Con en s lis s a ailable a ScienceDi ec
In o ma ics in Medicine Unlocked
jou nal homepage: www.else ie .com/loca e/imu
Machine lea ning o diffe en ia e diseased ca diomyocy es om heal hy
con ol cells
Ma i Juhola
a,∗
, Hen y Jou sijoki
a
, Ki si Pen inen
b
, Ka iina Aal o-Se älä
b,c
a
Facul y o In o ma ion Technology and Communica ion Sciences, Tampe e Uni e si y, Finland
b
Facul y o Medicine and Heal h Technology, Tampe e Uni e si y, Finland
c
Hea Cen e , Tampe e Uni e si y Hospi al, 33520, Tampe e, Finland
ARTICLE INFO
Keywo ds:
Calcium ansien p ofiles
Gene ic ca diac diseases
Machine lea ning
Diffe en ia ion o he diseased om con ols
ABSTRACT
Human induced plu ipo en s em cell-de i ed ca diomyocy es (iPSC-CMs) ha e been shown o be use ul o
imp o e echniques ha a e de eloped o he s udy o ca diac disease. Abno mali ies in Ca
2+
ansien s a e
commonly p esen in iPSC-CMs de i ed om indi iduals wi h a ca diac disease. We p e iously obse ed ha
Ca
2+
ansien signals o heal hy CMs can be dis inguished om ansien s o CMs de i ed om indi iduals
ha ing diffe en gene ic ca diac diseases. Machine lea ning was used o dis inguish diffe en diseases om each
o he as well as om con ols. We wan ed u he o in es iga e whe he we a e able o sepa a e iPSC-CM Ca
2+
signals o any gene ic ca diac disease as one g oup om hose o heal hy indi iduals by u ilizing machine
lea ning me hods. A o al numbe o 593 CM ansien signals om heal hy indi iduals and om pa ien s we e
analyzed. We ob ained a bes classifica ion accu acy o 87% be ween he disease g oup and con ols. This finding
p o ides e idence ha machine lea ning me hods a e efficien o iden i ying iPSC-CMs de i ed om indi iduals
wi h a disease pheno ype, and ha iPSC-CMs may be use ul o iden i y indi iduals a isk o a ca diac e en .
1. In oduc ion
Gene ic ca diac diseases p esen a wide ange o symp oms, anging
om comple ely asymp oma ic o se e e a hy hmias, and e en sudden
ca diac dea h [1,2]. Addi ionally, mos i no all o hese diseases ha e
an inc eased isk o a hy hmia, in addi ion o s uc u al o o he
ca diac abno mali y, e.g., in a ious ca diomyopa hies. I he mu a ion
causing disease is known in he amily, i is easy o ocus on he mu-
a ion ca ie s o ollow-up and p ima y p e en ion o po en ial a -
hy hmias. Howe e , his is o en p oblema ic when he disease is
p esen ed by sudden dea h in he amily, bu no mu a ions a e ound. In
such si ua ions, induced plu ipo en s em cell (iPSC) -de i ed ca dio-
myocy es may p o ide a use ul al e na i e o p edic a hy hmic isk,
and o iden i y hose amily membe s wi h inc eased isk o clinical
symp oms including a hy hmias.
On a cellula le el, ca diac unc ionali y can be s udied wi h he
help o CMs diffe en ia ed om human plu ipo en s em cells [3,4]. The
induced plu ipo en s em cell (iPSC) echnology offe s a way o ep o-
g am diffe en ia ed cells back o he plu ipo en s a e –and, he e o e,
i is a use ul ool o s udying he pa hophysiology o a ious diso de s
and d ug esponses in human cells [5]. Addi ionally, cellula diffe -
en ia ion and ma u a ion can be s udied wi h hese cells [6].
Thus a , iPSC-CMs ha e success ully been used o model gene ic
ca diac diseases such as ca echolamine gic polymo phic en icula
achyca dia (CPVT) [3,4,7–12], long QT synd ome (LQT) [13–16] and
hype ophic ca diomyopa hy (HCM) [17–19]. iPSC-de i ed CMs ha e
e ealed conside able abno mali ies and di e si y in in acellula Ca
2+
cycling ea u es compa ed o heal hy con ol CMs. Ca
2+
cycling plays
an impo an ole in ca diac unc ionali y by linking elec ical ac i a-
ion and con ac ion, and he cha ac e iza ion o Ca
2+
cycling is i al in
o de o acili a e in es iga ions o ca diac diso de s and dys unc ions,
as well as o s udy disease managemen wi h diffe en compounds.
He e o o e, machine lea ning has a ely been applied o he da a
associa ed wi h induced plu ipo en s em cell-de i ed ca diomyocy es.
Machine lea ning has howe e been applied o he mechanis ic ac ion
o ca dioac i e d ugs [20]. We demons a ed in p e ious a icles ha
he sepa a ion o calcium ansien signals o abno mally and no mally
g own ca diomyocy es can be accu a ely done wi h machine lea ning
[21,22]. In hese pape s, abno mali y is defined as de o med Ca
2+
peak
o ms a ying in ampli udes (sizes) and du a ions o Ca
2+
ansien s
and no mali y as ha monious ansien s o app oxima ely he same size
and o m h oughou en i e calcium ansien signals. To he bes o ou
knowledge, ou ecen s udy [23] was he fi s one in which diffe en
gene ic ca diac diseases we e sepa a ed acco ding o hei calcium
h ps://doi.o g/10.1016/j.imu.2019.01.006
Recei ed 17 Decembe 2018; Recei ed in e ised o m 16 Janua y 2019; Accep ed 31 Janua y 2019
∗
Co esponding au ho .
E-mail add ess: Ma i.Juhola@ uni.fi(M. Juhola).
In o ma ics in Medicine Unlocked 14 (2019) 15–22
A ailable online 02 Feb ua y 2019
2352-9148/ © 2019 Published by Else ie L d. This is an open access a icle unde he CC BY-NC-ND license
(h p://c ea i ecommons.o g/licenses/BY-NC-ND/4.0/).
T
ansien signals by using classifica ion pe o med wi h machine
lea ning me hods.
In he p esen esea ch, isually no mal and abno mal Ca
2+
an-
sien signals and peak a iables o h ee gene ic ca diac diseases and
heal hy con ol CMs we e used. Disease-specific CMs we e gene a ed
om pa ien s suffe ing om CPVT, an exe cise-induced malignan a -
hy hmogenic diso de [3,8], LQT ype 1, an elec ic diso de o he
hea ha p edisposes pa ien s o a hy hmias, and HCM, a diso de
ha affec s he s uc u e o hea muscle issue leading o a hy hmias
and p og essi e hea ailu e [19]. All o he ansien s o diseased cells
we e pooled and compa ed o all o he ansien s ob ained om con ol
cells, and diffe en machine lea ning algo i hms we e designed and
used o analyze his echnology o au oma ically dis inguish he g oups.
