Академический Документы
Профессиональный Документы
Культура Документы
' '
0 0
0!2G6!136 !0&123106
Pre #radP
0!000' 0!000G
0!13629
0!0001
+
=
=
Figure 16. +ost/rad +.
0!223G&22 0!&&2110&
$
0!001 0!0011
+
=
0!21'36'2
Post #radP
0!0001
21'&
=
=
0his "o$putation to )ind pre-#radient and post-#radient o)
other a,es applied the sa$e $ethod as abo,e!
Polarity
De)inition: 6) the a,e)or$ is positi,e, then "orrespond-
1& 0ar$i4i 6!7 et al.: 7 %ournal o) real pea8 re"o#nition o) ele"tro"ardio#ra$ (e"#) si#nals usin# neural netor8
in# pea8 is set as 1! 6) the a,e)or$ is ne#ati,e, then the
pea8 is set as 0! Positi,e is re)erred to sa$e pattern as the
nor$al a,e)or$, hile the shape oppose the "orrespond-
in# nor$al a,e)or$ is denoted as ne#ati,e! 6n this "ase,
polarity )or ea"h a,e)or$ assi#ned as belo
P J 1
@;( J 1
0 J 1
Data then been arran#ed usin# -i"roso) *A"el and a)ter
nor$ali4ation, the data then be )ed as an inputs to neural
netor8 pattern re"o#nition syste$! 0his nor$ali4ation is
i$portant to ensure the ,alues o) the data are in beteen the
"ertain ran#e! 0hus, it is easier to train the neural netor8s
e))i"iently as ell as easier )or the netor8 syste$ to learn!
0he netor8 ill learn input "hara"teristi"s and attributes o)
the "orrespondin# a,e)or$s! 6) )or eAa$ple the "hara"te-
risti"s o) P a,e are #i,en to the netor8 )or trainin#, the
tar#et o) P a,e ill be put as 1 and the other is 0! Ln"e the
P a,e is "orre"tly dete"ted, thus the real pea8 o) P is easily
obtained by the hi#hest pea8 o) that a,e)or$! 0his is sa$e
#oes to other a,e)or$s!
&# 'esult
Figure 1!. Network (erformance 12SE ,s num!er of e(oc"s3.
/etor8Cs "on)i#uration has been detail identi)ied ith '0
hidden neurons, 2 layers )eed )orard, trained ith s"aled
"on%u#ate #radient ba"8 propa#ation (trains"#) ith si#$oid
type o) hidden nodes and output neurons! &9 sa$ples ha,e
been used ith 2 input )eatures, des"ribed as a$plitude,
duration, pre-#radient, post-#radient and polarity! 0he per-
)or$an"e o) the netor8 is e,aluated in $ean s>uared error
and "on)usion $atri"es! /etor8 ill be retrained to a"hie,e
as per desired tar#et!
(i$ulation result shoin# that trainin# and testin#
per)or$an"e 8eep de"reasin#! 6n other ords, it pro,ed that
the netor8 is learnin#! Durin# testin#, it tried to a"hie,ed
the tar#et approAi$ately as hat has been trained! 7 part
)ro$ that, it is seen that best ,alidation per)or$an"e, hi"h
is the s$allest di))eren"e o) desired tar#et ith the netor8
output is at 1!1103A10
-1
a)ter #oin# throu#h 2G iterations! 7t
this point, the netor8 possessed the ability to #enerali4e
,ery ell a)ter per)or$an"e be"o$es $ini$i4ed to the #oal!
Figure 1#. Network trainin state.
;esult abo,e presented trainin# state o) the netor8! 0he
plot depi"ted the trainin# state o) the netor8 )ro$ a trainin#
re"ord! 0he $ini$u$ #radient rea"hed at epo"hs 33 at a
,alue o) G!GG11G A 10
-2
! =alidation "he"8s are at 6 also o"-
"urred durin# 33th iteration! 0he netor8 has been ell
trained and learnin# to "lassi)y respe"ti,e inputs to the "or-
respondin# tar#et! 6t stops hen $ini$u$ #radient has been
rea"hed to a,oid netor8 )ro$ o,er)ittin# and be"o$es
un"ontrolled! 6n neural netor8 "o$putin#, ,alidation is
used to ensure the netor8 able to #enerali4e respe"ti,e
inputs $appin# to the "orrespondin# tar#et! =alidation ill
halt the trainin# hen #enerali4ation stops i$pro,in#!
