Resea ch A icle
Gene ic Algo i hm Based App oach in A ibu e
Weigh ing o a Medical Da a Se
Ki si Va pa, Ka i Il anen, and Ma i Juhola
Compu e Science, School o In o ma ion Sciences, Uni e si y o Tampe e, 33014 Tampe e, Finland
Co espondence should be add essed o Ma i Juhola; [email p o ec ed] a. i
Recei ed 28 May 2014; Re ised 30 July 2014; Accep ed 6 Augus 2014; Published 3 Sep embe 2014
Academic Edi o : Ma in J. Mu phy
Copy igh © 2014 Ki si Va pa e al. This is an open access a icle dis ibu ed unde he C ea i e Commons A ibu ion License,
which pe mi s un es ic ed use, dis ibu ion, and ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly ci ed.
Gene ic algo i hms ha e been u ilized in many complex op imiza ion and simula ion asks because o hei powe ul sea ch me hod.
In his esea ch we s udied whe he he classi ica ion pe o mance o he a ibu e weigh ed me hods based on he nea es neighbou
sea ch can be imp o ed when using he gene ic algo i hm in he e olu ion o a ibu e weigh ing. The a ibu e weigh s in he
s a ing popula ion we e based on he weigh s se by he applica ion a ea expe s and machine lea ning me hods ins ead o andom
weigh se ing. The gene ic algo i hm imp o ed he o al classi ica ion accu acy and he median ue posi i e a e o he a ibu e
weigh ed k-nea es neighbou me hod using neighbou ’s class-based a ibu e weigh ing. Wi h o he me hods, he changes a e
gene ic algo i hm we e mode a e.
1. In oduc ion
One o he mos commonly used simple classi ica ion me h-
ods is he nea es neighbou (NN) me hod ha classi ies a
new case in o he class o i s nea es neighbou case [1]. The
nea es neighbou me hod is an ins ance-based lea ning
me hod ha sea ches o hemos simila caseo he es case
om he ainingda abysomedis ancemeasu e,usuallywi h
he Euclidean dis ance. A na u al ex ension o NN is he k-
nea es neighbou (k-NN) me hod ha assigns he majo i y
class o he knea es aining cases o he es case [2].
Di e en e inemen s and ex ensions ha e been p oposed o
k-NN in o de o imp o e classi ica ion esul s and o e come
classi ica ion p oblems, o example, dis ance-weigh ing o
neighbou s [2], ex ensions using p ope ies o he da a se [3],
weigh ing o a ibu es [2,4,5], and a ibu e weigh op i-
miza ion wi h gene ic algo i hms (GA) [6–11].
Gene ic algo i hms [12,13]ando he e olu ionalgo-
i hms [14,15]ha ebeenu ilizedin a iouscomplexop i-
miza ion and simula ion p oblems because o hei powe ul
sea ch and op imiza ion capabili ies. A sea ch me hod o a
gene ic algo i hm is a combina ion o di ec ed and s ochas ic
sea ch and he sea ch can be done mul idi ec ionally because
GA main ains a popula ion o po en ial solu ions om he
sea ch space [14]. The basics o he sea ch me hod o GA
unde lie in na u al selec ion and gene ic inhe i ance [12];
indi idualso hepopula iona eusedin he ep oduc iono
new solu ions by means o c osso e and mu a ion. Gene ic
algo i hms ha e been used wi h a ious machine lea ning
me hods o op imize weigh ing p ope ies o he me hod.
Since ou esea ch is based on he nea es neighbou sea ch
applying machine lea ning me hods, we concen a e on
ela ed wo ks whe e GAs ha e been applied only wi h he k-
nea es neighbou me hod. Kelly and Da is [6]combined he
GA wi h a weigh ed k-nea es neighbou (wk-NN) me hod
in he algo i hm called GA-WKNN in o de o ind a single
a ibu eweigh ec o ha wouldimp o e heclassi ica ion
esul s o he wk-NN. A simila kind o app oach was used in
[7] whe e GA was combined wi h he wk-NN and a pa allel
p ocessing en i onmen in o de o op imize classi ica ion o
la ge da a se s. In bo h s udies, a se o eal- alued weigh s o
a ibu es o disc imina e all classes o da a we e achie ed as
a esul a e GA uns.Thes udyo Husseine al.[8]showed
ha GA can be applied success ully in se ing a eal- alued
weigh se o 1-NN classi ie bu he imp o emen o accu-
acy happened a he expense o inc ease in p ocessing ime.
Resul s showed ha GA me hods combining he wk-NN
ou pe o med he basic k-NN [6–8]. Howe e , a single se o
weigh s o allclassesisno always hebes solu ionbecause
a ibu es ha e a di e en e ec on classes [11]. The e o e,
Hindawi Publishing Co po a ion
Jou nal o Compu a ional Medicine
Volume 2014, A icle ID 526801, 11 pages
h p://dx.doi.o g/10.1155/2014/526801
2Jou nal o Compu a ional Medicine
solu ions o sea ching o a weigh o each class and
a ibu e ha e been de eloped. Lee e al. [9]combined he
GA-based a ibu e weigh ing me hod wi h a modi ied k-
NN, hus, o ming an adap i e ea u e weigh ing me hod
A3FW-MNN ha used di e en se s o a ibu e weigh s
o di e en classes. Also, Ma eos-Ga c´
ıa e al. [10]assigned
di e en weigh s o e e y a ibu e depending on each class in
hei e olu iona y algo i hm called Label Dependen Fea u e
Weigh ing (LDFW) algo i hm.
In his esea ch we s udied whe he he classi ica ion pe -
o mance o he a ibu e weigh ed machine lea ning me h-
ods based on he nea es neighbou sea ch can be imp o ed
when using he gene ic algo i hm in he e olu ion o
a ibu e weigh ing based on he expe s and machine lea n-
ing me hods when uns we e made wi h a medical da a se .
This medical da a has been ou es da a in ou p e ious
esea ches [16,17].
2. Ma e ial
In his esea ch an o oneu ological da a se ha ing 951 cases
om se en di e en e igo diseases (classes) (Table 1)was
used. The da a was collec ed o e a decade s a ing om he
1990s in he Depa men o O o hinola yngology a Helsinki
Uni e si y Cen al Hospi al, Finland, whe e expe ienced
specialis s con i med all he diagnoses. The dis ibu ion o
he disease classes is imbalanced; o e one- hi d o he cases
belong o he Meni`
e e’s disease class (36.8%), whe eas he
smalles disease class benign ecu en e igo has only 2.1%
o he cases.