2. Ma e ials
The cu en s udy was app o ed by he E hics Commi ee o
Pi kanmaa Hospi al Dis ic in es ablishing, cul u ing and diffe -
en ia ing human iPSC lines (R08070). The pa ien -specific iPSC lines
we e es ablished and cha ac e ized as desc ibed ea lie , as well as he
CM diffe en ia ion and dissocia ion o bea ing a eas [22]. The s udied
cell lines included six CPVT lines gene a ed om CPVT pa ien s ca -
ying ca diac yanodine ecep o (RyR2) mu a ions, ou HCM cell lines
gene a ed om HCM pa ien s ca ying ei he α- opomyosin (TPM1) o
myosin-binding p o ein C (MYBPC3) mu a ions, wo LQT ype 1 cell
lines gene a ed om pa ien s ca ying po assium ol age-ga ed channel
sub amily Q membe 1 (KCNQ1) mu a ions, and one cell line gene a ed
om a heal hy con ol indi idual. Thus, he e we e 13 subjec s al o-
ge he .
Ca
2+
imaging was conduc ed in spon aneously bea ing, 4 μM Fu a-2
AM (In i ogen, Molecula P obes) o 4 μM Fluo-4 AM (Li e
Technologies L d) loaded dissocia ed CMs as desc ibed ea lie [3].
Du ing he measu emen s, CMs we e pe used wi h 37 °C HEPES based
pe usa e consis ing o (in mM) 137 NaCl, 5 KCl, 0.44 KH2PO4, 20
HEPES, 4.2 NaHCO3, 5 D-glucose, 2 CaCl2, 1.2 MgCl2, and 1 Na-py -
u a e ( he pH was adjus ed o 7.4 wi h NaOH). Ca
2+
measu emen s
we e conduc ed on an in e ed IX70 mic oscope wi h a UApo/340 x20
ai objec i e (bo h Olympus Co po a ion, Hambu g, Ge many) o wi h
Axio Obse e .A1 mic oscope wi h Objec i e Flua 20x/0.75 M27 (bo h
Ca l Zeiss Mic oscopy GmbH, Gö ingen, Ge many). Images we e aken
wi h an ANDOR iXon 885 CCD came a (Ando Technology, Bel as ,
No he n I eland) and synch onized wi h a Polych ome V ligh sou ce
by a eal ime DSP con ol uni o wi h Lambda DG-4 Plus (Su e In-
s umen , Cali o nia, USA) wa eleng h swi che and TILL isION, Li e
Acquisi ion (TILL Pho onics, Munich, Ge many) o ZEN 2 blue edi ion
so wa e (Ca l Zeiss Mic oscopy GmbH, Gö ingen, Ge many) so wa e.
Fo Ca
2+
analysis, egions o in e es we e selec ed o spon aneously
bea ing cells and backg ound noise was sub ac ed be o e u he p o-
cessing. Each Ca
2+
signal co esponded o a eco ding om one cell.
3. Da a compu ed om Ca
2+
ansien signals
Human induced plu ipo en s em cell-de i ed CMs we e he da a
sou ce om which cycling Ca
2+
ansien signals we e ob ained. Da a
used in he compu a ion we e based on he peaks o Ca
2+
ansien
signals. Cycling peaks we e ecognized, and da a a iables o ea u es
we e ex ac ed om e e y peak. P e iously Ca
2+
ansien signals we e
ca ego ized using ou ecogni ion algo i hm [22], which classified hem
in o ei he no mal ype o abno mal signal ype on he basis o no mal
o abno mal peaks o he signals, and whe e we obse ed ha i was
possible o sepa a e no mal om abno mal signals up o he accu acy o
app oxima ely 90% when compa ed o a human expe 's classifica ion
decisions.
Fig. 1 p esen s 10 s segmen s o ou signals as examples. The signals
we e sho , hei du a ions being om 7.7 s o 46.5 s and 19.0 s on
a e age. An en i e signal was de e mined o be abno mal i e en a
single peak was obse ed as abno mal. In he cu en esea ch, we did
no diffe no mal and abno mal signals om each o he , bu ad anced
o sepa a e Ca
2+
ansien signals o diseased induced plu ipo en s em
cell-de i ed CMs om hose o con ol subjec s. I was in e es ing o
s udy whe he disease ansien signals can be diffe en ia ed om hose
o con ols, al hough bo h classes con ained bo h no mal and abno mal
signals. The numbe o he abno mal con ol signals was only 12.6% o
all con ol signals. The disease ansien signals o igina ed om he
g oup o he h ee abo e-men ioned diseases: LQT1, HCM and CPVT.
These we e used join ly as he disease class. The o he class was o med
by he signals o he con ols (wild ype, WT).
The da a used comp ised 394 disease ansien signals and 199
con ol ansien signals. These con ained, espec i ely, 179 no mal and
215 abno mal signals, and 174 no mal and 25 abno mal signals. I is
no iceable ha he e we e only ela i ely ew abno mal signals in
con ols, since hese we e a mo e in equen compa ed o hose o
disease CMs.
Sampling o he ansien signals con ained h ee diffe en e-
quencies, because he da a we e eco ded a diffe en imes and he
sampling equency was inc eased in he mean ime. The app oxima e
sampling equencies we e 8 Hz, 11 Hz and 23 Hz. In his o de , 35%,
26% and 39% o he diseased signals we e eco ded, and, co espond-
ingly, 5%, 15% and 80% o he con ol signals.
In o de o de ec indi idual peaks o a ansien signal, alues o i s
fi s de i a i e we e compu ed in sho sequen ial segmen s, whe e
slopes o linea eg ession compu ed wi h sequen ial signal segmen s o
a ew samples we e used o es ima e fi s de i a i e alues (Fig. 2). To
de e mine he beginning o a peak, i s fi s de i a i e alues had o
emain smalle han a small posi i e h eshold alue de e mined ex-
pe imen ally du ing a ew sequen ial slope alues. The ea e , slope
alues became g ea e while p oceeding o wa d along a ypically s eep
le side o a peak p oducing la ge posi i e fi s de i a i e slope alues.
