Listen to this manuscript’s

audio summary by

JACC Editor-in-Chief

Dr. Valentin Fuster.

J O U R N A L O F T H E A M E R I C A N C O L L E G E O F C A R D I O L O G Y VO L . 7 1 , N O . 1 5 , 2 0 1 8

ª 2 0 1 8 B Y T H E A M E R I C A N C O L L E G E O F C A R D I O L O G Y F O U N D A T I O N

P U B L I S H E D B Y E L S E V I E R

Prediction of Abnormal MyocardialRelaxation From Signal ProcessedSurface ECG

Partho P. Sengupta, MD, DM,a,b Hemant Kulkarni, MD,c Jagat Narula, MD, PHDa

ABSTRACT

ISS

Fro

Un

An

de

Dr

the

Ma

BACKGROUND Myocardial relaxation is impaired in almost all cases with left ventricular diastolic dysfunction (LVDD)

and is a strong predictor of cardiovascular and all-cause mortality.

OBJECTIVES This study investigated the feasibility of signal-processed surface electrocardiography (spECG) as a

diagnostic tool for predicting the presence of abnormal cardiac muscle relaxation.

METHODS A total of 188 outpatients referred for coronary computed tomography (CT) angiography underwent an

echocardiogram for assessment of LVDD. The use of 12-lead spECG for predicting myocardial relaxation abnormalities as

identified using tissue Doppler echocardiography was validated with machine-learning approaches.

RESULTS A total of 188 subjects underwent diagnostic testing, with 133 (70%) showing abnormal myocardial relaxation

on tissue Doppler imaging. A 12-lead spECG showed an area under the curve of 91% (95% confidence interval: 86% to

95%) for prediction of abnormal myocardial mechanical relaxation with a sensitivity and specificity of 80% and 84%,

respectively. The spECG demonstrated more accurate diagnostic performance in individuals age $60 years as well as

those with obesity or hypertension, compared with their respective counterparts. Prediction of low early diastolic

relaxation velocity (e0) also correctly identified concomitant significant underlying coronary artery disease in 23 of 28

cases (82%). Furthermore, a superior integrated discrimination and net reclassification improvement was observed for

spECG over clinical features and traditional ECG.

CONCLUSIONS The spECG provides a robust prediction of abnormalmyocardial relaxation. These data suggest a potential

role for spECG as a novel screening strategy for identifying patients at risk for LVDD who would benefit undergoing echo-

cardiographic evaluations. (J Am Coll Cardiol 2018;71:1650–60) © 2018 by the American College of Cardiology Foundation.

L eft ventricular diastolic dysfunction (LVDD)appears early during any cardiovascular dis-ease and is recognized in approximately 20%

to 30% of the general adult population (1). Clinicalstudies demonstrate that early stages of LVDD canprogress to heart failure and are predictive of all-cause mortality, even after controlling for othercomorbidities (1,2). The current guidelines, therefore,recommend stratifying heart failure clinically into 4stages (stages A to D) (3). Stage A refers to patientswith risk factors with no structural or functional car-diac changes, whereas stage B includes patients

N 0735-1097/$36.00

m the aDepartment of Cardiology, Icahn School of Medicine at Mount Si

iversity Heart and Vascular Institute, West Virginia University, Morgant

tonio, Texas. This research was funded by an investigator-initiated gran

signing, acquisition or interpretation of data. Dr. Sengupta has served as a

. Kulkarni has served as a statistical consultant to HeartSciences. Dr. Narula

contents of this paper to disclose. Bijoy Khanderia, MD, served as Guest

nuscript received November 16, 2017; revised manuscript received Febru

with structural heart disease with no current or priorsymptoms of heart failure. The largest proportion ofpatients with stage B that were evaluated in clinicaltrials have an ischemic origin or subclinical myocar-dial dysfunction related to conditions such as dia-betes and hypertension. Nearly all conditions withassociated structural and functional cardiac changeshave concurrent impairment of LVDD that can bereadily identified using echocardiography (4). How-ever, echocardiography remains expensive, androutine clinical screening of asymptomatic stage A/Bpatients is not currently recommended (5).

https://doi.org/10.1016/j.jacc.2018.02.024

nai University, New York, New York; bWest Virginia

own, West Virginia; and cM&H Research, LLC, San

t from HeartSciences. The sponsors had no roles in

consultant for HeartSciences and Hitachi Aloka Ltd.

has reported that he has no relationships relevant to

Editor for this paper.

ary 1, 2018, accepted February 2, 2018.

