y
the ke
y a
ttribu
t
es
which contr
i
bute m
o
r
e
towards the decision b
y
tak
i
n
g
the
13 indep
e
ndent
attr
ibutes as input.
Subsequentl
y
cl
assifica
tion
usi
ng
Multi
La
ye
r Percep
trons (
M
LP) also
performed to generate the accur
a
cy
of
predictio
n using all attr
ibutes. In the
second exp
e
riment, iden
tified k
e
y
attribu
t
es using associative classifier
ar
e
used as inputs for the feed fo
rward neural n
e
tworks for predicting
th
e
pres
ence
or
abs
e
nce of
CVD.
Keyword:
Artificial n
e
u
r
al n
e
two
r
k
s
Ass
o
ciative cla
ssifier
Classification
CVD
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
1.
INTRODUCTION
W
i
t
h
t
h
e e
v
er
-g
ro
wi
n
g
c
o
m
p
l
e
xi
t
y
i
n
rece
nt
y
ears,
h
u
g
e
am
ount
s
of
i
n
f
o
rm
at
i
on i
n
t
h
e area
o
f
m
e
di
ci
ne ha
ve
been
sa
ved
eve
r
y
day
i
n
di
ffe
r
e
nt
el
ect
ronic form
s
suc
h
as E
l
ectronic Health
Rec
o
rds (E
HRs)
and
regi
st
e
r
s
whi
c
h i
s
use
d
fo
r di
f
f
e
r
ent
p
u
r
p
oses. C
a
r
d
i
ova
scul
ar
di
se
ase (hea
rt
di
se
ase) [
1
]
-
[
3
]
ref
e
rre
d as
CV
D is th
e class of
d
i
seases th
at invo
lv
e t
h
e h
e
ar
t or
b
l
ood v
e
ssels.
I
t
is essen
tial to
ev
al
u
a
te th
e pr
esence or
abse
nce of cardiovasc
u
lar di
sease (CVD) risk. Seve
ra
l
m
e
t
h
o
d
s are di
s
c
usse
d by
res
earche
r
s to improve
cardi
ovasc
ul
ar
ri
sk
p
r
edi
c
t
i
o
n.
The
dat
a
of
t
h
e pat
i
e
nt
s c
o
l
l
ect
ed f
r
om
di
f
f
e
r
ent
s
o
urces
i
s
st
ore
d
i
n
regi
s
t
ers
and m
a
inly us
ed for m
onitoring a
n
d
analyz
ing
health c
o
nditions
. T
h
e exis
tence of accurate epidem
iological
reg
i
sters a
b
a
sic p
r
erequ
i
site fo
r m
o
n
ito
ri
ng
and
an
alyzing
h
ealth
an
d
so
cial co
nd
itio
ns in
th
e po
pu
l
a
tio
n
.
They
are
fre
qu
ent
l
y
used
fo
r
researc
h
, e
v
al
u
a
t
i
on,
pl
an
ni
n
g
and
ot
he
r
pu
r
pos
es by
a
vari
et
y
of use
r
s i
n
t
e
rm
s
of
anal
y
z
i
n
g a
n
d
p
r
e
d
i
c
t
i
ng t
h
e
heal
t
h
st
at
u
s
o
f
i
n
di
vi
dual
s
.
Dat
a
M
i
ni
ng a
i
m
s
at
di
scove
ri
n
g
k
n
o
w
l
e
d
g
e
out
o
f
dat
a
a
nd
pre
s
ent
i
n
g
i
t
i
n
a form
t
h
at
i
s
easi
l
y
com
p
ressi
bl
e t
o
h
u
m
a
ns. It
i
s
a pr
ocess t
h
at
i
s
de
vel
o
p
e
d t
o
e
x
am
i
n
e l
a
rge am
ount
s of
dat
a
r
o
ut
i
n
el
y
col
l
ect
ed.
Dat
a
m
i
ni
ng i
s
m
o
st
usef
ul
i
n
a
n
e
xpl
orat
ory
a
n
alysis scenari
o
i
n
whic
h
the
r
e are
no predetermine
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Hybri
d
A
p
pr
oa
ch f
o
r Pre
d
i
c
t
i
o
n
of
C
a
r
d
i
o
v
a
scul
ar
Di
se
ase
Usi
n
g C
l
ass
A
ssoci
at
i
o
n
Rul
e
s ...
. (
K
.
Sri
n
i
v
as)
1
801
n
o
tion
s
abou
t
wh
at
will con
s
titu
te an
"in
t
erestin
g
"
ou
tco
m
e. Data m
i
n
i
n
g
is th
e search
fo
r n
e
w, v
a
l
u
able, and
no
nt
ri
vi
al
i
n
f
o
rm
ati
on i
n
l
a
r
g
e
vol
um
es of dat
a
. B
e
st
r
e
s
u
l
t
s
are ac
hi
ev
ed
by
bal
a
nci
n
g t
h
e
k
n
o
w
l
e
d
g
e
of
h
u
m
an
exp
e
rts
in
d
e
scri
b
i
ng
p
r
ob
lem
s
an
d
g
o
a
ls with
t
h
e search
cap
a
b
i
lities o
f
co
m
p
u
t
ers. In
p
r
acti
ce, th
e
t
w
o
pri
m
ary
goal
s
o
f
dat
a
m
i
ni
n
g
t
e
n
d
t
o
be
cl
assi
fi
catio
n
an
d
p
r
ed
iction
.
Pred
iction
[4
]
in
vo
lv
es u
s
ing
so
m
e
v
a
riab
les or field
s
in
th
e
dataset to
p
r
edict u
n
kno
wn
o
r
fu
t
u
re
v
a
lues o
f
o
t
h
e
r
variab
les of in
t
e
rest.
Classificatio
n
[5
],[6
] refers to th
e task
o
f
analyzin
g
a se
t of
pre
-
classified data
ob
jects
to learn a
m
ode
l (or
a
fu
nct
i
o
n) t
h
at
can be
use
d
t
o
cl
assi
fy
unsee
n dat
a
o
b
j
ect
i
n
t
o
one
of se
v
e
ral
pre
d
efi
n
e
d
cl
asses. Desc
r
i
pt
i
on,
on
t
h
e
ot
her
h
a
nd
,
foc
u
ses
o
n
fi
n
d
i
n
g
pat
t
e
rns
de
scri
bi
ng
t
h
e dat
a
t
h
at
c
a
n
be i
n
t
e
rp
ret
e
d
by
h
u
m
a
ns.
Fo
r a
gi
ve
n a col
l
ect
i
on o
f
rec
o
r
d
s
(t
rai
n
i
n
g set
)
e
ach rec
o
r
d
contain
s
a set o
f
attrib
u
t
es, ou
t of wh
ich
on
e of th
e
attrib
u
t
e is th
e class attrib
u
t
e o
r
class
v
a
riab
le. Ot
h
e
r at
t
r
i
but
es ar
e o
f
t
e
n cal
l
e
d i
n
de
p
e
nde
nt
o
r
pre
d
i
c
t
o
r
attrib
u
t
es (o
r
variab
les). Th
e
set o
f
exam
ples used to learn the classificat
i
o
n
m
o
d
e
l is called
th
e train
i
ng
d
a
ta
set.
W
e
nee
d
to fi
nd a m
o
del for class
attribute as
a
fu
nct
i
o
n
o
f
t
h
e
val
u
es
o
f
ot
he
r at
t
r
i
b
ut
es.
F
u
rt
he
r
pre
v
i
o
usl
y
u
n
s
een
reco
rd
s s
h
oul
d
be assi
g
n
ed
a class as accu
rately as
p
o
ssib
le.
A test set is u
s
ed
to d
e
t
e
rm
in
e
the accuracy of the m
odel. Us
ually, the gi
ve
n data set is
di
vide
d int
o
training a
n
d test sets, training set
used to
b
u
ild th
e m
o
d
e
l an
d test set
u
s
ed
to v
a
lidate it.