Nex a peak maximum o op was ound when slope alues again de-
c eased less han he posi i e h eshold alue. A e he maximum, fi s
de i a i e slope alues changed nega i e along he dec easing igh
side o a peak (Fig. 2). Ul ima ely, he end o a peak was encoun e ed
when he fi s de i a i e slope alues again inc eased close o ze o. The
de ailed p ocedu e o he peak de ec ion was in oduced in ou p e-
ious esea ch [22]. Howe e , oscilla ions o e y small ampli udes
we e no accep ed as alid peaks in a signal, as ollows. A e he e-
mo al o a possible linea end in a signal, he ampli ude o la ge peaks
in a signal was es ima ed as a diffe ence om he a e age o he highes
sample (ampli ude) alues (15% o all) o he lowes sample alues in
he cu en signal. Such peak candida es ha had he ampli ude o he
le o igh side o a peak less han app oxima ely 8% om he abo e
ampli ude es ima e o he la ge peaks we e no accep ed as peaks, bu
we e suspec ed as being p obable noise [22]. The numbe s o he peaks
ex ac ed om he signals a ied om 1 o 61, and we e only 12.8 on
a e age.
A e he ecogni ion o he peaks om he ansien signals, alues
o 12 a iables o e e y peak we e compu ed as ollows. Fi s , he
ampli udes o he le and igh side o a peak we e compu ed - see
Fig. 2. Second, he du a ions o bo h peak sides we e compu ed om
loca ions a o cand om c o e. Thi d, he maximum o he fi s de i-
a i e om he le side o a peak and he absolu e minimum o he fi s
de i a i e om he igh side we e compu ed. Fou h, he maximum
and absolu e minimum o he second de i a i e we e compu ed om
he igh side only. The le side was no now applied, because e-
quen ly hese we e oo sho (as o he numbe o samples) o second
de i a i e alues o ha e been calcula ed. Fi h, he su ace de e mined
by a peak cu e and a line om he beginning o he end o a peak was
compu ed. Six h, he du a ion ( ime diffe ence) om he maximum a
loca ion c o he maximum o he p eceding peak was compu ed, o i
he cu en peak was he fi s peak o a signal, he du a ion was cal-
cula ed om he signal beginning. Se en h, he du a ion ( ime diffe -
ence) om peak beginning a o loca ion bo he fi s de i a i e
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
16
maximum o he le peak side and he du a ion om loca ion co he
peak maximum o loca ion do he fi s de i a i e minimum o he igh
peak side we e compu ed.
To isualize he da a, in o he wo ds, a iable alues compu ed
om 5290 ecognized peaks o he disease ansien signals and 2291
peaks o he con ol ansien signals, a e he no maliza ion o he
da a S ochas ic Neighbo Embedding algo i hm wi h he Euclidean
dis ance measu e in MATLAB was used o p esen he da a in wo di-
mensions. This is depic ed in Fig. 3. When a g ea pa o he cases in
wo diffe en classes a e apa om each o he , i is possible ha his
p edic s a success ul classifica ion o he classes.
The means and s anda d de ia ions o all 12 a iables a e p esen ed
in Table 1. Conside able diffe ences o he means o wo classes o
e e y a iable a e seen, which may deno e a a o able classifica ion
chance be ween he disease and con ol ansien signals.
Nex we e alua ed how efficien he 12 a iables a e o sepa a e o
classi y he wo classes. We an he elie F algo i hm in MATLAB o ou
da a. I unc ions on he basis o applying a nea es neighbo sea ching
me hod, in o de o measu e he diffe en ia ion powe o weigh ,
a iable by a iable. We chose nine k alues o he numbe s o nea es
neighbo s, con ol pa ame e o he algo i hm. They we e 3, 5, 7, 9, 11,
15, 21, 25 and 31, whe e odd alues only we e used o p e en possible
ies (equal numbe s om he wo opposi e classes) du ing nea es
neighbo sea ching. Fo each peak a iable, he median o weigh s
gi en by he elie F algo i hm o esul s o nine uns (Fig. 4) was
compu ed. The posi i e weigh s mean ha all a iables a e able o se-
pa a e he wo classes. Weigh s a e ela i e and hey exp ess anking o
a iables o diffe en ia ion o classes. Fo all nine uns, a iables 3, 4, 5
Fig. 1. (a) A 10 s segmen o a no mal CPVT signal in which he peaks ecognized o be no mal a e ma ked wi h g een ba s, and also he beginning and end o e e y
peak ma ked. (b) An abno mal CPVT signal in which all peaks we e ecognized as de o med o be abno mal and ma ked wi h s a s. (c) A 10 s segmen o a no mal
con ol ansien signal. (d) An abno mal con ol ansien signal, whe e wo abno mal peaks we e ma ked wi h s a s. (Fo in e p e a ion o he e e ences o colou in
his figu e legend, he eade is e e ed o he Web e sion o his a icle.)
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
17
and 11 ob ained he g ea es o bes weigh s and a iables 9, 10 and 12
esul ed in he smalles o poo es ins ances o sepa a e he wo classes.
Ins ead, a iables 1, 2, 6, 7 and 8 we e be ween he bes and poo es ,
and hei anking a ied sligh ly among he nine es uns, when hei
Fig. 2. The peak is he fi s one om he signal in
Fig. 1(a). The sho ed lines isualize imagina y
slopes (upwa d is posi i e, downwa d nega i e and
ho izon al ze o slope alue) o he fi s de i a i e
cu e o he signal du ing he cu en peak. The peak
begins om loca ion a; i s maximum is a cand end
a e. Loca ion bis o he maximum o he fi s de-
i a i e and d o i s minimum. (Fo in e p e a ion o
he e e ences o colou in his figu e legend, he
eade is e e ed o he Web e sion o his a icle.)
Fig. 3. The isualiza ion o he peak a iable alues o disease and con ol (WT)
ansien signals in wo dimensions.
Table 1
Means and s anda d de ia ions o 12 a iables o 5290 peaks o he disease ansien signals and 2291 peaks o he con ol ansien signals.
Va iable numbe Peak a iable Disease ansien signal peaks Con ol ansien signal peaks
1 Peak le side ampli ude 201.3 ± 134.4 272.5 ± 170.2
2 Peak igh side ampli ude 203.4 ± 135.4 275.4 ± 171.8
3 Le du a ion [s] 0.31 ± 0.181 0.492 ± 0.263
4 Righ du a ion [s] 0.597 ± 0.394 1.039 ± 0.601
5 Maximum o le side fi s de i a i e 1348 ± 985 2131 ± 1276
6 Absolu e minimum o igh side fi s de i a i e 780 ± 496.33 927 ± 635
7 Maximum o igh side second de i a i e 3397 ± 3110 4465 ± 3386
8 Absolu e minimum o igh side second de i a i e 2116 ± 2561 3938 ± 4359
9 Peak a ea 68 ± 76 132 ± 115
10 Du a ion om peak maximum o p eceding one (o beginning) [s] 1.039 ± 0.858 1.944 ± 1.58
11 Du a ion om peak beginning [s] o le side maximum o fi s de i a i e 0.201 ± 0.14 0.312 ± 0.198
12 Du a ion om peak maximum [s] o igh side minimum o fi s de i a i e 0.138 ± 0.075 0.156 ± 0.145
Fig. 4. Medians o weigh s we e calcula ed wi h he elie F algo i hm o he 12
a iables o he cu en da a by using 9 diffe en k alues (numbe o nea es
neighbo s). Va iables 3, 4, 5 and 11 a e he mos efficien and a iables 9, 10
and 12 he leas efficien o he diffe en ia ion o he disease ansien signals
om hose o con ols.
