Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
10
,
No.
4
,
A
ugus
t
2020
,
pp.
4023
~
40
34
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
4
.
pp
4023
-
40
34
4023
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Meth
odology f
or
detecti
on
of
p
aroxysm
al at
rial fib
rillati
on
based on
P
-
W
ave, HRV
an
d QR
electri
ca
l
alte
rnan
s featu
res
Henry
Cas
tro
1
, Ju
an D
avid
Ga
rci
a
-
Ra
ci
n
es
2
,
Al
va
r
o
Be
rnal
-
N
oreñ
a
3
1
Facul
t
y
of
Eng
i
nee
ring
,
Univ
ers
ida
d
San
ti
ago
de
Cali,
Colom
bia
1,
2
,3
El
ec
tr
ic
a
l an
d
Elec
tron
ic E
ng
ine
er
ing
School
,
Univer
sidad
d
el
Vall
e
,
Colom
bi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
24
, 201
9
Re
vised Feb
19
,
2020
Accepte
d
Fe
b 25
, 202
0
The
detec
t
ion
of
par
ox
y
sm
al
atrial
fibr
il
l
at
ion
(PA
F)
is
a
fai
rl
y
com
plex
proc
ess
per
form
ed
m
anua
lly
b
y
c
ard
iol
og
ists
or
el
ec
troph
y
s
iol
ogists
b
y
rea
ding
an
e
le
c
t
roc
ard
iogr
am
(E
CG).
Curre
ntly
,
computat
ion
al
te
chn
ique
s
for
aut
om
atic
d
e
te
c
ti
on
b
ase
d
on
fast
fouri
er
tra
n
sform
(FF
T),
B
ay
es
op
ti
m
al
cl
assifi
er
(BOC),
K
-
nea
rest
nei
g
hbors
(K
-
N
Ns
),
and
art
ifici
al
ne
ura
l
net
work
(AN
N)
have
been
proposed.
In
t
his
stud
y
,
six
fe
at
ure
s
wer
e
obt
a
ine
d
base
d
on
the
m
orpholog
y
of
th
e
P
-
W
ave
,
the
QRS
c
om
ple
x
and
the
hea
rt
r
ate
var
ia
b
il
i
t
y
(HR
V)
of
the
ECG.
The
p
erf
orm
an
ce
of
thi
s
m
et
h
odolog
y
wa
s
val
id
at
ed
using
cl
inica
l
ECG
signal
s
from
the
Ph
y
sione
t
arr
h
y
th
m
ia
dat
aba
s
e
MIT
-
BIH.
A
fee
dforward
neur
a
l
net
work
was
used
to
det
ec
t
t
he
pre
senc
e
of
PAF
rea
chi
ng
a
gene
ral
a
cc
ur
a
c
y
of
97.
4%
.
Th
e
result
s
obta
ined
show
tha
t
the
in
cl
usion
of
the
informa
ti
o
n
of
the
P
-
W
av
e,
HRV
and
Q
R
El
e
ct
ri
cal
al
t
ern
ans
inc
r
ea
s
es
the
a
cc
ura
c
y
t
o
ide
nti
f
y
the
P
AF
eve
nt
compa
red
to
other
works
tha
t
us
e
th
e
informat
ion
of
onl
y
one
or at
m
ost t
wo of
the
m
.
Ke
yw
or
d
s
:
Ar
ti
fici
al
n
e
ur
a
l netw
ork
Digital
sig
nal
processi
ng
Ele
ct
ro
ca
r
diog
ram
Feat
ur
es
ex
t
rac
ti
on
HRV
Me
thodo
l
og
y
f
or d
et
ect
io
n
Par
ox
ysm
al
atria
l fibr
il
la
ti
on
P
-
Wav
e
QR elec
tric
al
a
lt
ern
ans
Copyright
©
202
0
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Henry Ca
str
o
,
Faculty
of E
ngineerin
g, Bi
oe
ng
i
neer
i
ng
,
Un
i
ver
si
dad Sa
ntiago de
Cal
i
,
Ca
ll
e 5
# 62
-
00 Cal
i, Col
om
bia
.
Em
a
il
:
hecastro
1@
gm
ail.co
m
1.
INTROD
U
CTION
Atrial
Fib
rill
ation
(AF)
is
t
he
m
os
t
cl
inica
l
l
y
diag
no
se
d
ca
rd
ia
c
a
rrhyt
hmi
a,
both
in
out
patie
nts
a
nd
ho
s
pital
iz
ed
pa
ti
ents.
Its
pr
evalence
a
nd
incidenc
e
incr
ease
with
a
ge
reachi
ng
e
pi
dem
ic
char
act
erist
ic
s
in
seni
or
ci
ti
zens.
T
he
i
nd
i
cat
or
s
of
pro
gr
ess
of
pa
r
oxysm
al
a
tria
l
fibr
il
la
ti
on
(PAF)
to
a
persi
ste
nt
or
per
m
anen
t
on
e
hav
e
not
been
fu
ll
y
identifie
d,
the
refo
re,
de
te
ct
in
g
a
n
AF
in
it
s
ea
rly
fo
rm
is
i
mp
ort
ant
to avoid
the
r
is
ks
of a str
oke,
hear
t
fail
ur
e
and /
or m
or
ta
li
ty
[
1]
.
The
process
of
detect
in
g
a
n
AF
is
pe
rfo
r
m
ed
m
anu
al
ly
by
a
ca
rd
i
ologist
or
el
ect
r
ophysi
ologist
by
inter
pr
et
in
g
the
el
ect
r
oc
ard
i
ogram
(ECG)
r
eco
r
ds
.
This
process
is
highly
dem
and
i
ng
due
to
bot
h
the
num
ber
of
records
t
o
be
analy
zed
an
d
the
fact
that
so
m
et
i
m
es
i
t
i
s
necessa
ry
to
exam
ine
each
beat
ind
ivi
du
al
ly
to
ens
ur
e
the
c
orrect
ide
ntific
at
ion
of
the
card
ia
c
path
ol
og
y.
Th
us,
a
n
autom
at
ed
m
et
ho
d
for
cl
assifi
cat
ion an
d detec
ti
on
would i
m
pr
ove the
d
ia
gnos
t
ic
an
d p
re
ven
ti
on of a
n AF
[2
-
5]
.
To
date,
dif
fere
nt
auth
ors
ha
ve
pro
pose
d
m
et
hods
that
aut
om
at
e
the
detect
ion
of
P
AF
.
So
m
e
autho
r
s
hav
e
reac
hed
a
detect
ion
acc
uracy
betwee
n
70%
an
d
92
%
[6
-
9]
us
in
g
the
char
act
e
risti
cs
of
t
he
P
wav
e
[10]
,
oth
e
rs
pro
pose
the
us
e
of
hea
rt
rate
var
ia
bili
ty
ob
ta
ining
a
n
accuracy
betw
een
81.