In o al, he da a includes 176 a ibu es conce ning a
pa ien ’s heal h s a us: occu ing symp oms, medical his o y,
and clinical indings in o oneu ologic, audiologic, and imag-
ing es s [18,19]. Clinical es ing has no been done o e e y
pa ien and, he e o e, he e a e se e al es esul s ha ha e
missing alues o he a ibu es. A ibu es wi h low equen-
cies o a ailable alues we e le ou side his esea ch. A e
lea ing ou he a ibu es ha ing o e 35% missing alues, 94
a ibu es emained o be used in his esea ch: 17 quan i a-
i e (in ege o eal alue) and 77 quali a i e a ibu es (o
which 54 we e bina y (yes/no), 20 we e o dinal, and 3 we e
nominal). Gene ic algo i hm uns we e done wi h he da a
including missing a ibu e alues.
3. Gene ic Algo i hm
The basic idea o he gene ic algo i hm is he ollowing: in he
beginning, a popula ion o indi iduals is o med ei he
andomly o wi h in o ma ion abou he applica ion domain.
T adi ionally, a bina y ep esen a ion o he indi iduals has
been used bu in mul idimensional and nume ical p oblems
eal- alued ep esen a ion is nowadays used [14]. In each gen-
e a ion, he indi iduals o he popula ion a e e alua ed wi h
an objec i e e alua ion unc ion, hus, gi ing he indi idual
i s i ness a e. A selec ion me hod is used o ind he i es
indi iduals o a new popula ion. Some indi iduals o he new
popula ion unde go ep oduc ion by means o c osso e and
mu a ion. In he c osso e , he in o ma ion o he indi iduals
Table 1: The equency dis ibu ion o e igo disease classes.
Disease name Abb e ia ion F equency %
1 Acous ic neu inoma ANE 131 13.8
2 Benign posi ional e igo BPV 173 18.2
3Meni
`
e e’s disease MEN 350 36.8
4 Sudden dea ness SUD 47 4.9
5 T auma ic e igo TRA 73 7.7
6 Ves ibula neu i is VNE 157 16.5
7 Benign ecu en e igo BRV 20 2.1
To al 951 100
is swapped in hei co esponding elemen s. Mu a ion al e s
one o mo e elemen s o he indi idual a bi a ily. Eli ism is a
commonly applied su i o selec ion me hod. I keeps he
cu en i es indi idual unchanged in he popula ion so he
high-pe o mance indi iduals a e no los om one gene -
a ion o he nex [20]. The GA can be ended a e a ixed
numbe o i e a ionso i no u he imp o emen is obse ed
a e some numbe o gene a ions.
We u ilized he gene ic algo i hm in he e olu ion o he
a ibu e weigh alues. A pseudocode o he used gene ic
algo i hm is gi en in Pseudocode 1.Apopula ioncon ained
21 indi iduals ha used eal- alued ep esen a ion ins ead o
bina y p esen a ion because he a ibu e weigh alues we e
desc ibed wi h eal- alued numbe s, no jus wi h 0 and 1.
Each indi idual consis ed o se en di e en a ibu e weigh
se s o 94 a ibu es. The indi iduals o he s a ing pop-
ula ion we e based on he weigh s se by he expe s and
machine lea ning me hods. The s a ing popula ion is de ined
mo e accu a ely in Sec ion 3.1.Thegene icalgo i hmused
a oule e-wheel selec ion in pa en selec ion and a uni o m
c osso e wi h disc e e ecombina ion in o sp ing c ea ion.
The c osso e was done in 80.0% p obabili y (𝑝𝑐=0.8)and
he c osso e poin s we e selec ed andomly and indepen-
den ly o each gene (a ield on an indi idual). Mu a ion was
done in 1.0% p obabili y (𝑝𝑚=0.01) o hegeneandi was
done also in a uni o m manne : a andom alue was d awn
om he ange [0,1]whichwasse asanew aluein hecu -
en posi ion. In addi ion, eli ism was used in o de o keep
hebes indi idualwi hin hepopula iondu ing uns.Wedid
no wan o lose he bes pe o ming weigh se du ing he
e olu ion.I henumbe o heindi idualswashighe han
21 in he end o he gene a ion, a su i o selec ion was used.
The indi iduals we e o de ed by hei classi ica ion pe o -
mance and he indi iduals wi h he lowes accu acy we e
disca ded om he popula ion. The gene ic algo i hm ended
a e 20 gene a ions o i he bes classi ica ion accu acy
main ained he same du ing 10 successi e gene a ions. Fu -
he mo e, i all he indi iduals we e he same in he popula-
ion, he e alua ion ended. The pa ame e s used in he GA
uns a e desc ibed in Table 2.
The gene ic algo i hm uns we e done sepa a ely wi h
h ee di e en machine lea ning me hods used in he
popula ion e alua ion: wi h he nea es pa e n me hod
o he o oneu ological expe sys em (ONE), wi h he
Jou nal o Compu a ional Medicine 3
da a 𝐷= [𝐶𝑎𝑠𝑒1
.
.
.
𝐶𝑎𝑠𝑒951]=[[𝑐1,1,...,𝑐1,94]
.
.
.
[𝑐951,1,...,𝑐951,94]]
popula ion 𝑊𝑒𝑖𝑔ℎ𝑡𝑠= [𝑊𝑒𝑖𝑔ℎ𝑡1
.
.
.
𝑊𝑒𝑖𝑔ℎ𝑡21]=[[𝑤1,1,1,...,𝑤1,1,94;...;𝑤1,7,1,...,𝑤1,7,94]
.
.
.