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
18
weigh s we e app oxima ely equal as seen in Fig. 4. In any case, on he
basis o hese esul s as well hose in Table 1, all 12 a iables we e
ound o be use ul o he diffe en ia ion o disease and con ol calcium
ansien signals om each o he .
4. Classifica ion me hods used and design o expe imen s
We used a wide collec ion o classifica ion algo i hms in ou s udy,
anging om adi ional me hods o s a e-o - he-a me hods. Since ou
da ase consis s o a ue class label defined by he human expe o
each signal, we concen a ed only on supe ised classifica ion me hods
and, hence, semi-supe ised and unsupe ised me hods we e ou o he
scope o his pape . Me hods used in ou s udy we e mos ly he same as
used in ou p e ious s udy [23]. Howe e , in Re . [23] he classifica ion
ask was mul i-class by na u e, whe eas in his pape he classifica ion
ask is a wo-class p oblem. We do no p esen he de ailed desc ip ion
abou he ac ual classifica ion me hods, bu a eade can find ho ough
desc ip ions abou he algo i hms om he gi en e e ences. Compa ed
o Re . [23], he e we e now wo no el peak a iables: a iables 11 and
12 (Table 1).
As a fi s classifica ion algo i hm, we applied he kNea es Neighbo
(kNN) sea ching me hod [24–26] ha is one o he ea lies classifica-
ion algo i hms and mos used. The pe o mance o he kNN algo i hm
depends mainly on h ee ac o s: k alue, dis ance measu e, and dis-
ance weigh ing scheme. These ac o s a e da a-dependen and o each
da ase , a sui able combina ion mus be sea ched sepa a ely. Fo ou
s udy, we selec ed k alues o 1, 5, 7, 11, 13 and 17 o be examined, and
he e we ollowed he p inciple used in Re . [23]. We selec ed eigh
dis ance measu es o be es ed: Chebyshe , ci yblock (Manha an),
co ela ion, cosine, Euclidean, Mahalanobis, s anda dized Euclidean
and Spea man. We pe o med classifica ion wi h h ee dis ance
weigh ing schemes: no weigh ing (weigh ing equal o 1), ecip ocal,
and squa ed ecip ocal wi h espec o dis ance.
The second wholeness used was disc iminan analysis based algo-
i hms. Disc iminan analysis co e s a ious a ia ions, and om hem
we applied linea disc iminan analysis [27,28], quad a ic disc iminan
analysis [28,29], and Mahalanobis dis ance based disc iminan analysis
[30]. When mo ing o p obabili y based algo i hms, naï e Bayes clas-
sifie [24,31,32] canno be dismissed. Naï e Bayes is a classical widely
used me hod in many applica ions. I can be used wi h o wi hou
ke nel densi y es ima ion [24,31]. In his pape we used he naï e Bayes
algo i hm in bo h senses. When ke nel densi y es ima ion was used, we
examined Gaussian, box, iangle and Epanechniko ke nels. We also
applied he naï e Bayes classifie wi hou ke nel densi y es ima ion
when he no mal dis ibu ion assump ion o e he da ase is expec ed.
Besides naï e Bayes classifie , kNN and disc iminan analysis based
classifica ion me hods, we used mul inomial logis ic eg ession [33,34]
which e u ns o logis ic eg ession [31,32] in wo-class asks.
Decision ee-based solu ions a e commonly used al e na i es in
machine lea ning asks. Thei ad an ages a e easy in e p e a ion and
compu a ional efficiency. These issues a e impo an o ake in o ac-
coun when conside ing he end-use s o ou applica ion, who a e
pe sons no expe s in he machine lea ning a ea. In ou s udy, we in-
es iga ed CART [25,35] algo i hm and Random Fo es s [36–38]. Fo
Random Fo es s we a ied he numbe o ees om 1 o 100.
The Suppo Vec o Machine [39] has gained g ea popula i y since
he ea ly 1990s, and has been used in a ious applica ions. Howe e , in
ou s udy, we decided o use a a ian o SVM called he leas squa es
suppo ec o machine (LSSVM) [40–42] which diffe s om adi ional
SVM in such a way ha LSSVM sol es a sys em o linea equa ions
ins ead o a quad a ic op imiza ion p oblem. The pe o mance o
LSSVM is hea ily dependen on he selec ion o a ke nel unc ion and
(hype )pa ame e alues. Hence, i is always necessa y o pe o m a
ho ough sea ch o (hype )pa ame e alues in o de o ensu e he bes
possible esul . A common pa ame e o all ke nel unc ions in LSSVM
is C, also called box cons ain . We selec ed he linea , quad a ic, 3 d
deg ee o polynomial ke nel and he RBF ke nel o be examined in ou
pape . The pa ame e alue space o he box cons ain and hy-
pe pa ame e σ( he wid h o Gaussian unc ion) is {2
−12
,2
−11
,
…,2
16
,2
17
}. By his means polynomial ke nels we e es ed on 30
pa ame e alues and he RBF ke nel on 900 (C,σ) combina ions.
The classifica ion was pe o med based on he lea e-one-signal-ou
(LOSO) p ocedu e, which is a modifica ion compa ed o he lea e-one-
ou me hod. In signal classifica ion, a no iceable de ail mus be e-
membe ed. Va iables a e de e mined om peaks, and a signal consis s
o one o mo e peaks. Thus, he da a gained om one signal usually
includes se e al ows in an obse a ion ma ix. When defining aining
and es se s, one needs o ensu e ha he whole da a om he signal is
in ei he he aining o es se . Signal da a mus no be spli in o wo,
such ha one pa is in aining se and he o he is in es se . In LOSO,
he peak-based da a om each signal in u n o ms a es se , and he
es o he da a a e in he aining se .
When we ain a compu a ional model based on a classifica ion
me hod, we mus emembe ha he aining phase o an algo i hm is
pe o med on peak-based da a, no on signal le el da a. Hence, a
classifie lea ns i s model based on peaks and no on signals. A e
aining o a classifie , we gi e he es se as an inpu o he classifie .