AB BR E V I A T I O N S

AND ACRONYM S

A = late diastolic filling (atrial

contraction wave) velocity

BMI = body mass index

CAD = coronary artery disease

CT = computed tomography

E = peak early filling (diastolic

wave) velocity

eʹ = early diastolic relaxation

velocity

ECG = electrocardiogram

LVDD = left ventricular

diastolic dysfunction

ML = machine learning

spECG = signal-processed

surface electrocardiography

J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8 Sengupta et al.A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0 ECG Wavelets for Assessing Myocardial Relaxation

1651

The electrical and mechanical functions of cardiacperformance are closely coupled. A positive T-wavethat represents LV base-to-apex and transmuraladvancement of repolarization from the epicardiumto the endocardium is vital for normal LV mechanicalrelaxation (6,7). Subtle changes in the myocardialelectrical condition that may lead to myocardialrelaxation abnormalities, however, are not readilydiscerned on the surface electrocardiogram (ECG);therefore, routine application of ECG as a diagnostictool to assess LVDD is not commonly recognized. Aselectrical activity of the heart is highly dynamic,small changes in the surface ECG frequency spectrumare better discriminated by using signal-processingtechniques, and unique patterns can be extractedusing machine learning techniques (8–12). Such adevelopment may have a substantial effect on earlydetection of LVDD. Therefore, we explored the role ofa novel signal-processed surface ECG algorithm toextract electrophysiological signal patterns uniquelyassociated with abnormal myocardial relaxation.

SEE PAGE 1661

METHODS

We performed a prospective, cross-sectional study atthe Icahn School of Medicine at Mount Sinai (NewYork, New York). We initially recruited 200 unse-lected consecutive subjects in sinus rhythm, whowere referred from outpatient clinics for computedtomography (CT) coronary angiography. Subjectsunderwent 12-lead ECG, CT coronary angiography,and comprehensive 2-dimensional echocardiography(including tissue Doppler) in the same visit. Subjectswith arrhythmias, unstable angina, previous cardiacsurgery, a pacemaker, chest deformity, or an inabilityto express well-defined mitral annular velocities dueto severe mitral annular calcifications were excluded.Of the 200 subjects enrolled, we excluded 4 due toinadequate echocardiographic image quality, 3 due tosuboptimal electrocardiograms, and 5 others whowere unable to undergo CT scans. The resulting 188subjects included in the study had clinical, echocar-diographic, CT, ECG, and transformed ECG dataavailable. We also validated the normal ECG repolari-zation patterns with a comparison control cohort ofyoung sex-matched individuals with no known car-diac illness who had a normal echocardiogram andECG evaluation at the University of West Virginia(Morgantown, West Virginia). This comparison cohortwas recruited from an ongoing prospective investiga-tion that is evaluating the utility of signal-processedsurface ECG in detecting LVDD in populationwith a high prevalence of cardiac risk factors.

The institutional review board approved thestudy protocol, and all study participantsprovided written informed consent.

CLINICAL CHARACTERISTICS. We collectedand analyzed the following clinical charac-teristics of the study subjects: demographics,comorbidities, medications, body mass in-dexes, and laboratory data (including serumpotassium and renal function).