Figure 1*. Confusion (lot matri0.
7$eri"an :ournal o) /etor8s and .o$$uni"ations 2013, 2(1): 9-16 12
7 "on)usion $atriA depi"ted "learly the a"tual ,ersus
predi"ted "lass ,alues! 6t is also presented the "lasses hi"h
are "orre"tly "lassi)ied and $is"lassi)ied! 0hus, it is enablin#
us to see ho ell the $odel predi"ts the out"o$es E9F!
0here are three a,e)or$s that need to be re"o#ni4ed by
neural netor8s, hi"h is a,e)or$ o) P, @;( and 0! 0he
real pea8 is dete"ted by the hi#hest pea8 o) the a,e)or$!
0his is di))erent )or @;(! Ln"e the @;( a,e)or$ is "or-
re"tly identi)ied by neural netor8 syste$, thus it is 8non
that the hi#hest pea8 should be the real pea8 o) ;, the pea8
be)ore is @ and the pea8 a)ter the ; pea8 should be ( pea8!
32 dataset has been used )or trainin# purposes, G )or ,alida-
tion and G )or testin# phase! ?ased on the result abo,e, it
shoed that durin# trainin# session, there are 30 set o) data
are "orre"tly "lassi)ied and only 2 data are $is"lassi)ied! 6n
,alidation, 2 data are "orre"tly "lassi)ied out o) G hile
durin# testin# phase, 2 data too are "orre"tly "lassi)ied a"-
"ordin#ly hile ti"e o) $is"lassi)i"ation happened! (e,-
eral )a"tors "ontribute to netor8Cs i$pro,e$ent ill be
in"luded in a dis"ussion part! 0he per"enta#e a""ura"y o)
ea"h pro"ess (trainin#, ,alidation and testin#) is also "an be
seen in the "on)usion plot $atriA abo,e! L,erall netor8
per)or$an"e shoin# that &0 data are "orre"tly "lassi)ied as
desired hile 9 data are $is"lassi)ied! 6t per)or$ed #ood
"lassi)i"ation ith total a""ura"y o) '1!6B!
Figure 2,. .ecei,er 4(eratin C"aracteristic 1.4C3.
7bo,e is shon the out"o$e o) re"ei,er operatin# "ha-
ra"teristi" (;L.) )or this pro%e"t! 0he ;L. "ur,e a"ts as
)unda$ental tools )or dia#nosti" e,aluation )or positi,e test
hi"h plotted true positi,e rate ,s )alse positi,e rate E9F, E10F!
6t re)le"ts to a sensiti,ity and spe"i)i"ity o) the netor8! 0rue
positi,e rate re)erred to real pea8 si#nals "orre"tly identi)ied
as its real pea8 hile )alse positi,e rate re)erred to non real
pea8 o) the si#nals in"orre"tly identi)ied as real pea8!
(# )is!ussin
0he result )or this pro%e"t "annot a"hie,e 100B or ,ery
hi#h per"enta#e o) a""ura"y sin"e s$all dataset as used!
0hus, so$e re"o$$endations "ould be su##ested to i$-
pro,e netor8Cs per)or$an"e in order to obtain approA-
i$ately a""urate outputs! 0here are in"ludin# in"rease the
nu$bers o) hidden neurons and retrain the netor8 se,eral
ti$es! Further$ore, use lar#er data set so that the netor8
ill learn $ore and eApose to enou#h trainin#s! Do try a
di))erent trainin# al#orith$ as ell as ad%ust the initial
ei#hs and biases to ne ,alues! 0hen, train the netor8s
a#ain )or se,eral ti$es until it rea"hes the desired tar#et! 6n
addition, it is re"o$$ended to per)or$ auto$ated )eatures
eAtra"tion usin# respe"ti,e so)tare su"h as Hab=6*9 or
et" )or a""ura"y and >ui"8 eAtra"tion purposes!
