I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
,
pp.
570
~
583
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
1
.
pp
57
0
-
583
570
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
A
c
c
u
r
a
t
e
p
r
e
d
ic
t
io
n
of
c
h
r
on
i
c
d
is
e
ase
s
u
si
n
g
d
e
e
p
l
e
a
r
n
i
n
g
al
gor
ith
m
s
Ronal
d
S
.
Cor
d
ova
1
,
Rolou
L
yn
R.
M
aat
a
1
,
M
ali
k
Jawar
n
e
h
2
,
M
ar
wan
I
.
Als
h
ar
’
e
3
,
Ol
iver
C.
A
gu
s
t
in
4
1
D
e
pa
r
tm
e
nt
of
C
omput
in
g S
c
ie
nc
e
s
, G
ul
f
C
ol
le
ge
, M
u
s
c
a
t
,
O
ma
n
2
C
ol
le
ge
of
C
omput
e
r
S
c
ie
nc
e
a
nd I
nf
or
ma
ti
c
s
, A
mm
a
n A
r
a
b U
ni
ve
r
s
it
y, A
mm
a
n, J
or
da
n
3
F
a
c
ul
ty
of
C
omput
in
g a
nd
I
nf
or
ma
ti
on T
e
c
hnol
ogy
, S
oha
r
U
ni
ve
r
s
it
y, S
oha
r
,
O
ma
n
4
V
e
r
a
E
qui
nox T
e
c
hnol
ogi
e
s
I
nc
or
por
a
te
d,
C
a
ba
na
tu
a
n
C
it
y
,
P
hi
li
ppi
ne
s
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
F
e
b
11
,
2024
R
e
vis
e
d
J
ul
4
,
2024
Ac
c
e
pted
J
ul
26
,
2024
In
t
h
i
s
p
a
p
er,
t
h
e
res
earc
h
ers
s
t
u
d
i
ed
t
h
e
effect
s
o
f
d
i
fferen
t
act
i
v
at
i
o
n
fu
n
c
t
i
o
n
s
i
n
h
i
d
d
en
l
ay
er
s
a
n
d
h
o
w
t
h
ey
i
mp
ac
t
t
h
e
o
v
erf
i
t
t
i
n
g
o
r
u
n
d
erf
i
t
t
i
n
g
o
f
t
h
e
m
o
d
e
l
i
n
t
h
e
m
u
l
t
i
c
l
as
s
p
re
d
i
c
t
i
o
n
o
f
c
h
ro
n
i
c
d
i
s
e
as
es
.
T
h
i
s
p
ap
er
a
l
s
o
ev
al
u
at
e
d
t
h
e
effect
s
o
f
v
ar
y
i
n
g
t
h
e
n
u
m
b
er
o
f
l
a
y
ers
,
an
d
h
y
p
er
p
aramet
er
s
an
d
i
t
s
i
mp
ac
t
o
n
t
h
e
accu
rac
y
o
f
t
h
e
mo
d
e
l
an
d
i
t
s
g
en
er
al
i
za
t
i
o
n
cap
ab
i
l
i
t
i
es
.
It
w
as
fo
u
n
d
t
h
a
t
ex
p
o
n
e
n
t
i
a
l
l
i
n
ear
u
n
i
t
(E
L
U
)
d
o
e
s
n
o
t
h
a
v
e
a
s
i
g
n
i
f
i
can
t
ad
v
a
n
t
a
g
e
o
v
er
rec
t
i
f
i
ed
l
i
n
ear
u
n
i
t
(
ReL
U
)
w
h
e
n
u
s
ed
as
an
ac
t
i
v
at
i
o
n
fu
n
c
t
i
o
n
i
n
t
h
e
h
i
d
d
en
l
a
y
er.
A
d
d
i
t
i
o
n
al
l
y
,
t
h
e
p
erf
o
rman
ce
o
f
s
o
f
t
ma
x
fu
n
c
t
i
o
n
,
w
h
en
u
s
e
d
i
n
t
h
e
o
u
t
p
u
t
l
ay
er,
i
s
t
h
e
s
ame
a
s
a
c
l
as
s
i
c
s
i
g
m
o
i
d
o
u
t
p
u
t
act
i
v
at
i
o
n
fu
n
ct
i
o
n
.
In
t
erms
o
f
t
h
e
ab
i
l
i
t
y
o
f
t
h
e
mo
d
el
t
o
g
e
n
eral
i
ze,
t
h
e
res
earch
er
s
ach
i
ev
e
d
a
cl
as
s
i
f
i
cat
i
o
n
accu
ra
cy
o
f
1
0
0
%
w
h
e
n
t
h
e
t
ra
i
n
e
d
mo
d
e
l
w
as
u
s
ed
t
o
p
red
i
ct
u
n
s
ee
n
d
at
a.
T
h
ro
u
g
h
t
h
i
s
res
e
arch
,
t
h
e
res
earc
h
er
s
s
h
o
u
l
d
b
e
ab
l
e
t
o
as
s
i
s
t
med
i
ca
l
p
ro
fe
s
s
i
o
n
al
s
an
d
p
rac
t
i
t
i
o
n
er
s
i
n
O
man
i
n
t
h
e
v
al
i
d
a
t
i
o
n
an
d
d
i
ag
n
o
s
i
s
o
f
c
h
r
o
n
i
c
d
i
s
eas
e
s
i
n
cl
i
n
i
c
s
an
d
h
o
s
p
i
t
a
l
s.
K
e
y
w
o
r
d
s
:
C
hr
onic
dis
e
a
s
e
s
Da
ta
-
dr
iven
he
a
lt
hc
a
r
e
De
e
p
lea
r
ning
He
a
lt
hc
a
r
e
a
na
lyt
ics
P
r
e
dictive
a
na
lyt
ics
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
R
ona
ld
S
.
C
or
dova
De
pa
r
tm
e
nt
of
C
omput
ing
S
c
ienc
e
s
,
Gulf
C
oll
e
ge
M
us
c
a
t
-
133
,
Oma
n
E
mail:
r
ona
ld@gul
f
c
oll
e
ge
.
e
du.
om
1.
I
NT
RODU
C
T
I
ON
I
n
Oma
n,
ther
e
a
r
e
100
to
120
r
e
por
ts
o
f
medic
a
l
e
r
r
o
r
s
–
c
las
s
if
ied
a
s
ha
r
mf
ul,
li
f
e
-
thr
e
a
tening
a
nd
f
a
tal,
whic
h
wa
s
the
ke
y
dis
c
us
s
ion
dur
ing
the
8t
h
Oma
n
he
a
lt
h
e
xhibi
t
ion
a
nd
c
onf
e
r
e
nc
e
[
1]
.
I
n
2015,
the
Na
ti
ona
l
Ac
a
de
my
of
M
e
dicine
in
the
Uni
ted
S
tate
s
(
US
)
r
e
por
ted
that
20%
o
f
mos
t
pe
ople
will
ge
t
a
n
i
nc
or
r
e
c
t
diagnos
is
a
t
lea
s
t
onc
e
in
their
li
ve
s
[
2
]
,
thes
e
s
tatis
ti
c
s
a
c
c
ount
f
or
ove
r
12
m
il
li
on
a
dult
s
who
s
e
e
k
o
utpatient
medic
a
l
c
a
r
e
tr
a
ns
lating
to
a
bout
5
%
of
a
dult
s
,
o
r
1
out
of
20
a
dult
pa
ti
e
nts
[
3
]
.
I
n
S
ulaimaniya
h
C
it
y,
I
r
a
q,
it
wa
s
f
ound
that
non
-
phys
icia
n
h
e
a
lt
hc
a
r
e
wor
ke
r
s
a
r
e
the
pr
im
a
r
y
c
a
us
e
of
mi
s
diagnos
is
a
nd
mi
s
tr
e
a
tm
e
nt
of
C
OV
I
D
-
19
pa
ti
e
nts
whe
r
e
pa
ti
e
nts
we
r
e
f
a
ls
e
ly
diagnos
e
d
a
s
ha
ving
typhoi
d
(
63
%
)
,
in
f
luenz
a
(
14
%
)
,
pne
umoni
a
(
9%
)
,
g
a
s
tr
oe
nter
it
is
(
5%
)
,
c
omm
on
c
ol
d
(
4
%
)
,
b
r
uc
e
ll
os
is
(
4%
)
,
a
nd
meningit
is
(
1%
)
[
4]
.
I
n
2021
,
it
wa
s
f
ound
that
mi
s
s
e
d
va
s
c
ular
e
ve
nts
,
in
f
e
c
ti
on
s
,
a
nd
c
a
nc
e
r
s
a
c
c
ount
f
or
75%
of
s
e
r
ious
ha
r
ms
c
a
us
e
d
by
mi
s
diagnos
is
[
5]
.
