Improving the clinical understanding of hypertrophic cardiomyopathy by combining patient data, machine learning and computer simulations: A case study

Mo phologie (2019) 103, 169—179 Disponible en ligne su ScienceDi ec www.sciencedi ec .com ORIGINAL ARTICLE Imp o ing he clinical unde s anding o hype ophic ca diomyopa hy by combining pa ien da a, machine lea ning and compu e simula ions: A case s udy Combine données cliniques, machine lea ning e modélisa ion in o ma ique pou une meilleu e comp éhension clinique de la ca diomyopa hie hype ophique A. Lyona,b, A. Mincholéa, A. Bueno-O o ioa, B. Rod igueza,∗ aDepa men o Compu e Science, Uni e si y o Ox o d, Ox o d, Uni ed Kingdom bCa dio ascula Resea ch Ins i u e Maas ich (CARIM), Maas ich Uni e si y, Maas ich , Ne he lands A ailable online 27 Sep embe 2019 KEYWORDS Hype ophic ca diomyopa hy; Elec oca diog aphy; e-ca diology; Pheno yping; Compu a ional clus e ing; Compu e modeling; Pe sonalized simula ions; Ca diac magne ic esonance imaging; VPH Summa y Mos pa ien s wi h hype ophic ca diomyopa hy (HCM), he mos common gene ic ca diac disease, emain asymp oma ic, bu o he s may su e om sudden ca diac dea h. A be e iden ifica ion o hose pa ien s a isk, oge he wi h a be e unde s anding o he mechanisms leading o a hy hmia, a e c ucial o a ge high- isk pa ien s and p o ide hem wi h app op ia e ea men . Howe e , his cu en ly emains a challenge. In his pape , we p esen a success ul example o implemen ing compu a ional echniques o clinically- ele an applica ions. By combining elec oca diog am and imaging da a, machine lea ning and high pe o mance compu ing simula ions, we iden ified ou pheno ypes in HCM, wi h di e ences in a hy hmic isk, and p o ided wo dis inc possible mechanisms ha may explain he he e o- genei y o HCM mani es a ion. This led o a be e HCM pa ien s a ifica ion and unde s anding o he unde lying disease mechanisms, p o iding a s ep u he owa ds ailo ed HCM pa ien managemen and ea men . © 2019 The Au ho s. Published by Else ie Masson SAS. This is an open access a icle unde he CC BY license (h p://c ea i ecommons.o g/licenses/by/4.0/). Résumé La plupa des pa ien s a ein s de ca diomyopa hie hype ophique (CMH), une mal- adie ca diaque géné ique, son sou en asymp oma iques, mais ce e maladie es la cause p emiè e de mo subi e chez les jeunes. Une meilleu e iden ifica ion des pa ien s à isque, e une meilleu e comp éhension des mécanismes esponsables des a y hmies es essen ielle ∗Co esponding au ho : Depa men o Compu e Science, Uni e si y o Ox o d, Ox o d, Uni ed Kingdom. E-mail add ess: [email p o ec ed] (B. Rod iguez). h ps://doi.o g/10.1016/j.mo pho.2019.09.001 1286-0115/© 2019 The Au ho s. Published by Else ie Masson SAS. This is an open access a icle unde he CC BY license (h p:// c ea i ecommons.o g/licenses/by/4.0/). 170 A. Lyon e al. pou iden ifie ces pa ien s à hau isque e leu ou ni un ai emen app op ié. Cependan , cela es e un défi. Dans ce a icle, nous p ésen ons un exemple éussi de l’implémen a ion de echniques in o ma iques pou une applica ion clinique. En associan données d’image ie e d’élec oca diog aphie, machine lea ning e modélisa ion in o ma ique, nous a ons iden- ifié qua e phéno ypes pa mi la popula ion de CMH, p ésen an di é en s ni eaux de isque d’a y hmie e de mo subi e, e a ons p oposé deux mécanismes possibles pou explique l’hé é ogénéi é obse ée dans la CMH. Ce e é ude a donc amelio é la s a ifica