[𝑤21,1,1,...,𝑤21,1,94;...;𝑤21,7,1,...,𝑤21,7,94]]
popula ion size =21
𝑝𝑐=0.8 //C osso e a e
𝑝𝑚=0.01 //Mu a ion a e
di ide da a 𝐷in o 10 equally-sized subse s
o c ound =1 o10do
di ide aining da a D-𝑑𝑐V𝑟𝑜𝑢𝑛𝑑 in o ain (6 subse s) and es (3 subse s) da a
ini ialize me hods wi h ain da a:
cwk-NN and wk-NN OVA: HVDM ini ializa ion
ONE: i ness alue calcula ion o alues o a ibu es
e alua e s a ing popula ion Weigh s wi h es da a and ONE/cwk-NN/wk-NN OVA
while ending e ms o GA a e no ul illed do
//Su i o selec ion: Eli ism
sea ch o he indi idual wi h he highes i ness a e om he popula ion
//Pa en selec ion: Roule e-wheel selec ion wi h i ness-p opo iona e selec ion
o each indi idual in he popula ion do
calcula e indi idual’s i ness p opo iona e a e = indi idual‘s i ness a e/sum o indi iduals’
i ness a es
calcula e indi idual’s cumula i e i ness p opo iona e a e
end o
while n o indi iduals in he ma ing pool is smalle han popula ion size do
gene a e a andom numbe om [0,1]
sea ch o he j h indi idual ha has smalle cumula i e i ness p opo iona e a e han 𝑟
add he j h indi idual in he ma ing pool
end while
//C osso e : Uni o m c osso e wi h disc e e ecombina ion
o each indi idual in he ma ing pool do
gene a e a andom numbe s om [0,1]
i 𝑠is smalle han 𝑝𝑐 hen
add he indi idual in he pa en pool
else
add he indi idual in he new popula ion (o sp ing is a di ec copy o i s pa en )
end i
end o
while wo indi iduals can be aken om he pa en pool do
i wo indi iduals a e exac ly he same hen
add he i s indi idual in o he new popula ion
ake new indi idual om he pa en pool o use in he c osso e
end i
o each disease class weigh se do
selec he c osso e poin s andomly
swap in o ma ion o wo indi iduals in he co esponding c osso e poin s (c ea e child en)
end o
add child en in he new popula ion
end while
//Mu a ion: Uni o m mu a ion
o each indi idual in he new popula ion do
o each gene o indi idual do
gene a e a andom numbe 𝑡 om [0,1]
i 𝑡is smalle han 𝑝𝑚 hen
selec a andom alue V om he ange [0,1]
se he alue as a new alue o he gene
end i
end o
end o
Pseudocode 1: Con inued.
4Jou nal o Compu a ional Medicine
e alua e child en and mu a ed indi iduals in he new popula ion wi h es da a and
ONE/𝑐𝑤𝑘-NN/𝑤𝑘-NN 𝑂𝑉𝐴
add he eli e indi idual wi hou changes in o he new popula ion
//Su i o Selec ion
i n o indi iduals in he new popula ion is la ge han popula ion size hen
so cases descending by hei i ness a e
disca d he las cases in o de o ha e co ec n o indi iduals in he popula ion
else i n o indi iduals in he new popula ion is smalle han popula ion size hen
selec andomly missing cases om he old popula ion
end i
end while
ini ialize me hods wi h aining da a D-𝑑𝑐V𝑟𝑜𝑢𝑛𝑑:
cwk-NN and wk-NN OVA: HVDM ini ializa ion
ONE: i ness alue calcula ion o alues o a ibu es
e alua e he indi idual wi h he highes i ness a e a e GA wi h es ing da a 𝑑𝑐V𝑟𝑜𝑢𝑛𝑑 and
ONE/cwk-NN/wk-NN OVA
end o
Pseudocode 1: Pseudocode o he gene ic algo i hm used in he e olu ion o he a ibu e weigh alues wi h 10- old c oss- alida ion.
Table2:Pa ame e susedwi h hegene icalgo i hm.
Gene ic algo i hm pa ame e s
C osso e a e 0.8
Mu a ion a e 0.01
Popula ion size 21
Gene a ion 20 (and 100 o ONE)
Eli ism Yes (1 indi idual)
a ibu e weigh ed k-nea es neighbou me hod using neigh-
bou ’s class based a ibu e weigh ing (cwk-NN), and wi h
he a ibu e weigh ed k-nea es neighbou me hod using
one- e sus-all he o he (OVA) classi ie s (wk-NN OVA).
The e alua ion me hods a e de ined mo e accu a ely in
Sec ion 3.2. Du ing he gene ic algo i hm uns, o each
indi idual in he popula ion i s i ness a e was calcula ed
wi h he me hod a hand; ha is, he indi idual was e alua ed
agains he me hod. Wi hin he me hods cwk-NN and ONE,
he i ness a e o he indi idual was de ined wi h a o al
classi ica ion accu acy (ACC) and wi hin he wk-NN OVA
wi ha ueposi i e a e(TPR).The o alclassi ica ion
accu acy was used wi h he ONE and he cwk-NN because all
se en disease classes we e classi ied a he same ime whe eas
he wk-NN OVA concen a ed on one disease class (and i s
weigh se ) a a ime. Du ing GA wk-NN OVA uns, i was
mo e impo an o ind he weigh se ha sepa a ed well he
cases o he disease class a hand om he o he s han o
classi y he o he cases also well.
The o al classi ica ion accu acy showed he pe cen age o
all co ec ly classi ied cases wi hin he da a se :
ACC =100𝑡pos
𝑛cases
%,(1)
whe e 𝑡pos was he o alnumbe o casesco ec lyclassi ied
wi hin classes and 𝑛cases was he o al numbe o cases used
in heclassi ica ion.The ueposi i e a eexp essed hepe -
cen age o co ec ly in e ed cases wi hin he class as
TPR =100𝑡pos𝑐
𝑛cases𝑐
%,(2)
whe e 𝑡pos𝑐was he numbe o co ec ly classi ied cases in
class 𝑐and 𝑛cases𝑐was he numbe o all cases in class 𝑐.Wi h
he cwk-NN and wk-NN OVA me hods, he classi ica ion
pe o mance was calcula ed om he se en nea es neigh-
bou me hod (7-NN) esul s and wi h he ONE om he i s
diagnosis sugges ion (ONE1). Howe e , o disease class
benign ecu en e igo (BRV) wi h he wk-NN OVA
me hod i was necessa y o use he TPR o h ee nea es
neighbou s (3-NN) as he i ness a e because o he small size
o he disease class a hand. O he wise he TPR o classi ying
BRV would ha e always been ze o. None heless, i he e
occu ed a si ua ion whe e TPR o 3-NN was ze o wi h all
indi iduals in he s a ing popula ion, a new popula ion was
c ea ed andomly and e alua ed. Random new popula ion
was c ea ed a mos en imes and i he TPR did no change
du ing 10 uns, GA un was ended.
A 10- old c oss- alida ion (CV) [2]wasusedine alua ing
he classi ica ion pe o mance o he gene ic algo i hm. The
da a was andomly di ided in o 10 subse s o app oxima ely
equal size. The di ision was made in a s a i ied manne o
ensu e ha he class dis ibu ion o each subse esembled he
skewed class dis ibu ion o he en i e da a se . In he begin-
ning, one c oss- alida ion pa i ion (10% o he da a) was
le aside o es he pe o mance o he ound bes indi idual
a e gene ic algo i hm un. The nine c oss- alida ion pa i-
ions (90%) we e used du ing he aining p ocess. In o de
o calcula e he i ness a e o each indi idual in he popu-
la ion du ing gene ic algo i hm uns, he aining da a was
u he di ided in o wo pa s: six c oss- alida ion pa s we e
used o aining and h ee c oss- alida ion pa s we e used
o es ing hecu en machinelea ningme hodusedin he
i ness a e calcula ion. Thus, du ing he gene ic algo i hm
Jou nal o Compu a ional Medicine 5
un 60%–30% da a di ision was used. A e he gene ic
algo i hm un, he indi idual ha ing he highes i ness a e
was decla ed as a esul o weigh combina ion and i was hen
es ed wi h he le aside es da a subse . The 10- old c oss-
alida ion was epea ed en imes. In o al, he e we e 100 es
uns pe each e alua ion me hod used in he gene ic algo-
i hm. The same c oss- alida ion di isions we e used wi h all
he e alua ion me hods— ha is, each me hod had he same
aining and es ing se s used du ing he gene ic algo i hm
uns.
