A Novel end-to-end Digital Health System Using Deep Learning-based ECG Analysis

Preprint. This manuscript has been submitt ed to the International Journal of Information Managem ent Data Ins ights for possible publication. This version has not undergone peer revie w. 1 A Novel e nd - to -end Digital Health Syste m Using Deep Learning -based ECG Analysis Artemis Kontou 1 , Natalia Miroshnikova 2 , Costakis Matheou 2 , Sopho cles Sopho cleous 2 , Nicholas Tsekouras 2 , Kleanthis Malialis 1 , Panayiot is Kolios 1, 3 1 KIOS Research and Innovation Cent er of Excelle nce (K IOS CoE), University of Cyprus, Pane pistimi ou 1, Aglantzia 2109, Cyprus 2 MEDTL Medical Technologies Ltd., 91 Limassol Avenue, Aglantzia , O ffice 202, Nicosia 2121, Cyprus 3 Departm ent of Computer Science, University of Cyprus , Panepisti miou 1, A glantz ia 2109, Cyprus Corresponding author: A. Kontou (e - mail: kontou.artemis@ucy.ac.cy ). ABSTRACT This study present s AI-HEART, a cloud-base d information system for managing and analysing long -duration ambulatory electrocardiogram ( ECG) recordings and supporting clinician decision -making. T he platf orm operationalises an end - to -end pipeline that ingests multi-day three-lead ECG s, normalises inputs, performs signal preprocessing, and applies dedicated d eep neural networks for wa ve de lineation, noise /quality detection, and beat - a nd rhy thm-level mul ti-class arrhythmia class ification. To address class imbala nce and real-world signal variability, model development combines large c linically annotated dataset s with expe rt- in -the-loop curation a nd ge nerative augmentation for under -represented rhythms. Empirical evaluation on three - lead ambulatory EC G da ta shows that delineation acc uracy is sufficient for automated interval measurement, noise detection reliably flags poor-quality segments, and arrhythmia class ificati on achieve s high specificity with clinically us eful macro - averaged perfor mance a cross common and ra rer rhythms. Beyond pr edictive ac curacy, AI -HEART provides a scalable deployment approach for integrating AI into routin e ECG s ervices, e nabling traceable outputs, audit -friendly storage of recordings and derived annotations, and c linician review/editing t hat c aptures feedback for controlled model improvement. The findings demons trate the technical feasibility and operatio nal value of a no ise -aware AI-ECG platform as a d igital health information system. Keywords: Arrhythmia detection, Deep learning, Electrocardiography, Clinical decision support, Digital health information systems Acknowledgement: This work has been funded by the Rese arch and Innovation Foundation (RIF) of Cyprus under the RESTART 2016 – 2020 Programmes for Research, Technologic al Development and Innovation, through the «DISRUPT» Programme, Call DISRUPT/0123(C), P roject AI-H EART (Proj ect No. DISRUPT/0123(C)/0014), which is co-financed by the Recovery and Resilience Facility of the Eur opean Union and the Re public of Cyprus . 1. Introdu ction Cardi ovascu lar dis e ase s ( CVD s ) rem ain the lead ing cau se of mor tali ty glo bally , acco unt ing for an estim ated 18 .6 millio n d eath s annua lly [1]. Thi s burd en dispr op orti onat ely impa cts h ealth sy s tem s in lo w - a nd middl e - inc om e coun tri es, wher e acce ss to spe ciali st cardi olo gy servic es i s often cons tra in ed [2]. The exp andin g use of ambula tory electr ocard iog ra m (ECG) mon itoring furth er incr eases d ia gnos tic wor kl oad, m aking tim ely and s t anda rdi s ed re vi ew of lar g e ECG d a t aset s a pr actica l n ece ssit y [3]. In re spon se, d eep l earni ng h as r apidl y advan ced auto mate d ECG interpr etati on , enabl i ng scala ble det ectio n of arrh yth mi as and condu ction abnor malit ies fr om larg e volu mes of rea l-worl d re cor din gs. H owe ve r , dep loyme nt in rou tine clini cal wor kf lo ws rem ains con str ained by varia bl e sign a l quali t y, noise and artef acts, and the need for r ob ust perf or man ce acros s h eter oge neous devic es and rec ordi ng con dit ion s [4] . T hese ch all enge s a r e p art icular ly acu te in ambu la t ory moni toring , wher e recor din g s ar e typi call y a cquir ed cont inuo usly usin g two or three lead s over ext ended peri od s, a nd clin ic ally rel e vant event s m ay b e i ntermitt ent and e asi ly mi ssed in bri ef snap shot s. D evel oping noi s e-rob ust deep lear ning m od els tailor ed to t hre e -l ead a m bulat ory ECG ther ef ore addr e ss es a prac ti cal bottl ene c k f or r eal-w orld scr eenin g a nd m onitor ing: accur ate mul ti-cl ass d ete ctio n und er n on- id eal condit ions , wit h out puts th at c an s up port ti mely cl ini cal r evie w a nd prior itis ati on. Early ru le- bas ed alg orit hm s a nd shallo w ma c hi ne-l ear nin g method s have been l argel y super seded by conv olutio na l neu ra l networ ks (CNN s ) and recur re nt neu ral n etwor ks ( RNN s ), which lear n hierar chical fe atur e s dire ctl y fro m ra w or minim ally pro ces sed ECG data [5] . For exam pl e, Hann un et al. dem onstr at ed card iologi st-l evel mul ti- cl ass arr hy thm ia cla ssif icat ion fro m singl e-le ad ambu lato r y E CG re cor din gs usin g an e nd- to - e nd deep neur al net work, whi le Rib eiro e t al. show ed tha t r esi du al convo lution a l net wor ks (R esN et - style CNN s) can a c hie ve hig h -a ccura cy m ulti-l abel cla ssif icati on f ro m 12-le ad E CGs at scale [5 , 9]. S emi nal work by J i n e t a l. [ 7] in trodu c ed an i nterpr etabl e, kno wledg e -fu sed dee p n eural networ k fo r sin gle-le ad (lea d I I) ECG that ach ieve d p erfor mance compar able to o r e xce edi ng cardio lo gi sts fo r mul ti -cl ass arrhyt hmia d iagno sis, and s ub s equ e nt stud ies have a dapt ed archi te ctur es such a s ResNe t and De nseNet for mu lti -l ead ECG clas si fic ation , achiev ing st ate-of-t he-ar t resul ts on publ ic data set s incl uding PT B-XL [ 8, 9] . Bey on d rhythm disor d ers, dee p lear nin g mode ls have shown remar ka ble capab ility in extr acti ng "hi dden" biomar kers from stand ar d ECG s. Nota bly , an AI -E C G algor ithm for det ectin g asympto matic low ejecti on frac tio n (LEF ) achi eved an area under t he rec e iver oper ating ch aract erist ic cur ve (AUC) of ~0.93 and has rec ei ved U.S. F DA 510( k) cl eara nce, p a vi ng th e way f or r egul atory a ppr o val of AI - der iv ed d iagnos tic soft ware [1 0, 11] . Translational progress is evidence d by the emergence of commercial a nd near -commercial AI-ECG platforms integrated into clinical workflows. These include cloud-base Software as a Medical D evices (SaMD) for ambulatory recordings analysis such as DeepRhythmAI, C ardiologs and B eatLo gic that provid e auto mate d, h igh-a c cur acy report in g for m ajor a r rhy thm ia s in real- wor ld setti ngs [6 , 12, 13] . S uch pl atforms have mov ed beyo nd proof-of- concept to demonstrate pr actical utility in hos pital- based decision s upport. R egulatory bodie s, including the FDA and Eur opean CE marking authorities, ha ve established pathways for SaMD, with several AI-ECG algor ithms now cleared for clini cal use, unde rscoring the field's maturation [14]. Desp it e th e r emar kabl e pro gres s in algor it hmi c perf or ma nce and r egula tory ad vanc e m ent, the tran slat ion of AI-ECG re sear ch into robu st, ge ne r aliz able, an d clini cally ver s atil e tool s f ace s sever al int ertwin e d and per sist e nt l imi tat io ns. A si gnif i can t porti on of commer cial and re sear ch effor ts rem ains n arro wly optim ized f or singl e , high-pr eva le nce cond ition s such as AF [ 15, 16] . Whil e effe ctiv e for tar get ed scre en ing, th is focu sed dev elo pment leave s a sub stant ial diagno s ti c gap for th e broa de r spe ctru m of arrh yth mi as and condu ction dis ord er s i nclud ing comp lex ectopy, supr av e ntr ic ular tac hyc a rdi a, and va r iou s d egr ees of atri ov entr icul ar bl ock t hat ar e esse nti al for com pre hen s ive EC G interpr etati on in gener al pra c ti ce. Mod e ls tra in ed on limi ted rhyt hm set s oft en f ail t o gener aliz e to the wi der dif fe re nti al di agn ose s req uired in r outi n e cli nica l wor kflo ws, lim iting th e ir u tility as all- purp ose d eci s ion- suppor t t ool s. Furth e