ion du isque chez les pa ien s a ein s de CMH e une meilleu e comp éhension des mécanismes sous-jacen à la maladie, ou nissan une a ancée e s une p ise en cha ge indi iduelle de ces pa ien s e un ai emen pe sonnalisé. © 2019 Les Au eu s. Publi´ e pa Else ie Masson SAS. Ce a icle es publi´ e en Open Access sous licence CC BY (h p://c ea i ecommons.o g/licenses/by/4.0/). In oduc ion Ca dio ascula diso de s emain a majo bu den wo ldwide and a e esponsible o 30% o dea hs in he wo ld. Among hem, hype ophic ca diomyopa hy (HCM) is a gene ic ca - diac disease cha ac e ized by he hickening o he le en icula muscle o he hea , and i is a majo cause o sudden ca diac dea h (SCD), especially among young adul s and a hle es [1]. Mos pa ien s wi h HCM emain asymp- oma ic wi h no mal li e expec ancy, bu some o hem may die suddenly o ca diac a es wi h no p e ious signs. Ge ing a be e clinical unde s anding o his he e ogeneous clin- ical cou se and de ec high- isk pa ien s o p o ide hem wi h app op ia e ea men is he e o e a challenge and a p io i y in he managemen o HCM [2]. HCM hea s su e om s uc u al changes such as hype - ophy [3], ca diomyocy e disa ay [4], fib osis [5], as well as ion-channel dys unc ion [6], which may c ea e an uns a- ble elec ical milieu p edisposing o a hy hmia. In he pas , non-in asi e ools o assess he he e ogenei y o he HCM popula ion ha e been de eloped based on he elec oca - diog am (ECG) [7,8], bu hey lack specifici y o iden i y he pa ien s a highe isk [9,10]. In he absence o eliable ECG bioma ke s, con en ional isk ac o s (non-sus ained en- icula achyca dia, unexplained syncope, amily his o y o SCD, massi e le en icula hype ophy and abno mal exe cise blood p essu e esponse) a e also used o e al- ua e he a hy hmic isk o he pa ien s, and a alida ed HCM Risk-SCD p edic ion model has been p oposed in 2014 [11], bu s ill shows limi a ions [12]. Mo eo e , hese clin- ical obse a ions do no cap u e he amoun o unde lying myoca dial abno mali ies ha may lead o a hy hmia, and he pa hophysiological mechanisms ha may inc ease HCM isk a e s ill poo ly unde s ood. In his ansla ional case s udy, we p esen how he use o compu a ional me hods such as machine lea ning and high pe o mance compu ing simula ions helped imp o e he isk s a ifica ion o HCM pa ien s and shed ligh on he mechanisms unde lying he HCM disease. Fig. 1 p o ides an o e all summa y o he findings and po en ial clinical impac o he case s udy p esen ed in his pape . In a fi s pa , we epo how he de elopmen o a clus e ing algo- i hm based on no el mo phological bioma ke s de i ed om he ECG helped iden i y ou dis inc pheno ypes among he HCM popula ion, which exhibi ed di e ences in a hy hmic isk and dis ibu ion o le en icula hype ophy. In a second pa , we show how he use o high pe o mance compu ing simula ions based on clinical ca diac magne ic esonance (CMR) images p o ided mechanis ic unde s and- ing o he di e en ECG pheno ypes, and helped o be e unde s and he he e ogenei y o HCM. This pape desc ibes how, by de eloping no el compu a ional echnologies o analysis, in eg a ion and augmen a ion o clinical da a, we con ibu ed o ad ancing clinically- ele an insigh in o HCM and made a s ep owa ds indi idual pa ien managemen . Me hods and esul s ECG-based mo phological ma ke s iden i y ou dis inc HCM pheno ypes ha associa e wi h di e en a hy hmic isk In he fi s pa o his p ojec , we aimed a de eloping ma hema ical