Machine Lea ning o Find A eas o Ro o s Sus aining A ial Fib illa ion F om
he ECG
Gio gio Luongo1, Luca Azzolin1, Massimo W Ri ol a2, Tiago P Almeida3, Juan Pablo Ma ínez4,
Diogo C So iano5, Ola Dössel1, Robe o Sassi2, Pablo Laguna4, Axel Loewe1
1Ins i u e o Biomedical Enginee ing, Ka ls uhe Ins i u e o Technology (KIT), Ka ls uhe, Ge many
2Dipa imen o di In o ma ica, Uni e si à degli S udi di Milano, Milan, I aly
3Depa men o Ca dio ascula Sciences, Uni e si y o Leices e , Leices e , UK
4I3A, Uni e sidad de Za agoza, and CIBER-BNN, Za agoza, Spain
5Enginee ing, Modelling and Applied Social Sciences Cen e, ABC Fede al Uni e si y, São Be na do
do Campo, B azil
Abs ac
A ial ib illa ion (AF) is he mos equen i egula
hea hy hm due o diso ganized a ial elec ical ac i i y,
o en sus ained by o a ional d i e s called o o s.
The non-in asi e localiza ion o AF d i e s can lead
o imp o ed pe sonalized abla ion s a egy, sugges ing
pulmona y ein (PV) isola ion o mo e complex ex a-
PV abla ion p ocedu es in case he d i e is on o he
a ial egions. We used a Machine Lea ning app oach
o cha ac e ize and disc imina e simula ed single s able
o o s (1R) loca ion: PVs, le a ium (LA) excluding he
PVs, and igh a ium (RA), u ilizing solely non-in asi e
signals (i.e., he 12-lead ECG). 1R episodes sus aining
AF we e simula ed. 128 ea u es we e ex ac ed om he
signals. G eedy o wa d algo i hm was implemen ed o
selec he bes ea u e se which was ed o a decision
ee classi ie wi h hold-ou c oss- alida ion echnique.
All es ed ea u es showed signi ican disc imina o y
powe , especially hose based on ecu ence quan i ica ion
analysis (up o 80.9% accu acy wi h single ea u e
classi ica ion). The decision ee classi ie achie ed 89.4%
es accu acy wi h 18 ea u es on simula ed da a, wi h
sensi i i ies o 93.0%, 82.4%, and 83.3% o RA, LA, and
PV classes, espec i ely. Ou esul s show ha a machine
lea ning app oach can po en ially iden i y he loca ion o
1R sus aining AF using he 12-lead ECG.
1. In oduc ion
A ial ib illa ion (AF) is he mos common sus ained
a hy hmia in clinical p ac ice and a leading cause o
hospi aliza ion and dea h [1]. This a hy hmia is o en
sus ained by localized unc ional een an ci cui s called
o o s, cha ac e ized by cu ed wa e on s and wa e ails
ha mee each o he a a singula i y poin [2]. One
common he apy o e mina e AF is abla ion. Typically,
“ igge s” ha s a AF and/o he “subs a e” ha
pa icipa es in i s pe pe ua ion a e a ge ed du ing abla ion.
Howe e , i emains unclea which o he wo app oaches
is he mos e ec i e o ea ing AF, specially in ad anced
s ages o he disease. Na ayan e al. showed ha i
is impo an o localize and abla e o o s, ocal sou ces
d i e s o o ganizing sou ces o ib illa ion o e mina e
AF [3]. Addi ionally, igge s and sus aining mechanisms
a e o en localized in he pulmona y eins (PVs) [4]. Thus,
PV isola ion (PVI) is he i s abla ion p ocedu es applied
o y o e mina e AF.
In his p elimina y wo k, we sough o cha ac e ize and
iden i y single s able o o s (1R) loca ed nea he PVs,
on ex a-PV le a ium (LA) a eas, and on igh a ium
(RA) a eas by using 12-lead elec oca diog am (ECG) in
a simula ion s udy. This non-in asi e me hod could help
guide abla ion p ocedu es, highligh ing a ial egions ha
may be impo an in he AF pe pe ua ion, and hence a ge s
o abla ion. In case o o o s iden i ied wi hin he PVs,
he applica ion o a p io i in asi e and ime-consuming
elec ophysiologic mapping p ocedu es could be a oided,
p oceeding di ec ly wi h PVI.