Then he classifie gi es a p edic ed class label o each peak in a es
se . In his s age, esul s a e a peak le el. Howe e , since he aim o his
pape is signal classifica ion, we need o ans o m he peak le el esul s
in o a signal le el esul . This is done by aking a mode om he p e-
dic ed class labels o peaks in a es se . We can do his because we use
he LOSO p ocedu e, whe e a es se co e s da a only om one signal.
Howe e , mode is no always unambiguously defined and a ie may
occu . In ou pape we had only wo classes, so a ie could occu wi h
only wo classes. I a ie occu ed, we sol ed he p oblem in he ol-
lowing way.
1. Ex ac he aining da a o classes C
1
and C
2
occu ing in a ie, om
he aining se .
2. Find he p opo ions P
1
=(|C
1
|/(|C
1
|+|C
2
|))100% and P
2
=(|C
2
|/
(|C
1
|+|C
2
|))100% whe e |.| is he size o a se . Hence, an in e al
[0,P
1
] is o he class C
1
and in e al (P
1
,100] is o he class C
2
.
3. Gene a e a andom numbe R om he uni o m dis ibu ion U(0,1).
4. I R*100% belongs o in e al [0,P
1
], selec C
1
as final class label o
he signal. O he wise, selec C
2
.
A e finding a p edic ed class label o each signal in a da ase , we
can compa e he p edic ed class label wi h he ue label and cons uc
a con usion ma ix. F om he con usion ma ix, we can e alua e di -
e en kinds o measu es which desc ibe how well he classifica ion has
succeeded. Fo ou s udy, we selec ed accu acy ((TP + TN)/
(TP + TN + FN + FP)), ue posi i e a e (TP/(TP + FN)) o diseases
and ue nega i e a e (TN/(TN + FP)) o con ols. Fo he classifi-
ca ion me hods which equi e pa ame e uning, we epea ed LOSO
wi h all pa ame e alues examined, and selec ed a pa ame e alue
(combina ion) ha achie ed he highes accu acy.
5. Classifica ion esul s o disease o con ol ansien signals
The main a ge o he esea ch was o s udy how efficien ly wo
ansien signal g oups can be diffe en ia ed om each o he . Fo his
pu pose, se e al classifie s we e implemen ed as desc ibed abo e. Thei
esul s a e p esen ed in he ollowing.
Classifica ion esul s a e shown in Tables 2–4, in which ue posi i e
a es (sensi i i y) co espond o disease ansien signals, and ue ne-
ga i e a es (specifici y) o con ol signals. Accu acy equals he sum o
ue posi i e and nega i e cases di ided by he numbe o all cases.
Now knea es neighbo sea ching wi h ci yblock (Manha an) me ic
yielded he bes accu acy esul s o 86.0% in Table 2.InTable 3,k
nea es neighbo sea ching wi h Euclidean me ic and squa ed in e se
weigh ing was he bes me hod, ob aining a 84.5% le el. In Table 4, he
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
19

suppo ec o machine (LSSVM) wi h adial basis unc ion (RBF)
ke nel ha ing an accu acy o 84.7% and andom o es wi h 87.4%
we e he bes echniques. O e all, he classifica ion o da a in o wo
classes was e y success ul.
6. Discussion
This s udy was aimed a in es iga ing whe he machine lea ning
could, in gene al, sepa a e heal hy ca diomyocy es om diseased ones.
The pheno ype o he iPSC-de i ed CMs was de e mined by Ca
2+
imaging. Bo h con ol and diseased ca diomyocy es con ained cells
wi h no mal bea ing, as well as hose wi h abno mal bea ing beha io .
Wi h machine lea ning, e y high classifica ion accu acy alues (up o
87.4%) we e ob ained o dis inguish con ol and diseased cells, despi e
bo h ha ing mixed CM popula ions (con aining bo h no mal and ab-
no mal CMs), i.e., o dis inguish CMs de i ed om pa ien s ca ying a
mu a ion o a ca diac disease om con ol CMs.
The iPSC echnology has e olu ionized he s udy o gene ic ca diac
diseases [43]. I enables he in es iga ion o pa ien - and mu a ion-
specific cells in o de o unde s and he disease pa hophysiology o
in e es , as well as o p o ide a pla o m o s udy d ug esponsi eness in
a pe sonalized way [4]. Howe e , CMs de i ed om iPSCs ob ained
wi h cu en diffe en ia ion p o ocols s ill ha e se e al p oblems. They
a e fi s o all imma u e ca diomyocy es [44]. In addi ion, hey p esen
all ypes o ca diomyocy es including a ial, en icula and pacemake
cells. These issues make i p oblema ic o p oduce and de e mine a
disease pheno ype in a ep oducible way. In ou cu en s udy, hese
limi a ions we e e iden . The CMs we e imma u e, and bo h ou con ol
CMs as well as he diseased CMs we e mixed cell popula ions and wi h
Ca
2+
ansien signals, e.g., en icula o a ial cells canno be dis-
inguished om each o he . Howe e , despi e hese p oblems, ou
machine lea ning p ocedu e was success ul o diffe en ia e con ol CMs
om diseased CMs, sugges ing he p esence o cha ac e is ics o heal hy
o diseased cells al eady in he a al s a e and common o all ypes o
CMs.
Due o abo e men ioned p oblems wi h iPSC-de i ed CM, con ol
cells also included abno mal Ca
2+
ansien s. In ou s udy, 12.6% o
con ol cells p esen ed a ious ypes o abno mali ies. Howe e , he
amoun o abno mal ansien signals was much g ea e (54.6%) in CMs
ca ying any mu a ion o a ious gene ic ca diac diseases including
gene ic a hy hmias and ca diomyopa hies. We demons a ed ea lie
[23] ha i is possible o diffe en ia e gene ic ca diac diseases om
each o he based on machine lea ning echniques. In he cu en s udy,
he aim was o collec mo e ansien signals o con ol CMs, o dec ease
he diffe ence be ween he smalle signal numbe o con ol CMS and
he g ea e signal numbe o diseased CMs compa ed o he ea lie si-
ua ion [23], and pool all diseased ones as one g oup, o analyze
whe he his could be sepa a ed om con ol cells con aining also bo h
abno mally and no mally bea ing cells. Using diffe en algo i hms and
me hods o machine lea ning, we we e able o p oduce pa adigms wi h
high a classifica ion accu acy o up o 87.4%, sugges ing ha his
p ocedu e could ha e po en ial use in clinical applica ions in he u u e.