The
basi
c
o
r
ga
ni
zat
i
o
n
o
f
t
h
e
pa
per
i
s
as
f
o
l
l
o
ws:
Section 2 pr
esen
ts th
e r
e
v
i
ew
of related
works,
Sectio
n
3
d
e
scrib
e
s asso
ciativ
e classifier and
m
u
ltilayer
p
e
rcep
tro
n
s
(MLP),
Section
4
d
e
scrib
e
s
t
h
e
prop
o
s
ed
hy
b
r
i
d
cl
assi
fi
er a
p
p
r
oac
h
,
S
ect
i
on
5
pre
s
e
n
t
s
t
h
e
res
u
l
t
s
an
d
di
scus
si
o
n
, a
n
d t
h
e
c
o
n
c
l
u
si
o
n
s a
r
e
gi
ven
i
n
sect
i
on 6.
2.
RELATED WORK
Whe
n
processi
ng large
databases, one faces
two m
a
jor obstacles such as
num
erous sa
m
p
les
and
high
dim
e
nsionality of t
h
e fe
ature sp
ace. For exam
ple, the
doc
um
ents are
re
prese
n
ted by several
t
h
ousa
nds
of
words; im
ages
are com
posed
of m
i
llions
of
pixels, whe
r
e each word or pixe
l is here understood as a fe
ature
.
Cu
rren
tly, processin
g
ab
iliti
es are often
no
t ab
le to
h
a
n
d
l
e su
ch
h
i
gh
d
i
m
e
n
s
io
n
a
l
d
a
ta, m
o
stly
d
u
e
t
o
n
u
m
erical d
i
ffi
cu
lties in
pro
c
essin
g
, requ
iremen
ts in
st
orag
e an
d transm
i
ssio
n
with
i
n
a
reason
ab
le ti
me. To
reduce t
h
e c
o
mputational tim
e
,
it is comm
on
practice to
project the data ont
o
a
sm
aller, latent s
p
ace.
More
ove
r, s
u
c
h
a s
p
ace is
ofte
n be
ne
ficial for
furt
her investigation du
e
to noise re
duction
a
n
d
desire
d feature extraction
properties. Sm
aller dim
e
ns
ions are also advanta
g
eo
us w
h
en vi
sual
i
z
i
n
g
an
d
analyzing t
h
e
data. T
h
us, in orde
r to e
x
tra
c
t desira
ble inform
ation,
di
m
e
nsi
onality reduction m
e
thods
are
o
f
ten
ap
p
lied
.
Th
e ov
erall idea is to
d
e
termin
e th
e co
o
r
d
i
n
a
te syste
m
wh
ere th
e m
a
p
p
i
n
g
will create lo
w-
di
m
e
nsi
onal
co
m
p
act
represe
n
t
a
t
i
on o
f
t
h
e
da
t
a
whi
l
s
t
m
a
xi
m
i
zi
ng t
h
e i
n
f
o
rm
ati
on c
ont
ai
ned
wi
t
h
i
n
.
Th
ere are m
a
n
y
so
lu
tion
s
to th
is prob
lem
.
Sev
e
ral techniq
u
e
s
fo
r
d
i
m
e
n
s
ion
a
lity redu
ctio
n h
a
ve
been
devel
ope
d w
h
i
c
h use
bot
h l
i
n
ear a
nd
no
n-l
i
near
m
a
ppi
ng
s. A
m
ong t
h
em
are, l
o
w
-
di
m
e
n
s
i
ona
l
pr
o
j
ect
i
ons o
f
t
h
e
dat
a
, neu
r
a
l
net
w
o
r
ks
sel
f
-o
rga
n
i
z
i
n
g m
a
ps.
One ca
n
appl
y
sec
o
n
d
or
der m
e
t
hod
s
whi
c
h
use t
h
e
co
va
ri
ance st
r
u
ct
u
r
e i
n
det
e
rm
i
n
i
ng
di
rect
i
o
ns.
Pri
n
ci
pal
C
o
m
pone
nt
A
n
al
y
s
i
s
t
h
a
t
rest
ri
ct
s di
re
c
t
i
ons
to
th
o
s
e is ortho
gon
al. Factor An
alysis wh
ich
add
itio
n
a
lly
allo
ws th
e no
ise lev
e
l to
d
i
ffer alo
n
g
th
e d
i
rectio
n
s
and
In
de
pen
d
e
nt
C
o
m
pone
nt
Anal
y
s
i
s
f
o
r
whi
c
h t
h
e
di
rect
i
o
n
s
ar
e i
nde
pen
d
e
n
t
but
n
o
t
nec
e
ssari
l
y
ort
h
o
g
onal
.
Literatu
re presen
ts a lo
t o
f
tech
n
i
q
u
e
s fo
r C
VD
u
s
i
n
g m
a
chi
n
e l
ear
ni
n
g
t
echni
que
s. He
r
e
, we p
r
ese
n
t
som
e
of the
significant
resea
r
c
h
es a
v
ailable i
n
recent tim
e.
Sellap
p
a
n
Palan
i
app
a
n
et al. [7
] p
r
op
o
s
ed
a p
r
o
t
o
t
yp
e In
tellig
en
t Heart Disease Pred
ictio
n
Syste
m
(I
HD
PS)
by
m
eans o
f
dat
a
m
i
ni
ng a
p
pr
oac
h
es suc
h
as De
ci
si
on T
r
ees,
Naï
v
e B
a
y
e
s,
and
Ne
ural
Ne
t
w
o
r
k
.
Results ha
ve revealed that each appr
oac
h
has its unique
pote
n
cy in rea
lizing the obje
c
tives of t
h
e define
d
mining goals. The proposed syste
m
ha
s the
potential to for
ecast
the
possibility of
pat
i
ents getting a
hea
r
t
di
sease wi
t
h
t
h
e ai
d of m
e
di
cal
pro
f
i
l
e
s suc
h
as age
,
sex
,
bl
o
od
pre
ssu
re
, an
d bl
o
o
d
su
gar
.
Al
so
, t
h
e r
e
l
e
van
t
kn
o
w
l
e
d
g
e ha
s been est
a
bl
i
s
he
d, f
o
r e.
g.
pat
t
e
rns
,
rel
a
t
i
ons
hi
ps bet
w
e
e
n m
e
di
cal
fact
ors rel
a
t
e
d t
o
hea
r
t
disease.
C
a
rl
os O
r
d
o
n
e
z
et
al
. [8]
pro
p
o
sed E
v
al
u
a
t
i
ng a
ssociation rules and deci
sion
trees to
p
r
ed
ict
m
u
l
tip
le targ
et attrib
u
t
es an
d p
r
esen
ted
a detailed
co
m
p
ar
i
s
on bet
w
ee
n con
s
t
r
ai
ne
d as
soci
at
i
on r
u
l
e
s
and
d
ecision
trees to
p
r
ed
ict mu
ltip
le target
attrib
u
t
es
. Im
p
o
r
tan
t
d
i
fferences b
e
t
w
een
b
o
t
h
tech
n
i
q
u
es are
id
en
tified
for
su
ch
go
al.
Ex
ten
s
iv
e
e
xpe
rim
e
ntal evaluation was done on a
r
eal
m
e
d
i
cal d
a
ta set to
m
i
n
e
ru
les
pre
d
icting
disease on m
u
ltiple heart art
e
ries. The an
tecedent of a
ssociation
rules contai
ns
medical
m
easurem
ent
s
and
pat
i
e
nt
ri
s
k
fact
o
r
s,
whe
r
eas t
h
e conse
q
uent
re
fers t
o
t
h
e de
gree
of di
sease on
one a
r
t
e
ry
or
m
u
lt
i
p
l
e
art
e
ri
es. Pre
d
i
c
t
i
v
e r
u
l
e
s f
o
u
n
d
by
con
s
t
r
ai
ne
d as
soci
at
i
on
rul
e
m
i
ni
ng are m
o
re ab
u
nda
nt
an
d ha
v
e
h
i
gh
er reliab
ility th
an
p
r
ed
ictiv
e
ru
les i
n
du
ced
b
y
d
ecisi
o
n
trees.
Mo
h
a
mm
ed
Kh
alilia et al. [9
] p
r
opo
sed
p
r
ed
ictin
g
d
i
sease risk
s fro
m
h
i
g
h
l
y i
m
b
a
lan
ced
d
a
ta using
ran
d
o
m
forest
,
i
n
w
h
i
c
h t
h
ey
prese
n
t
e
d a
n
e
ffect
i
v
e
pr
oact
i
v
e ap
pr
oac
h
re
qui
res an
u
nde
rst
a
n
d
i
n
g o
f
di
sease
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
18
00
–
1
810
1
802
in
terd
ep
en
d
e
n
c
ies an
d
h
o
w
th
ey tran
slate in
to
a p
a
tient's fu
tu
re.