rm ore , th e cha lle ng e of re al- wor ld signal quali ty i s fre quent ly under- addr essed . Many rep ort e d high- pe rform ing E C G AI mod els are dev el op ed and valid a ted on rela tivel y clean or cur at ed d ata s ets w her e nois e is com monl y han dled indir ectly (e. g. , quali ty f ilt eri ng/ excl u sion) r at her t han being mod el ed a s a fir st-clas s p art of the p ipeli ne , whi c h can li mi t r eal-w orld pe rf orm ance [4, 1 7]. Thi s over sigh t i s criti cal, as m oti on artif acts, ba seli ne wa nde r, and el ectro de c ont act i ssue s are ubiqui to us in ambul atory, emer gen c y, and pri mar y car e sett ings. The lack of inte gra ted, lear nabl e no ise - dete c ti on m odu les le av es syste ms vul nera ble to perf orm a nce degr adatio n and erro neou s inter pre tati o ns when face d wit h poor-qual ity r ecordi ngs, u nd erm ini ng their reli abilit y out side cont r oll ed enviro nme nt s [18] . Equa lly conc e r nin g i s the stati c nat ure of m any AI-E CG solut ion s, whi ch ar e oft en d evelo ped a s fi xed pr oduct s wit hout b uilt - in mecha nisms for c onti nuo us l earnin g and life c ycl e man agem ent [1 9]. Post- deploym ent, mod e l s can s uff er from data set shift, str uggle with rare but cli ni cally criti cal arr hyt hmi as due t o class imbal anc e , a nd lack path wa ys to inc orpora te new exp ert kno wledg e . Few pl atf orm s de s crib e a regul ator y-gra de fr a me work that suppo rt s sy stem atic dat aset curat io n, e xper t - in - th e-loop feedb ack, and ver s io n-cont rolle d mode l update s capabil itie s that are essenti al for maint ainin g sa f ety and eff i cacy ov er time an d ali gning with e vol ving r egulato r y exp ectati ons f or ad aptiv e AI. Addit i onall y, many existi ng solut ions ar e tig htl y coup led to s pe c ifi c a cqu isiti on har dwar e, propri etary data form ats, or hi gh - re sourc e he althc a r e IT eco syst ems [20]. This hardw ar e and softwa re dep ende nce create s intero perabi lit y barri ers, limi ts scal abilit y acro s s d iver se clinic al setting s par ti cular ly in low-r esourc e env ironm e nt s and hinder s int egrati on into h eterog eneou s hospit al infor matio n sys te ms. The vi sion o f broad ly acce ssi bl e AI-a ssi sted E C G an a l ysis r equi r es platf orm- agno s ti c, cl oud - bas ed archi te c tur es th at can inge st s tan da r di zed d ata from d iv er se sourc es and scal e effi cien tly acro ss diff erent he a lt hcar e infr astruc ture s [21]. Thes e colle c ti ve gap s narr ow scope , noi s e sensi tivity , stati c de sign, and ecosy st em d epen dence curr entl y hi nder t he devel opme nt of AI-ECG syst ems that are truly fit for widespr ead, susta in able, and tru st ed inte grati on into gl obal healt hcar e deliv ery. Th is study is de signe d to addr ess these lim it ation s dire ctl y. It propo ses a clou d-b ase d Sa MD pla tfor m call ed AI-H e art that int egr ate s a suit e of dee p le a rn ing mod el s f or deli ne ation , nois e det ecti on, and mu lti- cla ss arrh yth mi a c l as sifi catio n into a coh esive, end- to - end workf l ow. It s archit ectur e expl icitly prior iti zes br oad rh ythm c over age, expli cit noi s e robu stnes s th rou gh dedic a t ed noi se-h a ndli ng compo nent s , a nd a r egul atory-gr ade f ra mewor k supp orting c ont inu ous l earni ng vi a gener ativ e dat a augm entati on and exp ert- in -th e-l oop cura tion tool s . By adop tin g a har dwar e -a gno stic , scal abl e cloud de plo ym ent mod el, the pla tfor m aim s to be s uit ab le f or di vers e clini cal sett ings, f r om hosp ita l s to u nder serve d r e gi ons. The ob jecti ve of t hi s pa per i s tw ofo ld. Fir s t , w e ex amine the tech nical fea s ibi lit y a nd in ter nal perfor m ance of t he pl a tf orm as a cloud- based system that int egr ates disti nct deep-l ear ni ng mod els for ECG deli neati on, noise det ection, beat-le vel clas sif icati on and rhythm- leve l arrhy thmia cla s sifi catio n into a s ingl e, end- to - end workfl ow. Seco nd, we ana ly se how t hi s ar chi tec tur e func tio n s as a digit al h ealth infor mati on sy s t em, foc using on its abil ity t o inge s t and m anag e long- dur ati on thre e -l ead ambul a tor y E CG data , suppo rt hum an- in -t he- loop de cisio n -m aking and int e rfa ce with existin g clin ical rep orti ng pro cess e s. We descr ib e the sys tem archi te c tur e and pr oce ssi ng pipeli ne, r epor t quant itati ve perf orm an ce for eac h mod el com pone nt on intern a l d atas ets a nd situ a te the pla tfor m w it hin c urr en t A I-E CG and digit al hea lth lit era tur e , wit h par ti cular empha sis on noise robu s tne ss, compr eh ensi ve rhyt hm co ver age and d eploym ent as a sca la bl e, clou d -hos ted de cision- suppor t ser vice. Th e remai nder of thi s pap e r is struct ured as foll ows: Se c ti on II d etail s the mat eri als and met hods, Secti on III descri bes the sys tem archi te ctur e, Sect ion I V pres ent s experi ment al r e sult s , S ect ion V di scu ss es im pli c ation s an d limi tation s, a nd S ectio n VI concl udes. 2. Data and Methods 2.1. Sy stem implem entation AI -HE ART is implem e nted as a cloud- bas ed s oft war e- as -a-m edical- d evic e (S a MD) plat form th at o perati on alis es a multi- stag e ECG ana ly sis pi pel ine within a se cur e w e b a pplic ation . In p racti ce, th e platf or m combi nes ( i) a bro wser-b ased cli nicia n int erf ace for uplo adin g r ecordi ngs and r evie wing out put s, (i i) an appl i cati on servic e t hat exp o ses RES T API s and orc he stra tes valid at ion , job exec uti on , per sisten c e an d audit logg ing, and ( iii) a cont ain e r ised A I in fer e nc e ser vice th at ex ecut es pr eproc ess ing and de ep- learni ng mod e l i nfe renc e i n an i s olat ed runt im e (G PU- enabl ed when a vai labl e). E C G fil es and d erive d art if a cts ar e man aged via an obj ect- stora ge l ay er, w hile s tr uctur ed entit ies, m et adat a, resu lt s and c lin icia n edits are stor e d in a re lati ona l d atab ase; authe ntic ation and r ol e-b ase d aut hor isat ion are enfor ced via a n identi ty and acce ss man ag emen t se r vi ce, fr onted by a rev er se prox y for secur e traf f ic ro uting and T LS termin ation . This impl e ment ation appr oach suppo rt s m odul a r updat es of mo del compo nent s unde r confi gur ati on contr ol and enabl es sc alabl e batch proce ssi ng, whil e the det ailed ar chite c tu re, comp onent re spon sib ilit ies a nd en d - to -end dat a flow ar e de scribe d in S ection 3. 2.2. Datas ets and annotation All d at a u sed for mod el dev elopm ent and t esti ng were de-id entifi ed pri or to pro ce ssing a nd han dled un der appr opr iate gov ernan ce proc e dur es . ECG rec ord ing s orig inat ed fr om r out i ne ambul ato ry m onito ri ng workf lo ws fro m more th an 10, 000 patie nts and wer e a cqu ired usin g mu lti-da y amb ul ator y moni tor s. Re cor ding s c onsi sted of lo ng-dur ation thre e-l ead elec tro c ar dio gr ams, typ ically sp an nin g m any hour s to sev e ral d ays per pati ent and co v erin g a wi de rang e of hear t rat es, ac tiv ity level s and s ign al - qu ality cond iti on s. The primar y traini ng resour ce is a propri etary MEDTL repo sito r y conta ining more than 1 .5 mil lio n manu ally an not ated beat s dra wn from l ong-t erm ECG re cor din gs and cover ing a tar get set of 35 b eat and rhy thm typ es, incl uding sin us and j uncti onal rhyt hm s, supra ve ntri cu lar and ventr ic ul ar ecto py, atr ial f ibr ill at i on and fl utt er, s upr aventr ic ul ar tac hycardi a, multi ple degr ees of atri ov entr icul ar blo ck and con ducti on abnor maliti es. This in - hous e data set is compl ement ed by publ ic ben ch mar k datab as es, inclu din g MIT-BIH [ 23] and AHA [24] , whi c h were inco rpor at ed to incre as e popu latio n div er sit y and signa l vari ab ili ty a nd t o red uce ov erfit ting t o any s ingl e de vi ce or acqui s it io n pr ot ocol. Refer ence annota tions were gener a t ed by tr ai ned clin icia ns f ol lo wing guid eli nes align ed wit h c onventi ona l ECG i nter pr etati on stan dar d s. For deli neati on, annot ator s mark ed the ons et and off set of P wa ves, QRS comp lexes and T waves on s ele c t ed lea d s. For noi se d etec tio