2. Me hods
2.1. Simula ions
1R episodes sus aining AF we e simula ed using he
phase singula i y dis ibu ion me hod on a olume ic a ial
model buil om clinical da a, as epo ed in [5]. B ie ly,
he phase singula i ies we e placed in 300 uni o mly
dis ibu ed poin s in he a ia, and 3 s o ac i a ion we e
Compu ing in Ca diology 2020; Vol 47 Page 1 ISSN: 2325-887X DOI: 10.22489/CinC.2020.181
compu ed. Only he cases wi h 1R episodes ha kep going
o he whole simula ion ime we e conside ed o u he
analysis. This led o unbalanced da a gene a ion. As
esul o he monodomain simula ion, he ansmemb ane
ol age was used o calcula e he body su ace po en ial
map (BSPM) on 8 di e en o so models gene a ed om
segmen ed MRI da a o heal hy male and emale subjec s
(Fig. 1), [5]. F om he BSPM, he 12-lead ECG was
ex ac ed wi h a leng h o 3 s (Fig. 1). Only -wa es
wi hou he QRS-T complex composed he 12-lead ECG,
since he en icles we e no included in he simula ions.
440 se s o 12-lead ECG o med he inal da ase (40 ECGs
wi h 1R loca ed in he PVs, 112 in ex a-PV LA a eas, and
288 in he RA).
2.2. Fea u e ex ac ion
128 ea u es we e ex ac ed om he he signals using
se e al biosignal p ocessing me hods, such as: Hjo
desc ip o s o analyse he spec al momen s om he ime
signals [5]; ecu ence quan i ica ion analysis (RQA) on
ec oca diog am (VCG) [6], indi idual componen RQA
(icRQA), and spa ial educed RQA (s RQA) [7] o analyse
he opological s uc u e o mul idimensional dynamical
sys ems; a io o he p incipal componen analysis (PCA)
eigen alues, o ganiza ion index, and spec al en opy o
s udy he a iabili y and s abili y o hese mechanisms o e
ime and equency [5], [8], [9].
2.3. Fea u e selec ion
The ea u e se was selec ed wi h a g eedy o wa d
selec ion echnique. S a ing wi h an emp y ea u e se ,
his algo i hm added he ea u e which lead o he highes
accu acy inc ease o he se a each i e a ion. The
pe o mances we e based on he alida ion se . When he
pe o mance did no inc ease u he , he algo i hm was
s opped. Candida e ea u es wi h a co ela ion coe icien
>0.6 wi h any o he ea u es al eady included in he se
ha e been emo ed o a oid possible co ela ion be ween
ea u es and edundancy o in o ma ion in o he se .
2.4. Classi ica ion
Due o i s simplici y, a decision ee classi ie was
implemen ed o a 3 classes disc imina ion: PV o o s,
ex a-PV LA o o s, and RA o o s.
All ex ac ed ea u es we e indi idually e alua ed wi h a
decision ee classi ie and a lea e-one-ou c oss- alida ion
echnique. Subsequen ly, wi h he ea u e se selec ed by
he g eedy echnique, a mul i- ea u e classi ica ion wi h
hold-ou c oss- alida ion was pe o med (70%, 15%, and
Table 1. Th ee single ea u es wi h he highes accu acy
o PV s. ex a-PV LA s. RA classi ica ion
Fea u e Accu acy (%)
RR
V CG 80.9
EDL
s RQA 80.4
EV L
icRQA480.0
15% o he o al da ase was andomly di ided in o aining
se , alida ion se , and es se , espec i ely). Sensi i i y
and speci ici y we e calcula ed o each class conside ing
he class a hand as posi i e and he emaining wo classes
as nega i e.
2.5. S a is ical analysis
The abili y o he ea u es in sepa a ing he di e en
classes was assessed using he he K uskal-Wallis non-
pa ame ic one-way analysis o a iance o a mul i-class
e alua ion. p- alues o less han 0.01 we e conside ed
s a is ically signi ican .
3. Resul s
3.1. Fea u es e alua ion
All ea u es showed an indi idual and signi ican
disc imina o y powe . Among hem all, RQA’s pa ame e s
ha e s ood ou pa icula ly well. Indeed, he mos
disc imina ing 3 indi idual ea u es we e: he ecu ence
a e ex ac ed om VCG (RR
V CG); he diagonal en opy
ex ac ed wi h s RQA (EDL
s RQAd; and he e ical en opy
ex ac ed wi h icRQA (EV L
icRQA4). Table 1 shows he
accu acy singula ly eached. These 3 ea u es showed
signi ican ly highe alues o 1R loca ed in he PV class,
ollowed by he ex a-PV LA class, and he RA class,
espec i ely (Fig. 2).