Due to
co
mm
o
n
g
e
n
e
tic, m
o
lecu
lar,
en
v
i
ron
m
en
tal, an
d
lifestyle-based
ind
i
v
i
dual
risk factors, most diseases do
not
occ
u
r i
n
i
s
ol
at
i
on. S
h
a
r
ed
ri
sk
and
e
nvi
r
o
nm
ent
a
l
fact
o
r
s
ha
ve si
m
i
l
a
r cons
eque
nces
,
pr
o
m
pti
ng t
h
e c
o
-
o
ccu
rre
nce
o
f
r
e
l
a
t
e
d di
sease
s
i
n
t
h
e
sam
e
pat
i
e
nt
. There
f
o
r
e, a
pat
i
e
nt
di
ag
n
o
se
d fo
r a co
m
b
i
n
at
i
on of
di
seases an
d
exp
o
se
d t
o
sp
eci
fi
c
envi
ro
nm
ent
a
l
,
l
i
f
est
y
l
e
and
genet
i
c
ri
s
k
fa
ct
ors m
a
y
be
at
consi
d
era
b
l
e
ri
sk
of
de
ve
l
opi
n
g
se
ve
ral
ot
he
r
g
e
n
e
tically and env
i
ro
n
m
en
tally related
d
i
seases.
An
u
p
ri
y
a
et
al
. [1
0]
pr
op
ose
d
Enha
nce
d
Pre
d
i
c
t
i
on
of Hea
r
t
Di
sease wi
t
h
Feat
ure S
ubs
e
t
Sel
ect
i
on
u
s
ing
Gen
e
tic
Algo
rith
m
in
wh
ich
th
ey used
Gen
e
tic al
g
o
rith
m
s
to
d
e
termin
e th
e attrib
u
t
es wh
ich
contrib
u
t
e
m
o
re to
ward
s t
h
e
d
i
agn
o
s
is
of h
e
art ail
m
en
ts wh
ich
ind
i
re
ctly reduces t
h
e num
b
er of tests whic
h are neede
d
t
o
be t
a
ken
by
a pat
i
e
nt
.
Thi
r
t
een at
t
r
i
b
ut
es
are re
d
u
ced
t
o
6 at
t
r
i
b
ut
es
usi
n
g
genet
i
c
search
. S
u
bseq
uent
l
y
,
three classi
fiers like
Naive
Bayes, Classification
by cl
ustering a
n
d Decision
Tree
a
r
e
used to pre
d
ict the
diagnosis of
pa
tients with t
h
e
sam
e
accuracy as obtain
e
d
before
the reducti
o
n
of
num
b
er of attributes.
Th
is article is a sign
ifican
t ex
ten
s
i
o
n of
[11
]
,
whe
r
e
deci
sion trees a
n
d
MLP are
com
b
ine
d
for t
h
e
first ti
m
e
, in
th
e co
n
t
ex
t o
f
h
e
art d
i
sease pred
ictio
n
.
Th
is
wo
rk
also
sho
w
ed
th
at
decision trees
where numeric
attrib
u
t
es are
au
to
m
a
tical
ly
sp
lit d
o
n
o
t
pro
d
u
ce m
u
ch
be
tter ru
les. In
th
is n
e
w article we presen
t a
m
o
re
com
p
rehe
nsi
v
e
experi
m
e
nt
al
eval
uat
i
o
n co
m
p
ari
ng t
r
adi
t
i
onal classification techniques
w
ith
p
r
op
osed
h
ybr
i
d
approach.
We
analyze how
g
ood
p
r
ed
ictive attrib
u
t
es are iso
l
ated
by decision
tree and
use
d
for furt
her
classification of
CVD.
3.
DEFIN
I
TIO
N
S
Ass
o
ci
at
i
on r
u
l
e
s bri
n
g a st
ron
g
an
d i
n
se
p
a
rabl
e b
o
nd
be
t
w
een i
t
e
m
s
and t
h
i
n
g
s
t
h
at
feat
ure
wi
t
h
reg
u
l
a
ri
t
y
i
n
a
gi
ve
n dat
a
set
.
The r
u
l
e
s t
hus
obt
ai
ne
d t
o
fi
nd
depl
oy
m
e
nt i
n
real
l
i
f
e event
s
suc
h
as l
o
oki
ng
in
to
th
e pu
rchasin
g
p
a
ttern
of clien
t
s
or p
r
e
f
ere
n
ces of
sh
op
pe
rs
f
o
r
a p
a
rt
i
c
ul
ar pr
o
d
u
c
t
.
Suc
h
o
b
ser
v
at
i
o
n
s
com
e
in useful
when
determ
ining cross
-
m
a
rketing st
rategi
es, catalogue
design
a
n
d
product placem
ent. The
poi
nt
of
depa
rt
ure f
o
r associ
a
t
i
on r
u
l
e
s st
em
s from
dat
a
set
m
i
ni
ng d
one
o
n
a fre
que
nt
b
a
si
s. It
i
s
a poi
nt
o
f
comm
on consent that strong associatio
ns correlate to freque
nt patterns and class-sets. Because association
rul
e
s ex
pl
o
r
e
hi
g
h
l
y
con
f
i
d
e
n
t
associ
at
i
o
n
s
am
ong m
u
l
t
i
pl
e at
t
r
i
but
es,
t
h
i
s
app
r
oac
h
m
a
y
overcom
e so
m
e
constraints i
n
troduced by
decision-
t
r
ee ind
u
c
tion
,
wh
ich
co
nsid
ers
on
ly on
e attribu
t
e at a tim
e
.
Th
e
asso
ciativ
e classificatio
n
h
a
s
b
een id
en
tified to
b
e
m
o
re
p
r
ecise th
an so
m
e
trad
ition
a
l classifiers, lik
e C
4
.5
.
In
th
is p
a
p
e
r w
e
pr
opo
se an
ap
pro
ach
for ass
o
ciative classification.
The
ove
ral
l
pr
ocess
of t
h
e h
y
b
ri
d cl
assi
fi
e
r
i
s
di
vi
de
d i
n
t
o
t
w
o st
eps
,
s
u
ch a
s
1
)
I
d
e
n
t
i
f
y
i
ng key
attributes using
associative classifier
al
g
o
r
i
t
h
m
and 2
)
P
r
edi
c
t
i
on
usi
n
g M
L
P.
3.
1.
Ass
o
ciati
v
e Cl
assifier
Ass
o
ci
at
i
on an
al
y
s
i
s
, C
l
assi
ficat
i
on an
d C
l
ust
e
ri
ng ar
e t
h
re
e di
ffer
e
nt
Dat
a
M
i
ni
ng t
ech
ni
q
u
es. T
h
e
aim
of any cla
ssification algorithm
is the design a
n
d c
oncep
tio
n of a st
an
d
a
rd
m
o
d
e
l with
referen
c
e to
th
e
gi
ve
n i
n
p
u
t
.
T
h
e m
odel
t
hus
gene
rat
e
d m
a
y
t
h
en
be de
pl
oy
ed t
o
cl
assi
f
y
new e
x
am
pl
es or
ena
b
l
e
a
bet
t
e
r
com
p
rehe
nsion of available data. Classification is a two
step process c
o
nsists of
training phase and testing
p
h
a
se. Th
e set o
f
ru
les will b
e
g
e
n
e
rated
du
ri
n
g
th
e train
i
n
g
p
h
a
se fro
m
th
e train
i
n
g
d
a
ta.
Th
e test p
h
a
se
h
e
lp
s
us t
o
determ
ine the acc
uracy
of the
classifi
er.
Differe
n
t a
p
proaches
ha
ve bee
n
propos
ed to build
ac
curate
classifiers, s
u
c
h
as
, Support
Vector
Mac
h
ines
(SVM), na
ive Bayesian c
l
assification,
Decision T
r
ee
s bas
e
d
classification a
n
d so on.
The A
ssoci
at
i
v
e cl
assi
fi
cat
i
on i
s
a ne
w
pr
op
ose
d
cl
assi
fi
cat
i
on t
e
chni
que
[8]
.
It
per
f
o
r
m
s
classification
by usi
n
g ass
o
ciation rules.