n, short segme nt s w e r e la bell ed a s “c l ean ” or “noi s y” a ccor ding to th eir su it a bili t y f or di agno s t ic i nterpr etati on . For bea t-lev el c lassif ic atio n, e ach b eat re ceiv ed a mor p hol ogic a l lab el f ro m th e 35-type taxo nom y, w hile rh yt hm-l evel l abel s were assig ned to conti gu ou s 1 0- seco nd segmen ts, c aptu ring patt ern s su ch as atri a l fibr il lati o n, atria l flu tt er, AV block s, s upr ave ntri cular run s and ventr ic ular tac hy cardi a epi sode s. A subset of r ecor ds unde rwent doubl e r eadi ng, a n d dis agr eeme nts wer e re solv ed by cons ens us t o r educ e lab e l noi s e. For the exp eri me nts repor te d in thi s pap e r, we foc u s on two lab eled E CG subset s deriv ed from th e fu ll tax onom y. The beat clas sif i catio n subs et c onsi st s of beat- c enter ed anno tate d ECG se gme nt s with lab el assi gn ed to th is bea t. Rhyt hm clas sifi cati on subs et co nsist s of 10 secon ds f ra mes capt uring ar r hyth mia e pis odes. Con s equ e nt ly, s om e rh yth ms s uch a s atr ial fibr ill ation ( AF) and a trial f lutter (AFL) app ear in bo th task s: a t the bea t l ev el, l abel s are as si gned to ind ivi du al be a t s oc cur rin g w ith in arrhyt hmic ep i sod e s, wher ea s a t th e r hythm lev e l t he mod el cla ssifi es 1 0-sec ond wind ows as AF, AFL , and s o on. T he se subs ets were cho sen to b alan ce cli nical r elev an ce w ith suffi cien t samp le siz es, w hil e s till incl udi ng r ar e but impor tant cond uctio n d istur banc es. The data set was parti tione d into trai ning, valid a tio n a nd test set s at the pat ie nt leve l, so that no ECG from a give n patie nt app ear s in more th an on e split. An 80/1 0/1 0 split wa s used for tra ining , val id ation an d te sting, r esp ecti vely. The vali dati on se t was us ed for hyp e rp ara meter tuni ng and earl y s to pping , whe re as t he held -out test set wa s r eserv ed for the fi nal p erf orm an ce eval uati on rep ort ed in S ecti on IV . 2.3. Signal Pre-pr ocess ing and Delineation All ECG recor d ing s unde rgo stand a rd i sed pre pro ces sing pr i or to model infer e nc e . A band-pa ss dig it al filt er is appli ed to redu c e bas eli ne wand er , high -fre qu ency noi se and po wer- li ne inter f er ence , and sig na l s are resam pled or s egmen ted as ne eded to matc h the in put r equire me nts of t he d owns tre am m odel s. For wav e delin eati on , AI-HE AR T empl oys a de ep-le a rn ing mod el th at oper ate s on fixed- lengt h wind ows of m ult i-l ead ECG data a nd pred ic ts the posi tio n of i soel ectri c segm ent s and P , Q RS and T wav es. The m od el m aps raw or li ghtly proc ess ed sampl e s to a se qu ence of label s at the sampl e or small-f r ame le vel, and it s outpu ts ar e post- pro ces sed to der ive ons et and off set time s for each w ave. Fro m the se f iduci al p oints , cli nica lly re leva nt int er val s suc h as P R , Q R S and QT ca n be comput ed ( Fig ure 1) . Delin eatio n qu alit y is a sse ssed by comp ari ng pre dict ed bound ar y lo catio ns with refer ence ann ota tio ns on the inter nal test s et . Sub sequ en tly we re por t per- sampl e a ccur acy, toler ant ac cur acy within a small tempo ral wind ow , F1-scor e and are a un der the recei ver oper ating ch ara cter ist ic curve ( AUC) , demo nstr ating perfor mance ad equat e for autom ated interv a l measur e m ent an d down s tr eam r h ythm cl assifi c atio n. Fig. 1. Fi du cial point s an d mai n in ter val s of a ty pical hea rtb eat. 2.4. N oise Detection To re duce th e im pa c t of art efact s on down str eam a nal ysis, AI-HE ART inclu des a d edic a ted noise-d e t ectio n mod el th at cl as sifi es short ECG segm ent s ac cor ding to th e ir suita bil ity for dia gn osti c int er pr etatio n. The mod el is a pplie d in a s lid ing - wi ndow fashio n acro s s ea ch r ecord ing, an d seg ment s iden tifi ed as low quali t y ar e eith e r exclud e d fr om auto mate d a rrh yt hmi a anal ysis or cl ear ly flag ged to th e clini cian. 2.5. Arr hythmia and Beats Classification The arrhythmia classifier is des igned to recognise a broad se t of ECG rhythm c lasse s, including both common and relatively rare patterns. In the current implementation, the model operates on 10 -second ECG s egments derived from a s ingle diagnostic lead ( lead II ). These s egments are preprocessed and quality-screened by the up stream denoi sing, delineation a nd signal -quality modules, then z-normalised before being fed into the class ifier. The same architecture is used for both beat -level and rhythm- level cla ssification ta sks, with different hyperparameters, enabli ng cons istent treatment of local morphology a nd longer rhythm episodes . The cla ssifi er emp loy s a hybr id deep-le a r nin g ar chit ectur e t hat com bi nes convo lution al neur al ne two rk s (CN Ns) with Tran s for mer encod e r s. The C NN blo c ks act as f eatur e extra ctor s, lear ning loc al mor ph olo gi cal patt e r ns su ch as QR S s hap e, P - wave pre sen ce a nd ST – T ab nor ma liti es, and pr ovidin g robu stnes s to small t empor a l shift s and bas eli ne va riatio n s. Th e Tran s for mer enco de r then ap plie s multi-h ead self -at te nti on and positio nal en codi ng to mod e l lo ng-r ange temp or al dep endenci es acro s s t he 10 -secon d windo w, allo wing the n et wor k to int egr at e beat- to - beat infor m ation a nd captu r e r hythm-l evel charact erist ics such as irr egular RR int er val s or runs o f e cto py. The high-lev e l archi te ctur e c ons i sts of stac ked conv olu tio nal lay er s, fo llowe d by a Tran sf orm er encoder and one or m ore fu lly c onne c t ed lay e r s with a f in al softm ax outp ut . A hig h -l evel diagr am of th e hybri d CNN – Tr ansf ormer archi te c tur e used f or be a t-le ve l and rh ythm-le vel arrh ythmi a cl assifi ca t ion is sho wn in Fig. 2, wh er e convo lution a l block s e xtr act lo cal morphol ogic al fea tur es a nd the Tran sfor mer encod er m od els l ong-ra ng e temp or al dep en den c ies acro s s t he 10-s e cond E CG w indow . The outp ut lay er pro duce s a probab ility dis trib ution over t he pred ef ined set of rhyt hm cla ss es us ed in thi s study : atrial fi b ri lla tio n (AF), atria l fl utt er ( A FL), normal sin us b ea t s (N) , ju ncti ona l b eat s ( J), s upr ave ntri cular ecto pi c be a ts (S) , v e ntr ic ular ecto pic beats (V), sinu s a rrh yt hmi a (S A) , supr avent ricul a r tac hycar dia ( SVT) , sev eral types of atrio ventr ic ul ar block (AV 1, AV2 1, AV22 , AV3) , ab erran cy bea ts ( A) and a n unsp ecifi ed/no ise cl ass ( X). Duri ng trai nin g, t he mod els are using c ate goric al c ro ss entro py los s fu ncti on with foc al l oss . Fo cal lo ss is use d to in crea se th e imp act of r are ar rhy thm ias and to emph asi s e hard e r exampl es. Optim is er is A dam W. T r aining is pe rf ormed on larg e a nno tate d data s et s com prisin g both inter nal M ED TL data and publi c benc hmar k d ata ba se s as emph asi zed above , w ith s tand a rd reg ular i satio n t echn ique s s uch as drop out a app lied to r edu ce ove rfitti ng. During inference, the classifier ge nerates class probabilities for ea ch analysed 10 -se cond window. These probabilities are then aggregated across the full recording to derive a do minant rhythm and to identi fy additional arrhythmia episodes, suc h as paroxysmal AF, runs of SVT or periods of high v entricular ecto pic burden. A n online platform with visual interf aces prese nts these results to the clinical user as machine-generated interpretations, with associated confidence scores, which clinicians can review, confirm, modif y or override before finalising the report. Fig. 2. Hybrid CNN – Transformer architecture used for beat - level and rhythm - level arrhythmia classification in AI - HEART. 