Single-Feature Method for Fast Atrial Fibrillation Detection in ECG Signals

Single-Fea u e Me hod o Fas A ial Fib illa ion De ec ion in ECG Signals Lucie Ma sano a1, And ea Nemco a1, Rado an Smisek1,2, Ma in Vi ek1, Lukas Smi al1 1 Depa men o Biomedical Enginee ing, B no Uni e si y o Technology, B no, Czech Republic 2 Academy o Sciences, B no, Czech Republic Abs ac A ial ib illa ion (AF) is he mos common a hy hmia in adul s and is associa ed wi h a highe isk o hea ailu e o dea h. He e, we in oduce simple and e icien me hod o au oma ic AF de ec ion based on symbolic dynamics and Shannon en opy. This me hod comp ises o h ee pa s. Fi s ly, QRS complex de ec ion is p o ided, han he aw RR sequence is ans o med in o a sequence o speci ic symbols and subsequen ly in o a wo d sequence and inally, Shannon en opy o he wo d sequence is calcula ed. Acco ding o he alue o Shannon en opy, i is decided, whe he AF is p esen in he cu en ca diac bea . We achie ed sensi i i y Se=96.32% and speci ici y Sp=98.61% on MIT-BIH A ial Fib illa ion da abase, Se=91.30% and Sp=90.8% on MIT-BIH A hy hmia da abase, Se=95.6% and Sp=80.27% o Long Te m A ial Fib illa ion da abase and Se=93.04% and Sp=87.30% o CinC Challenge da abase 2020. The achie ed esul s o ou one- ea u e me hod a e compa able wi h o he au ho s o mo e complica ed and compu a ionally expensi e me hods. Ou ECG expe s ound ha public da abases con ain e o s in anno a ions (in sense o AF). I means ha esul s a e a ec ed by e o s in anno a ions. Many e o s we e ound in Long-Te m AF da abase, se e al also in MIT-BIH AF da abase and MIT-BIH A hy hmia da abase. Tes ing algo i hms on poo ly anno a ed da abases canno b ing eliable esul s and algo i hms use ul in eal medical p ac ice. The examples o such anno a ions a e epo ed in his s udy. 1. In oduc ion Ca dio ascula diso de s a e s ill he mos common cause o dea h wo ldwide. Due o he ease o use, non- in asi eness and cheapness, elec oca diog am (ECG) is nowadays s ill he mos a ailable and widely used me hod o he ca dio ascula sys em examina ion [1]. A ial ib illa ion (AF) is he mos common a hy hmia in adul s and is associa ed wi h a highe isk o hea ailu e o dea h. AF is a sup a en icula achya hy hmia which is ep esen ed by incons an a ial ac i a ion and, he e o e, dys egula ion o a ial con ac ions. This cause uncomple ed blood ans e om a ia o en icles and dec ease he e iciency o hea unc ioning. This can esul in se ious complica ions such as ischemia, s oke, o ea ly mo ali y [1]. The e o e, ea ly de ec ion o AF is c ucial o e ec i e ea men . Au oma ic de ec ion o AF in ECG is s ill p oblema ic, as was shown by he esul s o p e ious s udies. He e, we in oduce simple and e icien me hod o au oma ic AF de ec ion based on symbolic dynamics and Shannon en opy. 2. Me hod This me hod comp ises o h ee pa s. Fi s ly, QRS complex de ec ion is p o ided by de ec o based on phaso ans o m, and sequence o RR in e als is compu ed. In he second pa , he aw RR sequence is ans o med in o a sequence o speci ic symbols and subsequen ly in o a wo d sequence. Finally, Shannon en opy o he wo d sequence is calcula ed. Acco ding o he alue o Shannon en opy, i is decided, whe he AF is p esen in he cu en ca diac bea . 