These
rules
ar
e st
rai
g
ht
f
o
r
w
ar
d a
n
d
si
m
p
l
e
t
o
u
nde
rst
a
nd
. I
n
associative classification, the
m
odel consis
ts of cl
ass association rules
where eac
h rule conse
que
nt is
restricted t
o
a
class attribut
e. R
ecent
studies show that
the a
p
proac
h
has ac
hieve
d
higher accurac
y
than
trad
itio
n
a
l
app
r
o
ach
es.
The
steps of associative
classifi
cat
i
on t
e
c
hni
que a
r
e r
u
l
e
gene
rat
i
o
n,
cl
assi
fi
er co
ns
t
r
uct
i
o
n an
d
classificatio
n
an
d
are sh
own in
Fig
u
r
e 5.1. Ru
le g
e
n
e
ratio
n
ph
ase will g
e
n
e
rate Class Asso
ciation
Ru
les
(C
AR
s)
by
usi
n
g
ass
o
ci
at
i
o
n
rul
e
m
i
ni
ng t
e
chni
que
s. C
l
as
si
fi
er i
s
c
o
nst
r
uct
e
d
f
r
om
t
h
e r
u
l
e
s
obt
ai
ne
d
i
n
t
h
e
pre
v
i
o
us
st
ep.
C
l
assi
fi
cat
i
on
pha
se assi
gns
a
cl
ass l
a
bel
t
o
t
h
e
gi
ve
n
ob
ject
.
We are
propo
sin
g
an
ap
pro
ach to
p
e
rfo
r
m
Cl
assificatio
n
b
a
sed
o
n
Po
sitiv
e an
d
Neg
a
tiv
e
Asso
ciation
Ru
les wh
ich
are k
n
o
w
n
as Class Asso
ciation
Ru
les. Prim
a
r
ily Class Association Rules of the form
X c are
mined whe
r
e
X is a
set
of at
tributes
and c
is a category
or
a
class of
an object.
He
re X
c nee
d
not be
only
p
o
s
itiv
e asso
ci
atio
n
ru
le rather it
m
a
y b
e
a
n
e
g
a
tiv
e as
sociatio
n
ru
le. Th
en
a classifier is co
n
s
tru
c
ted
b
y
co
nsid
eri
n
g stro
ng
ru
les,
wh
ich
is called
an
Asso
ciativ
e Cl
assifier.
It take
s an
object
as i
n
pu
t th
en
it attach
es
a
cl
ass l
a
bel
f
o
r
t
h
e
gi
ve
n i
n
p
u
t
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Hybri
d
A
p
pr
oa
ch f
o
r Pre
d
i
c
t
i
o
n
of
C
a
r
d
i
o
v
a
scul
ar
Di
se
ase
Usi
n
g C
l
ass
A
ssoci
at
i
o
n
Rul
e
s ...
. (
K
.
Sri
n
i
v
as)
1
803
Figure
1. Ass
o
ciative Classifier
3.
2.
Multilayer Pe
rceptr
ons (MLP)
A
m
u
ltila
yer feed
forw
ard neu
r
al
n
e
two
r
k
is an
in
terconnectio
n
o
f
p
e
rcep
tro
n
s
in
wh
ich
d
a
ta and
cal
cul
a
t
i
ons fl
ow i
n
a si
ngl
e
di
rect
i
o
n, f
r
o
m
t
h
e i
nput
da
t
a
t
o
t
h
e out
pu
t
s
. The n
u
m
b
er o
f
l
a
y
e
rs i
n
a
neu
r
a
l
net
w
or
k i
s
t
h
e num
ber o
f
l
a
y
e
rs of
perce
p
t
r
o
n
s. M
L
P ne
ura
l
net
w
o
r
ks c
o
n
s
i
s
t
of u
n
i
t
s
arr
a
nge
d i
n
l
a
y
e
rs
[1
2]
-
[1
4]
. Eac
h
l
a
y
e
r i
s
c
o
m
pose
d
of
n
o
d
e
s a
n
d i
n
t
h
e
f
u
l
l
y
co
nnect
e
d
ne
t
w
o
r
k
co
nsi
d
e
r
ed
he
re eac
h
n
ode
connects to every node in subse
que
nt layers. Each MLP
is com
posed of a
m
i
nim
u
m
o
f
th
ree layers con
s
isti
ng
of an i
n
p
u
t
l
a
y
e
r, one
or m
o
re hi
dde
n l
a
y
e
rs an
d an o
u
t
put
l
a
y
e
r. The
i
nput
l
a
y
e
r di
st
ri
but
es t
h
e i
n
put
s t
o
sub
s
eq
ue
nt
l
a
y
e
rs.
In
p
u
t
no
de
s ha
ve l
i
n
e
r
ac
t
i
v
at
i
on
fu
nct
i
ons
an
d
n
o
t
h
r
e
sh
ol
ds.
Eac
h
hi
d
d
en
u
n
i
t
n
o
d
e a
n
d
each
out
put
node
have
thresholds a
ssociate
d
with t
h
em
in addition to the
weights. The
hidde
n
unit
nodes ha
ve
nonlinea
r activation functions
and th
e outputs have linear
activation func
tions. Hence
,
each signal feedi
ng
in
to
a nod
e i
n
a sub
s
eq
u
e
n
t
l
a
yer h
a
s th
e orig
in
al i
n
pu
t m
u
ltip
lied
b
y
a
weigh
t
with
a
th
resh
o
l
d
add
e
d
an
d
t
h
en i
s
passed
t
h
r
o
u
g
h
an act
i
v
at
i
on f
u
nct
i
o
n
t
h
at
m
a
y
be linear o
r
n
onl
i
n
e
a
r (hi
dde
n u
n
i
t
s
). Ne
ural
net
w
o
r
k
s
allo
ws flex
i
b
ility in
m
o
d
e
lin
g
real wo
rl
d
com
p
lex
relatio
n
s
h
i
p
s
an
d
ab
le to
esti
m
a
te
th
e p
o
s
teri
o
r
p
r
o
b
ab
ility,
wh
ich
p
r
ov
id
es
th
e b
a
sis for
estab
lish
i
ng
classificati
on rul
e
and
pe
rf
o
r
m
i
n
g
st
a
tistical analysis [15],[16]
.
4.
PROP
OSE
D
HYBRID
MO
DEL
The
pr
o
pose
d
hy
b
r
i
d
m
odel
fo
r t
h
e
p
r
edi
c
t
i
on
of
C
V
D c
o
m
p
ri
ses o
f
t
w
o
p
h
ases:
(
1
) a
ssoci
at
i
v
e
cl
assi
fi
er and
(2
) pre
d
i
c
t
i
o
n
usi
ng M
L
P
.
The Ass
o
ci
ative classification
perform
s
c
l
assification by using
associ
at
i
on
r
u
l
e
s. Ass
o
ci
at
i
o
n
rul
e
s ar
e use
f
u
l
t
o
i
d
ent
i
f
y
t
h
e associ
at
i
on e
x
i
s
t
am
ong a g
i
ven set
of at
t
r
i
but
es
.
Gene
ral
l
y
, t
h
e
num
ber
o
f
at
t
r
i
but
es
o
n
t
h
e
ri
ght
si
de
o
f
as
s
o
ci
at
i
o
n
r
u
l
e
s
m
a
y
be o
n
e
o
r
m
o
re. I
f
c
o
ns
eque
nt
of a
n
ass
o
ciation
rule contains single attribute such a
ssocia
tion rules are
c
a
lled class ass
o
ciation
rules
.
These
ru
les are u
s
efu
l
to
i
d
en
tify
k
e
y
attrib
u
t
es fo
r t
h
e
p
r
ed
iction
of CVD usi
n
g
MLP. T
h
e
pha
s
e1 is
a classifi
cation
of at
t
r
i
b
ut
es
w
h
i
c
h a
r
e si
g
n
i
f
i
cant
f
o
r
pre
d
i
c
t
i
on o
f
C
V
D.
The
phase i
s
a
col
l
ect
i
on
of
f
i
ve m
e
t
hod
s n
a
m
e
l
y
:
PCR ( ), NCR
1
(
),
NCR2( ),
NCR3( ) a
n
d
CNOP
NAR(
). The m
e
thod
P
CR( ) ide
n
tifie
s positive ass
o
ciation
a
m
o
n
g
th
e attrib
u
t
es , t
h
e m
e
t
h
od
s NCR
1
( ), NCR2(
),
NC
R3
( )
i
d
en
tifies
asso
ciation
i
n
th
e
ev
en
t of
absen
ce
of
at
t
r
i
b
ut
es.