Globa
ll
y,
wor
ld
he
a
lt
h
or
ga
niza
ti
on
(W
HO
)
s
tate
s
that
diagnos
ti
c
e
r
r
or
s
oc
c
ur
in
a
bout
5%
of
a
dult
s
in
outpatient
c
a
r
e
s
e
tt
ings
,
mo
r
e
than
ha
l
f
of
whic
h
ha
ve
the
potential
to
c
a
us
e
s
e
ve
r
e
ha
r
m
[
6]
.
T
he
r
e
s
ult
of
thi
s
s
tudy
will
be
a
ve
r
y
r
e
leva
nt
a
nd
ind
is
pe
ns
a
ble
tool
in
hos
pit
a
ls
a
nd
he
a
lt
hc
a
r
e
pr
ovide
r
s
in
t
he
e
a
r
ly
diagnos
is
of
dis
e
a
s
e
s
whic
h
c
ould
ha
ve
be
e
n
pr
e
ve
nted
if
a
c
c
ur
a
tely
de
tec
ted
e
a
r
ly.
T
he
dis
e
a
s
e
pr
e
diction
s
of
twa
r
e
is
idea
l
f
o
r
us
e
in
r
e
mot
e
a
nd
r
ur
a
l
a
r
e
a
s
whe
r
e
the
a
va
il
a
bil
it
y
o
f
doc
tor
s
is
s
c
a
r
c
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
c
c
ur
ate
pr
e
diction
of
c
hr
onic
dis
e
as
e
s
us
ing
de
e
p
lear
ning
algor
it
hms
…
(
R
onald
S.
C
or
dov
a
)
571
I
n
thi
s
pa
pe
r
,
the
r
e
s
e
a
r
c
he
r
s
de
ve
loped
a
de
e
p
lea
r
ning
model
f
or
the
im
pleme
ntation
of
a
n
a
r
ti
f
icia
l
int
e
ll
igenc
e
(
AI
)
-
a
s
s
is
ted
c
hr
onic
dis
e
a
s
e
pr
e
diction
in
the
tr
a
c
king
s
ys
tem
f
o
r
c
hr
onic
dis
e
a
s
e
s
(
T
S
C
D)
pr
ojec
t.
T
a
ble
1
p
r
e
s
e
nts
the
de
e
p
lea
r
ning
f
r
a
me
wor
ks
tool
kit
s
that
the
r
e
s
e
a
r
c
he
r
s
ha
ve
c
ons
ider
e
d
in
thi
s
pr
ojec
t
f
o
r
e
va
luation.
T
he
objec
ti
ve
o
f
thi
s
p
r
o
pos
e
d
pr
ojec
t
is
to
de
ve
lop
a
c
hr
onic
dis
e
a
s
e
mul
ti
c
las
s
pr
e
diction
model
to
be
us
e
d
in
T
S
C
D.
T
his
pr
e
dicti
on
model
is
c
r
uc
ial
to
T
S
C
D
to
he
lp
doc
to
r
s
in
the
d
iagnos
is
an
d
ve
r
if
ica
ti
on
of
pa
ti
e
nt
il
lnes
s
e
s
in
hos
pit
a
ls
,
h
e
a
lt
h
c
e
nter
s
a
nd
c
li
nics
.
S
pe
c
if
ica
ll
y,
thi
s
r
e
s
e
a
r
c
h
pa
pe
r
a
im
s
to:
‒
Ana
lyze
the
im
pa
c
t
of
dif
f
e
r
e
nt
a
c
ti
va
ti
on
f
unc
ti
ons
on
the
pe
r
f
o
r
manc
e
of
hidden
laye
r
s
in
ne
ur
a
l
ne
twor
k
s
pe
c
if
ica
ll
y,
e
xpone
nti
a
l
li
ne
a
r
un
it
(
E
L
U)
a
nd
r
e
c
t
if
ied
li
ne
a
r
unit
(
R
e
L
U)
.
‒
I
nve
s
ti
ga
te
the
r
e
lative
pe
r
f
or
manc
e
of
the
s
igm
oid
a
nd
s
of
tm
a
x
f
unc
ti
ons
a
s
output
laye
r
a
c
ti
va
ti
on
f
unc
ti
ons
in
ne
ur
a
l
ne
twor
ks
.
‒
C
ompar
e
the
d
if
f
e
r
e
nt
pe
r
f
or
manc
e
a
nd
a
c
c
u
r
a
c
y
of
the
c
las
s
if
ica
ti
on
models
with
va
r
yi
ng
hype
r
pa
r
a
mete
r
s
e
tt
ings
.
T
a
ble
1.
M
a
c
hine/
de
e
p
lea
r
ning
f
r
a
mew
or
k
F
r
a
me
w
or
k
P
r
ogr
a
mm
in
g l
a
ngua
ge
s
J
upyt
e
r
l
a
b
P
la
tf
or
m
O
pe
n
-
s
our
c
e
D
e
e
p l
e
a
r
ni
ng
P
yT
or
c
h
P
hyt
on, C
+
+
, J
ul
ia
,
C
#
Y
e
s
L
in
ux, ma
c
O
S
, W
in
dow
s
,
A
ndr
oi
d
Y
e
s
Y
e
s
T
e
ns
or
F
lo
w
P
hyt
on (
ke
r
a
s
)
, C
/C
+
+
, J
a
va
, G
o,
J
a
va
s
c
r
ip
t,
R
, J
ul
ia
, S
w
if
t,
C
#
Y
e
s
L
in
ux, ma
c
O
S
, W
in
dow
s
,
A
ndr
oi
d
Y
e
s
Y
e
s
A
pa
c
he
M
X
N
E
T
C
+
+
, P
hyt
hon, G
o, R
, S
c
a
la
, P
e
r
l,
C
lo
ju
r
e
, C
#
Y
e
s
L
in
ux, ma
c
O
S
, W
in
dow
s
,
A
W
S
, A
ndr
oi
d, i
O
S
,
J
a
va
S
c
r
ip
t
Y
e
s
Y
e
s
L
ig
ht
G
B
M
P
hyt
hon, R
, C
+
+
, C
,
C
#
Y
e
s
L
in
ux, ma
c
O
S
, W
in
dow
s
Y
e
s
No
Out
of
numer
ous
de
e
p
lea
r
ning
f
r
a
mew
or
ks
e
va
luate
d
[
7]
,
the
r
e
s
e
a
r
c
he
r
s
na
r
r
owe
d
down
the
c
hoice
to
T
e
ns
or
F
low
be
c
a
us
e
of
it
s
matur
it
y,
s
c
a
labili
ty,
a
nd
e
a
s
y
-
to
-
de
ve
lop
model
s
,
wide
ly
us
e
d
a
t
the
pr
oduc
ti
on
leve
l
in
the
indus
tr
y,
c
ompar
e
d
with
the
other
s
.
P
ytor
c
h
[
8
]
is
a
mac
hine
lea
r
n
ing
f
r
a
mew
or
k
ba
s
e
d
on
the
T
or
c
h
li
b
r
a
r
y,
us
e
d
f
or
a
ppli
c
a
ti
ons
s
uc
h
a
s
c
omp
uter
vis
ion
a
nd
na
tu
r
a
l
langua
ge
pr
oc
e
s
s
ing.
P
y
to
r
c
h
wa
s
r
e
lea
s
e
d
in
S
e
ptembe
r
2016.
T
e
ns
or
F
low
[
9
]
is
a
n
e
nd
-
to
-
e
nd
ope
n
-
s
our
c
e
mac
hine
lea
r
ning
platf
or
m.
R
e
lea
s
e
d
on
Nove
mber
9,
2015,
it
f
e
a
tur
e
s
a
c
ompr
e
he
ns
ive,
f
lexible
e
c
os
ys
tem
of
tool
s
,
li
b
r
a
r
ies
,
a
nd
c
o
mm
unit
y
r
e
s
our
c
e
s
that
lets
r
e
s
e
a
r
c
he
r
s
pus
h
the
s
tate
-
of
-
th
e
-
a
r
t
in
mac
hine
lea
r
ning
a
nd
de
ve
loper
s
e
a
s
il
y
buil
d
a
nd
de
ploy
mac
hine
lea
r
ning
-
powe
r
e
d
a
ppli
c
a
ti
ons
.
Apa
c
he
’
s
ope
n
-
s
our
c
e
s
pin
on
a
de
e
p
lea
r
ning
f
r
a
m
e
wor
k
c
a
ll
e
d
M
XN
e
t
whic
h
s
uppor
ts
bui
ldi
ng
a
nd
tr
a
ini
ng
de
e
p
lea
r
ning
ne
twor
ks
in
mul
ti
ple
langua
ge
s
[
10]
,
[
1
1]
wa
s
r
e
lea
s
e
d
on
M
a
y
10,
2022.