2%
an
d
94
.
7%
[
7,
11
-
18]
,
finall
y,
[9
]
pro
po
s
es
the
us
e
of
QR
el
ect
rical
al
te
rn
at
ion
r
eachin
g
an
ac
cur
acy
of
70%
.
Accord
i
ng
to
this
,
the proble
m
o
f
appr
opriat
el
y detec
ti
ng
a
FAP is n
ot f
ully
s
olv
e
d
ye
t, due
to r
es
ults achie
ved
by these m
et
hods
are
not
def
init
ive
an
d
ca
n
st
il
l
be
i
m
pr
ove
d.
S
o,
in
this
pap
e
r
a
ne
w
m
et
ho
dolo
gy
is
pro
posed
t
o
address
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4023
-
4034
4024
this
pro
blem
integrati
ng
m
ultiv
ariat
e
sta
ti
sti
cs
[
19
]
on
t
he
c
har
act
erist
ic
s
of
the
P
wa
ve,
he
art
rate
va
riab
il
ity
and Q
R
elec
tric
al
alt
ern
at
ion.
Th
e
accu
racy
ob
ta
ine
d usi
ng
this ne
w
m
et
hod was
97.4%.
A
P
AF
is
c
ha
r
act
erized
by
irr
egu
la
r
m
ov
em
ent
of
the
le
ft
a
trium
that
pr
ev
ents
the
pro
per
bl
oo
d
fl
ow
into
the
ci
rc
ulatory
syst
em
and
al
so
by
a
re
duct
ion
of
the
ti
m
e
that
the
ve
ntricl
es
val
ves
hav
e
to
receiv
e
an
d
sen
d
blood
to
the
lun
gs.
I
n
an
ECG
signa
l,
these
two
char
act
e
risti
cs
hav
e
an
im
pact
in
the
m
or
phol
og
y
of
t
he
P
-
W
a
ve
an
d
in
t
he
distance
bet
w
een
the
P
-
Wav
e
an
d
t
he
R
-
W
a
ve
s
ee
Fig
ur
e
1
,
t
her
e
fore,
it
is
i
m
po
rtant
to
l
ocate
the
c
har
act
erist
ic
points
P
-
O
ns
et
,
P
-
O
ffset
,
P
W
i
dth
a
nd
P
Hei
ght,
as
well
as
the
P
R
segm
ent,
hear
t
rate va
riabil
it
y
(H
R
V)
a
nd
QR
elec
tric
al
alt
ern
at
ion t
o f
ully
d
escri
b
e a
PAF
.
7.3
7.35
7.4
7.45
7.5
7.55
7.6
7.65
T
i
m
e
(
s
)
-
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
A
m
p
l
i
t
u
d
e
(
m
V
)
P
-
O
n
s
e
t
P
-
O
f
f
s
e
t
P
Wi
d
t
h
P
H
e
i
g
h
t
Q
R
E
l
e
c
t
r
i
c
a
l
a
l
t
e
r
n
a
n
s
P
R
S
e
gm
e
n
t
R
Q
S
Figure
1
.
Cha
r
act
erist
ic
o
f
a
n EC
G
si
gn
al
Diff
e
re
nt
auth
or
s
reli
ed
on
on
e
or
t
wo
c
har
act
erist
ic
s
f
or
the
detect
ion
of
P
AF
as
il
lustrate
d
in
Ta
ble
1
.
In
this
pa
per,
unli
ke
ot
he
r
works
re
po
rted
i
n
the
l
i
te
ratur
e,
i
t
is
pro
pose
d
t
o
us
e
the
inf
or
m
at
ion
of
the
three
char
act
e
risti
cs
to
cover
al
l
the
sy
m
pto
m
s
of
the
PAF
an
d
extract
six
rel
evan
t
featur
e
s
to
im
pro
ve
the
dete
ct
ion
rates.
Se
ns
it
ivit
y,
sp
eci
fici
ty
and
acc
ur
acy
wer
e
us
ed
as
pe
rfor
m
ance
m
et
rics
fo
r
the
evaluati
on
of
t
he
m
et
ho
dolo
gy
pro
po
se
d
he
r
e.
Ta
ble
1
li
sts
so
m
e
works
t
hat
ad
dress
t
he
sam
e
them
e and
the
char
act
e
risti
cs u
se
d.
Table
1
.
C
har
a
ct
erist
ic
s u
se
d for the
d
et
ect
io
n of PA
F
Ref
erence
P
-
W
av
e
HRV
QR Elect
rical
alter
n
an
s
[
1
1
]
✗
✓
✗
[
6
]
✓
✗
✗
[
7
]
✓
✓
✗
[
8
]
✓
✗
✗
[
9
]
✓
✗
✓
[
1
2
]
✗
✓
✗
[
1
6
]
✗
✓
✗
[
1
5
]
✗
✓
✗
[
1
4
]
✗
✓
✗
[
1
3
]
✗
✓
✗
[
1
7
]
✗
✓
✗
[
1
8
]
✗
✓
✗
Prop
o
sed
m
e
th
o
d
✓
✓
✓
2.
RESEA
R
CH MET
HO
D
In
the
propose
d
m
e
tho
dolo
gy
,
a
pr
evi
ou
sl
y
dig
it
iz
ed
ECG
signa
l
is
received
as
in
put.
The
sign
a
l
is
pr
oce
ssed
i
n
f
our
m
ai
n
sta
ges
(Prep
r
oc
essing,
cha
ra
ct
erist
ic
po
int
s
extracti
on,
featur
e
s
extra
ct
ion
,
detect
i
on)
a
nd
it
is d
et
erm
ined wh
et
her o
r n
ot a PAF
ex
ist
s
, as
il
lustrate
d
i
n
Fi
gure
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Met
hodolo
gy
f
or
detect
ion of
Paroxys
ma
l At
rial Fi
br
il
lati
on B
as
e
d o
n P
-
Wave,
H
R
V… (
Hen
r
y Castr
o
)
4025
P
R
EP
R
O
C
ES
S
I
N
G
R
e
s
a
m
p
l
e
Mo
v
i
n
g
a
v
e
r
a
g
e
F
i
l
t
e
r
i
n
g
D
E
T
E
C
T
I
O
N
A
N
N
I
N
P
U
T
S
I
G
N
A
L
E
C
G
R
E
S
U
L
T
S
CH
A
R
A
CT
E
RIS
T
IC
P
O
IN
T
S
D
E
T
E
CT
IO
N
R
-
W
a
v
e
p
e
a
k
P
-
W
a
v
e
p
e
a
k
Q
-
W
a
v
e
p
e
a
k
S
-
W
a
v
e
p
e
a
k
P
-
O
n
s
e
t
P
-
O
f
f
s
e
t
Q
-
O
n
s
e
t
F
E
A
T
U
RE
S
E
X
T
R
A
CT
IO
N
Q
R
E
l
e
c
t
r
i
c
a
l
a
l
t
e
r
na
ns
H
R
V
P
-
Wa
v
e
a
r
e
a
P
R
s
e
g
m
e
n
t
P
-
W
a
v
e
w
i
d
t
h
P
-
W
a
v
e
h
e
i
g
h
t
Figure
2
.