modeling and machine lea ning me hods o iden i y ECG bioma ke s ha may help imp o e he unde - s anding and cha ac e iza ion o he he e ogenei y o he HCM popula ion [13]. To his end, we analyzed high-fideli y ECG eco dings om 85 HCM pa ien s and 38 heal hy olun ee s ec ui ed as pa o a p ospec i e s udy om he John Radcli e Hospi al in Ox o d, UK. 12-lead ECGs, measu ing he ca - diac elec ical ac i i y om wel e di e en pe spec i es on he body su ace, we e eco ded o 24 hou s using Hol e moni o s. CMR imaging was also pe o med o hese pa ien s and p o ided in o ma ion on he ex en and dis- ibu ion o he hype ophy. Finally, gene ic in o ma ion, con en ional isk ac o s and amily his o y we e ob ained as pa o ou ine examina ion. We hen de eloped signal p ocessing and ma hema ical modeling ools o compu e bioma ke s om he di e en ECG wa es in o de o cha - ac e ize hei mo phology. We ocused specifically on he QRS complex, ep esen ing he elec ical ac i a ion o he en icles, and he T wa e, cha ac e izing en icula elax- a ion. S anda d bioma ke s such as ampli ude and wid h o he wa es we e measu ed. The QRS shape was also cha ac- e ized by a combina ion o He mi e basis, well-es ablished ma hema ical unc ions able o p o ide a compac desc ip- ion o he QRS complex. Fou He mi e unc ions allowed o cap u e HCM he e ogenei y, and each pa ien was cha- ac e ized by a ec o o mo phological QRS and T wa e Imp o ing he clinical unde s anding o hype ophic ca diomyopa hy 171 Figu e 1 App oach and clinical impac o he case-s udy. Using signal p ocessing, ma hema ical modeling and clus e ing, we iden ified ou di e en pheno ypes in hype ophic ca diomyopa hy based on he ECG. This had a clinical impac by p o iding imp o ed pa ien isk s a ifica ion. High pe o mance simula ions hen in es iga ed he po en ial mechanisms unde lying hese pheno ypes. This p o ided new op ions o indi idual pa ien managemen and di e en he apeu ic app oaches. bioma ke s. We used an unsupe ised ea u e selec ion app oach combined wi h a clus e ing algo i hm o in es i- ga e and ex ac subg oups om he HCM popula ion. This was pe o med blinded o he clinical da a. S a is ical anal- ysis was finally pe o med o compa e he isk ma ke s be ween he g oups. Fig. 2 summa izes he me hodological app oach aken. Based on QRS mo phology only, h ee HCM subg oups we e iden ified. Pa ien s in G oup 1 (52% o he popula- ion) displayed no mal QRS mo phology. G oup 2 (22% o pa ien s) showed di e ences in he fi s h ee He mi e bases in lead V4 compa ed o heal hy olun ee s and G oup 1, bu no di e ences in V6. On he ECG, his was obse ed as a sho e R wa e du a ion and deepe S wa es in lead V4 compa ed o G oup 1. G oup 3 (26% o pa ien s) exhib- i ed la ge di e ences in lead II and V4—V6 compa ed o he o he HCM g oups and heal hy olun ee s, mo e specifically sho R wa e du a ion and ampli ude, and longe S wa e du a ion and ampli ude. Among hese h ee g oups, ECG ea u es we e he e o e significan ly di e en , bu clinical ma ke s and ma ke s o a hy hmic isk we e no . This sug- ges ed ha QRS bioma ke s alone we e no su ficien o HCM isk s a ifica ion. We hen combined bo h QRS mo phology and T wa e bioma ke s in he clus e ing algo i hm. This led o he iden ifica ion o ou dis inc subg oups. G oups 2 and 3 emained he same han he ones iden ified wi h QRS ea u es only. Howe e , he addi ion o he pola i y o he T wa e as a bioma ke sepa a ed G oup 1 in o G oup 1A, wi h in e ed T