W
it
h
e
f
f
icie
nc
y
a
nd
s
c
a
labili
ty
in
mi
nd
,
M
XN
e
t
is
ve
r
y
f
a
s
t,
li
g
htwe
ight
a
nd
por
table
[
12]
,
[
13]
.
L
ight
GB
M
[
14
]
or
li
ght
gr
a
dient
-
boos
ti
ng
mac
hine,
is
a
f
r
e
e
a
nd
ope
n
-
s
our
c
e
dis
tr
ibut
e
d
gr
a
dient
-
boos
ti
ng
f
r
a
mew
or
k
f
or
mac
hine
lea
r
ning
ba
s
e
d
on
de
c
is
ion
tr
e
e
a
lgo
r
it
hms
a
nd
us
e
d
f
or
r
a
nking
,
c
las
s
if
ica
ti
on,
a
nd
othe
r
mac
hine
lea
r
ning
tas
ks
.
L
ight
GB
M
wa
s
r
e
lea
s
e
d
in
2016.
Ac
c
or
ding
to
the
be
nc
hmar
k
r
e
s
ult
s
publi
s
he
d
by
vis
o.
a
i
[
15]
,
T
e
ns
or
F
low
a
nd
P
yT
o
r
c
h
s
how
e
qua
l
a
c
c
ur
a
c
y,
the
memor
y
us
a
ge
of
T
e
ns
or
F
low
is
low
e
r
but
the
tr
a
ini
ng
t
im
e
is
s
ubs
tantially
higher
.
Ac
c
or
ding
to
the
r
e
s
ult
s
,
T
e
ns
or
F
low
is
o
f
ten
us
e
d
f
or
pr
oduc
ti
on
a
ppli
c
a
ti
ons
due
to
it
s
s
pe
e
d
a
nd
s
c
a
labili
ty.
I
t
ha
s
a
n
e
xtens
ive
li
br
a
r
y
of
p
r
e
-
buil
t
models
a
nd
whic
h
m
a
ke
s
it
idea
l
f
or
the
de
ve
lopm
e
nt
of
mul
ti
c
las
s
models
.
T
his
r
e
s
e
a
r
c
h
is
r
e
leva
nt
to
the
p
r
ojec
t
due
mainly
to
th
e
f
a
c
t
that
the
opti
mi
z
e
d
model
a
s
the
r
e
s
ult
of
thi
s
pr
ojec
t,
will
be
us
e
d
in
the
im
pleme
ntation
of
the
AI
-
a
s
s
is
t
e
d
c
hr
onic
dis
e
a
s
e
e
va
luation
s
ys
tem.
T
h
is
will
be
the
c
or
e
pa
r
t
of
the
T
S
C
D
to
a
s
s
is
t
medic
a
l
pr
a
c
ti
ti
one
r
s
in
Oma
n
to
be
a
ble
to
va
li
da
te
pa
ti
e
nts
’
diagnos
is
us
ing
A
I
.
2.
M
E
T
HO
D
T
his
s
e
c
ti
on
de
s
c
r
ibes
the
method
that
wa
s
a
dopt
e
d
in
thi
s
r
e
s
e
a
r
c
h,
c
ompr
is
e
d
of
the
pr
oc
e
dur
e
in
s
ub
s
e
c
ti
on
2.
1
.
I
n
s
ub
s
e
c
ti
on
2.
2
,
the
methods
in
b
uil
ding
the
pr
e
diction
models
we
r
e
de
s
c
r
ibed,
how
to
a
void
ove
r
f
it
ti
ng
a
nd
unde
r
f
it
ti
ng
,
a
nd
identif
ying
the
o
pti
mal
lea
r
ning
r
a
te.
T
he
f
ull
s
our
c
e
c
ode
to
r
e
pli
c
a
te
the
s
im
ulations
a
nd
e
xpe
r
im
e
nts
in
thi
s
pa
pe
r
is
a
va
il
a
ble
with
the
a
utho
r
s
[
16]
.
2.
1.
Re
s
e
ar
c
h
d
e
s
ign
I
n
the
de
ve
lopm
e
nt
o
f
the
de
e
p
lea
r
ning
model
f
or
the
pr
e
diction
o
f
c
hr
onic
dis
e
a
s
e
,
the
f
r
a
mew
or
k
in
F
igur
e
1
wa
s
e
s
tablis
he
d.
I
n
the
modeling
s
tage
,
we
c
r
e
a
ted
5
models
with
va
r
ying
hype
r
p
a
r
a
mete
r
s
,
r
a
ndomi
z
e
d
the
da
tas
e
t,
opti
mi
z
e
d
the
lea
r
ning
c
ur
ve
,
a
nd
the
r
e
gular
iza
ti
on
pa
r
a
mete
r
s
.
M
ode
l
tr
a
ini
ng
is
c
ompl
e
ted
whe
n
the
s
toppi
ng
c
r
it
e
r
ia
ha
ve
be
e
n
a
c
hieve
d
or
the
number
o
f
e
poc
hs
ha
s
be
e
n
r
e
a
c
he
d
without
a
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
57
0
-
583
572
c
ons
ider
a
ble
de
c
r
e
a
s
e
in
e
r
r
or
r
a
te.
T
his
pr
oc
e
s
s
is
pe
r
f
or
med
with
the
other
4
models
.
I
n
the
pe
r
f
or
manc
e
e
va
luation
s
tage
,
we
s
e
lec
t
the
two
models
with
the
highes
t
pr
e
diction
a
c
c
ur
a
c
y
on
uns
e
e
n
da
ta.
W
e
us
e
d
J
upyter
L
a
b
[
17]
in
Ana
c
onda
Na
vigator
a
nd
Google
C
olab
[
18]
to
wr
it
e
python
s
c
r
ipt
c
ode
s
to
pe
r
f
or
m
da
ta
pr
oc
e
s
s
ing,
modeling
a
nd
pe
r
f
or
manc
e
e
va
luation
of
the
r
e
s
ult
ing
models
in
thi
s
pa
pe
r
.
F
igur
e
1
.
F
r
a
mew
or
k
f
o
r
p
r
e
diction
models
f
o
r
c
hr
onic
dis
e
a
s
e
s
2.
1.
1.
Dat
a
p
r
e
p
ar
at
ion
Da
ta
c
lea
nup
a
nd
pr
e
p
r
oc
e
s
s
ing
a
r
e
f
unda
menta
l
s
teps
in
mac
hine
lea
r
ning
tas
ks
.
P
r
e
diction
models
r
e
ly
on
da
ta
to
lea
r
n
a
nd
make
p
r
e
dictions
.
R
a
w
da
ta
of
ten
c
ontains
incons
is
tenc
ies
,
e
r
r
or
s
,
mi
s
s
ing
va
lues
,
e
xtr
a
c
olum
ns
,
a
nd
r
e
dunda
nt
da
ta
that
c
a
n
dr
a
s
ti
c
a
ll
y
a
f
f
e
c
t
the
pr
e
diction
c
a
pa
bil
it
y
of
the
model
.
I
n
da
ta
pr
e
pa
r
a
ti
on
,
a
loca
li
z
e
d
da
tas
e
t
f
oc
us
e
d
on
t
he
M
iddl
e
E
a
s
t
r
e
gion
is
not
r
e
a
dil
y
a
va
il
a
ble.
T
he
da
tas
e
t
f
ound
in
[
19
]
wa
s
us
e
d
in
thi
s
pa
pe
r
,
pe
r
f
or
med
da
ta
c
lea
nup
a
nd
p
r
e
-
pr
oc
e
s
s
ing
on
it
,
a
nd
s
pli
t
the
da
tas
e
t
int
o
tr
a
ini
ng,
va
li
da
ti
on,
a
nd
tes
t
s
e
ts
.
T
he
t
hr
e
e
s
e
ts
of
da
ta
will
then
be
us
e
d
in
the
modell
ing
s
tage
.
2.
2.
2.
M
od
e
li
n
g
T
he
modelli
ng
s
tage
uti
li
z
e
s
the
r
e
s
ult
ing
da
ta
pr
od
uc
e
d
by
the
da
ta
pr
e
pa
r
a
ti
on
s
tage
.
I
n
thi
s
s
tage
,
5
mul
ti
c
las
s
models
we
r
e
c
r
e
a
ted
with
va
r
ying
nu
mber
of
hidden
laye
r
s
,
identif
ied
the
opti
mal
lea
r
ning
r
a
te,
r
a
ndomi
z
e
d
the
da
ta
s
e
ts
,
a
nd
s
e
t
the
r
e
gular
iza
ti
on
pa
r
a
mete
r
s
a
nd
s
toppi
ng
c
ondit
ions
.