F
i
nal
l
y
C
NOP
N
A
R
(
)
m
e
t
hod
p
r
o
d
u
ces
use
f
u
l
attribu
t
es
for th
e pred
ictio
n
o
f
CVD. The ru
les
gene
rat
e
d
by
t
h
e al
go
ri
t
h
m
s
in t
h
e associ
at
i
v
e cl
assi
fi
cat
i
on fo
cus
on t
h
e
m
a
jor c
ont
ri
b
u
t
i
ng at
t
r
i
but
es
. From
th
is, we can
i
d
en
tify th
e attribu
t
es with
m
o
re con
t
ri
bu
tion
toward
s t
h
e
d
i
sease id
en
tificatio
n
and
att
r
ibu
t
e
associ
at
i
on
fo
r
t
h
e devel
o
pm
ent
of di
sease
i
n
hum
an bo
dy
. Out
o
f
al
l
t
h
e generat
e
d rul
e
s t
h
e i
m
po
rt
ant
attributes for prediction of dis
ease
are selected.
Furth
e
r th
e
selected
attrib
u
t
es fro
m
th
e p
r
ev
iou
s
step
will b
e
co
nsid
ered as inp
u
t
to the m
u
lt
ilayer
p
e
rcep
tro
n
s (MLP). MLP co
n
s
ists
o
f
3
layers in
th
e m
o
de
l. Th
e fi
rst layer is an
inp
u
t
layer, secon
d
i
s
h
i
dd
en
layer an
d
th
e th
ird
layer is an
o
u
t
pu
t layer wh
ich
is u
s
ed
for p
r
ed
icting
th
e p
r
esen
ce
o
r
ab
sen
ce of th
e CVD.
Th
e trai
n
i
ng
and
testing
o
f
M
L
P is carried ou
t using
PASW18
(Pred
i
ctive A
n
alytics Software
).
4.
1.
A
l
go
r
i
t
h
m fo
r A
s
so
c
i
at
io
n amon
g the Attr
ibutes
In
t
h
is section, we presen
t alg
o
rith
m
fo
r an
Ass
o
ciative Classifier
ca
lled
CPNAR
(Cl
a
ssificatio
n
b
a
sed
on
Po
sitiv
e and
Neg
a
ti
ve Asso
ciatio
n
Ru
les) to
d
ecide th
e classes to wh
ich
th
e
n
e
w o
b
j
ects
b
e
lon
g
s to
.
D
a
t
a
b
a
s
e
(
D
B)
,
mi
n
i
mu
m s
u
p
p
o
r
t
(
m
s
)
a
n
d
mi
n
i
mu
m
confide
n
ce
(mc) are
the i
n
pu
ts to
t
h
e algo
rith
m
CPN
A
R
. I
t
consists o
f
5 pro
c
ed
ur
es
n
a
m
e
ly
PCR (
)
,
NCR1 (
),
NCR2 ( ),
NCR3
( )
a
n
d
CNOP
N
A
R
(
)
.
PCR (
)
g
e
n
e
rates Po
sitiv
e Cl
ass Asso
ciation
R
u
les of t
h
e
form
X
c,
NCR1
( ) g
e
n
e
rates
Neg
a
tiv
e Class
Asso
ciatio
n
Ru
les of
th
e
fo
rm
￢
X
c,
NCR2
( ) g
e
n
e
rates
Neg
a
tiv
e Class
Asso
ciatio
n
Ru
les of
th
e
fo
rm
￢
XY
c,
NCR3( )ge
n
e
r
ates
Negative Class
Associat
ion Rules
of the form
￢
X
￢
Y
c
,
a
n
d
C
N
O
P
N
A
R
(
)
i
s
t
h
e
actual Ass
o
ciat
ive Classifier.
Train
i
ng
Dat
a
Se
t of Cla
ss Assoc
i
ation Rule
s
Cla
ssifie
r
Cla
ssifica
tion
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
18
00
–
1
810
1
804
Apriori-b
a
sed
i
m
p
l
e
m
en
tatio
n
is
u
s
ed
in th
is
work
as
Ap
ri
o
r
i is sim
p
le and
efficien
t
to
g
e
n
e
rate
associ
at
i
o
n
r
u
l
e
s. T
o
det
e
rm
ine t
h
e
val
i
di
t
y
of
AR
s,
t
h
e
su
pp
o
r
t
an
d c
o
nfi
d
ence
m
easure
s
ha
ve
bee
n
us
ed.
Alg
o
rithm
:
C
P
NA
R (
)
{
Call Pro
c
edu
r
e PCR ( )
Call P
r
oce
d
ure
NCR
1
(
)
Call P
r
oce
d
ure
NCR
2
(
)
Call P
r
oce
d
ure
NCR
3
(
)
Call Pro
c
edu
r
e CNOPNAR (
)
}
Procedure P
C
R ( )
{
Pcr =
Φ
Fin
d
L1
-
Fr
equen
t
1-
item
s
ets
L = L
∪
L1
f
o
r (
K = 2;
L
K-
1
≠
Ø; K++
)
{
/* Ge
ne
rating PCK
*/
for each l1,l2
∈
LK-
1
{
if(l1[1]=l2[
1]^……
…………l
1[k-2]=l2[k-
2]^l1[k-1]<l2[k-1])
PC
K
= PCK
∪
{{l1
[
1
]…….
l1
[k
-2],
l1
[
k
-1
],
l2
[k
-1]}
}
/* Prun
ing
u
s
ing
Apriori
prop
erty*
/
fo
r eac
h
(K
-1
)-
su
bsets s
o
f
I
∈
PC
K
{
if s
∉
L
K-
1
PC
K =
PCK
– {
I
}
}
/*
Pru
n
i
n
g
usin
g Suppo
r
t
C
o
un
t*
/
Sca
n
the
dat
a
base a
n
d fi
nd
su
pp
(I
)
fo
r all
I
∈
PCK
for each I
∈
PCK
{
if s
upp
(I
)
≥
m
s
L
K
=
LK
∪
{I
}
}
L
=L
∪
LK
}
}
/*
Gen
e
rating Po
sitiv
e Class
Asso
ciatio
n
R
u
les o
f
th
e
fo
rm
I(=XY)
c*
/
for each I(=XY)
∈
L
{
for eac
h
c
∈
C
{
i
f
c
o
nf
(I
c)
≥
mc
Pcr
=
Pcr
∪
{I
c}
}
}
}
In
itially th
e set PCR (Po
s
itiv
e Classificatio
n
R
u
le
s) is
e
m
p
t
y. In
itially it fin
d
s
L1-freq
u
e
n
t
1
-
ite
m
s
ets. In
t
h
e abov
e algorith
m
Lin
e
4
-
25 g
e
n
e
rates
all
p
o
s
itiv
e frequen
t
item
s
ets. Lin
e
6
-
10
g
e
nerates
p
o
s
itiv
e can
d
i
date ite
msets (PCK). Th
e g
e
n
e
rated
cand
i
da
te ite
m
s
e
t
s are p
r
u
n
e
d
u
s
ing
Aprio
r
i Prin
cip
l
e
(Lin
e
1
2
-1
6) an
d su
pp
ort co
un
t (Lin
e
1
8
-2
3).
Line 29
-34
g
e
n
e
rates po
sitiv
e class asso
ciatio
n
ru
les
b
y
co
nsid
eri
ng
positive
fre
que
n
t ite
m
s
et (I) a
nd a
class labe
l (c).
If t
h
e
c
o
nfi
d
ence
of the aforem
entio
ned rule is m
o
re than
min
i
m
u
m
co
n
f
id
en
ce
(m
c) th
en it is co
nsid
ered
as a
v
a
lid
Po
sitiv
e Cl
ass Asso
ciation
Ru
le and
will b
e
in
clu
d
e
d
i
n
PC
R o
t
h
e
rwise it will b
e
d
i
scarded
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Hybri
d
A
p
pr
oa
ch f
o
r Pre
d
i
c
t
i
o
n
of
C
a
r
d
i
o
v
a
scul
ar
Di
se
ase
Usi
n
g C
l
ass
A
ssoci
at
i
o
n
Rul
e
s ...
. (
K
.