wa es in leads V4—V6, and G oup 1B, wi h up igh T wa es in hese leads (Fig. 3, Panel A). In e es ingly, G oup 1A, wi h no mal QRS mo phology bu in e ed T wa es, showed he highes median HCM Risk-SCD sco e among he ou g oups (Fig. 3, Panel C). G oup 1A also exhibi ed he mos pa ien s wi h a mixed hype ophy dis ibu ion, wi h combined sep al and apical hype ophy, while G oups 2 and 3 had mos ly sep al hype ophy only. G oup 1B pa ien s exhibi ed li le o no hype ophy (Fig. 3, Panel B). This s udy he e o e iden ified ou dis inc HCM phe- no ypes based on newly de eloped ECG mo phological bioma ke s ex ac ed om high-fideli y eco dings. Such pheno ypes showed di e ences in a hy hmic isk sco es and dis ibu ion o le en icula hype ophy. Thus, ou s udy showed he po en ial o using ECG pheno yping combined wi h machine lea ning o dissec he he e ogenei y o HCM and help imp o e indi idual pa ien managemen . High pe o mance compu e simula ions based on CMR images p o ide di e en mechanisms o ECG abno mali ies in HCM The p e ious s udy iden ified ou subg oups in he HCM popula ion ha showed di e ences in a hy hmic isk and 172 A. Lyon e al. Figu e 2 Me hodological app oach o he iden ifica ion o ou HCM pheno ypes ( om [13]). dis ibu ion o hype ophy. Following up on his wo k, ou nex objec i e was o unde s and he mechanisms behind such a pheno ypic he e ogenei y, and p o ide po en ial explana ions o he di e en subg oups iden ified [14]. Clinically, imp o ing he mechanis ic unde s anding o HCM may yield a key impac on he indi idual managemen o hese pa ien s, including po en ial ea men and ailo ed he apies. To his end, we de eloped a high pe o mance com- pu e simula ion amewo k based on CMR imaging da a. We selec ed ep esen a i e pa ien s om G oup 1A (wi h no mal QRS mo phology and in e ed T wa es, and high- es isk sco e o sudden ca diac dea h), G oup 1B (wi h no mal ECG mo phology) and G oup 3 (wi h abno mal QRS, no mal T wa e, and second highes isk sco e). F om he CMR images o hese pa ien s, we compu ed a pe sonal- ized olume ic mesh o each pa ien ’s hea and o so. The elec ical ac i i y ac oss he en icles was hen defined by implemen ing a cellula compu a ional ac ion po en ial model [15] a each node o he olume ic mesh, o simu- la e p opaga ion o he elec ical signal. Vi ual elec odes we e addi ionally modelled o eco d he simula ed ECG on he i ual pa ien s. Wi h his simula ion amewo k, ou aim was o ep oduce he ECG abno mali ies iden ified in he di - e en HCM pheno ypes o he p e ious s udy, and p o ide po en ial mechanis ic explana ions o his ECG he e ogene- i y using compu e simula ions. We he e o e e alua ed he influence o a ious abno mali ies epo ed in HCM on he ECG mo phology, including inc eased wall hickness, ca - diomyocy e fibe disa ay, changes in conduc ion eloci y, ionic emodeling o abno mal coupling be ween he Pu kinje as conduc ion laye and he myoca dium. These di e en simula ions allowed us o in es iga e he indi idual e ec s o hese HCM abno mali ies on he ECG and iden i y hose ha may be esponsible o he HCM phe- no ypes iden ified. The dis ibu ion o hype ophy and he Imp o ing he clinical unde s anding o hype ophic ca diomyopa hy 173 Figu e 3 Fou pheno ypes we e iden ified in HCM, exhibi ing di e ences in ECG mo phology (A), dis ibu ion o hype ophy (B) and a hy hmic isk (C) ( om [13]). ana omy o he pa ien yielded simila QRS mo phology and no mal T wa es in all cases (Fig. 4). The e