T
r
a
ini
ng
o
f
the
models
a
r
e
then
pe
r
f
o
r
med
unti
l
a
s
toppi
ng
c
r
it
e
r
ion
ha
s
b
e
e
n
r
e
a
c
he
d
.
a)
Ove
r
f
it
ti
ng
a
nd
unde
r
f
it
t
ing
Ove
r
f
it
ti
ng
a
nd
unde
r
f
it
ti
ng
is
a
f
unda
menta
l
is
s
ue
[
20]
in
mac
hine
lea
r
ning.
I
t
r
e
s
ult
s
to
a
poor
ge
ne
r
a
li
z
a
ti
on
c
a
pa
bil
it
y
of
the
model
to
a
c
c
ur
a
tel
y
c
las
s
if
y
uns
e
e
n
da
ta.
T
his
dif
f
iculty
wa
s
a
ls
o
e
nc
ounter
e
d
in
the
pr
e
viou
s
r
e
s
e
a
r
c
h
uti
li
z
ing
ne
u
r
a
l
ne
twor
k
models
[
21]
a
nd
s
uppor
t
ve
c
tor
mac
hines
[
22]
,
[
2
3]
in
the
mul
ti
c
las
s
c
las
s
if
ica
ti
on
of
mi
ll
e
d
r
ice
.
W
he
n
tr
a
ini
ng
a
model
f
o
r
s
e
ve
r
a
l
e
poc
hs
,
the
a
c
c
ur
a
c
y
of
the
model
on
the
va
li
da
ti
on
s
e
t
would
pe
a
k
then
a
f
ter
s
ome
ti
m
e
it
will
s
tar
t
to
de
c
r
e
a
s
e
or
s
tag
na
te.
T
his
s
c
e
na
r
io
e
xe
mpl
if
ies
the
ove
r
f
it
ti
ng
of
th
e
model
to
the
tr
a
ini
ng
da
ta.
T
he
oppos
it
e
of
thi
s
s
c
e
na
r
io,
u
nde
r
f
it
ti
ng,
oc
c
ur
s
whe
n
the
model
c
a
n
be
im
pr
ove
d
on
the
tr
a
ined
da
ta
but
the
model
is
not
given
e
nough
t
im
e
to
lea
r
n
the
r
e
leva
nt
pa
tt
e
r
n
in
the
tr
a
ini
ng
da
ta.
T
o
de
mons
tr
a
te
thi
s
pr
oblem
in
T
e
ns
or
F
low
[
9]
,
the
Higgs
da
tas
e
t
[
24]
wa
s
us
e
d
to
de
ve
lop
mul
ti
-
c
la
s
s
if
ica
ti
on
models
to
s
e
e
how
ove
r
f
it
ti
ng
a
nd
unde
r
f
it
ti
ng
a
f
f
e
c
t
the
a
c
c
ur
a
c
y
of
the
model
f
or
thi
s
da
tas
e
t.
I
n
T
a
b
le
2,
f
our
models
with
va
r
ying
number
s
of
hidden
laye
r
s
a
r
e
s
hown.
F
or
thes
e
models
,
the
number
of
outpu
t
c
las
s
e
s
is
f
ixed
a
t
17
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
c
c
ur
ate
pr
e
diction
of
c
hr
onic
dis
e
as
e
s
us
ing
de
e
p
lear
ning
algor
it
hms
…
(
R
onald
S.
C
or
dov
a
)
573
T
a
ble
2.
De
e
p
lea
r
ning
model
f
o
r
higgs
da
tas
e
t
wit
h
va
r
ying
numbe
r
of
pa
r
a
mete
r
s
#
M
ode
l
n
a
me
L
a
ye
r
s
O
ut
put
s
h
a
pe
T
ot
a
l
p
a
r
a
ms
1
2
3
4
5
1
T
N
Y
464
16
481
2
S
M
L
928
1,056
32
2017
3
M
D
M
1,856
4,160
4,160
65
64
10,241
4
L
G
E
14,848
262,656
262,656
262,656
513
512
803,329
I
n
de
e
p
lea
r
ning,
ne
ur
a
l
ne
twor
k
a
r
e
tr
a
ined
us
ing
a
n
opti
mi
z
a
ti
on
a
lgor
it
hm
c
a
ll
e
d
s
tocha
s
ti
c
gr
a
dient
de
s
c
e
nt
to
mi
nim
ize
the
los
s
f
unc
ti
on,
ther
e
by
r
e
duc
ing
the
e
r
r
or
be
twe
e
n
the
ne
twor
k’
s
pr
e
dictio
n
a
nd
the
de
s
ir
e
d
r
e
s
ult
s
.
T
he
lea
r
ning
r
a
te
a
c
ts
a
s
a
c
r
uc
ial
hype
r
pa
r
a
mete
r
withi
n
thi
s
opti
mi
z
a
ti
on
pr
oc
e
s
s
.
W
e
know
th
a
t
lar
ge
r
lea
r
ning
r
a
te
may
r
e
s
ult
s
to
mi
s
s
ing
th
e
opti
mal
s
olut
ion
but
[
25]
ha
s
a
s
olut
ion
f
o
r
it
.
M
ode
ls
in
T
a
ble
2
us
e
d
the
s
a
me
c
onf
igur
a
ti
on
a
nd
lea
r
ning
r
a
te
a
s
s
hown
in
F
igur
e
2.
T
his
lea
r
ning
r
a
te
is
e
xpe
c
ted
to
hype
r
boli
c
a
ll
y
de
c
r
e
a
s
e
by
½
of
the
ba
s
e
r
a
te
a
t
1,
000
e
poc
hs
,
1/3
a
t
2,
000
e
poc
hs
,
a
nd
s
o
on.
F
igur
e
2
.
L
e
a
r
ning
r
a
te
F
igur
e
3
s
hows
the
p
lot
be
twe
e
n
tr
a
ini
ng
a
nd
va
li
da
ti
on
los
s
.
L
owe
r
va
li
da
ti
on
los
s
ind
ica
tes
be
tt
e
r
ge
ne
r
a
li
z
a
ti
on
to
p
r
e
dict
uns
e
e
n
da
ta.
T
he
t
r
a
ini
ng
e
nds
a
f
ter
200
e
poc
hs
if
ther
e
is
no
pr
ogr
e
s
s
.
I
n
th
is
s
im
ulation,
the
T
iny
ne
ur
a
l
ne
twor
k
(
T
NY
)
mode
l
ha
s
a
voided
ove
r
f
it
ti
ng
the
da
ta
while
the
mo
r
e
c
ompl
e
x
models
(
i
.
e
.
,
s
upe
r
vis
e
d
mac
hine
lea
r
ning
(
S
M
L
)
,
mul
ti
-
dim
e
ns
ional
model
(
M
DM
)
,
lar
ge
gr
a
dient
e
s
ti
mation
(
L
GE
)
)
,
tend
to
ove
r
f
i
t
the
da
ta
mor
e
quickly.
F
r
o
m
the
r
e
s
ult
in
F
igu
r
e
3
,
it
wa
s
obs
e
r
ve
d
th
a
t:
‒
I
t
is
no
r
mal
f
o
r
b
ot
h
t
r
a
i
ni
ng
a
n
d
va
l
ida
t
io
n
l
os
s
t
o
ha
ve
a
s
ma
l
l
d
i
f
f
e
r
e
n
c
e
(
T
I
NY
)
.
‒
I
t
is
no
r
mal
t
o
ha
ve
t
r
a
in
in
g
a
nd
va
l
i
da
t
i
on
los
s
to
be
mo
v
in
g
i
n
t
he
s
a
me
di
r
e
c
ti
on
(
T
I
N
Y)
.
‒
W
he
n
the
t
r
a
i
ni
ng
l
os
s
c
o
nt
in
ue
s
t
o
im
p
r
o
ve
bu
t
th
e
va
li
da
ti
on
l
os
s
s
ta
r
t
s
to
de
te
r
io
r
a
te
,
t
r
a
i
ni
ng
is
m
os
t
l
i
ke
ly
c
l
os
e
to
ove
r
f
it
ti
ng
(
S
M
L
,
M
DM
,
L
G
E
)
.
‒
I
f
t
he
va
li
da
ti
on
los
s
is
inc
r
e
a
s
in
g
(
go
i
ng
upw
a
r
d
)
,
th
e
m
od
e
l
is
o
ve
r
f
it
ti
ng
(
S
M
L
,
M
DM
,
L
G
E
)
.
b)
S
tr
a
tegie
s
to
a
void
ove
r
f
it
ti
ng
of
da
ta
Va
r
ious
s
tr
a
tegie
s
to
a
void
ove
r
f
it
ti
ng
t
ha
t
we
r
e
us
e
d
in
thi
s
pa
pe
r
a
r
e
dis
c
us
s
e
d
in
s
ub
s
ub
s
e
c
ti
on
2.