B
loc
k diag
ram
o
f
pro
posed
m
et
ho
d
2.1.
Prepr
oce
ssing
The
first
pa
rt
of
the
al
gorithm
is
pr
ep
ar
ed
to
receive
as
input
the
le
ad
II
of
a
sta
nd
a
rd
12
-
le
ad
EC
G
sign
al
.
D
ue
to
the
var
ia
ble
natu
re
of
the
sam
pling
fr
eq
uen
cy
of
the
ECG
sign
al
,
an
1170
Hz
res
a
m
plin
g
is
per
f
or
m
ed
to
ens
ur
e
a
sta
nd
a
r
d
fr
e
quen
cy
fo
r
the
s
ubseq
uen
t
ap
plica
ti
on
of
a
low
-
pass
fi
nite
i
m
pu
ls
e
respo
ns
e
dig
it
al
filt
er
(F
IR
)
a
nd
t
o
al
lo
w
ea
ch
of
t
he
cha
ra
ct
erist
ic
po
ints
of
t
he
sig
nal
t
o
be
est
ablishe
d
m
or
e
pr
eci
sel
y.
T
his
sta
ge
c
om
pr
ise
s thr
ee
steps: R
esam
pling
, m
ov
in
g
a
ver
a
ge
a
nd f
il
te
rin
g.
2.1.1
.
Res
ampl
ing
The
pro
po
se
d
m
et
ho
dolo
g
y
use
s
six
featu
re
s
f
or
t
he
rec
og
niti
on
of
at
rial
fib
rill
at
ion
th
at
are
ba
se
d
o
n
t
he
m
or
phol
og
y
of
the
si
gnal
.
T
her
e
fore
it
is
ver
y
i
m
po
rta
nt
to
prese
rv
e
t
he
f
reque
ncy
co
ntent
as
well
as
the
sh
ape
of
the
sig
nal
durin
g
proce
ss
ing
.
F
or
this
reason,
it
is
re
qu
i
red
to
be
rep
re
sent
eac
h
beat
by a s
uffici
ent
nu
m
ber
of poi
nts that e
nsure
a good
detect
io
n
a
nd a
good
f
eat
ur
e e
xtracti
on.
The
hei
gh
t
an
d
widt
h
of
the
P
-
W
a
ve
are
f
eat
ur
es
that
ne
ed
a
good
m
or
phologica
l
representat
io
n,
thu
s
,
in
t
his
pap
e
r
we
c
onsidere
d
us
in
g
a
resam
ple
frequ
e
ncy
to
en
su
re
that
the
P
-
Wav
e
has
a
t
le
ast
50
sam
ples.Con
si
der
i
ng
tha
t
there
a
re
docum
ented
case
s
of
patie
nts
with
P
AF
at
the
age
of
22
[20]
,
the
m
axi
m
u
m
hear
t
rate
(
)
c
on
si
der
e
d
i
n
this
m
et
ho
do
l
ogy
was
cal
c
ula
te
d
us
i
ng
t
he
pro
po
se
d
(
1
).
The res
ult o
btained was
16
8 b
eat
s p
er
m
inu
te
(
BPM
)
.
=
(
220
−
)
∗
85%
(1)
Subseque
ntly
, i
t was fo
und
t
ha
t t
he
du
rati
on
of the
P
-
Wav
e
is 4
3
m
s u
sin
g
on
(2
).
=
60
∗
12%
(2)
Finall
y,
a
resa
m
pl
ing
f
re
qu
e
ncy
(
)
of
ap
pro
xim
a
te
ly
11
70
Hz
wa
s
obta
in
ed
th
rou
gh
(3
)
by
relat
ing
t
he
50
sam
ples that rep
rese
nt the
P
-
Wav
e
w
it
h i
ts
durati
on.
=
50
(3)
2.1.2.
M
ov
in
g av
er
age
ECG
sig
nals
norm
al
ly
hav
e
a
baseli
ne
wa
nder
that
m
us
t
be
correct
ed
to
re
fer
e
nce
the
vol
ta
ge
le
vels
of
th
e
sig
nal
to
a
zero
D
C
le
ve
l.
The
m
ov
in
g
aver
a
ge
giv
e
n
in
(4
)
is
com
m
on
ly
use
d
t
o
do
this
w
hic
h
re
qu
i
res
sp
eci
fyi
ng
a
w
indow
siz
e
(
).
This
pa
per
pro
po
s
es
t
o
obta
in
M
base
d
on
t
he
m
os
t
c
ommon
hear
t
rate
va
lue
pr
ese
nt
in
the
sign
al
.
To
fin
d
this
val
ue,
we
obta
in
the
f
re
qu
e
ncy
with
th
e
highest
e
nergy
val
ue
in
t
he
powe
r
sp
ect
ral
den
sit
y
of
the
sig
nal
bounde
d
betwe
en
60
bpm
and
200
bp
m
.
The
refor
e
,
M
is
de
fine
d
as
the
i
nvers
e
of
t
he
f
reque
nc
y
with
the
hi
gh
e
st
energy
value
rou
nd
e
d
to
the
nea
rest
even
value
.
This
is
sho
wn
in
(5
).
Said
f
reque
nc
y
was
obta
ine
d
ap
plyi
ng
t
he
fast
Four
ie
r
transfo
rm
(F
FT)
to
t
he
sign
al
a
uto
c
orre
la
ti
on
giv
e
n
in
(
6)
.
̂
(
)
=
1
∑
(
+
)
2
=
−
2
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4023
-
4034
4026
=
1
a
rg
max
|
(
(
ℓ
)
)
|
∗
60
60
<
<
∗
200
60
(5)
(
ℓ
)
=
∑
(
)
(
−
ℓ
)
−
1
=
0
(6)
2.1.3.