2.1. QRS complex de ec ion Fi s ly, signal is il e ed by bandpass il e wi h cu -o equencies 12 and 19 Hz. In his equency ange lies he mos o QRS complex ene gy. Using his il e supp esses P and T wa es and also high equency a i ac s. On he o he hand, QRS complexes a e highligh ed. A e ha , phaso ans o m (PT) is applied. PT ans o ms each sample o he signal in o a complex alue p ese ing he signal in o ma ion [2]. Con e ing each ECG sample in o a phaso enhances changes in he ECG signal ( he wa es). The deg ee wi h which ECG wa es a e enhanced in phaso signal is de e mined by alue RV. The alue o RV is always wi hin he in e al 0-1. A cons an alue RV is conside ed as a eal pa , whe eas he imagina y componen is he o iginal alue o he ECG sample: 𝑦(𝑛)= 𝑅𝑉+𝑗𝑥(𝑛), (1) whe e y(n) is he phaso ial signal and x(n) is he o iginal sample o signal. The magni ude M(n) is compu ed as 𝑀(𝑛)= √𝑅𝑉 2+𝑥(𝑛)2, (2) and phase (phaso ) 𝜑(𝑛) is compu ed as 𝜑(𝑛)= 𝑡𝑎𝑛−1 (𝑥(𝑛) 𝑅𝑉). (3) Compu ing in Ca diology 2020; Vol 47 Page 1 ISSN: 2325-887X DOI: 10.22489/CinC.2020.335 In phaso signal 𝜑(𝑛), he QRS complexes ha e in all cases highe ampli ude han he o he ECG componen s. This applies also in a case o o iginal ECG signals wi h smalle ampli ude o QRS complex han T wa e. Due o his ac , he QRS complexes in phaso signal a e easie de ec able [2]. In he case o QRS de ec ion, phaso ans o m o ECG signal wi h RV = 0.001 is pe o med. The example o phaso ans o m (g ey cu e) o ECG signal (black cu e) p epa ed o QRS de ec ion is in Fig. 1. Than we sea ch o maxima in sliding window wi h 300 ms size. The nex s ep is a check, whe eas he ound maxima a e highe han used adap i e h eshold. A he end i is checked, whe he he in e al be ween wo subsequen QRS complexes (RR in e al) is 1.75x highe han p e ious one. I i is accomplished, backwa d sea ching is pe o med. This de ec o was used and es ed also in ou p e ious s udies [3], [4], [5]. The example o phaso ans o m o ECG signal is shown in Fig. 1. In phaso signal 𝜑(𝑛), he QRS complexes ha e in all cases highe ampli ude han he o he ECG componen s. This applies also in a case o o iginal ECG signals wi h smalle ampli ude o QRS complex han T. Figu e 1 An example o (a) an o iginal ECG signal; (b) phase signal φ(n); (c) de ail o he op pa o he phase signal (illus a ed by ed box in (b)), he g een line ep esen s he h eshold o QRS de ec ion. [3] 2.2. Symbolic dynamic and Shannon en opy The pu pose o using symbolic dynamics (symbols and wo ds) is o desc ibe he dynamic o hea a e. Du ing a ial ib illa ion, high a iabili y o RR in e als is p esen and hus also high alue o Shannon en opy [6]. Fi s ly, hea a e h (n) is de i ed om RR in e als (RR(n)). RR(n) a e compu ed om leng h wo consecu i e hea bea s (R(n)). The sequence o aw RR in e als is ans o med in o he hea a e (h (n)) acco ding o he eg. (4) and hen quan i ied in o symbol sequence (Sy(n)) [6] acco ding o he equa ion (4) ℎ𝑟(𝑛)=60/𝑅𝑅(𝑛), (4) 𝑆𝑦(𝑛)= { 63 𝑖𝑓 ℎ𝑟(𝑛)>315 ⌊ℎ𝑟(𝑛) 5⌋ 𝑜𝑡ℎ𝑒𝑟 , (5) The sequence o symbols Sy(n) ep esen ins an aneous s a e o hea a e ans o med in o he 64 possible symbols. To acili a e he analysis o Sy(n) is used 3-symbol empla e o examina ion o en opic p ope ies o sequence. This ans o ma ion is compu ed acco ding o he eg. (6) and a e ha we ob ain sequence o wo d (w (n)) [7]. Each wo d akes in o accoun 3 successi e symbols (3 hea bea s). 𝑤𝑣(𝑛)=(𝑠𝑦(𝑛 −2).212)+(𝑠𝑦(𝑛 −1).26)+ 𝑠𝑦(𝑛), (6) Finally, Shannon en opy (SH) is compu ed (7). SH is a s a is ical ool ha quan i ies a ime se ies (in ou case leng h o hea bea s) in e ms o he in o ma ion size [8]. A i s , we de ine he disc e e p obabili y space o a dynamic sys em as A = (A|P). A ep esen se o cha ac e is ic elemen s A = {a1, · · · , ak}, and P = {p1, · · · , pk}(1≤ k ≤N) is ele an p obabili y. Each elemen a(i) has p obabili y p(i) = N(i)/N (0<p(i)<1,∑𝑘 𝑖=1 𝑝(𝑖)= 1 whe e N is numbe o all elemen in A and N(i) is o al numbe o elemen a(i) in A. Thus, he SH o A is de ined as [26], 𝑆𝐻(𝐴)= − 𝑘 𝑁.