3.2. Ro o s loca ion classi ica ion
The decision ee classi ie achie ed 89.4% es
accu acy wi h a ea u e se o 18 ea u es, wi h sensi i i y
o 93.0%, 82.4%, and 83.3%, and a speci ici y o 95.2%,
77.8%, and 83.3% o RA, ex a-PV LA, and PV class
espec i ely. 5 selec ed ea u es ha e been calcula ed using
RQA me hods (including he 3 bes ea u es showed in
sec ion 3.1). 11 selec ed ea u es ha e been ex ac ed om
he a io o he PCA eigen alues app oach. Finally, 2
selec ed ea u es belonged o he Hjo desc ip o s. Table 2
shows he es -se con usion ma ix ob ained om he
decision ee using he ea u e se (class LA ep esen s
class ex a-PV LA).
Page 2
Figu e 1. A.1: Simula ed PV o o . B.1: Simula ed ex a-PV LA o o . C.1 : Simula ed RA o o . The ed a ows show he
o o posi ion and di ec ion. A.2-B.2-C.2: BSPMs o one o he 8 o so models gene a ed om MRI. The o so po en ial
was ob ained by sol ing he o wa d p oblem o elec ophysiology om he simula ed TMV on he a ia. A.3-B.3-C.3:
Example o he -wa e o lead I, II, and V1 om he 12-lead ECG signals ex ac ed om he BSPMs.
Table 2. Tes -se con usion ma ix o RA, ex a-PV LA,
and PV o o s classi ica ion
T ue class
RA LA PV
P edic ed class
RA 40 2 0
LA 3 14 1
PV 0 1 5
4. Discussion and Conclusions
Simula ions p o ide ideal and con olled scena ios
whe e he g ound u h o AF pe pe ua ion sus ained by
1R is known in all he cases. This allows he analysis
o each simula ion wi hou he in luence o seconda y, o
unknown, mechanisms, e.g., o he simul aneous o o s.
The RQA’s pa ame e s showed o be key ea u es o his
classi ica ion (Table 1). P obably due o hei sensi i i y
in de ec ing changes in he dynamic beha io o hese
mechanisms. In ac , looking also a he example ECGs in
Fig. 1A-B-C.3, ou simula ions ha e shown ha he ECG
signals a e mo e i egula in cases when 1R is no in he
PVs a ea. This in o ma ion was also quan i ied by some
RQA pa ame e s, ha ing signi ican ly highe alues o
he RA class, ollowed by he ex a-PV LA class, ending
wi h lowe alues o he PV class (Fig. 2). This can be
seen as con i ma ion o wha was sugges ed in a p e ious
s udy [5].
As men ioned abo e, he ECG signals in he case o 1R
no loca ed in he PVs a eas a e mo e i egula . All ea u es
ex ac ed we e aimed a de ec ing hese i egula i ies and
di e ences be ween classes. The esul s ob ained wi h
he hold-ou c oss- alida ion showed ha an au oma ic
Page 3
Figu e 2. Boxplo s o he 3 single ea u es wi h he
highes accu acy o PV ( ed) s. ex a-PV LA (blue) s.
RA (g een) o o loca ion classi ica ion. All ea u es a e
s a is ically di e en be ween he classes wi h p <0.01
classi ie wi h he ea u es ex ac ed in his wo k can
po en ially iden i y he a ea whe e a 1R is loca ed using
he 12-lead ECG.
The high sensi i i y and speci ici y alues ob ained
o all classes show ha his au oma ic classi ica ion
me hod ca ego izes mos o he cases in analysis in o he
co ec class. The e o e, i 1R was classi ied as a PV
case, doc o s could p oceed di ec ly wi h a PVI by c yo-
abla ion, wi hou using a p io i mapping sys em. In he
o he cases, a adio equency abla ion p ocedu e wi h a
p e ious mapping o he elec ical ac i i y o he a ium o
in e es would be equi ed.