o e, inc eased wall hickness could no explain he QRS and T wa e abno - mali ies obse ed in G oups 1A and 3. We hen ocused on he impac o issue mic os uc u e on he ECG, and e al- ua ed he e ec o fibe disa ay and al e ed conduc ion eloci y due o fib osis o hype ophy in a ious egions o he myoca dium (such as he sep um, o he apex). These led o abno mal QRS complexes, bu did no ansla e in o he specific deep S wa es in lead V6 ha cha ac e ized G oup 3 (Fig. 5). We hen ocused on he influence o al e ing he conduc ion sys em by modi ying he ac i a ion sequence o he en icles, and he coupling be ween he Pu kinje endoca dial laye and he myoca dium. Simula ing a ious conduc ion blocks a ec ed he QRS complexes mo phology bu did no lead o he abno mali ies o G oup 3 (Fig. 6). 174 A. Lyon e al. Figu e 4 E ec o he ana omy on he ECG. G oups 1B, 3 and 1A exhibi simila ECG mo phologies despi e di e ences in ex en and dis ibu ion o hype ophy. Ana omy alone does no explain G oup 3 and 1A specific abno mali ies. Taken om [14]. Howe e , modeling an abno mal coupling be ween he as endoca dial laye and he myoca dium was he only way o simula e he deep S wa es in leads V4—V6 iden ified in G oup 3 (Fig. 7, Panel D), by c ea ing a pa chy elec ical ac i a ion wi h a eas o la e ac i a ion (Fig. 7, Panels A o C). Finally, we modelled he HCM ionic emodeling (Fig. 8), including an inc ease o he la e sodium and he L- ype calcium cu en s, a educ ion o he po assium cu en s, and emodeling o he calcium handling subsys em, in hype ophic a eas. This led o a p olonged du a ion o he ac ion po en ial in hese egions and ansla ed in o in e sed T wa es on he la e al leads o he ECG, explaining he pheno ype o G oup 1A. These simula ions he e o e iden ified wo dis inc po en- ial mechanisms o he ECG abno mali ies associa ed wi h an inc eased isk o SCD in HCM. They also sugges ed ha he HCM ionic emodeling exp essed in G oup 1A, subg oup wi h he highes SCD isk sco e, may play a key ole in a hy hmogenesis. The na u e o hese mechanisms is e y di e en : one is based on conduc ion abno mali ies, while he o he in ol es ionic emodeling. This has implica ions on he di e en possible he apies o hese g oups o pa ien s. Indeed, while G oup 1A may benefi om a pha macological ea men a ge ing he exp ession o ion channels, G oup 3 may no espond o such he apy. Discussion Combining compu a ional me hods imp o es he clinical unde s anding o ca diac diseases In his pape , we epo an example o success ul imple- men a ion o compu a ional echnology o a clinical applica ion in ca diology. The compu a ional me hods conside ed he e combine signal p ocessing, ma hema ical modeling, machine lea ning and high pe o mance com- pu ing, and hey con ibu e o un angle he he e ogenei y o HCM and p o ided mo e insigh in i s mechanisms. As illus a ed in ou case s udy, he powe o hese echniques is wo old. Fi s ly, hey a e able o make sense o mul i- a ia e, complex and he e ogeneous da ase s and de ec di e ences ha migh be challenging o he human eye Imp o ing he clinical unde s anding o hype ophic ca diomyopa hy 175 Figu e 5 E ec o conduc ion changes in a ious egions o he myoca dium on he QRS complex ( om [14]). [16]. These echniques make ewe assump ions by selec ing disc imina o y ea u es om he whole ECG da a. Wi h he g owing amoun o eco ded da a in clinical se ings, he in eg a ion o hese me hods in he clinic may be c ucial o aid heal hca e decisions and imp o