𝑙𝑜𝑔2∑𝑘 𝑖=1 𝑝(𝑖)𝑙𝑜𝑔2𝑝(𝑖) (7) In ou wo k, he dynamic sys em A consis s o 95 consecu i e wo d elemen s om w (n-47) o w (n+47). The alue o SH lies in in e al <0,1>. I means ha on compu ing o SH o R(n) is needed in o ma ion o RR(n) o 95 consecu i e hea bea s. Acco ding o he alue o SH is de e mined i a ial ib illa ion is p esen o no in ac ual hea bea s. In ou wo k, we ound ha he bes h eshold he alue o disc imina ion be ween AF and non AF is T=0.733. In Figu e 2 p ocess o classi ica ion is illus a ed. Subg aph b) shows alues o Shannon en opy ans o med om leng h o RR in e als and also h eshold (blue line) o decision o p esence o a ial ib illa ion (o e line) o o he hy hm (below line). I is clea isible ha alue o Shannon en opy co ela ed wi h anno a ion o a ial ib illa ion c) ( alue 1 ep esen s AF, alue 0 ep esen s no AF). Page 2 Figu e 2 a) Leng h o RR in e als, b) Shannon en opy o RR in e als – blue line is eeshold o decision o p esence AF, c) annno a ion – alue 1 ep esen s AF, 0 ep esen s nonAF 3. Resul s We used ou publicly a ailable da abases o es ing o ou algo i hm. The esul s a e summa ized in Table 1. The exac alues o ue posi i e (TP), ue nega i e (TN), alse posi i e (FP), alse nega i e (FN) a e epo ed in Table 2. We achie ed sensi i i y Se=96.32% and speci ici y Sp=98.61% on MIT-BIH A ial Fib illa ion da abase, Se=91.30% and Sp=90.80% on MIT-BIH A hy hmia da abase, Se=95.6% and Sp=80.27% o Long Te m A ial Fib illa ion da abase and Se=93.04% and Sp=87.30% o CinC Challenge da abase 2020. Table 1 The pe o mance o AF de ec ion algo i hm signals om MIT- BIH a hy hmia da abase, MIT-BIH a ial ib illa ion da abase, CinC chalange da abase (1s ed) and Long e m a ial ib illa ion da abase (Se – sensi i i y; PP – posi i e p edic i i y). Da abase SE [%] PP [%] MIT-BIH AF 96,32 98,61 MIT-BIH AR 98,42 90,78 CinC Ch. (1s ed) 93,04 87,30 Long Te m AF 95,60 80,27 Table 2 The alues o ue posi i e (TP), ue nega i e (TN), alse posi i e (FP), alse nega i e (FN) hea bea s o AF de ec ion algo i hm on signals om MIT-BIH a hy hmia da abase, MIT-BIH a ial ib illa ion da abase, CinC chalange da abase (1s ed) and Long e m a ial ib illa ion da abase. Da abase TP TN FP FN MIT-BIH AF 80964655 12502921 5 1764240 3093610 MIT-BIH AR 2891746 23887217 2424716 46321 CinC Ch. 1s ed) 1145 4908 76 748 Long Te m AF 43758987 2 28439682 5 6990500 3 2014541 9 The achie ed esul s o ou one- ea u e me hod a e compa able wi h o he au ho s o mo e complica ed and compu a ionally expensi e me hods [9], [10]. Fas a simplici y o ou me hod is use ul o example in applica ion whe e he ime is limi a ion and also o mobile de ices, whe e he low compu a ional complexi y is impo an . Biosignals p ocessing is nowadays an ac ual opic. In Figu e 3, examples o e o s caused by ou de ec ion algo i hm, a e shown. The blue lines indica e anno a ion om da abase, g een lines indica e ou esul s, alue 1 indica es AF, alue 0 indica es nonAF. In a) is shown mis ake caused by ea lie e mina ion AF o ou de ec o – p obably due o he ac , ha SH is compu ed om 95 RR in e als and ollowing RR in e als wi hou AF dec eased alue o SH. These mis akes o inaccu a e e mina ion and beginning o AF is ela i ely equen . In b) is shown ha ou de ec o poin ed o he le pa o he signal, whe e sup a en icula a hy hmia is p esen , as a ial ib illa ion, p obably due o he ac o highe a iabili y o RR in e als caused by his ype o a hy hmia as wi h AF is also p esen . Figu e 3 Examples o e o s caused by ou de ec ion algo i hm (blue line – anno a ion, g een line - ou esul , 1-AF, 0-nonAF), a) signal 100 om Long e m a ial ib illa ion da abase, b) signal 05261 om MIT-BIH a ial ib illa ion da abase. 