Fil
terin
g
An
ECG
sig
na
l
is
re
pr
ese
nte
d
by
(7)
,
w
here,
(
)
is
the
sig
na
l
ge
ner
at
e
d
by
car
diac
act
ivit
y
wit
h
a
fr
e
quency
ra
ng
e
of
2.5
Hz
and
45
Hz
[
21]
,
r
(n)
is
el
ect
ri
cal
no
ise
a
nd
wh
it
e
no
ise
wi
th
freq
ue
ncies
gr
eat
er
than 4
5 Hz a
nd b
(
n)
is
b
asel
i
ne no
ise
w
it
h f
reque
ncies le
ss
than 2
.5 Hz
[
22]
.
(
)
=
(
)
+
(
)
+
(
)
(7)
No
ise
(
)
was
al
r
eady
rem
ov
ed
us
in
g
m
ov
ing
aver
a
ge
in
the
la
st
ste
p.
I
n
th
is
ste
p,
a
lo
w
-
pass
filt
er
with
a cutoff
freq
ue
ncy of
45
Hz w
as
desig
ne
d
t
o rem
ov
e the
noise
(
)
.
The pre
proces
s
ing
sta
ge
is
sum
m
arized in
A
lgor
it
hm
1
.
Algorithm 1. Preprocessing.
Begin
Load
Signal, Fs
Initialize
S_res, S_norm, S_ac, Sf, M, Sm, S_filter
S_res
Resample (
Signal,
1170)
S_norm
Normalize
(
S_res,
-
1,1)
S_ac
Autocorrel
ation (
S_norm
)
Sf
FFT (
S_ac
)
M
arg max
f
(
|
Sf|
) s.t.
60 bpm < f
< 200 bpm
Sm
MovingAver
age (
S_res
,
M
)
S_filter
lowpass(
Sm
, 45 Hz)
End
2.2.
C
ha
r
ac
te
ri
stic po
in
t de
tection
In
t
he
sec
ond
sta
ge
of
t
he
m
et
hodo
l
og
y
t
he
pea
ks
P
,
Q
,
R
,
S,
P
-
O
ns
et
,
P
-
O
ff
set
a
nd
Q
-
On
set
we
re
fou
nd
on
eac
h
beat
of
t
he
ECG
sig
na
l
.
These
points,
sh
ow
n
in
Fig
ur
e
3
,
will
la
te
r
be
use
d
to
extrac
t
the f
eat
ur
es
of
the b
eat
.
7.3
7.35
7.4
7.45
7.5
7.55
7.6
7.65
T
i
m
e
(
s
)
-
0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
A
m
p
l
i
t
u
d
e
(
m
V
)
P
-
O
n
s
e
t
P
-
O
f
f
s
e
t
Q
-
O
n
s
e
t
P
-
Wa
v
e
p
e
a
k
Q
-
Wa
v
e
p
e
a
k
S
-
Wa
v
e
p
e
a
k
R
-
Wa
v
e
p
e
a
k
Figure
3
.
Cha
r
act
erist
ic
p
oi
nts of a
n
EC
G
si
gn
al
2.2.1
R
-
W
av
e
peak
In
this
ste
p,
a
m
ov
ing
wi
ndow
f
our
tim
es
the
siz
e
M
fo
un
d
in
sect
ion
2.1.2
was
use
d
to
fin
d
the
R
-
Wav
e
pe
ak.
T
he
wind
ow
m
ov
es
thr
oughout
the
E
CG
sig
nal
fi
ndin
g
peak
s
th
at
exceed
0.6
tim
es
the
m
axi
m
u
m
a
m
plit
ud
e
in
the
wind
ow
an
d
ha
ve
a
separ
at
ion
bet
wee
n
the
m
of
at
least
35
3
m
s,
that
is,
the h
ea
rt ra
te
does
not ex
cee
d t
he
m
axi
m
u
m
value
c
hose
n
i
n
this
m
et
ho
dolog
y
of 17
0 bpm
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Met
hodolo
gy
f
or
detect
ion of
Paroxys
ma
l At
rial Fi
br
il
lati
on B
as
e
d o
n P
-
Wave,
H
R
V… (
Hen
r
y Castr
o
)
4027
2.2.2.
P
-
W
ave
peak
The
P
-
Wav
e
pe
ak
was
fou
nd
base
d
on
the
locat
io
n
of
t
wo
co
ns
ec
utiv
e
R
-
W
a
ve
pea
ks
.
As
se
e
n
in
Figure
4
,
The
P
-
Wa
ve
pe
ak
is
the
m
axim
u
m
value
found
withi
n
a
de
fine
d
searc
h
area
betwee
n
70%
an
d
90%
of
t
he dist
ance
betwee
n
t
wo consec
utiv
e R
-
W
a
ve pea
ks
.
7
0
%
2
.
2
2
.
4
2
.
6
2
.
8
3
3
.
2
3
.
4
3
.
6
3
.
8
T
i
m
e
(
s)
-
0
.
2
0
0
.
2
0
.
4
0
.
6
0
.
8
A
m
p
l
i
t
u
d
e
(
m
V
)
R
-
W
a
v
e
p
e
a
k
R
-
W
a
v
e
p
e
a
k
P
-
W
a
ve
pe
a
k
s
e
a
rc
h
a
re
a
P
-
W
a
v
e
p
e
a
k
R
R
(
1
0
0
%
)
10%
Figure
4
.
P
-
Wa
ve peak
searc
h area
2.2.3.
Q
-
Wave
peak
The
Q
-
Wav
e
peak
is
c
ha
racteri
zed
by
a
neg
at
ive
peak
l
ocated
j
ust
befo
re
th
e
app
ea
ra
nce
of
t
he
R
-
Wave,
f
or
this
rea
so
n,
a
de
rivati
ve
was
us
e
d
as
a
searc
h
m
et
hod
f
or
t
his
peak.
Acc
ord
ing
to
the
pro
po
s
ed
(
8)
,
the
valu
e
of
the
de
rivati
ve
is
cal
culat
ed
on
eac
h
sam
ple
on
e
at
a
tim
e
befor
e
the
R
-
Wav
e
.
This
process
i
s
done
unti
l
a
der
i
vative
with
a
negat
i
ve
va
lue
is
f
ound
as
seen
i
n
Fi
gure
5
.
I
n
this
pape
r
we
pro
pose
a
distance
of
ei
gh
t
sam
ples
t
o
be
use
d
in
order
to
a
void
s
m
al
l
var
ia
ti
on
s
t
hat
cou
l
d
hav
e
a n
e
gative
der
i
vative in
the
p
a
th.
(
)
=
(
+
4
)
−
(
−
4
)
8
(8)
0
.
1
9
0
.
2
0
.
2
1
0
.
2
2
0
.
2
3
0
.
2
4
0
.
2
5
0
.
2
6
T
i
m
e
(
s
)
-
0
.