4.
1
to
2
.
4.
6
.
T
he
s
tr
a
tegie
s
de
s
c
r
ibed
in
thi
s
s
ubs
ub
s
e
c
ti
on
a
void
ove
r
f
it
ti
ng
o
f
t
r
a
ini
ng
da
ta
lea
ding
to
im
pr
ove
d
ge
ne
r
a
li
z
a
ti
on
c
a
pa
bil
it
y
of
the
tr
a
ined
model
f
or
a
c
c
ur
a
te
pr
e
diction
on
uns
e
e
n
da
ta.
‒
T
r
a
ini
ng
with
mor
e
da
ta
One
of
the
mos
t
c
omm
on
s
tr
a
tegie
s
is
to
include
th
e
f
ull
r
a
nge
o
f
input
s
that
the
model
is
e
xpe
c
ted
to
ha
ndle
f
or
e
a
c
h
c
las
s
if
ica
ti
on
output
.
Additi
ona
l
tr
a
ini
ng
da
ta
may
only
be
r
e
leva
nt
to
c
ove
r
ne
w
c
la
s
s
e
s
a
nd
int
e
r
e
s
ti
ng
c
a
s
e
s
.
A
model
tr
a
ined
thi
s
wa
y
tends
t
o
ge
ne
r
a
li
z
e
be
tt
e
r
.
Gütte
r
e
t
al.
[
26
]
,
it
wa
s
s
hown
that
the
tr
a
ini
ng
s
e
t
s
ize
doe
s
play
a
r
ole
in
a
f
f
e
c
ti
ng
the
r
o
bus
tnes
s
of
the
model
a
ga
ins
t
nois
e
.
How
e
ve
r
,
thi
s
s
tr
a
tegy
may
not
wor
k
e
ve
r
y
ti
me.
I
f
nois
y
da
t
a
we
r
e
a
dde
d,
thi
s
tec
hnique
wouldn’
t
he
lp.
T
his
is
the
r
e
a
s
on
why
da
ta
mus
t
be
c
lea
ns
e
d
[
27]
.
‒
R
e
duc
ing
the
c
ompl
e
xit
y
of
the
ne
twor
k
Anothe
r
s
tr
a
tegy
is
to
r
e
duc
e
the
c
ompl
e
xit
y
o
f
th
e
ne
twor
k
s
im
il
a
r
to
wha
t
wa
s
done
in
thi
s
pa
pe
r
[
28]
.
As
s
hown
in
T
a
ble
2
,
the
L
GE
model
is
c
ompr
is
e
d
of
a
whopping
803
,
329
pa
r
a
mete
r
s
,
the
r
e
d
da
s
h
li
ne
in
in
F
igur
e
3
s
hows
the
va
li
da
t
ion
los
s
is
s
hooti
ng
in
a
n
upwa
r
d
di
r
e
c
ti
on.
I
t
is
e
nough
to
s
a
y
that
in
t
his
L
GE
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
57
0
-
583
574
model,
ove
r
f
it
ti
ng
c
ould
ha
ve
be
e
n
a
voided
if
a
les
s
c
ompl
i
c
a
ted
ne
twor
k
ha
d
be
e
n
a
dopted
a
s
s
ho
wn
in
the
T
NY
model
in
T
a
ble
2.
‒
E
a
r
ly
s
toppi
ng
E
a
r
ly
s
toppi
ng
is
a
r
e
gular
iza
ti
on
tec
hnique
f
or
a
de
e
p
ne
ur
a
l
ne
twor
k
that
s
tops
tr
a
ini
ng
whe
n
pa
r
a
mete
r
s
do
not
r
e
s
ult
in
a
ny
im
pr
ove
ments
on
va
li
da
ti
on
s
e
ts
a
f
ter
a
number
of
e
poc
hs
[
29]
.
T
he
r
e
s
e
a
r
c
he
r
s
a
r
e
us
ing
thi
s
s
tr
a
tegy
in
thi
s
pa
pe
r
by
s
toppi
ng
the
tr
a
ini
ng
a
f
te
r
100
e
poc
hs
of
no
im
pr
ove
ments
.
I
n
f
a
c
t,
mos
t
if
not
a
ll
mac
hine
lea
r
ning
f
r
a
mew
or
ks
s
uppor
t
thi
s
s
tr
a
tegy.
‒
R
e
gular
iza
ti
on
R
e
gular
iza
ti
on
is
a
n
e
xc
e
l
lent
tec
hnique
to
a
ddr
e
s
s
ove
r
f
it
ti
ng
pr
ob
lems
.
S
ome
e
xa
mpl
e
s
of
thes
e
tec
hniques
us
e
li
ne
a
r
r
e
gr
e
s
s
ion
[
30]
,
L
1
r
e
gular
i
z
a
ti
on
[
31]
,
a
nd
L
2
r
e
gular
iza
ti
on
[
32]
.
F
igur
e
4
s
hows
a
tr
a
ini
ng
a
nd
va
li
da
ti
on
los
s
f
or
the
L
GE
model
tha
t
uti
li
z
e
s
L
2
r
e
gular
iza
ti
o
n.
T
his
r
e
gular
ize
d
ve
r
s
ion
of
the
L
GE
model
is
c
ompetit
ive
with
the
T
NY
model
,
wi
th
mor
e
r
e
s
is
tanc
e
to
ove
r
f
it
ti
ng
than
the
L
GE
mod
e
l
it
wa
s
de
r
ived
f
r
om
.
‒
Dr
opout
Dr
opout
is
one
o
f
the
s
im
ples
t
wa
ys
[
33
]
to
a
voi
d
ove
r
f
it
ti
ng
us
ing
r
e
gular
iza
ti
on
tec
hniques
that
dis
c
a
r
d
r
a
ndom
node
s
dur
ing
tr
a
ini
ng
by
e
ns
ur
ing
that
no
va
r
iable
s
a
r
e
not
c
o
-
de
pe
nde
nt
with
one
a
n
other
.
I
t
mea
ns
that
the
model
will
not
lea
r
n
r
e
dunda
nt
de
ta
il
s
of
the
input
.
An
e
xa
mpl
e
is
pr
ov
ided
in
F
igu
r
e
5
whe
r
e
L
GE
uti
li
z
e
s
a
dr
opout
r
e
gular
iza
ti
on
tec
hnique.
F
igur
e
3.
T
r
a
ini
ng
vs
va
li
da
ti
on
los
s
e
s
f
or
Higgs
da
tas
e
t
F
igur
e
4.
T
r
a
ini
ng
vs
va
li
da
ti
on
los
s
f
or
L
GE
model
with
L
2
r
e
gula
r
iza
ti
on
tec
hnique
F
igur
e
5.
L
GE
M
ode
l
with
dr
opout
r
e
gular
iza
ti
on
c)
Optim
a
l
lea
r
ning
r
a
te
e
s
ti
mation
H
yp
e
r
pa
r
a
me
te
r
s
de
c
i
de
th
e
ti
me
a
n
d
c
om
pu
ta
ti
on
a
l
c
o
s
t
o
f
tr
a
i
ni
ng
n
e
u
r
a
l
ne
t
wo
r
k
mo
de
l
s
.
T
h
e
y
h
e
a
vi
ly
i
n
f
l
ue
nc
e
t
he
s
t
r
uc
t
ur
e
,
pr
e
d
ic
ti
on
a
c
c
u
r
a
c
y
,
a
nd
ge
ne
r
a
l
iza
ti
on
c
a
pa
bi
li
t
y
o
f
t
he
mo
de
l
.
T
he
lea
r
n
i
ng
r
a
te
i
s
o
ne
o
f
th
os
e
hy
pe
r
p
a
r
a
met
e
r
s
t
ha
t
r
e
q
ui
r
e
s
t
un
in
g
.
I
f
t
he
le
a
r
n
in
g
r
a
t
e
is
i
nc
r
e
me
nt
a
l
ly
s
ma
ll
,
the
n
t
r
a
in
in
g
is
mo
r
e
r
e
li
a
b
le
,
b
u
t
op
ti
m
iza
t
io
n
wi
ll
ta
ke
m
o
r
e
t
i
me
be
c
a
us
e
o
f
th
e
s
ma
ll
e
r
s
te
p
s
ize
t
o
ob
ta
i
n
the
mi
n
i
mu
m
o
f
th
e
los
s
f
u
nc
ti
o
n
.
O
n
th
e
ot
he
r
ha
nd
,
i
f
t
he
l
e
a
r
ni
ng
r
a
te
wa
s
s
e
t
a
b
i
t
to
o
h
ig
h
,
t
he
n
th
e
t
r
a
in
in
g
ma
y
n
ot
c
o
nve
r
ge
o
r
d
iv
e
r
ge
.