The use o a non-in asi e echnique (i.e., 12-lead ECG),
in combina ion wi h machine lea ning app oaches, may
di ec ly sugges o he doc o he a ial egions ha may
be impo an in he AF pe pe ua ion, and hence a ge s
o abla ion. Fu he es s on clinical da a, labelled by
inspec ing he local ac i a ion maps, a e necessa y o
e ec i ely assess he p oposed app oach. A subsequen
s udy o p edic he ou come o PVI in cases whe e he AF
d i e is in PV is ongoing.
In conclusion, his wo k could be ex ended wi h a
p io cha ac e iza ion o di e en AF d i e mechanisms
and AF complexi y analysis. Se e al and mo e
obus classi ica ion algo i hms can be es ed and mo e
simula ions can be gene a ed wi h di e en a ial models.
Acknowledgmen s
The au ho s hank Debo ah Nai n o he aluable
sugges ions. Resea ch suppo ed by he Eu opean Union’s
Ho izon 2020 esea ch and inno a ion p og amme unde
he Ma ie Sklodowska-Cu ie g an ag eemen No.766082
(MY-ATRIA). TPA ecei ed suppo om he B i ish
Hea Founda ion (PG/18/33/33780 and BHF Resea ch
Accele a o ). All au ho s con i m ha hey ha e no o he
ela ionships ele an o he con en s o his pape o
disclose.
Re e ences
[1] Calkins H, e al. 2012 h s/eh a/ecas expe consensus
s a emen on ca he e and su gical abla ion o a ial
ib illa ion: ecommenda ions o pa ien selec ion,
p ocedu al echniques, pa ien managemen and ollow-
up, de ini ions, endpoin s, and esea ch ial design. Hea
Rhy hm 2012;9(4):632–696.
[2] Pandi S, Jali e J. Ro o s and he dynamics o ca diac
ib illa ion. Ci c Res 2013;112:849–862.
[3] Na ayan SM, Baykane T, Clop on P, Sch icke A, Lalani
GG, K ummen DE, Shi kuma K, Mille JM. Abla ion
o o o and ocal sou ces educes la e ecu ence o a ial
ib illa ion compa ed o igge abla ion alone. J Am Coll
Ca diol 2014;63(17):1761–1768.
[4] Haissague e M, Jais P, Shah DC, Takahashi A, Hocini M,
Quiniou G, Ga igue S, Mou oux AL, Me aye PL, Clémen y
J. Spon aneous ini ia ion o a ial ib illa ion by ec opic bea s
o igina ing in he pulmona y eins. N Engl J Med 1998;
339:659–666.
[5] Luongo G, Azzolin L, Ri ol a MW, Sassi R, Laguna
P, Dössel O, Loewe A. Non-in asi e iden i ica ion o
a ial ib illa ion d i e loca ion using he 12-lead ECG:
Pulmona y ein o o s s. o he loca ions. IEEE 2020 EMB
Con e ence 2020;410–413.
[6] Yang H. Mul iscale ecu ence quan i ica ion analysis o
spa ial ca diac ec oca diog am signals. IEEE T ans Biomed
Eng 2011;58(2):339–347.
[7] Luongo G, Schule S, Luik A, Almeida TP, So iano DC,
Dössel O, Loewe A. Non-in asi e cha ac e iza ion o a ial
lu e mechanisms using ecu ence quan i ica ion analysis
on he ECG: a compu a ional s udy. IEEE T ans Biomed
Eng 2020.;.
[8] Ja man JWE, Wong T, Kojodjojo P, Spoh H, Da ies
JER, Rough on M, F ancis DP, Kanaga a nam P, Dphil M,
Ma kides V, Da ies DW, Pe e s NS. O ganiza ional index
mapping o iden i y ocal sou ces du ing pe sisen a ial
ib illa ion. J Ca dio asc Elec ophysiol 2014;25(4):355–
363.
[9] Vakku i A, Yli-Hankala A, Talja P, Mus ola S, Tol anen-
Laasko H, Sampson T, Vie iö-Oja H. Time- euqency
balanced spec al en opy as a measu e o anes he ic d ug
e ec in cen al ne ous sys em du ing se o lu ane, p opo ol,
and hiopen al anes hesia. Ac a Anaes hesiol Scand 2004;
48(2):145–153.
Add ess o co espondence:
Gio gio Luongo, Ka ls uhe Ins i u e o Technology (KIT)
F i z-Habe -Weg 1, 76131 Ka ls uhe, Ge many
[email p o ec ed]
Page 4