1
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
A
m
p
l
i
t
u
d
e
(
m
V
)
R
-
W
a
ve
pe
a
k
Q
-
W
a
ve
pe
a
k
Figure
5
.
Q
-
Wav
e
peak.
r
ed
dots re
present l
oc
at
ion
s
w
he
re t
he deri
vative
was
e
valuate
d
2.2.4.
S
-
Wave
peak
The
i
den
ti
ficat
ion
of
the
S
-
Wav
e
pea
k
wa
s
carrie
d
out
f
ollow
i
ng
a
pro
cedure
sim
il
ar
to
that
us
e
d
with
the
P
-
Wa
ve
pea
k.
T
his
tim
e,
the
m
ini
m
um
value
was
so
ug
ht
within a
def
ine
d
area b
et
wee
n
0%
a
nd
10%
of the
distance
betwee
n
tw
o
c
on
s
ecuti
ve
R pe
aks
as
is s
how
n
in
Fig
ure
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4023
-
4034
4028
2
.
2
2
.
4
2
.
6
2
.
8
3
3
.
2
3
.
4
3
.
6
3
.
8
T
i
m
e
(
s)
-
0
.
2
0
0
.
2
0
.
4
0
.
6
0
.
8
A
m
p
l
i
t
u
d
e
(
m
V
)
R
-
W
a
v
e
p
e
a
k
R
-
W
a
v
e
p
e
a
k
S
-
W
a
v
e
p
e
a
k
se
a
r
c
h
a
r
e
a
S
-
W
a
v
e
p
e
a
k
R
R
(
1
0
0
%
)
1
0
%
Figure
6
.
S
-
Wa
ve peak
searc
h
area
2.2.5.
P
-
O
nset
The
P
-
On
set
point i
s
def
in
ed a
s the sam
ple w
he
re th
e
P
-
W
ave starts a
nd i
deall
y has a
va
lue of
0
m
V.
This
po
i
nt
wa
s
f
ound
by
e
va
luati
ng
eac
h
on
e
of
the
sa
m
ples
pr
ior
t
o
the
P
-
W
a
ve
peak
one
by
one
un
ti
l
the
c
onditi
on
set
in
the
pro
pos
ed
(9
)
was
m
et
.
Thi
s
eq
uatio
n
c
onside
rs
t
he
fa
ct
that,
in
pract
ic
e,
the
P
-
O
ns
et
ha
s
a
posit
ive
value
higher
than
t
he
baseli
ne,
t
her
e
fore,
a
value
of
0.1
5
ti
m
es
the
a
m
pl
it
ude
of the
P
peak
was use
d
to
f
i
nd it
.
1
5
∑
(
+
)
2
=
−
2
<
0
.
15
(
9)
2.2.6.
P
-
O
ffset
This
c
har
act
er
ist
ic
po
int
is
def
i
ned
a
s
the
sam
ple
wh
er
e
the
P
-
W
a
ve
en
ds
.
T
o
fin
d
this
point
,
we
procee
ded
in
a
sim
il
ar
way
to
the
m
et
h
od
use
d
to
fin
d
the
P
-
O
ns
et
with
the
diff
e
r
ence
th
at
the
s
a
m
ples
evaluate
d
are
locate
d
a
fter
th
e P
-
W
a
ve peak
.
2.2.7.
Q
-
Onse
t
Q
-
Onset
is
the
sam
ple
wh
ere
t
he
Q
-
Wa
ve
be
gin
s
.
T
o
fi
nd
this
ch
aracte
rist
ic
po
i
nt,
a
sim
i
la
r
m
et
ho
d
us
e
d
in
sect
ion
2.
2.3
was
co
nsi
der
e
d.
Eac
h
of
the
sam
ples
befor
e
the
Q
-
Wav
e
pea
k
is
cal
culat
ed
one
by
on
e
on
the
der
i
vative
descr
i
b
ed
by
the
pro
pose
d
(10
)
unti
l
a
posit
ive
va
lue
i
s
f
ound.
In
thi
s
case,
a
sensi
ti
vity
gr
eat
er
t
ha
n
th
at
required
t
o
fin
d
the
Q
-
Wa
ve
pea
k
is
re
quire
d,
th
us
the
distance
was
reduce
f
r
om
eig
ht
to
four sam
ples.
(
)
=
(
+
2
)
−
(
−
2
)
4
(10)
The
c
har
a
ct
eris
ti
c p
oin
t
detect
ion
sta
ge
is
sum
m
arized in
A
lgorit
hm
2
.
Algorithm 2. Characteristic point detection.
Begin
Load
S_filter, M
Initialize
limMin, limMax, tempWin, Vmax, Rpeaks
//Find R
-
Wave peaks
For
i
0..
l
en
gt
h(
S_filter
), +
M
limMin
ma
x
(
i
-
2*
M
)
limMax
mi
n
(
i
+2*
M
)
tempWin
S_
fi
lt
er
[
limMin..limMax
]
Vmax
max(
t
em
pW
in
)
While
True
Find the amplitude and location of the HighestPeak in
tempWin
If
HighestPeak.
amplitude
< 0.6*
Vmax
Break
Else If
Distance be
tween HighestPeak
.location
and any peak in
Rpeaks.location
< 353ms
Break
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Met
hodolo
gy
f
or
detect
ion of
Paroxys
ma
l At
rial Fi
br
il
lati
on B
as
e
d o
n P
-
Wave,
H
R
V… (
Hen
r
y Castr
o
)
4029
Else
Insert HighestPeak in
Rpeaks
Delete HighestPeak from
tempWin
End If
End While
End For
//Find P
-
Wave peaks
Initialize
Peak1, Peak2, RR, Peak, Ppeaks
For
each two consecutive p
eaks in
Rpeaks
Peak1
Fi
rs
t
pe
ak
Peak2
Se
co
nd
p
ea
k
RR
Pe
ak
2.
l
oc
at
io
n
–
Peak1.location
Peak
max(
S_filter
[
Peak1+
0.7*
RR
..
Peak1
+0.9*
RR
])
Insert
Peak
in
Ppeaks
End For
// Find Q
-
Wave peaks
Initialize
dQ, j, Qpeaks //dQ means the derivative at Q
Fo
r
each Peak in
Rpeaks
dQ
1
j
0
While
dQ
> 0
j
j
+1
dQ
(
S_
fi
lt
er
[Peak
.location
–
j
+4]
–
S_filter
[Peak
.location
–
j
–
4])/8
End While
Insert
S_filter
[Peak
.location
–
j
] in
Qpeaks
End For
// Find S
-
Wave peaks
Initialize
Peak1, Peak2, RR, Peak,
Speaks
For
each two consecutive peaks in
Rpeaks
Peak1
Fi
rs
t
pe
ak
Peak2
Se
co
nd
p
ea
k
RR
Pe
ak
2.
l
oc
at
io
n
–
Peak1.location
Peak
min(
S_filter
[
Peak1
..