3.1. S a ing Popula ion. The s a ing popula ion consis ed
o 21 indi iduals. Each indi idual included se en di e en
a ibu e weigh se s (weigh s o 94 a ibu es), one se o
each disease class. Ins ead o selec ing he s a ing indi iduals
a andom, we decided o use good “guesses” as a s a ing
poin . The e o e, he s a ing indi iduals we e based on he
a ibu e weigh s de ined by he domain expe s ( h ee di e -
en weigh se e sions) and lea n by h ee machine lea n-
ing me hods ( he Sca e me hod [21–23]and heweigh -
ing me hod o he ins ance-based lea ning algo i hm IB4
[24] and i s a ian IB1w). Based on he weigh se s de ined
by he expe s and he machine lea ning me hods, wo
di e en modi ica ions we e c ea ed om weigh se s wi h
50% andom mu a ion, hus ha ing 18 weigh se s in o al. In
addi ion o hese, h ee o ally andom weigh se s we e
c ea edin o hes a ingpopula ion.
The weigh alues we e compu ed wi h he machine
lea ning me hods om he impu ed da a se , ha is, om he
da a se whe e he missing alues o a ibu es we e subs i-
u ed wi h he class-wise modes o he quali a i e and he
class-wise medians o he quan i a i e a ibu es. In o al,
10.1% o he alues o a ibu es we e missing in he da a se .
The impu a ion was done class-wise on he basis o he whole
da a p io o da a di ision in o aining and es ing se s. The
calcula ion o he weigh s was epea ed 10 imes o each CV
aining se in he Sca e , IB4, and IB1w me hods and he
mean weigh s o he 10 epe i ions we e used in he classi ica-
ion o handle he andomness in hese me hods. The weigh s
de ined by he applica ion a ea expe s we e he same o each
CV aining se .
The expe s’ weigh s we e based on h ee di e en com-
bina ions. The i s weigh se included he o iginal a ibu e
weigh s de ined by a g oup o expe ienced o oneu ological
physicians o he decision suppo sys em ONE made in he
1990s [25]. The second and he hi d weigh se s we e de ined
by wodomainspecialis sdu ing heupg adep ocesso he
decision suppo sys em in he 2000s [16].
The Sca e me hod is no mally used o a ibu e impo -
ance e alua ion [21–23]. I calcula es a sca e alue o
an a ibu e ha exp esses he a ibu es’ powe o sepa a e
classes in he da a se . Fo a ibu e weigh ing pu poses, he
sca e alues we e calcula ed o each a ibu e in di e en
class e sus o he classes’ si ua ions. In o de o use he sca e
alues as a ibu e weigh s, i was necessa y o ake in e ses o
sca e alues.
The weigh calcula ion me hod o he IB4 classi ica ion
me hod compu es a ibu e weigh s independen ly o each
class wi h a simple pe o mance eedback algo i hm [24]. The
a ibu e weigh s o IB4 e lec he ela i e ele ancies o he
a ibu es in he class. The di e ence be ween IB4 and i s sim-
ple e sion IB1w is ha IB1w sa es all p ocessed cases in i s
class desc ip ions and does no disca d any cases om he
class desc ip ions du ing uns. Also, he cases wi h poo
classi ica ion eco ds a e kep in class desc ip ions wi h IB1w
whe eas IB4 disca ds hese cases based on hei pas pe o -
mance du ing classi ica ion.
Mo e de ailed desc ip ion o he machine lea ning me h-
ods Sca e , IB4, and IB1w and hei use in weigh o ma ion
will be gi en in he pape [17].
In o de o ha e di e en weigh se s compa able o each
o he du ing he gene ic algo i hm uns, he a ibu e weigh s
we e no malized in o ange [0,1]. The alues o each weigh
se we e di ided by he highes weigh alue occu ing in he
weigh calcula ion me hod a issue.
3.2. E alua ion Me hods
3.2.1. Nea es Pa e n Me hod o ONE. The i s me hod used
wi hin he gene ic algo i hm o e alua e he pe o mance o
he indi iduals in he popula ion was he in e ence mecha-
nism o he o oneu ological decision suppo sys em ONE
[26]. I s in e ence mechanism esembles he nea es neigh-
bou me hodso pa e n ecogni ion.Ins eado sea ching o
he nea es case om he aining se , i sea ches o he mos
i ing class o a new case om i s knowledge base.
In heknowledgebaseo ONE,apa e nisgi en oeach
class ha co esponds o one e igo disease. The pa e n can
be conside ed a p o ile o a disease as i desc ibes i s ela ed
symp oms and signs. Each class in he knowledge base is
desc ibed wi h a se o a ibu es wi h weigh alues exp ess-
ing hei signi icance o he class. In addi ion, a i ness alue
o each a ibu e alue is gi en o desc ibe how i i s he class.
The i ness alues o a ibu e alues we e compu ed on he
basis o he 60% pa o aining da a. Fi ness alues can ha e
aluesbe ween0and100.The i ness alue0means ha he
a ibu e alue does no i he class, whe eas he i ness alue
100 shows ha he alue i s he class pe ec ly. The weigh
alues o a ibu es we e gi en in he popula ion in he GA;
hus, he weigh alues a ied om 0 o 1. The g ea e he
weigh alue is, he mo e impo an he a ibu e is o he
class.
The in e ence mechanism calcula es sco es o he classes
om he weigh and i ness alues o he a ibu es. The sco e
𝑆(𝑐) o a class 𝑐is calcula ed in he ollowing way:
𝑆(𝑐)=∑𝐴(𝑐)
𝑎=1 𝑥(𝑎)𝑤(𝑐,𝑎)𝑓(𝑐,𝑎,𝑗)
∑𝐴(𝑐)
𝑎=1 𝑥(𝑎)𝑤(𝑐,𝑎),(3)
whe e 𝐴(𝑐)is he numbe o he a ibu es associa ed wi h
class 𝑐,𝑥(𝑎)is 1 i he alue o a ibu e 𝑎is known and o he -
wise 0, 𝑤(𝑐,𝑎)is he weigh o he a ibu e 𝑎 o class 𝑐,and
𝑓(𝑐,𝑎,𝑗)is he i ness alue o he alue 𝑗o he a ibu e 𝑎
o class 𝑐[26].In hecaseo quan i a i ea ibu es, he i ness
alues a e in e pola ed by using he a ibu e alues in he
knowledge base as in e pola ion poin s. The i ness alues a e
al e ed o he ange o 0 o 1 du ing he in e ence p ocess.