Sri
n
i
v
as)
1
805
Procedure NCR1
(
)
{
Ncr
1
=
Φ
/* Neg
a
tiv
e Class
Asso
ciatio
n
R
u
le set 1*
/
NL =
Φ
/* Negativ
e Frequ
e
n
t
Item
se
t*
/
fo
r eac
h
I
∈
L
{
if
1-
supp(
I)
≥
ms
N
L
=NL
∪
{
I }
}
/*
Gen
e
ratin
g
Neg
a
tiv
e Class Asso
ciation
Ru
les of t
h
e fro
m
￢
I(=
XY
)
c*
/
fo
r eac
h
I
∈
NL
{
for each
c
∈
C
{
i
f
c
o
nf
(
￢
I
c )
≥
mc
Nc
r1 = Nc
r
1
∪
{
￢
I
c}
}
}
}
In
itially Ncr1
an
d
NL set to
Φ
. It
ge
nerat
e
s
negat
i
v
e cl
ass
associ
at
i
on
ru
l
e
s of t
h
e
fo
rm
￢
I (=
X
Y
)
c. First it gen
e
rates
n
e
g
a
ti
v
e
frequ
e
n
t
item
s
e
t
s fro
m
p
o
s
itiv
e frequ
en
t
ite
m
s
e
t
s g
e
n
e
rated
in
th
e
p
r
ev
iou
s
p
r
o
c
ed
ure PC
R b
y
find
ing
1
-
sup
p
(I).
If i
t
is
m
o
re th
an
m
i
n
i
m
u
m
su
pp
ort (m
s) th
en it is in
clu
d
e
d in
NL.
Oth
e
rwise, it is n
o
t
in
clud
ed
i
n
NL.
It is sh
own
in
lin
e 3
-
7. Fo
r each
I in
NL and
a class lab
e
l c i
t
g
e
n
e
rates a
rule
￢
I
c.
If the confi
d
ence is
m
o
re than
m
c
th
en
it is in
clu
d
e
d
in
Ncr1. Oth
e
rwise it is d
i
scard
e
d. It is
sh
own
in lin
e
9 -1
5.
Procedure NCR2
( )
{
Ncr
2
=
Φ
/
*
Neg
a
tiv
e Class
Asso
ciatio
n
R
u
le set 2*
/
NNL
=
Φ
/
*
Neg
a
tiv
e
Negativ
e Freq
u
e
n
t
item
s
e
t
*
/
NNC
2
= Set of n
e
g
a
tiv
e cand
itate ite
m
s
e
t
s
o
f
th
e
form
￢
{i1 }
￢
{i2 } where |i1 , i
2
∈
L1 and i1
≠
i2
fo
r (K
=
2;
N
N
C
K
≠
Ø;
K +
+
)
{
f
o
r
all I =
￢
X
￢
Y
∈
NN
CK
{
if s
u
pp(I
)
≥
ms
N
N
L
K
= N
N
LK
∪
{ I
}
else
{
for all i
∉
XY
{
/
*
Gene
rating Ca
ndi
dates
*/
Cand ={
￢
( X
∪
{ i
})
￢
Y,
￢
X (
￢
Y
∪
{ i }
)}
/
*
Pruni
n
g Cand*/
f
o
r eac
h item
in Cand
{
if
(
X
{i}
∉
L
or
￢
X1
￢
Y1
∈
N
N
L
whe
r
e X1
⊆
X{i} a
n
d
Y
1
⊆
Y)
Ca
nd= Cand – {
X
Y {i}}
N
N
CK+
1
=
NNC
K+1
∪
C
and
}
}
}
}
/*
Gen
e
ratin
g
Neg
a
tiv
e class asso
ciation
R
u
les
o
f
t
h
e
from
I(=
￢
X
￢
Y)
c*
/
fo
r eac
h
I
∈
NN
L
{
f
o
r
ea
ch c
∈
C
{
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
18
00
–
1
810
1
806
i
f
conf
(I
c)
≥
mc
N
c
r2 =
Ncr
2
∪
{
I
c }
}
}
}
In
itially NCR2 (Neg
ativ
e Class Asso
ciation Ru
le
s
o
f
secon
d
typ
e
) and
NNL (Neg
ative Neg
a
tiv
e
Freq
u
e
n
t
Item
set) set to
Φ
.
Li
ne
3
gene
rat
e
s
Negat
i
v
e
Ne
ga
t
i
v
e can
di
dat
e
2
i
t
e
m
s
et
. It
pr
o
duce
s
ca
ndi
dat
e
2
-
ite
m
s
et by taki
ng t
w
o
positive fre
que
nt 1- ite
m
s
ets from
L1
and the
n
a
pplied negation t
o
each item
.
Li
ne 4-25
g
e
n
e
rates all
v
a
lid
Neg
a
tiv
e Neg
a
tiv
e Freq
u
e
n
t
item
s
et
s.
Li
ne 8-
9 ge
nerat
e
s ne
gat
i
v
e negat
i
v
e
f
r
e
que
nt
i
t
e
m
s
et
s.
Li
ne
1
5
ge
nerat
e
s negat
i
v
e
negat
i
ve
ca
ndi
dat
e
i
t
e
m
s
et
s for t
h
e ne
xt
l
e
vel
.
I
t
gen
e
rat
e
s
ne
gat
i
v
e
n
e
g
a
tiv
e cand
i
d
a
te ite
m
s
ets b
y
add
i
ng
a
po
sitiv
e freq
u
e
nt 1
-
item
,
i.e., b
y
add
i
ng
freq
u
e
n
t
item
s
et
i to
an
in
frequ
en
t item
s
e
t
￢
X
￢
Y.
W
e
will o
b
tain
two
n
e
g
a
t
i
v
e
cand
i
d
a
te ite
m
s
ets
￢
X{
i}
￢
Y,
￢
X
￢
Y{i}.
Li
ne 1
7
-
2
1 pe
rf
orm
s
pru
n
i
n
g o
n
t
h
e ge
ne
rat
e
d can
di
dat
e
i
t
e
m
s
et
s. Li
ne 2
7
-
3
4 ge
ne
rat
e
s Neg
a
t
i
v
e
C
l
ass
Ass
o
ci
at
i
on R
u
l
e
s fo
r
Ne
gat
i
v
e Ne
gat
i
v
e F
r
e
que
nt
i
t
e
m
s
et
s
obt
ai
ne
d i
n
t
h
e
p
r
evi
ous
st
eps
.
Procedure NCR3
( )
{
Nc
r3
=
Φ
/
*
Neg
a
tiv
e Class
Asso
ciatio
n
R
u
le Set3
*
/
NPL=
Φ
/
*
Neg
a
tiv
e an
d Po
si
tiv
e Frequ
e
n
t
Ite
m
set*
/
NPC
1
,1
= Set
of
ne
gat
i
v
e i
t
e
m
s
et
s of t
h
e f
o
rm
￢
{i1 }{i
2
} where
i1
, i2
∈
L1 a
n
d i
1
≠
i2
/*
Neg
a
tiv
e
and
Po
sitiv
e Cand
id
ate item
s
et*
/
f
o
r (K
= 1; N
C
K,1
≠
Ø;
K +
+)
{
f
o
r (P = 1;
N
C
K,P
≠
Ø;
P
+ +)
{
f
o
r
all I
∈
NC
K,P
{
if s
u
pp(I
)
≥
ms
N
P
LK
,P =
NPL
K
,P
∪
{ I
}
}
/*Generating Candidates*/
f
o
r all
I
1
,
I
2
∈
NPLK,P
/*
I1
an
d I2
are
j
o
i
n
ab
le item
s
e
t
s*
/
{
X = I1
.neg
ativ
e,
Y = I1 .po
s
itiv
e
∪
I2 .po
s
itiv
e
I =
￢
XY
if
(
(
∄
X1
⊂
X
)
(
s
upp
(
￢
X1
Y
)
≥
ms
)
a
n
d
(
∄
Y1
⊂
Y
)
(
s
up
p(
￢
XY
1 )
<
m
s
))
NP
CK,P+
1
=
NP
CK,P+
1
∪
{ I }
}
}
f
o
r
all X
∈
LK
+1
, i
∈
L1
{
if
(
∄
X1
⊂
X)
(
￢
X1 {i}
∈
NPL
)
NPCK+
1
,
1
=
NPCK+
1
,
1
∪
￢
X{i}
}
}
/*
Gen
e
rating
Neg
a
tiv
e Class Asso
ciation
Ru
les of t
h
e fro
m
￢
XY
c *
/
f
o
r eac
h
I
∈
NPL
{
for each
c
∈
C
{
if
conf (
I
c )
≥
m
c
Ncr
3
=
Nc
r3
∪
{I
c}
}
}
}
/*
I1 and
I2 are jo
in
ab
le if
I1
≠
I2
,
I1 .n
eg
ativ
e = I2
.n
eg
ativ
e, I1
.po
s
itive and
I2
.po
s
itiv
e sh
are t
h
e
sam
e
k
−
1
item
s
, an
d
I1
.po
s
itiv
e
∪
I
2
.posi
tive
∈
L(
P1
)
p
+1
*
/
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Hybri
d
A
p
pr
oa
ch f
o
r Pre
d
i
c
t
i
o
n
of
C
a
r
d
i
o
v
a
scul
ar
Di
se
ase
Usi
n
g C
l
ass
A
ssoci
at
i
o
n
Rul
e
s ...