Peak1
+0.1*
RR
])
Insert
Peak
in
Speaks
End For
// Find POnset
Initialize
j, temp, POnset
For
each Pe
ak in
Ppeaks
temp
Peak
.
lo
ca
ti
on
While
temp
>= 0.15*Peak.
amplitude
j
j
+ 1
temp
me
an
(
S_filter
[Peak
.location
-
j
-
2.. Peak
.location
-
j
+2]
End While
Insert
S_filter
[Peak
–
j
] in
POnset
End For
// Find POffset
Initialize
j, temp, POffset
For
each
Peak in
Ppeaks
temp
Peak
.
lo
ca
ti
on
While
temp
>= 0.15*Peak
.location
j
j
+ 1
temp
me
an
(
S_filter
[Peak
.location
+
j
-
2.. Peak
.location
+
j
+2])
End While
Insert
S_filter
[Peak +
j
] in
POffset
End For
// Find QOnset
Initialize
dQ, j, QOnset
For
each
Peak in
Qpeaks
dQ
1
j
0
While
dQ
< 0
j
j
+1
dQ
(
S_
fi
lt
er
[Peak
.location
–
j
+2]
–
S_filter
[Peak
.location
–
j
–
2])/4
End While
Insert
S_filter
[Peak
.location
–
j
] in
QOnset
End For
End
2.3.
Fe
at
ure
s
ext
r
act
i
on
On
ce
the
ch
ar
act
erist
ic
po
ints
hav
e
bee
n
id
entifi
ed,
the
si
x
featu
res
pr
e
s
ented
on
the
thir
d
sta
ge
of
the
m
et
ho
do
l
ogy
des
cribe
d
i
n
Fi
gu
re
2
a
re
extracte
d
f
or
e
ach
beat
of
t
he
ECG
si
gnal
.
T
he
first
th
ree
f
eat
ur
es
P
-
Wav
e
hei
ght,
P
-
Wav
e
wi
dth
a
nd
PR
se
gm
ent
are
the
m
agn
it
ud
es
of
the
P
-
Wa
ve
peak,
the
dif
f
eren
ce
betwee
n
P
-
Off
set
an
d
P
-
On
s
et
,
an
d
t
he
diff
e
ren
ce
bet
w
een
Q
-
O
ns
et
and
P
-
O
ff
set
resp
ect
ively
.
As
for
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4023
-
4034
4030
the
f
ourt
h
fe
at
ur
e
P
-
Wa
ve
A
rea,
it
is
de
fin
ed
as
the
area
unde
r
t
he
c
urve
betwee
n
P
-
On
set
an
d
P
-
Offset
.
Con
si
der
i
ng
t
ha
t
the
ECG
sig
nal
is
discrete,
a
trapez
oid
al
nu
m
erical
integrati
on
is
use
d
as
an
a
pproxi
m
at
ion
to the i
nteg
ral
of the si
gn
al
be
tween t
hese t
w
o po
i
nts.
T
he
(
11)
desc
ribes
t
his c
onditi
on.
∫
(
)
≈
1
2
∑
(
)
+
(
+
1
)
=
(11)
The
fifth
featu
re
cal
le
d
Hear
t
Ra
te
Var
ia
bili
ty
(H
RV
)
is
t
he
num
ber
of
be
at
s
per
m
inu
te
(bpm
)
that
would
be
gen
e
rated
acc
ordin
g
to
the
distan
c
e
betwee
n
tw
o
co
ns
ec
utive
R
-
W
a
ve
pea
ks.
The
(12
)
des
crib
e
s
this p
ro
ce
ss.
ℎ
(
)
=
60
(
)
−
(
+
1
)
∗
(12)
In
(12)
,
is
the
beat
num
ber
.
(
)
is
the
locat
ion
of
the
R
-
W
a
ve
peak.
is
the
resam
pling
fr
e
qu
e
ncy,
1170
Hz
in
this
case.
The
six
th
and
la
st
fe
at
ur
e
cal
le
d
Q
R
el
e
ct
rical
a
l
te
rn
a
ns
is
def
i
ned
a
s
the
dif
fer
e
nce
betwee
n
the
a
m
pl
it
ud
e
of
th
e
R
-
W
a
ve
pea
k
an
d
t
he
Q
-
Wav
e
pea
k.
T
he
feat
ur
es
e
xt
racti
on
sta
ge
is s
umm
a
rized i
n Alg
or
i
thm
3
.
Algorithm 3. Features extraction.
Begin
Load
S_filter, Rpeaks, Ppeaks, Qpeaks, Speaks, POnset, POffset, QOnset
Initialize
PWaveHeight, PWaveWidth, PRsegment, PWaveArea, HRV, QRelectricalAlternans
For
i
0.
.l
en
gt
h(
Ppeaks
), +1
PWaveHeight
[
i
]
Ppeaks
[
i]
.amplitude
PWaveWidth
[
i
]
POffset
[
i
]
.location
–
POnset
[
i
]
.location
PRsegment
QOnset
[
i
]
.location
–
POffset
[
i
]
.location
For
j
PO
ns
et
[
i
]..
POffset
[
i
]
PWaveArea
(
S_filter
[
j
] +
S_filter
[
j
+1]) / 1170
End For
HRV
ro
un
d(
6
0/
(
Rpeaks
[
i
+1]
.location
–
Rpeaks
[
i
]
.location
) )
QRelectricalAlternans
Rpeaks
[
i
]
.amplitude
–
Qpeaks
[
i
]
.amplitude
End For
End
2.4.
De
tecti
on
Detect
ion
is
the
final
sta
ge
of
the
pro
pose
d
m
et
ho
do
l
ogy.
To
dete
rm
i
ne
the
pr
ese
nc
e
of
a
P
AF
in
the
ECG
,
a
feedfo
rw
a
r
d
neural
netw
ork
with
tw
o
hidden
la
ye
rs
each
with
10
ne
uro
ns
wa
s
us
e
d
as
a
cl
assifi
er
[
23]
.
T
his
ne
ur
al
netw
ork,
whose
trai
ning
was
ca
rr
ie
d
ou
t
us
in
g
60%
of
t
he
in
f
orm
ation
in the dat
a
base
show
n
i
n
Ta
bl
e 2
,
to
i
den
ti
fy
the prese
nce
or not o
f
a
P
AF
i
n
eac
h beat
of
t
he
EC
G.