e pa ien s a ifica ion in la ge coho s. Secondly, compu a ional echniques, such as compu e modeling, allow he independen assessmen o he influence o indi idual pa ame e s o ming a sys em. This is a key s eng h compa ed o s anda d expe imen al echniques, o which isola ing pa ame e s o s udy hei e ec emains a challenge. As a consequence, compu- a ional echniques a e ecei ing a g owing a en ion 176 A. Lyon e al. Figu e 6 E ec o ac i a ion blocks in a ious egions o he myoca dium on he QRS complex ( om [14]). o analyze medical da a and add ess clinical p oblems [17—19], and despi e ha se e al challenges emain, mos ly due o he na u e o eal-wo ld da a (p in -ou ECGs, incomple e da ase s, need o expe consensus), hey can help unco e new mechanisms, imp o e disease knowledge, guide he apies and aid clinical decisions. Clinical impac o he wo k As men ioned p e iously in he ex , his wo k showed clin- ical implica ions in wo ways. Fi s , i p o ided a new classifica ion o HCM pa ien s based on ECG bioma ke s, p o- iding insigh in he he e ogenei y o he disease. We showed Imp o ing he clinical unde s anding o hype ophic ca diomyopa hy 177 Figu e 7 Abno mal Pu kinje-myoca dium coupling led o a pa chy ac i a ion (A, B) wi h a eas o la e ac i a ion (C), and ansla ed in o deep S wa es in lead V6, explaining G oup 3 abno mali ies. Taken om [14]. ha HCM pa ien s wi h a p ima y T wa e in e sion (and no mal QRS) ha e a g ea e isk o SCD and a hy hmia, com- pa ed o pa ien s wi h solely QRS abno mali ies, highligh ing he key ole o epola iza ion in a hy hmogenesis in HCM. We also showed ha he loca ion and dis ibu ion o hype - ophy was associa ed o highe SCD isk, a he han he ex en o hype ophy. Secondly, he use o compu e simu- la ions p o ided wo dis inc mechanisms ha may explain he di e en pheno ypes we iden ified in HCM. Abno mal Pu kinje conduc ion can explain he QRS abno mali ies o G oup 3, while HCM ionic emodeling in he egion o hype - ophy may be esponsible o in e ed T wa es in G oup 1A. The ac ha G oup 1A displayed he highes SCD isk sco e sugges ed ha ionic emodeling in HCM may play a key ole in a hy hmogenesis, while he conduc ion abno - mali ies in he Pu kinje sys em may be less p oa hy hmic. Finally, his o e all be e unde s anding o HCM he e o- genei y may ha e implica ions in pa ien managemen , as pe sonalized ea men o HCM pa ien s could be p o ided depending on he mechanisms unde lying hei ECG abno - mali ies. Indeed, a pa ien om G oup 1A may benefi om a pha macological ea men imp o ing ion channels unc- ion, while his would ha e no e ec on a pa ien o G oup 3. Cu en and po en ial de elopmen s o he wo k De elopmen s o he wo k p esen ed in his pape a e cu - en ly ongoing. A new da abase o HCM ECG eco dings is being collec ed and will enable he applica ion o he mod- eling and clus e ing echnology on a la ge da ase . This will allow us o e alua e he ECG c i e ia p esen ed in he s udy on a la ge coho wi h a la ge numbe o ca - dio ascula end-poin s and e en s. This could lead o he defini ion o a new classifica ion c i e ion used in he clinic o isual inspec ion o ECGs in HCM. Mo eo e , his wo k gene a ed s ong mo i a ion o deepe clinical in es iga- ions, such as he use o endoca dial o epica dial mapping s udies o e i y he hypo hesis p esen ed he e o he mech- anisms esponsible o he HCM pheno ypes. The use o new echnologies such as ECG imaging may add e en mo e in o - ma ion o hese findings.