4. Discu ions Du ing ou wo k, ou ECG expe s made a e y impo an inding. Public da abases con ain many e o s in anno a ions (in sense o AF) and also in o he pa hologies. I means ha esul s o all au ho s, who used hese da abases, a e a ec ed by e o s in anno a ions. Many e o s we e ound in Long-Te m AF, se e al also in MIT- BIH AF da abase and also in MIT-BIH A hy hmia da abase. Tes ing algo i hms on poo ly anno a ed da abases canno b ing eliable esul s and algo i hms Page 3 use ul in eal medical p ac ice. The examples o such anno a ions a e epo ed in Figu e 4. In he le pa o subg aph a) is p esen AF, bu in anno a ion om da abase (blue line) is no ma ked (signal no. 05261 om MIT-BIH a ial ib illa ion da abase), co ec anno a ion is ma ked by ed line. In subg aph b) is p esen AF all he ime, bu acco ding o he anno a ion is i no ue. The pa o he signal whe e he ex asys oles a e, is no ma ked as AF. Figu e 4 Examples o e o s in anno a ion o AF (blue - anno a ion in da abase, ed - co ec anno a ion, , 1-AF, 0- nonAF); a) signal 05261 om MIT-BIH a ial ib illa ion da abase, b) signal 100 om Long e m a ial ib illa ion da abase. 5. Conclusion In his wo k, highly e icien single ea u e algo i hm o a ial ib illa ion is p oposed. The achie ed esul s o ou one- ea u e me hod a e compa able wi h o he au ho s o mo e complica ed and compu a ionally expensi e me hods [9], [10]. Fas a simplici y o ou me hod is use ul o example in applica ion whe e he ime is limi a ion and also o mobile de ices, whe e he low compu a ional complexi y is impo an . In addi ion, ou ECG expe s ound ha in o en used es ing da abases o ECG signals a e e o s in anno a ion (in sense o AF) and also in o he pa hologies. Tes ing algo i hms on poo ly anno a ed da abases canno b ing eliable esul s and algo i hms use ul in eal medical p ac ice. The e is a place o hei co ec ion o c ea ion o new ones. Acknowledgmen s This wo k has been unded by he Uni ed S a es O ice o Na al Resea ch (ONR) Global, awa d numbe N62909-19- 1-2006. Re e ences [1] B. F. Kusumo o, “ECG in e p e a ion: om pa hophysiology o clinical applica ion,” Sp inge , pp. 107, 2009. [2] A. Ma ínez, R. Alca az, J.J. Rie a, “Applica ion o he phaso ans o m o au oma ic delinea ion o single- lead ECG iducial poin s,” Physiological Measu emen 31, pp. 1467-85, 2011. [3] L. Ma šáno á, e al., “Au oma ic de ec ion o P wa e in ECG du ing en icula ex asys oles,” P oceedings o he Wo ld Cong ess on Medical Physics and Biomedical Enginee ing, pp. 381-85, 2018. [4] L. Ma šáno á, A. Němco á, R. Smíšek, e al., “Ad anced P wa e de ec ion in ECG signals du ing pa hology: E alua ion in Di e en A hy hmia Con ex s,” Scien i ic Repo , no. 9, 2019. [5] L. Ma šáno á, e al., “Au oma ic de ec ion o P wa e in ECG du ing en icula ex asys oles,” Wo ld Cong ess on Medical Physics and Biomedical Enginee ing 2018, IFMBE P oceedings, pp. 381-385, 2018. [6] X. Zhou, e al., “A Real-Time A ial ib illa ion de ec ion algo i hm based on he ins an aneous s a e o hea a e,” PLoS ONE, ol. 10, no. 9, 2015. [7] X. Zhou, e al., “Au oma ic online de ec ion o a ial ib illa ion based on symbolic dynamics and Shannon en opy”, BioMedical Enginee ing OnLine, ol. 13 no. 1, 2014. [8] A. A dala, e al., “Au oma ic De ec ion o a ial ib illa ion using basic Shannon en opy o RR In e al ea u e,” Jou nal o Physics: Con e ence Se ies, ol. 795, 2017. [9] Y. Li, X. Tang, A. Wang,. e al., P obabili y densi y dis ibu ion o del a RR in e als: a no el me hod o he de ec ion o a ial ib illa ion. Aus alas Phys Eng Sci Med, ol. 40, pp. 707–716, 2017. [10] M. Lown, e al., Machine lea ning de ec ion o a ial ib illa ion using wea able echnology, PLoS ONE, ol. 15, no. 1, 2020. Add ess o co espondence: Ing. Lucie Ma šáno á Depa men o Biomedical Enginee ing Facul y o Elec ical Enginee ing and Communica ion B no Uni e si y o Technology Technická 12 B no 616 00 Czech Republic. [email p o ec ed] Page 4