Table
2
. Fea
tu
r
es
of the
af
db a
nd n
s
r
db d
at
a
ba
ses
Beat
P
-
W
av
e heig
h
t
P
-
W
av
e width
PR s
eg
m
en
t
P
-
W
av
e ar
e
a
HRV
QR elect
rical
alter
n
an
s
1
0
.20
7
1
0
.08
1
0
0
.06
1
0
8
.67
1
3
100
3
.92
3
9
2
0
.20
2
7
0
.15
0
0
0
.05
2
0
1
1
.89
3
6
99
3
.79
8
1
3
0
.21
1
5
0
.08
4
0
0
.06
7
0
9
.32
8
4
99
3
.84
1
1
4
0
.19
5
3
0
.08
7
0
0
.05
2
0
9
.16
8
5
97
3
.62
8
0
5
0
.22
5
8
0
.09
2
0
0
.04
9
0
9
.84
7
8
98
3
.65
2
0
6
0
.23
9
3
0
.16
0
0
0
.04
9
0
1
3
.60
7
1
97
3
.68
5
3
7
0
.21
2
9
0
.10
5
0
0
.05
2
0
9
.79
5
5
100
3
.68
6
5
8
0
.21
1
3
0
.15
2
0
0
.04
4
0
1
2
.69
4
1
99
3
.74
8
8
9
0
.20
3
1
0
.10
7
0
0
.04
6
0
8
.87
2
1
98
3
.70
7
4
10
0
.23
2
6
0
.16
2
0
0
.04
9
0
1
4
.56
7
3
97
3
.67
1
7
···
···
···
···
···
···
···
9
9
0
0
0
0
.32
0
2
0
.22
4
0
3
.98
0
0
3
5
.70
1
9
3
1
.13
4
4
9
9
0
0
1
0
.27
9
6
0
.30
3
0
0
.17
1
0
5
7
.28
3
4
34
0
.58
0
4
9
9
0
0
2
0
.40
0
0
0
.29
9
0
1
.67
9
0
7
9
.24
3
7
5
0
.17
0
4
The det
ect
ion s
ta
ge
is
s
umm
ar
iz
ed
in
Algo
rithm
4
.
Algorithm 4. Detection.
// Training
Begin
Load
afdb, nsrdb
Initialize
PAF_features
[1.
.99002, 1..6],
input
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Met
hodolo
gy
f
or
detect
ion of
Paroxys
ma
l At
rial Fi
br
il
lati
on B
as
e
d o
n P
-
Wave,
H
R
V… (
Hen
r
y Castr
o
)
4031
For
each signal in
afdb
and
nsrdb
Obtain
PWaveHeight, PWaveWi
dth, PRsegment, PWaveArea, HRV,
QRelectricalAlternans
End For
PAF_features
[
PWaveHei
ght, PWaveWidth, PRsegment, PWaveArea, HRV,
QRelectricalAlternans
]
Set
input
as the 60%
of
PAF_features
selected randomly
Train
ANN
using
input
End
2.4.1.
Perf
orm
an
ce
metric
s
Sens
it
ivit
y
(SN),
s
pecifici
ty
(S
P)
a
nd
ac
cur
acy
(
ACC)
,
show
n
in
(
13
-
15
)
res
pect
ively
,
wer
e
cal
culat
ed
sinc
e
these
a
re
t
he
m
os
t
widely
us
e
d
perform
a
nce
m
et
rics
to
asses
s
the
pr
obabili
ty
of
suc
cess
of
a cla
ssifie
r
[24
]
. Tab
le
3
s
ho
ws
th
e re
su
lt
s
of these m
et
rics in
dif
fer
e
nt wo
r
ks
re
ported
i
n
the
li
te
ratur
e
.
=
+
(13)
=
+
(14)
=
+
+
+
+
(15)
Table
3
. C
om
par
iso
n of m
et
h
od
s
for d
et
ect
ion o
f
P
AF.
Ref
erence
SN (
%
)
SP (
%)
ACC
(
%)
[
1
1
]
8
0
.0
9
6
.0
8
8
.0
[
6
]
-
-
8
1
.5
[
7
]
8
2
.1
-
-
[
8
]
9
6
.0
8
8
.0
9
2
.0
[
9
]
-
-
7
0
.0
[
1
2
]
4
6
.5
9
8
.6
8
1
.2
[
1
6
]
9
0
.4
9
5
.2
9
2
.8
[
1
5
]
9
2
.9
9
6
.3
9
4
.6
[
1
4
]
9
1
.5
9
6
.1
9
4
.7
[
1
3
]
9
1
.5 / 93
.3 / 94
.1
9
6
.9 / 92
.8 / 93
.4
-
[
1
7
]
9
4
.5
9
6
.5
-
[
1
8
]
-
-
9
3
.1 / 93
.1 /9
2
.5
Propo
sed
m
ethod
9
6
.7
9
7
.4
9
7
.4
3.
RESU
LT
S
AND A
N
ALYSIS
To
e
valuate
th
e
pro
po
s
ed
m
et
hodo
l
og
y,
th
e
At
rial
Fib
rila
ti
on
(a
f
db)
a
nd
N
or
m
al
Si
nu
s
Rhyt
hm
(n
s
rdb
)
databa
ses
from
Ph
ysi
on
et
[25]
were
us
ed.
Eac
h
on
e
has
ECG
sign
al
sam
ples
fr
om
bo
th
sic
k
a
nd
healt
hy
patie
nts.
Eac
h
si
gnal
is
pr
ocesse
d
us
in
g
t
he
m
eth
od
ology
desc
ribe
d
befor
e
.
Fi
gure
7(a
)
shows
a
n
or
i
gin
al
ECG
sign
al
from
th
e
database,
w
hile
Fi
gu
r
e
7
(
b)
s
hows
the
sign
al
after
pr
eprocessi
ng.
Finall
y
,
Fi
gure
7
(c
)
s
ho
ws
th
e sig
nal
with it
s c
har
act
erist
ic
p
oi
nts
obta
ined
.
The
e
xtracti
on
of
cha
racteri
st
ic
s
was
a
pp
li
e
d
to
ea
ch
of
th
e
recor
ds
in
bo
th
data
bases.
We
ob
ta
ine
d
six
feat
ur
es
of
a
total
of
99
,002
beats
as
il
lustrate
d
in
Table
2
.
T
o
e
ns
ure
t
he
li
ne
ar
in
dep
e
nden
ce
o
f
t
he
featu
res,
t
he
degree
of
c
orrelat
ion
bet
we
en
each
of
the
m
was
determ
i
ned
t
hro
ugh
th
e
correla
ti
on
m
at
rix.
As
it
is
sho
wn
in
Ta
ble
4,
th
e
relat
ion
betw
een
the
six
fea
tures
is
l
ow
in
al
l
cases
exce
pt
bet
ween
P
-
Wa
ve
area,
P
-
Wa
ve
heig
ht
and
P
-
Wav
e
width
w
hich
is
m
od
erate.