. (
K
.
Sri
n
i
v
as)
1
807
The associ
at
i
o
n rul
e
s
obt
ai
ne
d by
pr
oce
d
u
r
e
s
PC
R
(
), NC
R
1
( ),
NC
R
2
( ) and
NC
R
3
( )
are or
dere
d
base
d o
n
t
h
e s
u
p
p
o
rt
an
d c
o
n
f
i
d
e
n
ce val
u
es. It represents a
n
actual classifi
er. The
proces
s of classification i
s
lo
ok
ed in
to fo
r th
e set
o
f
ru
les fo
r t
h
ose classes th
at
are
rel
e
vant
t
o
t
h
e
ob
j
ect
prese
n
te
d for classification.
Assign
m
e
n
t
o
f
class lab
e
l to
th
e n
e
w obj
ect will b
e
ex
p
l
ain
e
d
b
y
th
e fo
llo
wi
n
g
pro
c
ed
ure called
CNOPNAR
(C
lassificatio
n
of
New Ob
j
ect based
o
n
Po
s
itive an
d
Neg
a
tiv
e Asso
ciatio
n
R
u
les).
Th
e i
n
puts to
the CNOPNAR are Associat
ive Classifi
er, Confide
n
ce Margi
n
(CM) and a ne
w
object
O to be classified. It
produces
a cat
egory attache
d
to the
ne
w
obje
ct.
Procedure
CNOPNAR ( )
{
S =
Φ
/
*
set
of r
u
les t
h
at m
a
tch
o*/
for each rule
R in t
h
e s
o
rte
d
set of
rules
{
if
( R
⊂
O)
/* O is an ob
j
ect
wh
ich is to
b
e
Classifi
ed
*
/
{c
ount++
}
if
(count
= = 1)
fr.c
o
nf
= R
.
conf
/*
keep the
fir
s
t r
u
l
e
confi
d
ence
*/
S =
S
∪
R
else if
( R.c
o
nf >
f
r
.c
onf
- CM)
S= S
∪
R
}
S=
S1
∪
S2
∪
S
3
………
……… Sn
f
o
r
each
sub
s
et S
1
,S
2,
. .
.
.
S
n
{
}
O = c
j
,
w
h
ere
c
j
=
m
a
x { Confi
d
ence
Score1, .
...
...
Confide
n
ceSc
o
r
e
n }
}
In
th
e abo
v
e
alg
o
rith
m
CN
OPNAR, t
h
e lin
es 2-11
sele
cts a set of appropriate rul
e
s within a
Confide
n
ce Margi
n
(CM). T
h
e selected
ru
les are in
th
e interv
al [ R.c
onfidence – CM , fr.c
o
nfide
n
ce]
. The
p
r
ed
ictio
n
o
f
ru
les starts at li
n
e
11
.
In
lin
e
12
, th
e set
o
f
sel
ected
ru
les is
div
i
d
e
d
b
a
sed
on
th
e classes.
In
lin
e
12
-
1
4
,
t
h
e
gr
o
ups
have
been
arra
nge
d
base
d
o
n
t
h
e a
v
er
a
g
e confi
d
enc
e
per class. In line 17 t
h
e classification
i
s
m
a
de by
assi
gni
ng
a cl
ass t
h
at
ha
s t
h
e
hi
g
h
est
C
o
n
f
i
d
e
n
c
e
Sco
r
e t
o
t
h
e
new
o
b
j
ect
.
5.
EX
PER
I
M
E
NTS AN
D R
E
SU
LTS
To
illu
strate t
h
e in
trod
u
c
ed ap
pro
ach
we u
s
e th
e
d
a
taset fro
m
th
e UCI Repo
sito
ry
o
f
Mach
in
e
Learni
ng
Dat
a
bases [
17]
. C
l
evel
and
heart
fou
n
d
at
i
o
n
dataset with
1
4
attrib
u
t
es an
d
3
0
3
d
a
ta
ite
m
s
,
Hung
arian
d
a
taset with
13
at
trib
u
t
es
and
294
d
a
ta item
s
, Switzerland
d
a
tase
t with
1
3
attrib
u
t
es and
12
3 d
a
ta
ite
m
s
u
s
ed
i
n
ou
r exp
e
rim
e
n
t
s. Th
is
d
a
ta consists o
f
m
u
ltiv
ariate attrib
u
t
es ou
t
o
f
wh
ich
th
e last 1 attribu
t
e is
a d
e
p
e
nd
en
t attrib
u
t
e and
th
e rem
a
in
in
g
attrib
u
t
es are
ind
e
p
e
nd
en
t attributes. Th
e d
e
p
e
nd
en
t attri
b
u
t
e sh
ows
whet
her t
h
e C
VD i
s
prese
n
t
or a
b
se
nt
i
n
t
h
e pat
i
e
nt
dat
a
.
The
depe
n
d
ent
at
t
r
i
but
e i
s
t
r
a
n
sf
orm
e
d i
n
t
o
bi
na
ry
data suc
h
as presence of dise
ase with a val
u
e 1
or a
b
se
nc
e of
disease wi
th a value
0. C
l
ass distributions are
54% heart dise
ase
abse
nt
a
n
d
46% heart dise
ase
prese
n
t.
We de
vel
o
ped
code i
n
M
A
TLAB
an
d ex
peri
m
e
nt
s are conducted.
We co
m
p
ared our propose
d
m
e
t
hod
wi
t
h
t
h
e
po
p
u
l
a
r al
g
o
ri
t
h
m
s
gen
f
i
s
2 [
1
8]
, F
u
zzy
+ deci
si
o
n
t
r
ee cl
assi
fi
er
[
19]
, cl
i
n
i
cal
d
eci
si
on
su
ppo
r
t
system [
2
0
]
, a h
y
br
id
pr
ac
tical swarm
o
p
timiz
atio
n
(PSO)
base
d fuzzy expert system
[21] and
Pre
d
i
c
t
i
on o
f
C
a
rdi
o
vasc
ul
ar
Di
sease-
A H
y
bri
d
Ap
pr
oac
h
[
22]
-
[
2
4
]
o
n
t
h
e real
dat
a
s
e
t
s
cal
l
e
d C
l
evel
an
d
D
a
taset,
H
ungar
i
an
d
a
taset an
d Sw
itzer
land d
a
taset. Th
e r
e
su
lts ar
e su
mmar
i
zed
as:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
18
00
–
1
810
1
808
Table 2.
Resul
t
s
sum
m
ary
Cleveland
Hungarian
S
w
itz
e
rland
Genfis2
Sensitivity
20.
2
66.
0
76.
5
Specifici
ty
68.
6
22.
2
44.
7
Accuracy
39.
2
39.
5
62.
3
Fu
zz
y + d
ecision t
ree classifier
Sensitivity
50.
549
55.
974
59.
375
Specifici
ty
62.
386
62.
386
55.
737
Accuracy
51.
793
36.
0
62.
5
Clinical
decision support syste
m
Sensitivity
52.
066
53.
766
61.
125
Specifici
ty
44.
875
47.
350
74.
256
Accuracy
46.
483
42.
417
58.
183
A hybrid par
t
icle
s
w
arm
opti
m
i
z
a
tion base
d fu
zz
y
expert
syste
m
Sensitivity
62.
00
60
95.
0
Specifici
ty
61.