6Jou nal o Compu a ional Medicine
In addi ion o he sco e, he minimum and maximum sco es
a ecalcula ed o heclassesusing helowes and hehighes
i ness alues o he a ibu es ha ing missing alues.
The classes a e o de ed p ima ily by he sco e and secon-
da ily by he di e ence o he minimum and maximum sco e.
I he classes ha e he same sco e bu one class has a smalle
di e ence be ween he minimum and maximum sco es han
he o he s, he class ha ing he smalles di e ence is placed
highe in o de . I he classes ha e he same sco e and he
minimum and maximum sco e di e ence, hei o de is sel-
ec ed andomly. The class ha ing he highes sco e is e e ed
o as he bes diagnosis sugges ion.
Some e igo diseases esemble each o he by ha ing a
simila kind o symp oms wi h o he diseases du ing some
phase o he disease and, in addi ion, some pa ien s can ac u-
ally ha e wo (o mo e) e igo diseases p esen concu en ly
[27]. The e o e, i is good o check he classi ica ion esul s o
ONEwi hmo e hanonediseasesugges ion.In heend, he
inal diagnos ic choice mus be made by he physician based
on he in o ma ion gi en on all al e na i e diseases [27].
3.2.2. A ibu e Weigh ed k-Nea es Neighbou Me hod Using
Neighbou ’s Class-Based A ibu e Weigh ing. The o he
me hod used in he popula ion e alua ion was he a ibu e
weigh ed k-nea es neighbou me hod using neighbou ’s
class-based a ibu e weigh ing (cwk-NN). The dis ance
measu e o he basic k-nea es neighbou me hod [1]was
expanded o ake he a ibu e weigh ing in o accoun [6].
Lee e al. [9] used a simila class-dependen a ibu e
weigh ing wi h hei modi ied k-nea es neighbou me hod
whe e di e en a ibu e weigh se s o di e en classes we e
de e mined wi h he adap i e-3FW ea u e weigh ing
me hod. Wi h ou cwk-NN he a ibu e weigh ing depends
on he disease class o he neighbou case. Thus, he e ough
o be as many a ibu e weigh s se s a ailable as he e a e
classes.
The dis ance measu e used wi h he cwk-NN was he He -
e ogeneous Value Di e ence Me ic (HVDM) [28] expanded
wi h he a ibu e weigh ing. HVDM was used because i can
handle bo h quali a i e and quan i a i e a ibu es in he da a
se . The a ibu e weigh ed HVDM is de ined as
weigh ed HVDM (𝑥,𝑦)=√𝑚
∑
𝑎=1 𝑤𝑐𝑎𝑑𝑎(𝑥𝑎,𝑦𝑎)2,(4)
whe e 𝑚is henumbe o a ibu es,𝑐is he disease class o
he case 𝑦,𝑤𝑐𝑎 is he weigh o he a ibu e 𝑎in class 𝑐,and
𝑑𝑎(𝑥𝑎,𝑦𝑎)is he dis ance be ween he alues 𝑥𝑎and 𝑦𝑎 o
a ibu e 𝑎.Thedis ance unc ion𝑑𝑎(𝑥𝑎,𝑦𝑎)is de ined as
𝑑𝑎(𝑥𝑎,𝑦𝑎)
={
{
{
{
{1, i 𝑥o 𝑦is unknown
no malized dm𝑎(𝑥𝑎,𝑦𝑎), i 𝑎is quali a i e
no malized di 𝑎(𝑥𝑎,𝑦𝑎), o he wise.(5)
Because HVDM compu es dis ances o quali a i e and o he
a ibu es wi h di e en measu emen anges, i is necessa y
oscale hei esul sin oapp oxima ely hesame angein
o de o gi e each a ibu e a simila in luence on he o e all
dis ance [28]. The no malized dis ance o a quan i a i e
a ibu e is calcula ed wi h (6):
no malized di 𝑎(𝑥𝑎,𝑦𝑎)=󵄨󵄨󵄨󵄨𝑥𝑎−𝑦𝑎󵄨󵄨󵄨󵄨
4𝜎𝑎,(6)
whe e 𝜎𝑎is he s anda d de ia ion o he nume ic alues o
a ibu e 𝑎in he aining se o he cu en classi ie , and o a
nominal a ibu e wi h (7):
no malized dm𝑎(𝑥𝑎,𝑦𝑎)=√𝐶
∑
𝑐=1
󵄨󵄨󵄨󵄨󵄨󵄨󵄨󵄨󵄨󵄨𝑁𝑎,𝑥,𝑐
𝑁𝑎,𝑥 −𝑁𝑎,𝑦,𝑐
𝑁𝑎,𝑦
󵄨󵄨󵄨󵄨󵄨󵄨󵄨󵄨󵄨󵄨
2,(7)
whe e 𝐶is henumbe o ou pu classesin hep oblem
domain (in his case 𝐶=7), 𝑁𝑎,𝑥(𝑦),𝑐 is henumbe o casesin
𝑇 ha ha e a alue 𝑥(o a alue 𝑦) o a ibu e𝑎and he
ou pu class 𝑐,and𝑁𝑎,𝑥(𝑦) is henumbe o casesin𝑇 ha ha e
a alue𝑥(o a alue 𝑦) o a ibu e𝑎[28]. In o he wo ds,
we a e calcula ing he condi ional p obabili ies o ha e he
ou pu class 𝑐when ha ing a ibu e 𝑎wi h he alue 𝑥(o
he alue 𝑦).
Thisapp oachallowedmodi ica ionso all heweigh sa
hesame ime.
3.2.3. A ibu e Weigh ed k-Nea es Neighbou Me hod Using
One- e sus-All Classi ie s. In addi ion o he neighbou ’s
class-based a ibu e weigh ing he a ibu e weigh ed k-
nea es neighbou me hod was es ed wi h one- e sus-all
classi ie s (wk-NN OVA). Wi hin his me hod, he mul iclass
classi ica ion p oblem was con e ed in o mul iple bina y
classi ie s— ha is, he 𝑚class p oblem was di ided in o 𝑚
bina y p oblems [29]. Each bina y OVA classi ie was ained
o sepa a e a class om all he o he classes by ma king he
caseso hisoneclassasmembe casesand hecaseso he
o he classes as nonmembe cases in he aining se .