7. Conclusions
Gene ic ca diac diseases a e clinically o en p oblema ic. Fi s , he
disease pheno ype is a iable e en wi hin a single amily [45]. Ad-
di ionally, i is s ill common despi e ad ances in molecula gene ics ha
he mu a ion causing he disease in he amily is no known. In his
si ua ion, i is impossible o dis inguish hose who a e po en ially a isk
o de eloping he disease pheno ype, which amily membe s should be
egula ly checked clinically, and who should be ad ised o li es yle
es ic ions o p e en i e medica ion. The esul s ob ained in his s udy
a e po en ially p omising o iden i y indi iduals a isk. iPSC-de i ed
CMs ca ying a disease causing mu a ion can accu a ely be sepa a ed
om cells de i ed om heal hy indi iduals, and hus po en ially iPSC-
de i ed ca diomyocy es combined wi h machine lea ning algo i hms
could in he u u e be used also clinically o iden i y indi iduals a isk,
Table 2
Classifica ion esul s o knea es neighbo (kNN) sea ching, wi h diffe en me ics o measu es wi h he bes k alue.
Classifica ion me hod T ue posi i e a es o diseases % T ue posi i e a es o con ols % Accu acy %
kNN wi h Chebyche me ic and equal weigh ing, k= 1 87.3 69.3 81.3
kNN wi h Chebyche me ic and in e se weigh ing, k= 1 87.3 69.3 81.3
kNN wi h Chebyche me ic and squa ed in e se weigh ing, k= 5 86.5 71.4 81.5
kNN wi h ci yblock me ic and equal weigh ing, k= 1 91.1 75.9 86.0
kNN wi h ci yblock me ic and in e se weigh ing, k= 1 91.1 75.9 86.0
kNN wi h ci yblock me ic and squa ed in e se weigh ing, k= 1 91.1 75.9 86.0
kNN wi h co ela ion measu e and equal weigh ing, k= 1 89.3 67.8 82.1
kNN wi h co ela ion measu e and in e se weigh ing, k= 5 89.3 71.4 83.3
kNN wi h co ela ion measu e and squa ed in e se weigh ing, k= 5 89.3 71.9 83.3
kNN wi h cosine measu e and equal weigh ing, k= 1 86.8 71.4 81.6
kNN wi h cosine measu e and in e se weigh ing, k= 5 87.8 74.4 83.3
kNN wi h cosine measu e and squa ed in e se weigh ing, k= 7 89.3 72.9 83.8
Table 3
Mo e classifica ion esul s o knea es neighbo (kNN) sea ching, wi h diffe en me ics o measu es wi h he bes k alue.
Classifica ion me hod T ue posi i e a es o diseases % T ue nega i e a es o con ols % Accu acy %
kNN wi h Euclidean me ic and equal weigh ing, k= 1 89.1 74.4 83.8
kNN wi h Euclidean me ic and in e se weigh ing, k= 1 89.1 74.4 84.1
kNN wi h Euclidean me ic and squa ed in e se weigh ing, k= 5 90.1 73.4 84.5
kNN wi h Mahalanobis me ic and equal weigh ing, k= 1 90.9 71.4 84.3
kNN wi h Mahalanobis me ic and in e se weigh ing, k= 1 90.9 71.4 84.3
kNN wi h Mahalanobis me ic and squa ed in e se weigh ing, k= 1 90.9 71.4 84.3
kNN wi h s anda dized Euclidean me ic and equal weigh ing. k= 1 89.1 74.4 84.1
kNN wi h s anda dized Euclidean me ic and in e se weigh ing, k= 1 89.1 74.4 84.1
kNN wi h s anda dized Euclidean me ic and squa ed in e se weigh ing, k= 5 89.8 73.4 84.3
kNN wi h Spea man measu e and equal weigh ing, k= 1 88.6 66.3 81.1
kNN wi h Spea man measu e and in e se weigh ing, k= 5 89.6 69.3 82.8
kNN wi h Spea man measu e and squa ed in e se weigh ing, k= 5 90.4 69.3 83.3
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
20
and make i possible o ocus p e en i e ac ions on hose, and elie e
he disease bu den om hose wi hou any signs o disease a he cel-
lula le el.
The high accu acy ob ained wi h ou bes machine lea ning algo-
i hm sugges s ha iPSC echnology combined wi h machine lea ning
could be used e en o diagnos ic pu poses in he u u e. We will
con inue o collec mo e Ca
2+
ansien signals o CMs de i ed om a
la ge collec ion o iPS cell lines ca ying diffe en mu a ions and pa-
ien popula ions, as well as signals o heal hy con ols, and also om
isogenic lines, and hus o imp o e hese me hods o make hem mo e
sui able o clinical pu poses.
Conflic s o in e es
None.
E hical s a emen
The cu en s udy was app o ed by he E hics Commi ee o
Pi kanmaa Hospi al Dis ic , Tampe e, Finland, in es ablishing, cul-
u ing and diffe en ia ing human iPSC lines (R08070).
Acknowledgmen s
None.
Appendix A. Supplemen a y da a
Supplemen a y da a o his a icle can be ound online a h ps://
doi.o g/10.1016/j.imu.2019.01.006.
Re e ences
[1] Geo ge AL. Molecula and gene ic basis o sudden ca diac dea h. J Clin In es
2013;123:75–83.
[2] Asa yan B, Medei os-Domingo A. T ansla ing eme ging molecula gene ic insigh s
in o clinical p ac ice in inhe i ed ca diomyopa hies. J Mol Med 2018;96:993–1024.
[3] Kujala K, Paa ola J, Leh i A, La sson K, Pekka inen-Ma ila M, Vii asalo M, Lah inen
AM, Toi onen L, Kon ula K, Swan H, Laine M, Sil ennoinen O, Aal o-Se älä K. Cell
model o ca echolamine cig polymo phic en icula achyca dia e eals ea ly and
delayed a e depola iza ions. PLoS One 2012;7(9)h ps://doi.o g/10.1371/jou nal.
pone.0044660.
[4] Pen inen K, Swan H, Vanninen S, Paa ola J, Lah inen AM, Kon ula K, Aal o-Se älä
K. An ia hy hmic effec s o Dan olene in pa ien s wi h ca echolamine gic poly-
mo phic en icula achyca dia and eplica ion o he esponses using iPSC models.
PLoS One 2015;10(7)h ps://doi.o g/10.1371/jou nal.pone.0125366.
[5] Takahashi K, Tanabe K, Ohnuki M, Na i a M, Ichisaka T, Tomoda K, Yamanaka S.
Induc ion o plu ipo en s em cells om adul human fib oblas s by defined ac o s.
Cell 2007;131:861–72.
[6] De alla HD, Passie R. Ca diac diffe en ia ion o plu ipo en s em cells and im-
plica ions o modeling he hea in heal h and disease. Sci T ansl Med 2018;10.
eaah5457.