09
60.
369
56.
37
accuracy
53.
93
45.
09
63.
45
proposed techni
q
u
e
Sensitivity
66.
90
72.
81
95.
93
Specifici
ty
63.
87
64.
69
58.
37
Accuracy
84.
967
71.
88
77.
59
The c
o
m
p
arison
of
res
u
lts over Cleveland
Dataset,
Hu
ng
ari
a
n
d
a
taset
and
Switzerland
d
a
taset
u
s
ing
genfis2, Fuzz
y + decision tree classifier, clinical
dec
i
si
on su
p
p
o
r
t
sy
st
em
, a hy
bri
d
p
r
act
i
cal
swarm
opt
i
m
i
zati
on b
a
sed f
u
zzy
e
x
pert
sy
st
em
and p
r
op
ose
d
t
e
chni
que
are s
h
ow
bel
o
w as
bar c
h
a
r
t
s
. F
r
o
m
t
h
e
gra
p
hs i
t
i
s
e
v
i
d
ent
t
h
at
t
h
e
p
r
op
ose
d
m
e
t
hod
pe
rf
orm
s
very
wel
l
.
Fig
u
re
2
.
Co
mp
ariso
n
of
Sensitiv
ity
Fig
u
re
3
.
Co
mp
ariso
n
of
Sp
ecificity
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Hybri
d
A
p
pr
oa
ch f
o
r Pre
di
ct
i
o
n
of
C
a
r
d
i
o
v
a
scul
ar
Di
se
ase
Usi
n
g C
l
ass
A
ssoci
at
i
o
n
Rul
e
s ...
. (
K
.
Sri
n
i
v
as)
1
809
Fi
gu
re
4.
C
o
m
p
ari
s
on
o
f
Acc
u
racy
6.
CO
NCL
USI
O
N
The accuracy
of
propose
d
model is
close t
o
the res
u
lts obtained
by
MLP with total attributes.
W
e
have int
r
oduce
d a new classifier using, ass
o
ciative cla
ssi
fier an
d M
L
P by
bl
endi
ng t
h
e o
u
t
p
ut
s of ass
o
c
i
at
i
v
e
classifier as the inputs of M
L
P.
T
h
e assoc
i
ative classifier does not
provide a m
a
thematical
m
odel for t
h
e
cl
assi
fi
cat
i
on
of f
u
t
u
re o
b
j
e
c
t
s
, whe
r
eas t
h
e p
r
o
p
o
se
d m
e
t
hod
pr
ovi
des a m
a
t
h
em
at
i
cal
m
odel
for t
h
e
classificatio
n
with
li
m
ited
n
u
m
b
e
r o
f
variab
les. Two
im
p
o
r
tan
t
step
s u
tilized
in
ru
le gen
e
ration
pro
c
ess are
sel
ect
i
ng t
h
e
im
port
a
nt
at
t
r
i
but
es
usi
n
g a
ssoci
at
i
on
rul
e
s and cl
assi
f
i
cat
i
on usi
n
g
M
L
P. Fi
nal
l
y
,
t
h
e
expe
ri
m
e
nt
at
i
o
n i
s
carri
ed o
u
t
usi
ng t
h
e C
l
ev
el
and,
Hu
n
g
ari
a
n an
d Swi
t
zer
l
a
nd dat
a
set
s
a
nd t
h
e pe
rf
o
r
m
a
nce
was analyzed
with sensitivity, specific
ity and acc
uracy. F
r
om
the above
study
it is observe
d that decision tree
co
rrectly classifies 77
%, M
L
P with all 13
attribu
t
es c
o
rrectly classifies 82% a
n
d the
propose
d
method
correctly classifies 85%
wi
t
h
l
i
m
i
t
e
d
n
u
m
b
er of
at
t
r
i
but
es.
The
p
r
o
p
o
se
d m
e
t
hod o
u
t
p
e
r
fo
rm
s
t
h
an dec
i
si
on
t
r
ee
an
d ot
he
r exi
s
t
i
n
g
m
e
t
hods.
REFERE
NC
ES
[1]
C. Becqu
e
t,
et a
l
.
, “Strong associationru
l
e minin
g
for large scale
gene expr
ession data an
aly
s
is:
a case stud
y
on
human SAGE data,”
Genom Bio
l
,
vol/issue: 3(12)
, 2002.
[2]
“Cardiovascular
diseases in
I
ndia Challenges
and
way
ahead,”
In
ternational Heart
Protection Sum
mit
, 2011
.
[3]
B. Shah and P.
Mathur, “Surveillan
ce of c
a
rdio
vas
c
ular d
i
s
eas
e
ris
k
factors
in I
ndia: Th
e ne
ed
& s
c
ope,
”
India
n
J
Med Res
, vo
l. 13
2, pp
. 634-642
,
2010.
[4]
F. L. Duff,
et a
l
.
, “
P
redi
cting
Survival C
a
uses A
f
ter Out
of Hosp
ital Cardiac Arrest usi
ng Data
Mining Method
,”
MEDINFO, Fieschi
et al. (
E
ds
)Amsterdam, IOS Press, 2004.
[5]
A. Freitas
,
“Understanding
the
cr
ucial diff
erences between
cl
as
s
i
fi
cat
ion and
as
s
o
cia
tion
rules –
a
position pap
e
r,”
SIGKDD
Explorations,
vo
l/issue: 2(1), pp. 65–69, 2000.
[6]
S
.
Özekes
and A. Y. Çam
u
rcu, “
C
las
s
i
fication
and prediction in a data Mining
application
,
”
Jo
urnal of Marma
r
a
for Pure and
Ap
plied
Sciences
, v
o
l. 18
, pp
. 159-1
74, 2002
.
[7]
S. Palaniapp
a
n and R. Awang, “
I
ntellig
ent Heart
Disease Prediction S
y
stem
,
”
Computer Systems and Applications, AICCSA 20
08. IEEE/ACS In
ter
national Conference,
2008
.
[8]
C. Ordonez and
K. Zh
ao, “Ev
a
l
u
ating
a
ssociati
on rules and
decision tr
ees
to
p
r
edict m
u
ltip
le
t
a
rget
at
tribut
es,”
Intell
igen
t Data
Analysis
, vol. 15
, pp
. 173–192
, 2
011.
[9]
M. Khali
lia
,
et al
., “Predicting
disease r
i
sks from highly
imbalanced
data usin
g random forest,”
BMC M
e
dic
a
l
Informatics and
Decision Ma
kin
g
,
vol/issue: 11(
51), 2011
. doi:1
0.1186/1472-69
47-11-51.
[10]
E. Anupriy
a,
et
al.
, “Enhan
c
ed Prediction of
Heart Dis
ease with
Feature
Subset Sele
ctio
n using Genetic
Algorithm,”
International
Journal of Engi
n
eering
Scien
c
e and Technology,
vol/issue: 2(10)
, pp
. 53
70-5376, 2010
.
[11]
K. Srinivas,
et
al
.
, “Analy
sis of Coronar
y
Heart
Disease and Pred
iction of Heart Attack in Coal
Mining Regions
Using Data Mining Techniqu
es,”
The
5
th
IEEE In
ternational Conf
erence on Comp
uter Scien
ce
&
Education
,
Hef
e
i
,
China.
[12]
S. Hay
k
in
, “Neu
ral Networks: A
Comprehensive
F
oundation,” 2n
d Edn., New
Jersey
, Pren
tice-Hall, 1999
.
[13]
J
.
M
.
Nazza
l,
et al
.
, “
M
ultil
a
y
er
Perceptron Neu
r
al Network (MLPs)For Anal
y
z
ing the Propert
i
e
s of Jordan Oil
S
h
ale,
”
World
Applied
Sciences
Journal
, vo
l/issu
e: 5(5)
, pp
. 546-
552, 2008
.
[14]
B. K. Wong an
d Y. Selvi, “Neural ne
twork ap
plications in f
i
n
a
nce: A revi
ew and anal
y
s
is of liter
a
tur
e
(1990-
1996),”
Information &
Management
, vol. 34
, pp
.
129-139,1998.
[15]
X.
Wu,
et al
., “Top 10 algorith
m
s in data mining,”
Knowl Inf
Syst
, vol. 14, pp. 1–37, 2008. DOI 10.1007/s10115-
007-0114-2.
Evaluation Warning : The document was created with Spire.PDF for Python.