The a ibu e weigh ed k-NN OVA is an ins ance-based
lea ning me hod ha sea ches o he kmos simila cases
(neighbou s) o a new case om each classi ie sepa a ely.
The e is one classi ie pe each class and each classi ie gi es a
o e o he case being a membe o nonmembe o he class
based on he majo i y class o he kneighbou s. The inal class
o he new case is assigned om a classi ie sugges ing he
case being a membe o a class. The e can occu a si ua ion in
which he new case ge s mo e han one membe o a class o e
(a ie si ua ion) o all o he classi ie s o e o he o he class
( he case o be a nonmembe o all he classes). In a ie si ua-
ion he class o he new case is de e mined by sea ching o
he mos simila membe case om he membe o ing
classi ie s. The case ge s he class o he membe case wi h he
sho es dis ance o i . When all he classi ie s o e o he case
o be a nonmembe , he basic 1-nea es neighbou classi ie
using he whole aining da a con aining he o iginal disease
classes is employed o ind he mos simila case (and i s class)
o he new case.
The dis ance measu e used in he wk-NN OVA was also
he HVDM measu e. The di e ence in he HVDM desc ip-
ion in (4)is ha he𝑐is he class o he classi ie a issue, no
Jou nal o Compu a ional Medicine 7
Table 3: Example e alua ion compu a ion ime o one popula ion (21 indi iduals, one gene a ion) in GA uns wi h di e en compu e s.
Example one popula ion e alua ion ime
Compu e GAONE GAcwk-NN GAwk-NNOVA Speci ica ions
C1 3 min 25 s 48 min 54 s 4 h 57 min 8 s W7 In el Co e i7-3540M 3.00 GHz, 16 GB RAM
C2 — 49 min 53 s 6 h 59 min 16 s I3-530 2.93 GHz, 12 GB RAM
C3 — 3 h 47 min 41 s 9 h 41 min 9 s Q6600 2.4 GHz, 8 GB RAM
C4 — 3 h 12 min 41 s 21 h 14 min 0 s HP P oLian DL580 G7 se e : 4∗In el Xeon X7560 2.26 GHz, 1 TB RAM
C5 — 3 h 4 min 58 s 7 h 22 min 52 s DL785 G5 se e : 8∗AMD Op e on 8360 SE 2.5 GHz, 512 GB RAM
C6 — — 10 h 34 min 55 s In el Co e2 Duo E6750 2.66 GHz, 2 GB RAM
Table 4: The ending ime o he gene ic algo i hm uns wi hin di e en e alua ion me hods.
Gene ic algo i hm GAONE GAcwk-NN GAwk-NNOVA GAONE100
Ended be o e 20 h gene a ion [%] 75.0 18.0 82.9 39.0∗
Ended on 10 h gene a ion [%] 48.0 6.0 54.9 12.0∗
Ended on 20 h gene a ion [%] 25.0 82.0 17.1 61.0∗
∗The ending gene a ions o he GA ONE100 uns was examined be o e 100 h gene a ion, on 50 h gene a ion and on 100 h gene a ion.
he class o he case 𝑦. In addi ion, in (7)wk-NN OVA has wo
ou pu classes (𝐶=2). The da a in he lea ning se 𝑇o he
classi ie is di ided in o he membe and nonmembe classes.
4. Resul s
The esul s o he GA uns wi h ONE and cwk-NN as an
e alua ion me hod we e he a e ages o he 10 imes epea ed
10- old c oss- alida ion whe eas he esul s wi h he wk-NN
OVA we e he a e ages o he 5 imes epea ed 10- old c oss-
alida ion. The 10- old c oss- alida ion was epea ed only i e
imes wi h he GA wk-NN OVA due o i s huge compu a ion
ime. Fo example, he e alua ion o a popula ion (21 indi id-
uals in one gene a ion in a GA un) in one c oss- alida ion
se wi h he GA ONE las ed 3 minu es and 25 seconds, wi h
he GA cwk-NN 48 minu es and 54 seconds, and wi h he
GA wk-NN OVA 4hou s,57minu es,and8secondswhen
unning he GA wi h he compu e C1 (Table 3). Wi h he
o he compu e s, he compu a ion was e en slowe . Thus,
a wo s , he compu a ion ime o one c oss- alida ion se
las ing 20 gene a ions wi h he compu e C1 and GA wk-NN
OVA was o e ou days (o e 12 days wi h C4) assuming
ha wi hin each gene a ion all indi iduals we e e alua ed. In
p ac ice, he numbe o e alua ed indi iduals a ied wi hin
gene a ions due o he c osso e and he mu a ion. No ice
ha compu e s C4 and C5 we e se e s ha ing se e al o he
use s simul aneously and, hus, we had only mino pa o
hei CPU in use. Du ing GA cwk-NN and GA wk-NN OVA
uns, he GA was un pa allel in i e compu e s, hus, ha ing
a bes 11 pa allel GA uns in p ocess. GA ONE was un only
wi h he compu e C1.
The numbe o gene a ions in he GA uns wi h all
used e alua ion me hods a ied om 10 o 20. In o al,
75.0%, 18.0%, and 82.9% o GA uns ended be o e he 20 h
gene a ion due o ha ing he same bes accu acy (GA ONE
and GA cwk-NN)o TPR(GA wk-NN OVA)in10consecu i e
GA uns wi h ONE me hod, cwk-NN, and wk-NN OVA,
espec i ely (Table 4). Wi h he GA wk-NN OVA,all he
GA uns wi h he disease classes sudden dea ness, auma ic
e igo, and benign ecu en e igo ended be o e he 20 h
gene a ion and wi h he o he classes om 58.0% o 88.0%
o he uns. I he numbe o ending gene a ion was 10, his
mean ha he bes ACC o TPR in he popula ion did no
change a all du ing he GA un and, he e o e, he un was
ended. GA cwk-NN ended a e 10 gene a ions only in 6.0%
o he GA uns whe eas GA ONE and GA wk-NN OVA ended
du ing he GA uns a ound hal o uns (in 48.0% and 54.9%
o uns, esp.). In he GA wk-NN OVA uns, his happened
especially wi h disease class auma ic e igo whe e all CV
uns ended a e 10 gene a ions and wi h sudden dea ness
(96.0%) and benign ecu en e igo (94.0%). The o he
disease classes ended du ing he GA wk-NN OVA uns a e
10 gene a ions om 12.0% (acous ic neu inoma) o 34.0%
( es ibula neu i is) o he uns. Mos o he GA cwk-NN uns
las ed 20 gene a ions (82.0%) whe eas only a ou h o he
GA ONE uns and 17.1% o he GA wk-NN OVA uns wen
h ough 20 gene a ions.