[7] Fa ima A, Xu G, Shao K, Papadopoulos S, Lehmann M, A naiz-Co JJ, Rosa AO,
Nguemo F, Ma zkies M, Di mann S, S one SL, Linke M, Zechne U, Beye V, Hennies
HC, Rosenk anz S, Klauke B, Pa wani AS, Ha e kamp W, Pfi ze G, Fa M,
Cleemann L, Mo ad M, Mil ing H, Heschele J, Sa ic T. In i o modeling o ya-
nodine ecep o 2 dys unc ion using human induced plu ipo en s em cells. Cell
Physiol Biochem 2011;28:579–92.
[8] Jung CB, Mo e i A, Mede os y Schni zle M, Iop L, S o ch U, Bellin M, Do n T,
Ruppen hal S, P eiffe S, Goedel A, Di schinge RJ, Sey a h M, Lam JT, Sinnecke
D, Gude mann T, Lipp P, Laugwi z KL. Dan olene escues a hy hmogenic RYR2
de ec in a pa ien -specific s em cell model o ca echolamine gic polymo phic
en icula achyca dia. EMBO Mol Med 2012;4:180–91.
[9] No ak A, Ba ad L, Lo be A, Ghe ghiceanu M, Rei e I, Eisen B, Eldo L, I sko i z-
Eldo J, Elda M, A ad M, Binah O. Func ional abno mali ies in iPSC-de i ed ca -
diomyocy es gene a ed om CPVT1 and CPVT2 pa ien s ca ying yanodine o
calseques in mu a ions. J Cell Mol Med 2015;19:2006–18.
[10] I zhaki I, Maizels L, Hube L, Geps ein A, A bel G, Caspi O, Mille L, Belhassen B,
No E, Glikson M, Geps ein L. Modeling o ca echolamine gic polymo phic en-
icula achyca dia wi h pa ien -specific human-induced plu ipo en s em cells. J
Am Coll Ca diol 2012;60:990–1000.
[11] Zhang XH, Ha iland S, Wei H, Sa ic T, Fa ima A, Heschele J, Cleemann L, Mo ad
M. Ca
2+
signaling in human induced plu ipo en s em cell-de i ed ca diomyocy es
(iPS-CM) om no mal and ca echolamine gic polymo phic en icula achyca dia
(CPVT)-afflic ed subjec s. Cell Calcium 2013;54:57–70.
[12] Di Pasquale E, Lodola F, Mi agoli M, Deneg i M, A elino-C uz JE, Buonoco e M,
Nakahama H, Po a a o P, Bloise R, Napoli ano C, Condo elli G, P io i SG. CaMKII
inhibi ion ec ifies a hy hmic pheno ype in a pa ien -specific model o ca echola-
mine gic polymo phic en icula achyca dia. Cell Dea h Dis 2013;4:e843.
[13] Mo e i A, Bellin M, Welling A, Jung CB, Lam JT, Bo -Flugel L, Do n T, Goedel A,
Hohnke C, Ho mann F, Sey a h M, Sinnecke D, Schomig A, Laugwi z KL KL.
Pa ien -specific induced plu ipo en s em-cell models o long-QT synd ome. N Engl
J Med 2010;363:1397–409.
[14] Ma sa E, Rajamohan D, Dick E, Young L, Mello L, S ani o h A, Denning C. D ug
e alua ion in ca diomyocy es de i ed om human induced plu ipo en s em cells
ca ying a long QT synd ome ype 2 mu a ion. Eu Hea J 2011;32:952–62.
[15] Lah i AL, Kujala VJ, Chapman H, Koi is o AP, Pekkanen-Ma ila M, Ke kela E,
Hy inen J, Kon ula K, Swan H, Conklin BR, Yamanaka S, Sil ennoinen O, Aal o-
Se ala K. Model o long QT synd ome ype 2 using human iPS cells demons a es
a hy hmogenic cha ac e is ics in cell cul u e. Dis. Model Mech. 2012;5:220–30.
[16] Ki iaho AL, Ahola A, La sson K, Pen inen K, Swan H, Pekkanen-Ma ila M,
Venäläinen H, Paa ola K, Hy inen J, Aal o-Se älä K. Dis inc elec ophysiological
and mechanical bea ing pheno ypes o long QT synd ome ype 1-specific ca dio-
myocy es ca ying diffe en mu a ions. IJC Hea & Vascula u e 2015;8:9–31.
[17] Han L, Li Y, Tchao J, Kaplan AD, Lin B, Li Y, Mich-Basso J, Lis A, Hassan N, London
B, Be CG, Tobi a K, Rasmusson RL, Yang L. S udy amilial hype ophic ca dio-
myopa hy using pa ien -specific induced plu ipo en s em cells. Ca dio asc Res
2014;104(2):258–69. h ps://doi.o g/10.1093/c /c u205. Epub 2014 Sep 10.
[18] Lan F, Lee AS, Liang P, Sanchez-F ei e V, Nguyen PK, Wang L, Han L, Yen M, Wang
Y, Sun N, Abilez OJ, Hu S, Ebe AD, Na a e e EG, Simmons CS, Wheele M, P ui
B, Lewis R, Yamaguchi Y, Ashley EA, Be s DM, Robbins RC, Longake MT, Wu JC.
Abno mal calcium handling p ope ies unde lie amilial hype ophic ca diomyo-
pa hy pa hology in pa ien -specific induced plu ipo en s em cells. Cell S em Cell
2013;12:101–13.
[19] Ojala M, P ajapa i C, Pölönen RP, Rajala K, Pekkanen-Ma ila M, Rasku J, La sson
K, Aal o-Se älä K. Mu a ion-specific pheno ypes in hiPSC-de i ed ca diomyocy es
ca ying ei he myosin-binding p o ein C o α- opomyosin mu a ion o hype -
ophic ca diomyopa hy. S em Cell In 2016:1684792h ps://www.hindawi.com/
jou nals/sci/2016/1684792/.
[20] Lee EK, T an DD, Keung W, Chan P, Wong G, Chan CW, Cos a KD, Li RA, Khine M.
Machine lea ning o human plu ipo en s em cell-de i ed enginee ed ca diac issue
con ac ili y o au oma ed d ug classifica ion. S em Cell Rep 2017;9:1560–72.
[21] Juhola M, Jou sijoki H, Va pa K, Saa ikoski J, Rasku J, Il anen K, Lau ikkala J,
Hyy ö H, Á alos-Salgue o J, Sii ola H, Pen inen K, Aal o-Se älä K. On compu a ion
Table 4
Resul s o disc iminan analysis, decision ee, mul inomial logis ic eg ession, naï e Bayes, andom o es and leas squa e (LS) suppo ec o machines (SVM).