2.6. Data Augmentation and Dataset C uration To impr ove perfor manc e, par ti cular ly for un der-re present e d arr hythm i a cla sse s, AI - HEA RT e m ploy s d at a augm e nt ation str ategi es b ase d on gen e ra tiv e a dver sari al net wor ks (GAN s). Synt he t ic ECG beat s or se gme nt s are gener ated for rare cla sse s us ing GANs tr aine d to appr oxima te the dis tri bution of rea l sign al s and ar e then inc orp ora te d into the tra ini ng se t under cont rol led condit ions . In ter nal exp erim ents h ave shown that this appr oa ch can i ncr ea se c l assif i catio n ac c ur acy and sensi tiv ity for rare rhyt hm s wit hout d egr adin g perf orman ce o n com mon cla ss e s. In addi tio n, a c us tom “Da tas et Tornado ” appli cati on has bee n d evelo ped to s uppo rt expla inabi lit y a nd dat aset optim isat ion . This to ol c om pute s lo w -di me nsio nal embed din gs of ECG segme nts u s ing manifo ld -le arning t echni ques a nd visual ises them in an int eract ive inter fac e. User s ca n ex plor e clu ster s cor respon ding to diff erent arr hyt hmi a c la s se s, id entif y r egio ns with low - conf id enc e or incon sis te nt pre di ction s and co rre c t mi slab e ll ed examp le s. Corr ec t ed lab e ls and cu rat ed sub set s are then fed b a ck into th e tr aining pi peli n e a s par t of a contin uou s l earni ng lo op. 2.7. Evaluation Metrics For evaluation purposes, s tandard classification metrics are computed at both class and aggregate levels. For the delineation model, per-sample accuracy, tolerant accuracy, F1-sc ore and AUC are reported. For the noise -detection model, accurac y, precision, recall, F1-score and AU C are calculated, and confusi on matrices were used to illustrate error patterns. For arrh yth mi a cla ssifi c atio n, p er-cla ss se nsiti vity (re call ), specif i city, pr e cis io n, F1-s core and overa ll acc ur acy a re rep ort e d for both the be at and rhyt hm sets . Macro-a v erage d and m icr o-av e r aged m etri cs are comput ed to s umm aris e perf orman ce acr oss clas se s w ith diff erent pr eval ences . Whe re r el evan t, model s train ed wi th an d w ithout GAN-ba sed a ugm entati on and cur ated labe ls are c ompar ed to qu antify t h e im pact of the se t e chniq ues on under-r epr ese nt ed cl a sses. We r eport 95% c onf id enc e i nt erval s ( CI s) for pr opor tio n-b ase d m etri cs (s ensit ivity, specif icity , pre cisio n, accura c y) using Wilso n scor e inter val s. F or F1 - scor e , CIs were estim a t ed vi a no n-par ametri c boo tstra p r esa mpl in g. 3. System Architecture and Software Platform AI -HE ART is a proprietary s ystem developed by M EDTL Medical T echnologies. For intellectua l property and regulatory reasons, only a high -level description of the model architecture and implementation is provided here. 3.1. Ov erall Architecture Figur e 3 summ arise s th e AI-HEA RT r e fer enc e archi te ctur e and the sep arati on of conc ern s betw een (i) the clini c ian-f acing w eb cli ent, (ii) th e appli cati on la yer res pon s ib le for requ est handlin g, v ali datio n, p ersi s ten c e and orch e stratio n, and (ii i) t he AI servi ces that exe c ute prepr oces sing and m odel infer ence. Dat a are s plit betw een obj ect stor age for ECG file s and deri ved ar tif acts and a rel ati onal datas tor e for us ers, meta data, re s ult s and audit trai l s; iden tit y and acce ss mana geme nt i s e nf or ced a t the pl atf or m bou ndary to suppor t rol e-b ased acc ess cont rol. ECG recor din g s and asso ci ated met adata are receiv ed from ext ernal acquis ition devic es or inter medi a t e s yst em s and uplo aded to the platf or m through se cur e endpo int s. The appli catio n lay er valid a t es and norm ali se s i nput s , s tor es the rec ord ing s in encr ypted form and schedu les AI a nal ysis job s . The AI servic es are encap sulat e d in dedic a t ed com po nent s th at c an b e sc aled ind epend ently , al lowin g th e pl atform to ha ndle varyi ng wor k load s witho ut af fecti ng t he user-f acin g i nt erfa ce. The s ystem is designed to support deployment in multiple geographic regions, s o that data can be hos ted close to the point of care and in c ompliance w ith local data-residency requirements. Internal routing mechanis m s en sure that requests are directed to the appropriate regional instance while maintaining a consistent user experience. Fig. 3. System architecture of the AI-HEART cloud platform. 3.2. Application and AI Services The applica tion layer impl e m ents the plat form’ s ext e rn al in ter face that med iate s all int erac tio n s be tween the user interface, the data s tore and the AI components. It implements business logic such as user and organisation manageme nt, case creation, credit a nd billing operations, and orchestration of analysis workf lows. By centralising these responsib ilities in a w ell-defined API, client applications can remain thin and independent of internal im plementation details. The AI l ayer enc ap sula tes the de ep-le arn ing mode ls de scrib ed in Sect ion 2. Model infer ence ru ns in a sep arat e s ervi ce that recei ves ECG dat a fr om the appl icati on l ayer, perf orms pr epro ce ssing an d call s th e deli neation , nois e -det ecti on and arr hy thm ia- clas sif i catio n model s. Resul t s ar e retur ned as struct ured annot ati ons and s umm a ry findin gs, whi ch the appli cati on l ayer per si sts and expo ses to the u ser int erf ace . This separ a ti on allow s mo del s to be upd ated, versi on ed and rol led back und er confi gurat ion contr o l wit hout modif ying th e r est of the p latf orm . Back gro und jobs handl e t ask s such as batc h re -an al ysi s wi th u p dated model s, gene rati o n of s ummar y stati stic s and period ic healt h che cks. All such jobs are lo gged and audi tabl e, in li ne with the qua lity-man agem ent r equir eme nt s for medi cal-d evic e softw are. 3.3. Data Manageme nt, Security and Access Control The platform m aintains a clear separation between c linical data, operational data and logs. E CG recordings and patient -related metadata are stored in se cure, acce ss -controlled repositories with enc ryption at rest and in transit. Application data s uch as user accounts, or ganisations, case status and bill ing information is ma intained in a tr ansac tional database , while tec hnical lo gs and audit trails a re s tored in de dicated logging systems. T hese design priorities align with hospital manage ment perspectives th at emphasise da ta quality mana gement, metadata manage ment , and da ta security as core domains for effective healthcare data governance [22]. User auth ent icat ion re li es on in du str y-st andar d m echa ni sms, incl udi ng s tr ong pas sword pol icie s an d mult i -fact or authe ntic ation. A uth or i satio n is im plem ented throug h ro le - bas ed acc e ss cont rol , with di stinc t rol es fo r sy stem admini s tr at or s, org ani sati on al manage rs, clini cal user s and tria l coordi nator s . Eac h ro le is asso ciate d wit h a def ined set of per mi ssio n s c ontr o lli ng acc ess to p atien t s, re c or din gs and conf igurat ion o pti on s. Every cli ni cally rel ev ant a ctio n, s uch as vi ewin g a cas e, editin g a report or e xpo rtin g d ata, is r ecord e d i n an audit log with u s er ident ity , ti mest amp and actio n det ails . The se l og s s up port tr a ce abil it y, incid ent i nv esti ga tio n an d regu lato r y compli ance. 3.4. C linical Workflow And User Interface AI -HE ART is inte gra ted into the c l ini cal wor kfl ow as a deci si on-suppo rt tool rat h er than a fully autom ated r ead e r . When a new ambul atory E CG re cord in g is uplo aded and associa ted with a pati ent a nd req uest in g c li ni cian, the s yst e m perfor ms a utom ated analy sis in the backgr ound, incl udin g pr epr oce ssin g, deli neat ion, nois e hand ling an d mult i -c las s arr hyt hmi a cla ssifi cati on. On ce pro ces sing is com plete , th e cas e ap pear s in a clini cian wor kl ist with bas ic m etad ata and a sum mar y o f t he main findi ngs. Clini cian s op en i ndi vidu al ca ses to rev iew both the und erly ing ECG and the corre s pond ing AI -g ener ated inter preta tio ns as show n in F ig ure 4. The m ai n c ase vi ew provid e s acce ss t o the ra w tr ace s, k ey in terva l m easu rem ent s and an over vie w of d ete c t ed rhyt hm s and epi sode s. User s can navi gate thr oug h the recor ding , adju s t the d ispl a y and exam in e segme nt s that the s ys tem has high lig ht ed as contai ning pote nti al arr hyt hmi as or other event s of intere st. Th e AI ou tpu ts are pre s ent ed as sug gest ion s that ca n be co nfirme d, mo difi ed or re ject e d by the cl ini cian befor e fin ali sati on. Aft er revi ew, the clini ci an can gen er at e a s tr uctur ed repor t that combi nes a utom ati cally deri ve d me asur e m ents and rhythm clas sif i catio ns wit h th eir own edit s a nd comme nts. Repor ts a r e s tor ed in t he platfor m for aud it and f ollow-up and can b e expor t ed or int e rfa ced with ext ern al e le c tro ni c health r ecor d system s as requ ire d. At an organ i satio nal l ev el, aggre gate vie ws of s yst e m usag e and case m ix (f or exam pl e, numbe r of analy sed re cord in gs and di stri but ion of report ed rhyth ms) suppor t m onito ri ng , qua l ity assur anc e and servic e pl annin g. Fig. 4. Event-analysis view in the AI-HEART clinical user interface. 