Wi hin he GA wk-NN OVA uns o he disease class
benign ecu en e igo occu ed si ua ions whe e he TPRs
in he s a ing popula ion we e ze o ega dless o using he
TPR o 3-NN ins ead in popula ion e alua ion. The TPR o
3-NNwasusedwi hBRVins eado 7-NNbecauseo he
smallsizeo hediseaseclass.TheTPRso s a ingindi iduals
we e ze o in 30 ou o 50 c oss- alida ion se s wi hin he
GA wk-NN OVA un concen a ing on he BRV. In his case,
new s a ing indi iduals we e c ea ed andomly. Random
indi idual c ea ion was epea ed in di e en c oss- alida ion
se s om one o i e and nine imes. The GA wk-NN OVA
un ended i he TPR o s a ing popula ion s ayed ze o en
imes. This happened in 14 (28.0%) c oss- alida ion se s only
wi h he disease class benign ecu en e igo.
In o de o see he e ec o gene ic algo i hm on he popu-
la ion, we examined he wo s and he bes o al classi ica ion
accu acies o indi iduals ( he a ibu e weigh ec o s) in he
beginning and in he end o he gene ic algo i hm un. The
meanwo s and hemeanbes o alaccu aciesand hei
s anda d de ia ions wi h GA uns using ONE and cwk-NN as
8Jou nal o Compu a ional Medicine
Table 5: The mean and i s s anda d de ia ion o he bes and wo s o al classi ica ion accu acies o indi iduals in he s a ing and ending
popula ions occu ing du ing di e en GA uns wi hin 10 imes (in GA wk-NN OVA 5 imes) epea ed 10- old c oss- alida ion.
Me hod Popula ion Bes accu acy [%] Wo s accu acy [%]
Mean S d de . Mean S d de .
GA ONE
(ONE1)
s a 74.0 0.8 49.8 1.6
end 73.8 0.7 61.4 2.8
end 100 73.9 0.9 66.5 2.0
GA cwk-NN
(7-NN)
s a 63.6 1.6 27.9 2.2
end 68.3 1.9 56.2 2.2
GA wk-NN OVA
(7-NN)
s a 79.2 0.5 75.3 0.5
end 78.6 0.9 78.7 0.8
Table 6: The s a ing poin o he gene ic algo i hm using ONE in e ence (GA ONE), he a ibu e weigh ed 𝑘-nea es neighbou me hod
wi h neighbou ’s class-based a ibu e weigh ing (GA cwk-NN) and wi h OVA classi ie s (GA wk-NN OVA) as e alua ion me hod. The ue
posi i e a es (TPR) o se en disease classes and he o al classi ica ion accu acies o he bes indi idual om he s a ing popula ion a e gi en
in pe cen ages (%) om 10 imes ( i e imes wi h GA wk-NN OVA) epea ed 10- old c oss- alida ion.
Disease ANE BPV MEN SUD TRA VNE BRV Median TPR To al accu acy
Cases 131 173 350 47 73 157 20 951
GA ONE
ONE1 63.5 55.0 91.1 67.4 84.0 67.5 37.0 67.4 74.0
ONE12 76.0 84.7 96.6 97.0 96.3 75.4 69.5 84.7 87.5
ONE123 88.1 94.7 98.1 99.6 99.9 84.6 86.0 94.7 93.8
GA cwk-NN
1-NN 47.6 50.2 75.7 28.7 59.0 55.0 10.5 50.2 58.8
3-NN 48.9 52.5 82.2 24.0 58.9 57.0 9.0 52.5 61.9
5-NN 49.0 54.4 85.1 21.1 57.0 56.5 8.5 54.4 62.9
7-NN 48.9 55.0 86.6 19.6 56.3 57.8 5.5 55.0 63.6
9-NN 49.2 56.0 87.8 16.4 53.4 57.5 3.5 53.4 63.7
GA wk-NN OVA
1-NN 70.4 73.5 85.0 67.2 62.7 78.2 19.0 70.4 75.8
3-NN 71.1 75.8 91.8 73.2 61.1 79.4 18.0 73.2 79.2
5-NN 70.7 75.7 92.8 74.5 62.5 79.5 15.0 74.5 79.6
7-NN 69.9 74.7 93.0 73.2 60.0 80.1 15.0 73.2 79.2
9-NN 68.9 73.2 93.2 71.9 58.1 80.5 16.0 71.9 78.7
an e alua ion me hod we e calcula ed om 10 imes epea ed
10- old c oss- alida ion and wi h GA uns using wk-NN OVA
om 5 imes epea ed 10- old c oss- alida ion (Table 5). The
mean bes accu acies s ayed app oxima ely he same wi h he
GA ONE,whe eas hemeanbes accu acyinc eased4.7%
wi h he GA cwk-NN and dec eased 0.6% wi h he GA wk-
NN OVA.Theimp o emen canbeseen om hemeanwo s
classi ica ion accu acies: he wo s accu acy occu ing in he
popula ion inc eased du ing GA uns, especially wi h he GA
cwk-NN (28.3%). Wi h he GA ONE, he mean wo s accu acy
imp o ed 11.6% when using a mos 20 gene a ions and 16.7%
when using a mos 100 gene a ions. Wi h he GA wk-NN
OVA, he imp o emen was mode a e (3.4%) bu one mus
no ice ha i s mean wo s classi ica ion accu acy was al eady
o e 75% in he s a ing popula ion, which was be e han
he mean bes accu acies o he o he me hods.
The mo e de ailed esul s o he GA ONE, heGA cwk-NN,
and he GA wk-NN OVA uns in he beginning and in he end
wi h he bes indi idual occu ing in he popula ion a e gi en
in Tables 6and 7.The ueposi i e a eso hediseaseclasses
a e shown wi h GA ONE o he i s (ONE1), he i s and
second (ONE12), and he i s , second, and hi d (ONE123)
diagnosis sugges ions o ONE and wi h GA cwk-NN and
GA wk-NN OVA o one, h ee, i e, se en, and nine nea es
neighbou s (1-NN–9-NN). Du ing c oss- alida ion uns in
GA, he indi iduals we e e alua ed by he o al classi ica ion
accu acy o he ONE1 wi h he GA ONE and o he 7-NN wi h
he GA cwk-NN andby he ueposi i e a eo he7-NNwi h
he GA wk-NN OVA (excep wi h disease class BRV ha used
he TPR o 3-NN).The ue posi i e a e was used as a i ness
a e wi h he GA wk-NN OVA ins ead o he o al accu acy
because i concen a ed on classi ying one disease class a a
ime whe eas GA ONE and GA cwk-NN classi ied all se en
disease classes a he same ime.