These
res
ults
ensu
r
e
that
the
featu
res
obt
ai
ned
thr
ough the
pr
opos
e
d
m
et
ho
dolo
gy are
su
it
a
ble for
the t
rainin
g of a
neura
l netw
ork.
The
P
AF
was
detect
ed
th
r
ough
a
fee
dfo
rward
ne
ur
al
net
work
w
hose
tr
ai
nin
g
data
c
orrespo
nd
e
d
t
o
60%
of
the
in
for
m
ation
pro
vid
e
d
by
the
99,00
2
beats
ob
ta
ine
d
be
f
ore.
The
netw
or
k
was
trai
ne
d
on
10
diff
e
re
nt
occa
sion
s
a
nd
wa
s
ob
ta
ine
d
th
e
SN
,
SP
an
d
ACC
in
each
trai
ning.
The
cal
culat
io
ns
of
the m
axi
m
u
m
, mi
ni
m
u
m
, av
er
age a
nd stan
da
rd d
e
viati
on of
each
perform
a
nce m
et
ric
are
sh
ow
n
in
Ta
ble
5.
A
c
om
par
at
ive
analy
sis
of
the
perform
ance
m
e
tric
s
between
dif
fer
e
nt
cl
assifi
ers
use
d
in
sim
il
ar
works
an
d
th
e
propose
d
m
et
hodo
l
og
y
w
as
done.
T
he
resu
lt
s
are
s
how
n
in
Tabl
e
3
.
The
pro
po
s
ed
m
et
h
odology
ob
ta
ine
d
a
m
i
nim
u
m
SN
of
96
.
4%
that
i
s
higher
tha
n
the
oth
e
rs.
On
the
oth
e
r
hand,
the
SP
reache
d
a
m
axi
m
u
m
value
of
98.1%
being
s
urpa
ssed
only
by
[8
]
,
howe
ver,
the
ACC
excee
ds
al
l
the r
e
ported
works e
ve
n wit
h i
ts
m
ini
m
u
m
v
al
ue
of
96.
3%.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
10
, No
.
4
,
A
ugus
t
2020
:
4023
-
4034
4032
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
S
am
p
le
(
n
)
-
0.5
0
0.5
1
A
m
p
l
it
u
d
e
(
m
V
)
x(
n
)
=
y(
n
)
+
r
(
n
)
+
b
(
n
)
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
S
am
p
l
e
(
n
)
-
0.4
-
0.2
0
0.2
0.4
0.6
0.8
1
A
m
p
l
it
u
d
e
(
m
V
)
y(
n
)
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
S
am
p
le
(
n
)
-
0.4
-
0.2
0
0.2
0.4
0.6
0.8
1
Am
p
l
i
t
u
d
e
(
m
V)
y(
n
)
P
-
Wave
p
e
ak
R
-
Wave
p
e
ak
Q
-
Wave
p
e
ak
S
-
Wave
p
e
ak
(
a
)
(
b
)
(
c
)
Q
S
R
c
om
p
le
x ar
e
a
Figure
1
.
(a
) O
r
iginal si
gnal
fro
m
the d
at
a
ba
se, (b
) p
re
proc
essed
si
gnal
,
(c
) d
et
ect
ion
of c
har
act
erist
ic
points
Table
4
.
C
or
rel
at
ion
m
at
rix
of
the
feature
s
P
-
W
av
e heig
h
t
P
-
W
av
e width
PR seg
m
en
t
P
-
W
av
e ar
e
a
HRV
QR elect
rical
alter
n
an
s
P
-
W
av
e heig
h
t
1
0
.20
7
3
0
.00
9
0
0
.60
8
9
0
.21
9
1
-
0
.03
1
5
P
-
W
av
e width
0
.20
7
3
1
0
.01
8
2
0
.62
4
9
0
.04
9
5
-
0
.12
0
2
PR seg
m
en
t
0
.00
9
0
0
.01
8
2
1
0
.01
2
6
-
0
.01
8
5
-
0
.02
2
2
P
-
W
av
e ar
e
a
0
.60
8
9
0
.62
4
9
0
.01
2
6
1
0
.17
5
9
-
0
.20
8
2
HRV
0
.21
9
1
0
.04
9
5
-
0
.01
8
5
0
.17
5
9
1
-
0
.07
5
1
QR elect
rical
alter
n
an
s
-
0
.03
1
5
-
0
.12
0
2
-
0
.02
2
2
-
0
.20
8
2
-
0
.07
5
1
1
Table
5
.
SN, S
P and ACC
m
e
tric
s of the
propo
s
ed
m
et
ho
d
Metr
ic
Maxi
m
u
m
Mini
m
u
m
Mean
Co
ef
f
icien
t o
f
variatio
n
SN
9
7
.2%
9
6
.4%
9
6
.7%
0
.38
%
SP
9
8
.1%
9
6
.4%
9
7
.4%
0
.49
%
ACC
9
7
.5%
9
6
.3%
9
7
.4%
0
.42
%
4.
CONCL
US
I
O
N
To
date,
dif
fe
ren
t
a
uthors
ha
ve
pr
opos
e
d
m
et
ho
ds
t
hat
autom
at
e
the
detect
ion
of
PA
F
us
in
g
the
cha
racteri
sti
cs
of
t
he
P
wa
ve,
hear
t
rate
va
riabil
it
y
or
Q
R
el
ect
rical
al
t
ern
at
io
n
.
The
a
ccur
acy
reache
d
by
these
m
et
ho
ds
var
y
betwe
en
70%
and
94.7%.
Th
us
,
the
pro
blem
of
ap
propriat
el
y
detect
ing
a
FA
P
is
not
fu
ll
y
so
lve
d
ye
t,
due
to results ac
hi
eved by t
hese
m
et
ho
ds are
no
t def
i
n
it
ive and
can st
il
l be im
pro
ved.
This
pa
per
pro
po
s
es
a
m
et
ho
do
l
og
y
to
ide
nt
ify
the
pr
esen
ce
of
a
PA
F
in
patie
nts
by
an
al
yz
ing
their
ECG.
T
he
m
eth
od
ology
incl
ud
e
s
both
t
he
identific
at
ion
of
t
he
c
har
act
e
risti
c
po
i
nts
of
the
ECG
sig
na
l
and
the
m
e
tho
ds
t
o
ext
ract
six
fea
tures
that
al
lo
w
a
PA
F
t
o
be
detect
ed
thr
ou
gh
a
cl
assifi
er
.
The
res
ults
obta
ined
Evaluation Warning : The document was created with Spire.PDF for Python.