3.5. De velopment Process And Quality Management The software development process for AI -HEART follows a documented lifecycle aligned with medical -device s oftware standards. Requirements, design specifications, t est cases and ri sk analyses are maintained in contr olled repositories. Ch ang es to the applica tion and AI c omponents are managed through ve rsion control and unde rgo peer review and automated tes ting before deployment to staging and production environments. A qua lity-m anagem e nt sy st em al igned w ith IS O 134 85 pro vid es the fr a mewor k for ma nagin g do cumen tatio n, d esign revi ews , verif ication and vali datio n acti vit ies , incid ent h andli ng, and post-de ploym ent m onito r ing . The platf orm arc hit ectur e and t ool ing are chosen t o sup port tr ace ab le link s bet ween r equi r ement s, impl ement ati on and t est evide nce, which is ess entia l for fut ure reg ulator y su bmi ssi on s. 4. Experimental Results 4.1. Tr aining Performance The ove rall training perf ormance is s ummarized in Table I. GPU NVIDIA A100-SXM4-40G B was used for training the algorithms. TABLE I O VERALL TRAINING PE RFORM ANCE Algorithm Number of model parameters (MB) Number of epochs Delineation 21.92 80 Beat cla ssification 26.01 40 Rhythm class ification 26.01 40 4.2. De lineation Performance The delineation model was evaluated on the in ternal test set using the reference annotations described in S ection 2.3. For each lead and annotated wave , predicted onset and offset positions we re c ompared w ith ground t ruth at the s ample level. Ta ble II summarises the overall perform ance, reported as per-sample accuracy, tolerant accuracy withi n a fixed temporal window, F1- score and AUC . Acr oss all annot ated segmen ts, t he mode l achie ved a p e r- sampl e ac cur acy of 93. 9%, wit h a t oler ant a ccur acy of 98.4% withi n ±5 samp le s. The corr es pon di ng F1- scor e a nd AU C were 0.94 an d 0.99 7, re spe ctiv ely (Ta ble I I) . These re s ult s indi cat e tha t the mod el is abl e to locate P, QRS and T wave s with an accur acy su ffici ent for aut omat ed in ter val m easur ement and for s uppor t ing down s tr eam r hythm c l assif i catio n. Vis ual in specti on of test case s conf irmed t hat r esi dual er ror s are t ypicall y confi ned to bord erl in e r egion s b e tw een w aves or to se gm ents aff ected by sub s tant ial n oi se. TABLE II P ERFORMANCE OF THE ECG D ELINEATION MODEL ON THE INTERNAL TEST SET Metric Value Accuracy (95% CI), % 93.9 (93.8-94.1) Tolerant accuracy (%) (±5 sample s) 98.4 Precision (95% CI), % 93.8 F1 -score (95% CI) 0.94 (0.99-0.99) AUC 0.997 4.3. N oise Detection The nois e-d ete c tio n m odel was a ssess ed on th e t est set using s eg men t- lev e l la be l s for “clea n” and “noi s y” r e cor ding s. T abl e III rep ort s ac cur acy, pre c i sion , recall , F1-scor e and AUC. T he model ac hi eved an accur acy of 99.0 5%, pre c i sion of 99.20 %, r ecal l of 99 .01% an d an F1- scor e of 0.99 1, indi cating that low- qual ity segm ents ca n be reli abl y ident ifi e d and fla gg ed for ex clu s io n or manua l r evi ew. In pr actic e, this redu c es the n umb er of mis cla ssifi cati on s attr ibu tabl e to obvio us artef acts, suc h as electr od e deta c hm ent, motio n n oi se or sever e baseli ne ins ta bil ity , and supp ort s mor e stabl e arr hyt hmi a analy sis acro ss het eroge neou s recor d ing c onditi on s. TABLE II I P ERFORMANCE OF THE NOISE DETECTION MODEL Metric Value Accuracy (95% CI), % 99.05 (98.98 - 99.11) Precision (95% CI), % 99.20 (99.11-99.27) Reca ll (95% CI), % 99.01 (95% CI 98.91- 99.09) F1 -score (95% CI) 0.991 (95% CI 0.9- 99.09) AUC 0.997 4.4. Beat Clas sification Beat classifier wa s first e valuated on a c ore se t of rhythm c lasses comprising atrial fibrillation (AF), atr ial flutter ( AFL), nor mal sinus rhythm (N), junctional rhythm (J), supraventricular ectopic bea ts (S), ventricular ec topic beats (V) and a n unspecified/noise c lass (X). Ta ble IV prese nts per-class sens iti vity, s pecificity, precis ion, F1 -sc ore a nd a ccuracy. Values a re reported as % (95% CI) in Table IV. For th e be a t l evel set, pe r- cl as s sens iti vitie s wer e gen e r all y in t he high ninet ie s, wi th spe cifi c it ies clo se t o or a t 100% f or mo st clas se s. Overa ll accur acies for AF, AFL , N, J , S , V and X wer e approx imate ly 99% or higher . Mi scla ssif ic ati ons wer e rar e and typi call y invo lv ed c onfu sion b e t ween supr aven tri cul ar and ventri cul ar ecto py in bor der li ne e xa mpl es or bet ween no isy s egm ent s and t he X cl ass in low-q ualit y re cor ding s. Macro- aver aged and micro- aver aged F 1- score s a cr oss t he c or e cl asses w er e bo th clo se to 0.99 , indi cati ng bala nced perf orm anc e . In t hi s subse cti on , A F, AFL, N, J, S, V and X r ef er to bea t -l evel l abel s as sign ed to indi vid ual QRS compl exe s. TABLE IV B EAT CLASSI FICAT ION RESULTS Metric AF (95% CI) AFL (95% CI) J (95% CI) N (95% CI) S (95% CI) V (95% CI) X(95% CI) Sensitivity (Recall, %) 98.7 (98.1 – 99.3) 96.5 (96.0 – 97.0) 99.4 (98.9 – 99.9) 98.5 (97.8 – 99.2) 98.4 (97.2 – 99.6) 99.0 (98.1 – 99.9) 97.2 (96.3 – 98.1) Specificity (%) 99.6 (99.3 – 99.9) 99.9 (99.8 – 1) 99.9 (99.8 – 1) 99.6 (99.3 – 99.9) 99.6 (99.3 – 99.9) 99.8 (99.7 – 99.9) 99.8 (99.7 – 99.9) Precision (%) 98.8 (98.1 – 99.5) 97.0 (96.4 – 97.6) 98.7 (98.0 – 99.4) 99.0 (98.6 – 99.4) 97.6 (97.0 – 98.2) 98.9 (98.1 – 99.7) 96.7 (96.0 – 97.4) F1 -score (%) 98.7 (98.1 – 99.3) 96.7 (96.1 – 97.3) 99.0 (98.6 – 99.4) 98.8 (98.2 – 99.4) 98.0 (97.6 – 98.4) 98.8 (98.3 – 99.3) 97.1 (94.4 – 96.8 ) Accuracy (%) 99.3 (99.1 – 99.5) 99.7 (99.6 – 99.8) 99.9 (99.8 – 99.9) 99.3 (99.2 – 99.4) 99.4 (99.3 – 99.5) 99.6 (99.4 – 99.8) 99.6 (99.5 – 99.7) AF = atrial fibrill ation ; AFL = atrial flutter; N = normal rhythm; V = ventricula r ectopic; S = supraventricular ectopic; J = junctional rhythm ; X = unspecified/noise . Entr ies are metric % (95% CI). 4.5. Arr hythmia Classification The rhythm- leve l c l assif i er was als o e val uated on a rhyth m lab el set th at incl ude s f irs t -, s eco nd- and thi rd-d e gr ee atri ov entr icul ar bloc k (AV 1, AV21, AV2 2, AV3) , sinu s arr hyth mia (S A) and s upr a ven tricul a r tachy c ar dia ( SV T), in addi tio n to AF, AFL, norm al sinu s r hyt hm (N) and an unsp ecif ied/ nois e clas s ( X). These clas se s ar e l es s fr equent in the inter nal dat aset, whi ch pr esent s a gre at er chal le ng e for m odel train ing. Desp it e clas s imbal ance, the rhyt hm-l evel mode l achi e ved v e ry high spec ifi c it ies (97 – 10 0%) acro ss all cla ss es in this rhyth m label set, with sen sit iv iti es betw een 76 .5% a nd 98 .9% and F1 - sco re s bet w een 80. 0% and 98.5% (Tabl e V). Sin u s arr hyt hmia an d SV T, whic h wer e bet ter repre sen te d i n t he data set , s how ed F1-s cor es of 88.9% and 88. 5% an d accur acies of 99.9% an d 99. 8%, re spec tiv ely, com para ble to th e core be at-le ve l cla sses. For th e AV block subt ype s , sensi tiviti es rang e d from 76.5% for secon d - degr ee AV bl ock typ e I ( AV21) to 97 .4% f or fir s t-de gree A V blo ck ( AV1) , w ith c orr espondi ng F1- scor es fr om 80.0% to 9 3.3%, indi cati ng r e l iabl e d etec tio n of bo th f irst-degr ee a nd hi gh er-d egr ee bl ock d esp ite th eir lo wer pr eval enc e . M acro-ave ra ged sen sit ivi ty and F 1-sc ore acros s th is ten-c las s rh yth m-l ev el set were a ppr oxi m ately 89 – 90%, with ma cro s pe c ifi cit y a bove 99% and a verage accur a cy a ro un d 99. 