Wi hin 20 gene a ions las ing GA, he bes imp o emen
be ween he s a popula ion and he end popula ion was
yielded wi h he GA cwk-NN ha imp o ed he o al classi i-
ca ion accu acies and he mean ue posi i e a es when using
one o nine nea es neighbou s in he classi ica ion. To al
classi ica ion accu acy o he GA cwk-NN ose a bes 5.1%
Jou nal o Compu a ional Medicine 9
Table 7: The end esul o he gene ic algo i hm using ONE in e ence (GA ONE), he a ibu e weigh ed 𝑘-nea es neighbou me hod wi h
neighbou ’s class-based a ibu e weigh ing (GA cwk-NN) and wi h OVA classi ie s (GA wk-NN OVA)ase alua ionme hodinpopula ion
e alua ion a e a mos 20 gene a ions. The ue posi i e a es (TPR) o se en disease classes and he o al classi ica ion accu acies o he
bes indi idual in he end popula ion a e gi en in pe cen ages (%) om 10 imes ( i e imes wi h GA wk-NN OVA) epea ed 10- old c oss-
alida ion.
Disease ANE BPV MEN SUD TRA VNE BRV Median TPR To al accu acy
Cases 131 173 350 47 73 157 20 951
GA ONE
ONE1 63.5 55.4 90.8 66.2 83.0 68.0 31.5 66.2 73.8
ONE12 77.0 82.7 96.4 93.6 96.2 76.2 62.0 82.7 87.0
ONE123 87.6 92.8 98.0 98.5 99.5 84.4 84.5 92.8 93.2
GA cwk-NN
1-NN 70.2 50.0 68.4 30.6 70.0 60.3 15.0 60.3 61.1
3-NN 70.8 53.9 78.1 27.7 72.5 62.9 14.5 62.9 65.9
5-NN 70.5 56.1 81.5 23.2 71.9 63.8 12.0 63.8 67.4
7-NN 69.5 56.6 84.7 21.1 71.0 63.9 8.5 63.9 68.3
9-NN 69.0 57.5 86.6 18.1 69.7 64.1 6.0 64.1 68.8
GA wk-NN OVA
1-NN 71.5 74.1 84.6 67.2 67.1 77.8 18.0 71.5 76.2
3-NN 71.6 75.3 91.7 74.9 66.8 78.7 16.0 74.9 79.5
5-NN 70.4 73.6 92.2 77.0 63.6 79.2 14.0 73.6 79.1
7-NN 70.4 71.8 92.6 77.0 59.5 79.6 13.0 71.8 78.6
9-NN 70.5 72.4 92.7 74.9 59.7 79.6 13.0 72.4 78.7
(in9-NN)andmedianTPR10.7%(in9-NN).TheGAhada
smalle e ec on he esul s o he GA ONE and he GA wk-
NN OVA. The esul s in he s a popula ion and in he end
popula ion s ayed qui e nea each o he . Small imp o emen
in he mean o al classi ica ion accu acy and he mean TPR
canbeseenwi h heGA wk-NN OVA using one o h ee
nea es neighbou s in he classi ica ion. O he wise, he o al
classi ica ion accu acies dec eased a bi when using he GA
ONE and wi h he GA wk-NN OVA using i e o se en nea es
neighbou s in he classi ica ion.
Changes wi hin he ue posi i e a es o disease classes
compa ed o he s a and end esul s a ied be ween me h-
ods. The GA cwk-NN mainly inc eased he TPRs. Du ing GA
uns,i inc eased hemos heTPRo acous icneu inoma
(22.6% in 1-NN) and auma ic e igo (16.3% in 9-NN).
Meni`
e e’s disease was he only class whe e he TPR dec eased
(a wo s −7.3% in 1-NN) du ing GA cwk-NN uns. Wi h he
GA ONE, he TPRs o classes mainly dec eased. I dec eased
hemos heTPRo benign ecu en e igo(−7.5 % in
ONE12) and sudden dea ness (−3.4% in ONE12). Howe e ,
smallinc easeinTPRcanbeseenwi hacous icneu inoma
(1.0% in ONE12) and wi h es ibula neu i is (0.8% wi h
ONE12). Wi h he GA wk-NN OVA,someTPRsinc easedand
some dec eased. The TPR inc eased he mos wi h auma ic
e igo (5.8% in 3-NN) and sudden dea ness (3.8% in 7-NN)
and dec eased he mos wi h benign ecu en e igo (−3.0%
in 9-NN) and benign posi ional e igo (−2.9% in 7-NN).
Because he compu a ion ime wi h he ONE me hod was
so much as e han wi h he k-nea es neighbou me hods,
he e olu ion o he popula ion wi h GA ONE uns was es ed
also wi h 100 gene a ions in addi ion o he 20 gene a ions.
The ending condi ion was also changed: he GA un ended i
hemaximumaccu acys ayed hesamein50successi e uns
o 100 gene a ions we e un. In o al, 39.0% o he GA ONE100
Table 8: The end esul o he gene ic algo i hm using ONE
in e ence in popula ion e alua ion a e a mos 100 gene a ions.
T ue posi i e a es and he o al classi ica ion accu acies o he bes
indi idual in he end popula iona e gi en in pe cen ages [%] om
10 imes epea ed 10- old c oss- alida ion.
Disease Cases GA ONE 100
ONE1 ONE12 ONE123
ANE 131 67.1 79.9 89.6
BPV 173 56.9 82.0 92.8
MEN 350 89.9 96.1 97.9
SUD 47 61.7 90.9 97.0
TRA 73 80.3 96.4 99.7
VNE 157 69.6 78.7 86.0
BRV 20 23.0 53.5 75.0
Median TPR 67.1 82.0 92.8
To al accu acy 951 73.9 87.3 93.5
uns ended be o e he 100 h gene a ion and wi hin 12.0% o
he uns he ewasnochangein hebes o alclassi ica ion
accu acy du ing 50 gene a ions (Table 4). The classi ica-
ion esul s o he GA ONE100 uns a e gi en in Table 8.
The inc ease o gene a ions om 20 o 100 did no a ec
much he mean o al classi ica ion accu acy no he mean
median TPR. Wi hin disease classes, benign ecu en e igo
su e ed he mos om he gene a ion inc ease: i s ue
posi i e a e dec eased a wo s −16.0% (ONE12) compa ed o
he s a ing popula ion and −9.5% (ONE123) compa ed o he
20 h gene a ion. The bes TPR inc ease was achie ed wi h
acous ic neu inoma: 3.9% om he s a ing popula ion and
3.6% om he 20 h gene a ion.