I
n
th
is
c
a
s
e
,
th
e
o
pt
im
iz
e
r
ma
y
n
ot
be
a
bl
e
to
a
c
h
ie
ve
t
he
m
in
i
mu
m
o
f
th
e
los
s
f
u
nc
ti
on
wh
ic
h
ma
ke
s
the
mo
de
l
wo
r
s
t
.
L
oc
a
l
qu
a
d
r
a
t
ic
a
p
p
r
ox
i
ma
ti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
c
c
ur
ate
pr
e
diction
of
c
hr
onic
dis
e
as
e
s
us
ing
de
e
p
lear
ning
algor
it
hms
…
(
R
onald
S.
C
or
dov
a
)
575
[
3
4
]
,
s
to
c
h
a
s
t
ic
g
r
a
d
ien
t
de
s
c
e
nt
[
3
5
]
a
n
d
th
e
me
tho
d
us
e
d
i
n
s
up
e
r
-
c
o
nv
e
r
ge
nc
e
[
2
5
]
a
r
e
s
o
me
o
f
t
he
l
e
a
r
n
in
g
r
a
te
e
s
t
i
ma
ti
on
s
t
r
a
te
g
ies
.
F
ig
ur
e
6
(
a
)
s
ho
ws
th
e
lea
r
ni
ng
r
a
t
e
us
e
d
,
F
i
gu
r
e
s
6
(
b
)
a
nd
6
(
c
)
s
h
ows
a
n
e
xc
e
l
len
t
c
on
ve
r
g
e
n
c
e
o
f
th
e
t
r
a
i
ni
ng
/v
a
l
id
a
t
io
n
l
os
s
/ac
c
u
r
a
c
y.
(
a
)
(
b)
F
igur
e
6.
L
e
a
r
ning
r
a
te
,
a
c
c
ur
a
c
y
a
nd
los
s
e
s
(
a
)
lea
r
ning
r
a
te
,
(
b
)
t
r
a
ini
ng
vs
va
li
da
ti
on
a
c
c
ur
a
c
y,
a
nd
(
c
)
tr
a
ini
ng
vs
va
li
da
ti
on
los
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
57
0
-
583
576
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
I
n
th
is
s
e
c
ti
on,
the
de
ve
lopm
e
nt
of
the
c
las
s
if
ica
ti
on
models
buil
t
us
ing
T
e
ns
or
F
low
is
e
xplaine
d
in
de
tail.
S
ubs
e
c
ti
on
s
3.
1
a
nd
3
.
2
de
s
c
r
ibe
the
c
ons
ider
a
ti
ons
whe
n
de
s
igni
ng
a
nd
tr
a
ini
ng
a
mul
ti
c
l
a
s
s
de
e
p
lea
r
ning
model
.
S
ubs
e
c
ti
on
s
3.
3
a
nd
3
.
4
de
s
c
r
ibe
i
n
de
tail
how
the
de
e
p
ne
u
r
a
l
ne
two
r
k
mo
de
l
wa
s
d
e
ve
loped.
F
inally,
s
ub
s
e
c
ti
on
3.
5
de
s
c
r
ibes
the
pe
r
f
or
manc
e
e
va
luation
a
nd
s
e
lec
ti
on
o
f
the
mul
ti
c
las
s
de
e
p
lea
r
ning
c
las
s
if
ica
ti
on
model
in
thi
s
pa
pe
r
.
3
.
1
.
D
a
t
as
e
t
c
l
e
a
n
u
p
an
d
p
r
e
p
ar
a
t
io
n
T
he
da
tas
e
t
[
19
]
is
us
e
d
in
thi
s
pa
pe
r
to
buil
d
the
de
e
p
lea
r
ning
mul
t
icla
s
s
model.
F
or
thi
s
da
tas
e
t,
ther
e
a
r
e
134
c
olum
ns
in
tot
a
l
–
132
c
olum
ns
r
e
pr
e
s
e
nt
dis
ti
nc
t
s
ympt
oms
,
c
olum
n
#133
r
e
pr
e
s
e
nts
the
dif
f
e
r
e
nt
diagnos
e
s
of
our
output
c
las
s
e
s
,
a
nd
c
olum
n
134
i
s
a
n
ir
r
e
leva
nt
c
olum
n
that
mus
t
be
r
e
moved.
F
ur
t
he
r
mor
e
,
the
output
c
a
tegor
ies
c
ons
is
ti
ng
o
f
41
diagnos
e
s
m
us
t
be
t
r
a
ns
late
d
int
o
numer
ica
l
va
lues
r
e
pr
e
s
e
nted
a
s
s
hown
in
F
igur
e
7.
F
igur
e
7.
I
ndice
s
of
output
c
a
tegor
ies
3
.
2
.
S
p
l
it
t
i
n
g
of
t
h
e
d
at
as
e
t
in
t
o
t
r
a
i
n
i
n
g
,
va
l
id
a
t
io
n
an
d
t
e
s
t
s
e
t
I
n
the
tr
a
ini
ng
da
tas
e
t,
ther
e
a
r
e
4,
920
r
ows
a
nd
132
f
e
a
tur
e
s
.
T
he
tr
a
ini
ng
da
tas
e
t
wa
s
r
a
ndoml
y
s
pli
t
int
o
80/20
pe
r
c
e
nt
whe
r
e
the
20%
va
li
da
ti
on
s
e
t
is
c
ompr
is
e
d
of
984
r
ows
.
T
he
t
r
a
ini
ng
a
nd
va
li
da
ti
o
n
s
e
t
will
be
us
e
d
f
or
tr
a
ini
ng
a
nd
f
it
ti
ng
the
model.
T
he
tes
t
da
tas
e
t
is
c
ompr
is
e
d
of
41
r
ows
a
nd
132
f
e
a
tur
e
s
whic
h
is
uns
e
e
n
da
ta
to
be
us
e
d
f
or
tes
ti
ng
the
pe
r
f
or
manc
e
of
the
model
.
T
he
number
of
f
e
a
tur
e
s
or
c
olum
ns
r
e
pr
e
s
e
nts
va
r
ious
s
ympt
oms
that
c
a
n
be
e
nc
ounter
e
d
by
a
pa
t
ient.
T
he
number
of
output
c
las
s
e
s
or
c
a
tegor
ies
is
41
whic
h
r
e
pr
e
s
e
nts
the
pos
s
ibl
e
c
hr
onic
dis
e
a
s
e
s
,
a
s
s
hown
in
F
igur
e
8.
F
igur
e
8.
C
hr
onic
dis
e
a
s
e
s
(
output
c
a
tegor
ies
or
c
l
a
s
s
e
s
)
3
.
3
.
B
u
i
l
d
i
n
g
a
m
u
l
t
ic
las
s
c
l
as
s
i
f
ic
at
i
on
m
od
e
l
T
his
s
e
c
ti
on
outl
ines
the
pr
oc
e
s
s
that
wa
s
f
oll
ow
e
d
f
or
buil
ding
a
mul
ti
c
las
s
pr
e
diction
model
f
o
r
c
hr
onic
dis
e
a
s
e
s
.
S
ubs
ubs
e
c
ti
on
3.
3.
1,
de
s
c
r
ibes
t
he
tr
a
ini
ng
c
onf
igu
r
a
ti
on
us
e
d
f
o
r
5
dif
f
e
r
e
nt
c
las
s
if
ica
ti
on
models
dis
c
us
s
e
d
in
thi
s
pa
pe
r
.
T
he
numbe
r
of
f
e
a
tur
e
s
is
132,
r
e
pr
e
s
e
nti
ng
dif
f
e
r
e
nt
s
ympt
oms
e
nc
ounter
e
d
by
a
pa
ti
e
nt.
T
he
da
tas
e
t
is
divi
de
d
int
o
tr
a
ini
ng,
va
li
da
ti
on,
a
nd
tes
t
da
ta
with
a
number
of
ins
tanc
e
s
e
quivale
nt
to
3936,
984
a
nd
41,
r
e
s
pe
c
ti
ve
ly.
T
h
is
s
e
c
ti
on
a
ls
o
pr
ovides
the
s
e
tt
ings
us
e
d
dur
ing
the
opt
im
iza
ti
o
n
of
the
model
s
uc
h
a
s
the
maximum
number
of
e
poc
hs
100
,
the
number
of
t
r
a
ini
ng
ins
tanc
e
s
to
pr
oc
e
s
s
(
ba
tch
s
ize
=
50)
a
t
a
ti
me,
a
nd
the
nu
mber
o
f
s
teps
pe
r
e
poc
h
62
.
S
ubs
ubs
e
c
ti
on
3.
3.
2
pr
ovides
the
c
onf
igur
a
ti
on
us
e
d
to
de
c
a
y
the
lea
r
ning
r
a
te.