5%. TABLE V A RRHYTHMIA CLASSIFIC ATION RESULTS Metric AF (95% CI) AF (95% CI) AV1(95% CI) AV21(95% CI) AV22 (95% CI) AV3(95% CI) N(95% CI) SA(95% CI) SVT (95% CI) X(95% CI) Sensitivity (Recall, %) 97.3 (97.1 – 97.5) 95.0 (94.88 – 95.2) 97.4 (97.2 – 97.6) 76.5 (76.3 – 76.7) 80.0 (79.88 – 80.2) 83.3 (83.1 – 83.5) 98.9 (98.7 – 99.1) 85.7 (85.5 – 85.9) 91.1 (90.9 – 91.3) 89.9 (89.7 – 90.1) Specificity (%) 99.3 (99.1 – 99.5) 99.6 (99.58 – 99.7) 99.9 (99.8 – 1) 99.9 (99.8 – 1) 99.9 (99.8 – 1) 99.8 (99.78 – 99.9) 98.1 (98.0 – 98.2) 99.9 (99.8 – 1) 99.8 (99.78 – 99.9) 99.7 (98.6 – 99.9) Precision (%) 98.1(98.0 – 98.2) 94.7 (94.5 – 94.9) 89.4 (89.2 – 89.6) 87.8 (87.6 – 88.0) 80.0 (79.98 – 80.1) 80.0( 79.8 – 80.2) 98.1 (98.0 – 98.2) 92.3 (92.1 – 92.5) 86.0 (85.8 – 86.2) 94.7 (94.5 – 94.9) F1 -score (%) 97.7 (97.5 – 97.9) 94.9 (94.78 – 95.1) 93.3 (93.1 – 93.5) 81.8 (81.7 – 81.9) 80.0 (79.9 – 80.1) 80.0 (79.8 – 80.2) 98.5 (98.4 – 98.6) 88.9 (88.78 – 89.1) 88.5 (88.4 – 88.6) 92.3 (92.1 – 92.5) Accuracy (%) 98.8 (98.7 – 98.9) 99.6 (99.58 – 99.7) 99.9 (99.8 – 1) 99.9 (99.8 – 1) 99.9 (99.8 – 1) 99.8 (99.78 – 99.9) 98.1 (98.0 – 98.2) 99.9 (99.8 – 1) 99.8 (99.78 – 99.9) 99.0 (98.9 – 99.1) AF = atrial fibrill ation ; AFL = atrial flutter; AV1 = first-degree AV block; AV21 = s econd-degree AV block type I; AV22 = sec ond-degree AV block type II; AV3 = third -degree AV block; N = nor mal rhythm; S A = sinus a rrhythmia; S V T = supraventricular tachycardia; X = unspec ified/noise. Entr ies are metric % (95% CI) 4.6. Impact of Data Augmentation and Curation To qua ntif y the e ff e ct of GA N -ba s ed augm ent ation and expert- in -the-l oop dat ase t c ur atio n, mod els train ed wit h and with out thes e te c hni que s were comp ar ed on th e same t est set. For u nd er-repr esent e d clas ses suc h as s pec ifi c A V blo ck s ubtyp es a nd SV T, the use of synt heti c exam pl es and r el abell ed dat a led to consis tent improv ement s in sensiti vity a nd F1 -s cor e, typi cally i n the ran ge of sever a l p erce nt age p oi nt s, wh ile l eavi ng per form a nc e on ab undant cla sses es senti ally un chan ged. The Dat aset Tor nad o cur ati on tool l ev erag e s low- dim ens ion al embeddi ngs comp uted usi ng Unif orm Manif old Appr ox ima tio n and Proj ectio n (UM AP) to vi s u alis e the str uct ure o f the feat ur e spac e. Emb eddi ng vi sua li satio ns gen erate d befor e a nd a f ter cur ation show e d that , f oll owing exp ert rel abell ing , clu sters corr espo ndi ng to each rhyt hm bec am e more comp act and bett er sep ara ted (F ig . 5). T his p attern s ugg ests th at a sig nifi cant por tion of the or igi nal error s s temm ed fr om labell ing noi se a nd ambig uou s ex a mp les r a t her th a n mod el limi tatio ns. T he se ob servati ons sup por t the i nclu s i on of expl aina bili ty and c ura tio n to ol s as p art of th e overa ll AI- develop ment wor kflo w and as en abler s of c onti nuo us im pro veme nt in a r egul ated set tin g. Fig. 5. UM AP of ECG embe ddings generated by the Da taset Tornado tool, illustrating class -cluster structure and supporting expert- in -the-loop dataset curation. 4.7. Runtime and Scalability Finally, the runtim e performance of the AI pipeline was measur e d in a r epresentative deployment configuration. For stand ard 10-second s egments extracted from three -lead ambulatory recordings, end- to -end inference time, including preprocessing, delineation, noise detec tion and arrhythmia classification, was on the order of a fr action o f a second pe r case on a c ommodit y server-class CPU, and subs tantially lo wer when a G PU was avai lable. Batch processing of la rger volumes of recordings scales linearly with the number of i nstance s assigned to the AI service . The se r esult s in dicat e t hat th e pl atform is cap ab le of h and lin g r ea l -time or n e ar-r eal- tim e analy sis i n t ypi cal clini cal wor kfl ow s and c an be scale d hor i zont ally i n th e clo ud to ac com mod ate high er wor klo ads. 5. Discuss ion 5.1. Sum mary of Findings This paper pr esen ts the de sign and inte rnal eva lu ation of AI -H EART, a cloud- based AI-ECG platf orm t hat in tegr ate s d i stinct deep-l earn ing mode ls for ECG del ineati on, noise dete c ti on and multi- clas s c l as sifi cati on at bot h beat and r hyt hm le vel . The delin eatio n mo del achiev e d a per- s ampl e accur acy of 94.3% an d a toler a nt ac cur acy of 98. 4% withi n ±5 samp les, with an F1 - scor e of 0.94 and an AU C of 0.99 7 on th e int ernal te st se t. The noi se - dete c ti on m odel rea ched 99. 05% accur acy, wit h prec isio n, recal l and F1-s cor e all clo se to 0.9 9, indic ating that low-q ualit y segm ents can be reli ably ident ifi ed and f lagg ed. For bea t-lev el clas sif i catio n, the dedi c ated beat cl assifi e r ev aluat ed on th e cor e be a t l abel set (AF , AFL , N , J, S, V, X) sho wed sen sit ivi t ies a nd spe cifi cit ie s pr e dom i nantl y i n the high n ineti es, wi th p er-c las s a ccurac ie s ar ound 99% and macr o- an d micr o- aver age d F 1- scor es clos e to 0.99. For r hy thm- leve l a rr hythmi a cla ssific ati on , a se par ate rhy thm c las sif i er eval uated on th e ten-cla ss rhyth m lab el set that inclu de s A F, AFL , mul tip le degre es of A V bl ock, sin us arr hy thmia, supr ave ntri cular ta chycard ia, nor ma l si nus r hyth m an d a noi se clas s achi e ved s pecif ici ties at or cl ose to 100% , with ma cr o -aver aged sen si tiv ity and F 1-sc ore ar ound 89 – 90%. AV bloc k subty pes a tt ained sen sitivi ties of at least 7 6% and F 1- score s b e tw een 80% and 93%, in di catin g cli ni cally u sefu l perf orm a nc e even for the s e comp ar ati vely r are rhy thm s. Toget her , the se resul t s in dicate tha t AI - HEA RT c an pr ovi de accur ate , noi se - awar e ECG int erpr et ati on a c ro ss a br oad spectr um of rhy thm s, s uit able f or integr ati on into cl inic al wor kflo ws. 5.2. C omparis on With E xisting AI-ECG Approaches The r eport ed perf orm an ce is con s ist e nt with , and in severa l as pect s c om plem entary to, exi sting de ep -le arn in g appr oa che s for ECG analy sis de scrib ed in th e lit e ratu r e. Pri or work has d emon strat ed t hat convo lutio nal a nd r ecurr ent neu ral net wor k s can achie v e high a c cur acy in arr hy thm ia detecti on, oft e n exceed ing tra dit ion al r ule-b ase d al gorithm s [5, 7]. Met a-analyses of AI-based atrial fibril lation de tection from photoplethysmography and singl e -lead ECG signals have reported pooled s ensitivities and specificities in the 90 – 96% a nd 95 – 97% range [16], which is broadly comparable to the AF performance observed in A I - HEART. Commer cial sy s tem s s uch a s mobil e ECG platf orm s and clou d -b ase d holt er analy sis s oluti o ns hav e show n high per form anc e for AF an d oth er maj or arr hythmi as in r eal-worl d co hort s [6, 12, 13]. H owev er, m any of the se syst e m s ar e op timis ed f or sp e cif ic devic e e c osy st ems or m onitori ng sc en ario s an d o fte n f ocus on a na rr ower set o f c on diti on s (f or exam ple, AF sc re ening or Hol te r rhyt hm ana ly sis). I n contr ast, t hi s study i s design ed as a mult i- cla ss cla ssifier cover ing b oth common r hyt hm s and rarer condit ions such as AV block subtyp es and suprav entri cular tachyc ar di a, with expli cit mec hani s ms f or noi s e h andlin g and sig na l qual ity quanti ficat ion. Th e c ombin ati on of de line ati on, noise- aw are prep roc essin g an d br oad rhythm cove rage di s ti ngu ish e s th e pl atform from AF- onl y or sin gle- in di catio n solut io ns and align s with the nee d for mor e g ener al ECG deci sion -s uppor t tool s in rout ine pra ctic e . The presen ce of a high- perf orm ing noi se-de te ction mode l is par tic ul arly rele vant in comp ari son wit h prior wor k, wher e the imp act of signal quali ty is ofte n man aged im pli citl y t hro ugh da taset cura tio n or exclu s ion criteri a. By m akin g noise det ection an expli cit c omp on ent o f the p ip eline , AI-HE A RT ca n prov id e tr ans par ent qua lity f eedba ck a nd re duce mi scl assifi c atio ns dr iv en b y art efact s, whi ch is c ri tic al for deplo yment in prim ary car e and resour ce-lim it ed sett ings wh er e acqui s iti on c ondit ions may be subop timal . 