*# +n!lusin
/eural netor8 pattern re"o#nition is suitable so)tare
ith hi#h ability to "lassi)y "ertain input patterns into a
"orrespondin# output tar#et ith o,erall a""ura"y o) '1!6B!
0he trainin# a""ura"y is '2!GB, ,alidation a""ura"y is G1!&B
hile the testin# a""ura"y is G1!&B too! 6t "an be "on"luded
that the real pea8 o) *.< si#nals "an be identi)ied by
trainin# the netor8 a""ordin#ly!
A!,n"ledgement
0he appre"iation #oes to absolutely $y $ain super,isor,
Dr! (yed (ahal /a4li 7lhady )or pro,ides endless help in-
"ludin# $oti,ation and #uidan"e and also not )or#et to
"o-super,isor as ell as )ield super,isor )or so$e supported
ideas dire"tly or indire"tly! -y deepest #ratitude then eA-
tends to -inistry o) ("ien"e, 0e"hnolo#y and 6nno,ation,
#o,ern$ent o) -alaysia )or illin# to sponsor $e
throu#hout $y study, ithout the s"holarship, this pro%e"t
$i#ht )a"e di))i"ulty and "ould not be a""o$plished in a
"o$)ortable $anner! 0han8 you ,ery $u"h to those in-
,ol,ed!
'e%eren!es
E1F 7 ebsite on
http://biolo#y!about!"o$/library/or#ans/heart/blatrionode!ht
$, i$a#e "ourtesy o) .arolina ?iolo#i"al (upply / 7""ess
*A"ellen"e!
E2F .hun#, -!3, and ;i"h, -!9! 6ntrodu"tion to the "ardio,as-
"ular syste$! 7l"ohol 1ealth and ;esear"h 9orld1&
(&):269-2G6, 1990!
E3F 6n)or$ation,
http://biolo#y!about!"o$/od/hu$ananato$ybiolo#y/ss/heart
anato$y2!ht$
E&F -a4har ?!0ayel, -oha$ed *!*l-?ouridy, *.< i$a#es
"lassi)i"ation usin# )eatures eAtra"tion based on a,elet
trans)or$ation and neural netor8, 76-H 06 6nternational
.on)eren"e, 13-12 :une 2006, (har$ *l (hei8h, *#ypt,
102-10G!
E2F ;a%esh <hon#ade, =isha8ar$a, Dr! ?abasaheb, 7 brie)
16 0ar$i4i 6!7 et al.: 7 %ournal o) real pea8 re"o#nition o) ele"tro"ardio#ra$ (e"#) si#nals usin# neural netor8
Per)or$an"e *,aluation o) *.< Feature *Atra"tion 0e"h-
ni>ues )or 7rti)i"ial /eural /etor8 ?ased .lassi)i"ation!
E6F .hristos (ter#iou and Di$itrios (i#anos, 7 ;eport on /eural
/etor8s!
EGF (i$on 1ay8in, 7 boo8 o) /eural /etor8s 7 .o$prehen-
si,e Foundation, (e"ond *dition, 1999!
E'F 6nternational %ournals on 9(*7( 0ransa"tions on (yste$s,
6ssue 1, =olu$e &, :anuary 2002, 6((/ 1109-2GGG, paper
titled P, @, ;, (, and 0 pea8s re"o#nition o) *.< usin#
-;?F ith sele"ted )eatures, 13G!
E9F -!P!( .hala, Depart$ent o) *le"tri"al *n#ineerin#, 6ndian
6nstitute o) 0e"hnolo#y, ;oor8ee, 2&G66G 6ndia, Para$ete-
ri4ation and ;-pea8 error esti$ations o) *.< (i#nals usin#
6ndependent .o$ponent 7nalysis, .o$putational and -a-
the$ati"al -ethods in -edi"ine, =ol! ', /o! &, De"e$ber
200G, 263-2'2!
E10F *Aer"ise on 7rti)i"ial /eural /etor8s, 6n)or$ation on
rad!ihu!edu!#r/)ilead$in/labs)iles/de"isionMsupportMsyste$s
/lessons/neuralMnets///s!pd)!