S
ubs
ubs
e
c
ti
on
3.
3
.
3
pr
e
s
e
nts
the
mode
ls
c
r
e
a
ted
whos
e
hype
r
pa
r
a
mete
r
s
e
tt
ings
r
a
n
ge
f
r
om
a
tot
a
l
of
481
up
to
803
,
309
hype
r
pa
r
a
mete
r
s
a
nd
a
number
o
f
laye
r
s
f
r
om
1
up
to
5
hidden
laye
r
s
.
3
.
3
.
1
.
T
r
a
i
n
i
n
g
c
o
n
f
ig
u
r
at
io
n
T
a
ble
3
p
r
e
s
e
nts
the
c
onf
igur
a
ti
on
f
or
a
ll
the
mod
e
ls
to
be
buil
t
in
thi
s
p
a
pe
r
.
T
he
tot
a
l
number
of
pos
s
ibl
e
s
ympt
oms
is
r
e
pr
e
s
e
nted
by
the
tot
a
l
num
be
r
of
f
e
a
tur
e
s
.
Us
ing
the
80:20
P
a
r
e
to
r
ule
,
the
r
e
s
e
a
r
c
he
r
s
s
pli
t
the
tr
a
ini
ng
s
e
t
int
o
tr
a
ini
ng
ins
tanc
e
s
e
qua
l
t
o
3
,
936
a
nd
va
li
da
ti
on
ins
tanc
e
s
e
qua
l
to
984.
T
he
f
inal
tes
t
da
t
a
whic
h
is
the
uns
e
e
n
da
ta
ins
tanc
e
s
is
41
will
b
e
us
e
d
to
e
va
luate
the
pe
r
f
or
manc
e
of
the
models
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
c
c
ur
ate
pr
e
diction
of
c
hr
onic
dis
e
as
e
s
us
ing
de
e
p
lear
ning
algor
it
hms
…
(
R
onald
S.
C
or
dov
a
)
577
T
a
ble
3.
De
e
p
lea
r
ning
model
tr
a
ini
ng
c
onf
igur
a
t
io
n
Set
t
i
n
g
s
O
u
t
p
u
t
s
h
ap
e
T
o
t
al
n
u
mb
er
o
f
feat
u
re
s
132
N
u
mb
er
o
f
v
a
l
i
d
at
i
o
n
i
n
s
t
a
n
ces
984
N
u
mb
er
o
f
t
ra
i
n
i
n
g
i
n
s
t
a
n
ces
3
,
9
3
6
N
u
mb
er
o
f
u
n
s
ee
n
d
at
a
i
n
s
t
a
n
ces
41
Bu
ffer
s
i
ze
100
Bat
ch
s
i
ze
(
e
p
o
ch
s
)
50
St
ep
s
p
er
ep
o
ch
62
3
.
3
.
2
.
L
e
a
r
n
i
n
g
r
a
t
e
I
nve
r
s
e
de
c
a
y
f
unc
ti
on
is
a
s
tr
a
tegy
f
or
s
c
he
duli
n
g
the
lea
r
ning
r
a
te
dur
ing
tr
a
ini
ng
.
I
t
a
djus
ts
the
lea
r
ning
r
a
te
ba
s
e
d
on
the
c
ur
r
e
nt
it
e
r
a
ti
on
s
tep
du
r
ing
tr
a
ini
ng
.
T
he
lea
r
ning
r
a
te
is
inc
r
e
a
s
e
d
a
s
the
in
ve
r
s
e
of
the
it
e
r
a
ti
on,
of
ten
wr
it
ten
a
s
(
1)
:
l
e
a
r
ni
ng
_
r
a
te
=
(
1
+
×
)
(
1
)
whe
r
e
r
-
ini
ti
a
l
lea
r
ning
r
a
te,
dr
–
de
c
a
y
r
a
te,
ds
–
de
c
a
y
s
tep
a
nd
s
–
s
tep.
F
or
our
models
,
pa
r
a
mete
r
s
a
r
e
s
e
t
in
a
c
c
or
da
nc
e
with
T
a
ble
4.
As
the
tr
a
ini
ng
pr
oc
e
e
ds
,
lea
r
ning
r
a
te
will
de
c
a
y
ba
s
e
d
on
(
1)
,
a
nd
pa
r
a
mete
r
s
in
T
a
ble
4.
T
a
ble
4.
L
e
a
r
ning
r
a
te
c
onf
igur
a
ti
on
S
e
tt
in
gs
V
a
lu
e
I
ni
ti
a
l
le
a
r
ni
ng r
a
te
(
)
0.001
D
e
c
a
y s
te
p
s
(
)
S
te
ps
pe
r
e
poc
h x 100
D
e
c
a
y r
a
te
(
)
0.6
3
.
3
.
3
.
H
yp
e
r
p
ar
a
m
e
t
e
r
m
o
d
e
l
s
e
t
t
i
n
gs
T
he
r
e
a
r
e
5
models
us
e
d
in
thi
s
pa
pe
r
in
whic
h
c
ompl
e
xit
ies
va
r
y
s
igni
f
ica
ntl
y
.
T
he
f
ir
s
t
model
-
M
OD
E
L
1
is
the
s
im
ples
t
with
a
tot
a
l
pa
r
a
mete
r
of
11,
177
while
the
mos
t
c
ompl
e
x
is
M
OD
E
L
5
with
93,
609
pa
r
a
mete
r
s
.
T
he
r
e
s
e
a
r
c
he
r
s
e
xpe
r
im
e
nted
with
va
r
ying
the
number
of
hidden
laye
r
s
a
nd
s
e
ve
r
a
l
pa
r
a
mete
r
s
by
us
ing
the
s
e
tt
ings
in
T
a
ble
5
in
dif
f
e
r
e
nt
t
r
ial
r
uns
.
T
a
ble
5.
Hype
r
pa
r
a
mete
r
s
s
e
tt
ings
#
M
o
d
e
l
n
a
m
e
L
a
y
e
r
s
O
u
tp
ut
s
h
a
p
e
T
o
t
a
l
p
a
r
a
m
s
1
2
3
4
5
1
M
O
D
E
L
1
64
41
11,177
2
M
O
D
E
L
2
128
256
41
60,585
3
M
O
D
E
L
3
128
128
256
41
77,097
4
M
O
D
E
L
4
128
128
128
256
41
93,609
5
M
O
D
E
L
5
64
64
64
64
256
41
48,169
I
n
thi
s
pa
pe
r
,
the
i
nput
laye
r
is
a
lwa
ys
s
e
t
to
us
e
d
e
ns
e
.
T
he
de
f
a
ult
output
laye
r
us
e
d
is
the
s
of
tm
a
x
[
36]
f
unc
ti
on
but
the
mos
t
c
omm
only
us
e
d
a
c
ti
va
ti
on
f
unc
ti
on
-
s
igm
oid
f
unc
ti
on
[
37]
,
wa
s
a
ls
o
tes
ted
to
identif
y
whe
ther
s
of
tm
a
x
will
pe
r
f
o
r
m
be
s
t
us
ing
the
hype
r
pa
r
a
mete
r
s
e
tt
ings
in
T
a
ble
5
.
T
he
a
c
ti
va
ti
on
f
unc
t
ion
us
e
d
in
laye
r
s
2
–
5
wa
s
E
L
U
[
38]
.
R
e
L
U
[
39]
,
[
40]
wa
s
a
l
s
o
us
e
d
a
s
a
n
a
c
ti
va
ti
on
f
unc
ti
on
in
hidden
laye
r
s
to
e
va
luate
whe
ther
c
ons
ider
a
ble
pe
r
f
or
manc
e
im
p
r
ove
ments
will
be
c
ompar
e
d
to
E
L
U.
3
.
4
.
T
r
a
in
i
n
g
t
h
e
m
o
d
e
ls
T
he
r
e
s
e
a
r
c
he
r
s
pe
r
f
or
med
va
r
ious
t
r
ial
r
uns
to
f
in
d
the
be
s
t
model
in
thi
s
pa
pe
r
.
T
he
lea
r
ning
r
a
te
in
F
igur
e
9
wa
s
us
e
d
in
a
ll
models
.
Dur
ing
the
c
our
s
e
of
t
r
a
ini
ng
the
models
de
s
c
r
ibed
in
thi
s
pa
pe
r
,
a
n
i
s
s
ue
ha
s
be
e
n
e
nc
ounter
e
d
whe
r
e
the
r
e
s
ult
s
obtaine
d
a
r
e
di
f
f
e
r
e
nt
e
a
c
h
ti
me.
T
his
pr
e
ve
nted
us
f
r
om
movi
ng
f
or
wa
r
d.