5.3. C linical Implications and Potential Use Cases The result s sugge st sever al p otent ial use ca ses for AI-HE ART. In prim ary car e or emerg ency setting s, the pla tform could be used to pro vide ra pi d, auto mated pr e-in terpre tatio n of multi- lead E CGs. In thi s study we eva lu at ed thr ee-l ead Holt er recor di ng s; exten sion and val idati on on s tand a r d 12-lea d ECG s repr esent an importa nt futur e step. In hosp it als or tele- cardi ology serv ices, AI -HE ART c oul d f unct ion as a deci sion- supp ort tool , pr e- scre ening ECG s and pri orit isin g ca ses tha t requi re urg ent exper t att entio n. The a bil ity to qua ntify s ign al qu alit y an d to cl early f lag seg ment s tha t are t oo noi sy for reli abl e int erpret ation m ay h elp avoid fals e al a rm s a nd supp ort mor e con s i stent r e por tin g stan dar ds. The mu lti- cl as s n a tur e of the cla s sifi er al so suppor t s l ong itu din al monitor i ng s ce nar io s, w h ere chan ges i n r hyt hm pat tern s ov er tim e, such as the emer gence of parox ysmal AF , progr ession of AV blo ck or in cr easi ng ectopi c bur den, can be tr acke d syst em atic a ll y. F urth erm ore, the cloud- ba sed a rc hit ectu re and confi gur abl e wor kf lo w make it pos sibl e to adapt th e pla tform to diffe re nt org an isat ion al cont ext s, fr om sin gle-clin ic d eploym ent s to m ulti- s ite networ k s. It i s impor tant to empha sis e t hat, in all the se scen ario s, AI-HE ART is in tend e d to a ugm ent rat he r than re pl ace clini cian judg emen t. The us er int erfac e is expl icitl y de signe d to allo w clin ician s to revi ew, mod ify and overrid e AI -g ener ated int erpr et ati ons, and to in cor por ate their own find ings into the final rep ort. This “hum a n - in - the- lo op” design is align ed with reg ulator y exp e ctati ons f or AI-b ase d m edi c al d evice s an d with pr acti ca l con s ide ratio ns f or clin icia n ac c eptan ce. 5.4. Robustne ss, Ge neralisation and Continuou s Learning A recur rin g chall enge for AI-E C G syst ems is robu s tne ss acros s devi ces , popu lati on s and re cordi ng condi tions . The e xpl i cit mod ellin g of noi se in A I-HE A RT addr esse s o ne asp ect of this chall enge by a ll owing lo w -qual ity seg ment s to be exclud ed or flag ged, there by r educ ing th e ri sk of er ro neou s cl assifi c atio ns dri ven by artef a ct s . Th e stron g perfor manc e of the n oise-det ectio n mod el sug ge sts t hat this ap proac h is eff e ctiv e in th e int ern al dat aset and pr ovi des a basi s f or rob ust depl oymen t in envir onm ent s with var iab le acq uisit ion q ualit y. Anoth er cha ll enge is the han dlin g of rare arrh ythmia s and ed ge cas es, wh ere lim it ed tr ainin g dat a can le ad to unst able perf orm a nce . The inco rpor ation of GAN- bas ed augm entat io n an d UMAP-b ased da tas et cur ati on is a pra gmati c res pon se to this prob le m. T he observ e d im pro veme nts in sensit ivi ty and F1-sc ore for A V blo c k subtyp es and SV T after augm enta tion and rel abell ing i ndic a te t ha t, wh en c ar efu lly c ontrol led, s ynthe tic da t a and expe rt- in -th e-loo p curati on can hel p m iti gate c las s imb alan ce with out d egr adi ng p er for mance on mor e pr eval ent rhy thm s. The se tool s als o c ont ri bu te to cont inu ou s lea rni ng. B y enab lin g s yst em atic ide ntif ication and corre ct ion of labelli ng error s or ambig uou s examp les , the pl atf orm can evol ve as more dat a a r e coll ecte d, while still ope ra tin g under a qua lit y -m anag ement fram ew ork c omp atibl e wit h medi cal-de vic e re gul ati ons . In fu ture pos t-m ark et sett ings , sim il ar me c hani sms coul d supp ort moni tor ed mo del u pdat es u nd er a pred et ermi ned c han ge contr ol pla n, con s ist e nt wit h e mergi ng r e gulat ory gu idance. 5.5. Inform ation Management Implications AI -HE ART is not only a set of deep-l earn ing model s but also an i nform a ti on syste m for man agin g long-dur ation amb ul ator y ECG data and deriv ed annot at ion s. The platf orm’ s mod ul ar, cloud -b a se d arch it ectur e allo ws ECG rec ordi ng s, interm ed iate mod el outp ut s an d c lin ician edit s to be s tor ed u nd er a u nifi ed dat a mo del wit h con s ist e nt ide ntif iers and aud it tr ails. T his d esig n support s tra ceab ili ty of each autom ated deci s ion back to the origi nal s ign al and assoc ia t ed model ver s ion, which is essen tial for qu al it y manag e m ent, post- mar ket surv eillan ce and reg ulator y audit s. At the same tim e, separ ating the us er -f acin g int erfa ce fr om back- end AI ser vi ces mak es it p ossibl e to upda te or s c ale i ndivid ual compon e nts wit hou t d i srupti ng clini cal workflo ws. From an inf orm ati on gov ernan ce per spe ct iv e, AI-HE ART i llu str ates how re gul at ory and org a ni sati on al requ ire ment s s hap e data f lows in AI- enab le d clin ical sy stem s . R equ ire ment s r el ated to ISO 1348 5 and anti cipat e d M D R/F DA c ertif icat ion inf lu ence deci s ion s about log gin g granu larity , reten tion of raw and proc essed dat a, and th e docum enta tion of train ing and valid ation dat asets. Rec ent evid e nce from hosp it al m anag em ent als o hi ghl ight s gover nan ce gap s aroun d dat a qua lity, docum entat ion /m et adat a, sec urit y, and e t hical acce ss, fact or s tha t direct ly shape how AI-e nable d clini cal data pipel ine s shoul d be des ign ed and oper at ed [22]. R egio nal de pl oym ent and da ta -r esi de ncy con str aint s furth er affec t wher e re cord ing s and rep ort s are stor e d and how cro ss - bord er acc ess is mana ged. The s ys tem ’s huma n - in -th e -l oop workf l ow, in which clini c ian s rev iew, corre c t and a ppro ve AI - gener ated annot ations be for e fi nal repor ting, also em bed s fee db ack i nto th e i nfor ma tio n li fe cycl e : cli nici an edit s ca n b e ca ptur e d as struct ur ed data for fu tur e model refin ement, whil e e nsu rin g that leg al res pon sibil ity for the diagno stic i nter pr etati on remai ns with t he hu man us er . For h ealth c ar e org ani sati ons, the se de sig n choi ces have pract ic al impli cati ons. I nte gr ati ng AI -HEA RT wit h existi ng ECG manag e m ent syst em s or ele ctroni c heal th rec ord s req uir es alignm ent of ide nti fie rs , consen t and acc es s -cont rol polic ies , rath e r tha n mer ely deplo yi ng a stand a lone a l gori thm . More br oadly, the cas e s ugg ests th at succe ssf ul deploy ment of AI -E CG servic e s depen ds as mu ch on ro bust inf orm ation m anag em ent, cover ing dat a captur e, s tor age , prov en anc e, acc ess and aud itabil ity, as on mod el accura c y. Platf orms that e xpl icit ly incor por at e t hese consi der ati ons ar e b e tt er posi tio ned to suppor t sust ai nabl e, lar ge -sc a l e use of AI i n rout in e clini cal p ra ctic e . 5.6. Limitations And Future Work The curre nt eval uati on has sever a l limit a ti ons. Fir st, all res ult s are ba s ed on inter nal dat a sets and inter nal te st spli ts. A ltho ugh patie nt-lev e l partit ionin g re duce s the ri s k of infor mat ion leaka ge, exte rn al val idati on o n ind e pen dent cohort s fr om d iffe rent in stitut ion s and po pul atio ns will be n ece ss ary to conf irm the generali sabilit y of the mod els. S econd , the data set use d for ra re arrh yth mi as rem ains comp arativ el y sma ll de spite au gm ent ation , and some perf orm anc e m etri cs for the rar e st cla ss es hav e wide uncert ainty ; pro spe c ti ve mu ltice ntr e data c oll ecti on wil l be r equir ed to furt her st abili se the se e stim ate s. Third , the pre sent st ud y f ocus es on rhy thm ana lysi s and si gn al qual ity; str uc tur al mar ker s su ch as lef t ven tri cul ar dysf unction or o th er E C G-d erive d ri sk score s , w hich hav e be e n e xplor ed in ot her AI-E C G work [ 10], are not yet int egr ate d into t he d eploy ed mod el set. Exte nd ing AI- HEA RT to in c or por at e such mar kers i s a n atural n ext s tep and coul d enha nce it s util ity for scre eni ng a nd ris k s tr atif ic ation . F in all y, while the sy stem ar chit ectu re and us er work flo w ha ve b een de s ign ed with reg ul ator y and clinic a l req uir ement s in min d, th is paper does not repo rt form al usab ilit y te s ting or prosp ecti ve cli