Classifica ion me hod T ue posi i e a es o diseases % T ue posi i e a es o con ols % Accu acy %
Linea disc iminan analysis 78.4 62.3 73.0
Mahalanobis disc iminan analysis 33.0 94.0 53.5
Quad a ic disc iminan analysis 76.1 59.3 70.5
Decision ee 89.1 74.4 84.1
Mul inomial logis ic eg ession 77.7 62.8 72.7
Naï e Bayes wi h no mal dis ibu ion 71.8 67.8 70.5
Naï e Bayes wi h no mal ke nel 68.0 78.4 71.5
Naï e Bayes wi h box ke nel 66.5 79.4 70.8
Naï e Bayes wi h Epanechniko ke nel 67.5 78.9 71.3
Naï e Bayes wi h iangle ke nel 68.8 78.9 72.2
Random o es , numbe o ees 18 92.4 77.4 87.4
LS-SVM wi h linea ke nel, C=2
−4
69.3 76.4 71.7
LS-SVM wi h quad a ic ke nel, C= 1 74.6 77.9 75.7
LS-SVM wi h cubic ke nel, C=2
−5
79.4 78.4 79.1
LS-SVM wi h RBF ke nel, C=2
11
,σ= 2 91.6 70.9 84.7
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
21
o calcium cycling anomalies in ca diomyocy es da a. 36 h Annual In . Con . IEEE
Eng. Med. Biol. Socie y. 2014. p. 1444–7. Chicago, Illinois, USA.
[22] Juhola M, Pen inen K, Jou sijoki H, Va pa K, Saa ikoski J, Rasku J, Sii ola H,
Il anen K, Lau ikkala J, Hyy ö H, Hy inen J, Aal o-Se älä K. Signal analysis and
classifica ion me hods o calcium ansien da a o s em cell de i ed ca diomyo-
cy es. Compu Biol Med 2015;61:1–7.
[23] Juhola M, Jou sijoki H, Pen inen K, Aal o-Se älä K. De ec ion o gene ic ca diac
diseases by Ca
2+
ansien p ofiles using machine lea ning me hods. Sci Rep
2018;8:9355h ps://doi.o g/10.1038/s41598-018-27695-5.
[24] Han J, Kambe M, Pei J. Da a mining: concep s and echniques. hi d ed. Mo gan
Kau mann; 2012.
[25] Duda RO, Ha PE, S o k DG. Pa e n classifica ion. second ed. John Wiley & Sons;
2001.
[26] Wu X, Kuma V, Quinlan R, Ghosh J, Yang Q, Mo oda H, McLachlan GJ, Ng A, Liu B,
Yu PS, Zhou Z-H, S einbach M, Hand DJ, S einbe g D. Top 10 algo i hms in da a
mining. Knowl In Sys 2008;14(1):1–37.
[27] Tha wa A, Ta ek G, Abdelhameed I, Aboul Ella H. Linea disc iminan analysis: a
de ailed u o ial. AI Commun 2017;30(2):169–90.
[28] Cios KJ, Ped ycz W, Swinia ski RW, Ku gan LA. Da a mining: a knowledge dis-
co e y app oach. Sp inge ; 2007.
[29] Wu W, Malle Y, Walczak B, Penninckx W, Massa DL, Heue ding S, E ni F.
Compa ison o egula ized disc iminan analysis, linea disc iminan analysis and
quad a ic disc iminan analysis, applied o NIR da a. Anal Chim Ac a
1996;329(3):257–65.
[30] G. Bohling, Classical no mal-based disc iminan analysis, Technical epo , Kansas
Geol Su . h p://people.ku.edu/gbohling/EECS833 Accessed 26.7.2018.
[31] Has ie T, Tibshi ani R, F iedmann J. The elemen s o s a is ical lea ning –da a
mining, in e ence, and p edic ion. second ed. Sp inge ; 2009.
[32] Ng AY, Jo dan MI. On disc imina i e s. gene a i e classifie s: a compa ison o
logis ic eg ession and naï e Bayes NIPS. 2002. p. 841–8.
[33] Ag es i A. Ca ego ical da a analysis. John Wiley & Sons; 1990.
[34] Kwak C, Clay on-Ma hews A. Mul inomial logis ic eg ession. Nu s Res
2002;51(6):404–10.
[35] Rokach L, Maimon O. Da a mining wi h decision ees: heo y and applica ions.
Wo ld Scien ific; 2015.
[36] B eiman L. Random o es s. Mach Lea n 2001;45(1):5–32.
[37] Chen W, Xie X, Wang J, P adhan B, Hong H, Bui DT, Duan Z, Ma J. A compa a i e
s udy o logis ic model ee, andom o es , and classifica ion and eg ession ee
models o spa ial p edic ion o landslide suscep ibili y. Ca ena 2017;151:141–60.
[38] Cou onné R, P obs P, Boules eix A-L. Random o es e sus logis ic eg ession: a
la ge-scale benchma k expe imen . BMC Bioin 2018;19:270.
[39] Co es C, Vapnik V. Suppo - ec o ne wo ks. Mach Lea n 1995;20(3):273–97.
[40] Suykens JAK, Van Ges el T, De B aban e J, De Moo B, Vandewalle J. Leas squa es
suppo ec o machines. Wo ld Scien ific; 1999.
[41] Suykens JAK, Vandewalle J. Leas squa es suppo ec o machine classifie s.
Neu al P ocess Le 1999;9(3):293–300.
[42] Van Ges el T, Suykens JAK, Baesens B, Viaene S, Van hienen J, Dedene G, De Moo
B, Vandewalle J. Benchma king leas squa es suppo ec o machine classifie s.
Mach Lea n 2004;54(1):5–32.
[43] Casini S, Ve ke k AO, Remme CA. Human iPSC-de i ed ca diomyocy es o in-
es iga ion o disease mechanism and he apeu ic s a egies in inhe i ed a hy hmia
synd omes: s eng hs and limi a ions. Ca dio asc D ugs The 2017;31:325–44.
[44] Vee man CC, Kosmidis G, Mumme y CL, Casini S, Ve ke k AO, Bellin M. Inma u i y
o human s em-cell-de i ed ca diomyocy es in cul u e: a al flaw o soluble p o-
blem. S em Cell De 2015;24(9):1035–52. h ps://doi.o g/10.1089/scd.2014.0533.
[45] Wilde AA, Beh ER. Gene ic es ing o inhe i ed ca diac disease. Na Re Ca diol
2013;10:571–83.
M. Juhola e al. In o ma ics in Medicine Unlocked 14 (2019) 15–22
22