T
he
r
e
s
e
a
r
c
he
r
s
we
r
e
a
ble
to
f
ix
the
p
r
oblem
by
s
e
tt
ing
a
f
ixed
r
a
ndom
s
e
e
d
globally
.
F
igur
e
9
s
h
ows
the
s
umm
a
r
y
of
the
a
c
c
ur
a
c
y
of
the
model
whe
n
a
ppli
e
d
to
the
t
r
a
ini
ng
s
e
t
a
nd
va
li
da
ti
on
s
e
t.
P
r
e
diction
a
c
c
ur
a
c
y
a
nd
los
s
e
s
on
uns
e
e
n
da
ta
a
r
e
de
s
c
r
ibed
in
mor
e
de
tail
in
the
ne
xt
s
e
c
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
1
,
F
e
br
ua
r
y
2025
:
57
0
-
583
578
F
igur
e
9.
Va
li
da
ti
on
vs
tr
a
ini
ng
a
c
c
ur
a
c
y
As
obs
e
r
ve
d
in
F
igu
r
e
6
,
the
gr
a
ph
of
a
ll
the
model
s
in
ter
ms
of
a
c
c
ur
a
c
y
ha
s
c
onve
r
ge
d
to
1
.
0
whic
h
is
nor
mal.
T
he
g
r
a
ph
tells
us
that
the
we
ighi
ng
va
l
ue
s
ha
ve
a
lr
e
a
dy
be
e
n
opti
mi
z
e
d
a
t
e
poc
h
<
100
b
e
c
a
us
e
it
ha
s
r
e
a
c
he
d
the
maximum
pos
s
ibl
e
a
c
c
ur
a
c
y
of
1.
0.
F
igur
e
10
pr
e
s
e
nts
the
f
inal
model
that
wa
s
tr
a
i
ne
d
a
nd
buil
t
in
thi
s
p
r
ojec
t.
F
o
r
s
im
pli
s
ti
c
im
pleme
ntation
in
a
pr
oduc
ti
on
e
nvir
onment,
M
OD
E
L
1
is
r
e
c
om
mende
d.
I
f
a
mor
e
r
obus
t
model
is
to
be
de
s
ir
e
d,
then
M
OD
E
L
5
is
highl
y
r
e
c
omm
e
nde
d.
I
ts
numer
ous
number
s
of
hype
r
pa
r
a
mete
r
s
a
r
e
not
to
be
de
s
ir
e
d,
but
the
los
s
is
take
n
int
o
a
c
c
ount,
M
OD
E
L
5
ou
tper
f
or
ms
M
OD
E
L
1
by
a
f
a
c
tor
o
f
213
ti
mes
.
F
igur
e
10
s
hows
the
numbe
r
o
f
laye
r
s
,
a
nd
type
of
a
c
ti
va
ti
on
f
unc
ti
ons
f
o
r
both
M
OD
E
L
1
a
nd
M
OD
E
L
2.
Hidde
n
laye
r
s
us
e
E
L
U
[
38]
a
nd
output
laye
r
s
uti
li
z
e
s
S
of
tm
a
x
[
36]
.
Note
that
the
number
of
input
s
c
or
r
e
s
ponds
to
the
number
of
pos
s
ibl
e
s
ympt
oms
a
nd
the
ou
tput
c
o
r
r
e
s
ponds
to
the
number
of
output
c
a
tegor
ies
whic
h
is
41.
T
he
c
o
r
r
e
c
t
output
c
las
s
if
ica
ti
on
is
de
t
e
r
mi
ne
d
by
taking
the
index
of
the
maximum
outp
ut
va
lue.
F
igur
e
10.
M
OD
E
L
1
vs
M
OD
E
L
5
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
c
c
ur
ate
pr
e
diction
of
c
hr
onic
dis
e
as
e
s
us
ing
de
e
p
lear
ning
algor
it
hms
…
(
R
onald
S.
C
or
dov
a
)
579
3
.
5
.
P
e
r
f
or
m
an
c
e
e
va
l
u
a
t
i
o
n
B
a
s
e
d
on
the
e
va
luation
of
the
model
that
wa
s
c
onduc
ted,
it
wa
s
e
vident
that
a
ll
of
the
models
pe
r
f
or
med
we
ll
on
tes
t
da
ta
with
pe
r
f
e
c
t
a
c
c
ur
a
c
y.
How
e
ve
r
,
W
he
n
the
model
wa
s
us
e
d
to
p
r
e
dict
da
ta
that
we
r
e
not
e
nc
ounter
e
d
be
f
or
e
,
M
OD
E
L
1
a
nd
M
OD
E
L
5
we
r
e
the
c
lea
r
victor
s
a
s
s
hown
in
T
a
ble
6
.
S
o
me
M
L
p
r
a
c
ti
ti
on
e
r
s
no
r
ma
ll
y
n
e
g
lec
t
a
c
c
u
r
a
c
y
v
e
r
s
us
los
s
m
e
t
r
ics
.
B
ut
in
th
is
pa
pe
r
,
t
he
r
e
s
e
a
r
c
h
e
r
s
e
q
ua
ll
y
t
r
e
a
t
t
his
m
e
tr
ic
be
c
a
us
e
it
p
r
o
vi
de
s
in
f
o
r
ma
ti
on
on
th
e
l
os
s
f
u
nc
t
io
n
us
e
d
a
n
d
h
ow
w
e
l
l
t
he
mo
de
l
ha
s
pe
r
f
o
r
med
in
te
r
ms
o
f
m
in
im
iz
in
g
th
is
c
a
l
c
u
la
ted
l
os
s
.
T
a
bl
e
7
s
ho
ws
the
pe
r
f
o
r
ma
nc
e
o
f
the
5
mo
de
ls
in
th
e
c
las
s
i
f
ic
a
t
i
on
of
u
ns
e
e
n
d
a
t
a
.
T
a
ble
6.
T
he
p
r
e
diction
a
c
c
ur
a
c
y
of
uns
e
e
n
da
ta
m
ode
l
D
a
ta
s
a
mpl
e
M
ode
l
1
2
3
4
5
T
e
s
t
d
a
t
a
1.0
1.000000
1.000000
1.000000
1.0
U
ns
e
e
n da
ta
1.0
0.97619
0.97619
0.97619
1.0
T
a
ble
7.
L
os
s
vs
a
c
c
ur
a
c
y
on
uns
e
e
n
da
ta
model
M
ode
l
1
2
3
4
5
L
os
s
0.02007
0.0335
0.08491
0.10055
0.00009
A
c
c
ur
a
c
y
1.00000
0.9761
0.97619
0.97619
1.00000
3
.
5
.
1
.
P
r
e
c
is
io
n
,
r
e
c
a
ll
a
n
d
F
1
-
s
c
or
e
P
r
e
c
is
ion,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
a
r
e
e
s
s
e
nti
a
l
metr
ics
f
or
e
va
luating
the
pe
r
f
or
manc
e
of
the
model.
P
r
e
c
is
ion
in
(
2)
r
e
f
e
r
s
to
the
a
bil
it
y
o
f
the
c
las
s
if
ie
r
not
to
labe
l
a
ne
ga
ti
ve
s
a
mpl
e
a
s
pos
it
ive
a
nd
r
e
c
a
ll
in
(
3
)
i
s
the
a
bil
it
y
of
the
c
las
s
if
ier
to
f
ind
a
ll
the
pos
it
ive
s
a
mpl
e
s
.
F
1
-
s
c
or
e
in
(
4)
,
on
the
othe
r
ha
nd,
is
a
n
e
v
a
luation
metr
ic
that
mea
s
ur
e
s
a
model's
a
c
c
ur
a
c
y.
I
t
c
ombi
ne
s
the
pr
e
c
is
ion
a
nd
r
e
c
a
ll
s
c
or
e
of
a
c
las
s
if
ica
ti
on
model.
T
he
a
c
c
ur
a
c
y
metr
ic
c
omput
e
s
how
many
ti
mes
a
model
made
a
c
or
r
e
c
t
pr
e
diction
a
c
r
os
s
the
e
nti
r
e
d
a
tas
e
t
.
=
+
(
2
)
=
+
(
3
)
1
−
=
2
+
(
4
)
whe
r
e
T
P
–
is
t
r
ue
pos
it
ive,
F
P
–
is
f
a
ls
e
pos
it
ive,
a
n
d
F
N
–
is
f
a
ls
e
ne
ga
ti
ve
.
F
igur
e
11
s
hows
the
pe
r
f
or
manc
e
of
the
model
s
howing
pr
e
c
is
ion=1.
0
,
r
e
c
a
ll
=
1.
0
,
a
nd
f
-
s
c
or
e
=
1.
0.
F
igur
e
11.
P
r
e
c
is
ion,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
Evaluation Warning : The document was created with Spire.PDF for Python.