ni cal imp act stud ie s, whic h will be crit ic al to demo n strat e real-w orld b e nef i t. Futur e work will t h eref ore focu s on ext ernal val idati on in di ver se clin ical envir on me nts, in cludi ng under-ser ved setting s; pro spec tiv e studi es as se ssing workfl ow imp a ct, diagn osti c yiel d and c l ini cian tru st; and the integ ratio n of additi onal predi c tive mod els for struct ura l and fun ctio nal cardi ac abnorm ali ti es. Fur t her dev el opm ent of post - depl oymen t monitor ing an d change- manag e m ent p roce ss es w ill als o be e ssenti al t o su ppor t safe mo d el ev olutio n in lin e wit h reg ul ator y expe c t ation s. CRediT authorship contrib ution statement Art emi s Kont ou : Wri tin g – or igi nal draft , Proj ect a dmin istr atio n, Conc e ptua lizat ion . Natal ia Mi roshni kova: Writi ng – re view & edit ing, Data cur atio n, Meth odolo gy, Form al a n alys is, Costa ki s M ath eo u: Writin g – re view & editi ng, Proj ect admin istra tio n, Res our ces . So pho cl es Soph ocl eous: Meth od ology, For mal ana l ysis , Dat a curat ion. Nichol as Ts eko ura s : Sof tw ar e. Kle anthis Malia lis : Writi ng – re view & edi tin g. P an ayi oti s Kol ios: Wri tin g – revi ew & edit i ng. Declaration of Competing Interest Sev eral aut hor s are empl oye d by MEDTL Medic al Tec hnol ogi es Ltd. , th e develop e r and own e r of the AI - HEA RT platf orm evalu a t ed in this stu dy. T he ECG data set s used for mo del dev elop ment and eva lu ati on are hel d by MED TL Medic a l Techn ol ogi es. The se relat ionship s ar e dis clo sed to en sure tr ansp ar ency . T he au thor s de c l are no oth er comp et ing fin a nci al int ere sts or per s onal rel ati onsh ip s th at co uld hav e inf l uenc ed t he wo rk r eport e d i n th is pape r. Data Availabilit y The E CG d ata set s u sed to dev el op an d ev al uate AI - HEA RT co nsist of de-id entif i ed lo ng- dur ati on th r ee-l e ad ambul atory elec tro c ar dio gr am re c or din gs coll ect e d in routin e clini cal care and i nter nal annot ated r epo sitor ies held by M ED TL Medic al Techno logie s. Du e t o pati ent priva c y con sider ation s a nd con tra ctu al con str ain ts with data provi ders, the se dat aset s cannot be made publ icly a vaila ble. Aggr e gat ed perf orm anc e metr ics a nd mod el-e valua tion summ a ri es ar e availa bl e fr om the corr esp ond ing aut hor upo n r eas onabl e requ est. Declaration of Generative AI Use No Ge nerati ve AI wa s u sed to pr ep are th is study . References [1] World Health Organization. (2025) . Car diovascular diseases (CVDs) ( Fact sheet). Geneva, Switzerland. [2] Roth, G . A., et al. (2020). Global burden of cardiovascular dise ases and risk factors, 1990 – 2019: Update from the G BD 2019 Study. Journal of the American College of Cardiology, 76 (25), 2982 – 3021. [3] Kushwaha, A. K., Kar, A. K., & Dwivedi, Y. K. (2021). Applications of big data in emerging management discipline s: A literature review using text mining. International Journal of Information Management Data Insights, 1 (2), 100017. https://doi.org/10.1016/j.j jimei.2021.100017 [4] Khunte, A., Sangha, V., Oikonomou, E. K., et al. (2023) . D etection of left ventricular systolic dysfunction from single - lead electrocardiography adapted for portable and wearable dev ices. npj Digital Medicine, 6 , Article 124. https://doi.org/10.1038/s41746-023-00869- w [5] Hannun, A. Y., Rajpurkar, P., Haghpanahi, M., et al. (2019). Cardiol ogist -level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, 25 , 65 – 69. https ://doi.org/10.1038/s41591- 018-0268-3 [6] Johnson, L. S., et al. (2025). Ar tificial intelligence for direct - to -physician reporting of ambulatory electrocardiography. Nature Medicine, 31 , 925 – 931. [7] Jin, Y., Li, Z ., Wang, M., et a l. (2024). Cardiologist-level interpretable knowledge -fused deep ne ural network for automatic arrhythmia diagnosis. C ommunications Medicine, 4 , 31. ht tps://doi.org/10.1038/s43856-024-00464-4 [8] Raghunath, S., Pfeif er, J. M., Ulloa -Cerna, A. E., et al. (2021). Deep neural networks can predict new -on set atrial fibril lation from the 12-lead EC G and help identify those at risk of atrial f ibrillation – related stroke. Cir culation, 143 (13), 1287 – 1298. https://doi.org/10.1161/CIRCULATIONAHA.120.047829 [9] Ribeiro, A. H., Ribeiro, M. H., Paixão, G . M. M., e t al. (2020). Automatic diagnosis of the 12 -lead ECG using a deep neural network. Nature C ommunications, 11 , 1760. https://doi. org/10.1038/s41467-020-15432-4 [10] Attia, Z . I., Kapa, S., Lopez-Jimenez, F., et al. (2019). Scree ning for cardiac contractile dysfunction using an artificial intelligence – enabled e lectrocardi ogram. Nature Medicine, 25 , 70 – 74. https://doi.or g/10.1038/s41591-018-0240-2 [11] U.S. F ood and Drug Administration. (2021, Janua ry). Artific ial intelligence/machine learning (AI/ML )-based s oftware as a medical dev ice (SaM D) action plan. [12] Sm ith, S. W., Walsh, B., Grauer, K., et al. (2019). A deep ne ural network learning algorit hm outperforms a conve ntional algorithm for emergency department electrocardiogram interpretation. Jour nal of Electrocardiology, 52 , 88 – 95. [13] Teplitzky, B. A., McRoberts, M., & Ghanbari, H. (2020). Deep learning for comprehensive ECG annotation. Hear t Rhythm, 17 (5 Pt B), 881 – 888. https://doi.or g/10.1016/j.hrthm.2020.02.015 [14] Muehlematter, U. J., Daniore, P., & Vokinger, K. N. (2021). Approval of artificial intelligence and mac hine learning - based medica l devices in the USA and Europe (2015 – 20): A comparative analysis. The Lancet Digital Health, 3 (3), e195 – e203. https://doi.org/10.1016/S2589-7500( 20)30292-2 [15] Ansari, Y., Mourad, O., Qaraqe, K., & Serpedin, E. (2023) . Deep learning for ECG arrhythmia detection and class ification: An overview of progr ess for period 2017 – 2023. Frontiers in Physiology, 14 , 1246746. https://doi.org/10.3389/f phys.2023.1246746 [16] Manetas-Stavrakakis, N., Sotiropoulou, I. M., Paraskevas, T., et al. (2023). Accuracy of artif icial intelligence -bas ed technologies for the diagnosis of atrial fibrillation: A sys temati c review and meta -analysis. Journal of Clinical Medicine, 12 (20), 6576. https://doi.or g/10.3390/jcm12206576 [17] Kim, A., Chatterjee, M., Iansavitchene, A., et al. (2025). Artif icial intelligence for electrocardiographic diagnosis of perioperative myocardial ischaemia: A scoping review. British Journal of A naesthes ia, 135 (3), 561 – 570. https://doi.org/10.1016/j.bj a.2025.05.037 [18] Satija, U., Ramkumar, B., & Manikandan, M. S. (2018). A r eview of signal process ing techniques for electrocardiogr am signal quality assessment. IEEE R eviews in Biomedical Engine ering, 11 , 36 – 52. https://doi.org/10.1109/RBME.2018.2810957 [19] Kelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G ., & King, D. (2019). Key challenge s for delivering clinical impact with artificial intellig ence. BMC Medicine, 17 , Article 195. https://doi.org/10.1186/s12916-019-1426-2 [20] L ehne, M., Sas s, J ., Essenwanger, A., Scheper s, J., & Thun, S. ( 2019). Why digital medicine depends on interoperability. npj Digital Medicine, 2 , Article 79. https://doi. org/10.1038/s41746-019-0158-1 [21] Bachtiger, P., Petr i, C. F., Scott, F. E., et al. (2022). Point- of-care screening for heart failure with reduced ejection fraction using artificial intell igence during ECG-enabled s tethoscope examination in London, UK: A prospective, observational, multicentre study. The Lancet Digital Health, 4 (2), e117 – e125. https://doi.org/10.1016/S2589- 7500(21)00256-9 [22] Oktaviana R, S., Handayani, P. W., Hidayanto, A. N., & Siswanto, B. B. (2025). Healthcare data governanc e assessment based on hos pital management persp ectives. International Journal of Information Management Data Insights, 5 (1), 100342. https://doi.org/10.1016/j.j jimei.2025.100342 [23] Moody, G . B., & Mark, R. G. (2001). The impac t of the MIT -BIH arrhythmia database. IEEE E ngineering in Medicine and Biology Magazine, 20 (3), 45 – 50. https://doi.or g/10.1109/5 1.932724 [24] Hermes, R. E., G eselowitz, D. B ., & Oliver, G . C. (1980). Deve lopment, distribution, and use of the American Heart Assoc iation dat abase for ventricular arr hythmia detector evaluation. I n Computer s in Cardiology 1980 (pp. 263 – 266). IEEE.