Hypertension was defined by systolic bloodpressure >140 mm Hg or diastolic blood pres-sure >90 mm Hg, physician-documented his-tory of hypertension, or by the use ofantihypertensive medications. Diabetes mel-litus was defined by the presence of physician-documented history of diabetes or use of oralhypoglycemic agents or insulin for the treat-

ment of hyperglycemia. Coronary artery disease (CAD)was defined by the presence of coronary stenosis of>50% on a coronary CT angiogram, history ofmyocardial infarction, or percutaneous intervention.Obesitywas defined as a bodymass indexes>30 kg/m2.

SIGNAL-PROCESSED SURFACE ECG. All subjectsunderwent baseline 12-lead surface ECG recording atthe time of the baseline echocardiogram. We used theUniversity of Glasgow Interpretive Analysis program(release 28.5, January 2014) for automated interpreta-tion of ECGs. This program has been extensively eval-uated and provides standard amplitude, duration, andaxes measurements, as well as a rhythm analysis anddiagnostic interpretation, and is well-suited for diag-nostic studies (13–15). The quantitative descriptionincludes average heart rate; the P, Q, R, and S waves(QRS); T axes; P and QRS durations; PR and QT in-tervals; and corrected QT (QTc) intervals. In addition,for qualitative summary descriptions (normal,borderline, abnormal), the software used the quanti-tative description to automatically classifythe electrocardiographic abnormalities according tothe Minnesota Code Manual.

Signal processing was performed using continuouswavelet transform mathematics (MyoVista hsECGInformatics, HeartSciences, Southlake, Texas). Usingwavelets for medical diagnostic purposes is a recentdevelopment, although the mathematical theory isnot new (8,16–18). The principles are similar to thoseof Fourier analysis, which were developed in theearly part of the 19th century. This processing con-verts an ECG signal into a normalized energy distri-bution in which frequency is shown on the y-axis (3 to200 Hz), time on the x-axis, and a colored spectrumrepresenting myocardial energy (colors scale 0 to255), with blue indicating the lowest energy and red

FIGURE 1 Continuous Wavelet Transform ECG

The original surface ECG (A) and its corresponding continuous wavelet transformation is

shown. The energy frequency in the wavelet transform ECG (B) is shown on the y-axis

(3 to 200 Hz), time on the x-axis, and the color spectrum represents energy (colors scaled

0 to 255), with blue representing the lowest energy and red the highest energy.

ECG ¼ electrocardiogram.

Sengupta et al. J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8

ECG Wavelets for Assessing Myocardial Relaxation A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0

1652

the highest energy (Figure 1). The continuous wavelettransform is time-aligned with the traditional ECGsignal for calculating multiple indices. For example,at a pre-defined time period prior to the T-wave peak,the maximum value of the MyoVista Color Waveform(normalized, unsigned continuous wavelet trans-form) across all frequencies (scales) is selected. Thisvalue is termed the ventricular index early measure(VIEM). Similarly, at a pre-defined time period afterthe T-wave peak, the maximum value of the MyoVistaColor Waveform across all frequencies is alsoselected. This value is termed the ventricular indexlate measure (VILM). The MyoVista energy algorithmcalled an Icon also classifies the patients into 3 cate-gories of energies (taking into account factors such assex and age) with grade 3 being the lowest and grade 1

referring to the highest energy levels. A list ofdifferent indices used in the analysis is available inthe Online Supplementary Data File.

ECHOCARDIOGRAPHY. All subjects underwent acomplete 2-dimensional Doppler echocardiogram andtissue Doppler echocardiographic examination usingan iE33 system (Philips Medical Systems, Andover,Massachusetts), with additional dedicated imaging ofmitral inflow using pulsed-wave Doppler echocardi-ography and pulsed-wave tissue Doppler echocardi-ography of the septal and lateral mitral annulusaccording to guidelines published by ASE (4). Atrained single reader, blinded to the subject’s ECG,Glasgow Interpretive Analysis, and clinical data,reviewed the echocardiograms. Utilizing the apical 2-and 4-chamber loops, the LV end-diastolic volume,end-systolic volume, and ejection fraction werecalculated using the biplane Simpson’s method ofdiscs, and the left atrial (LA) maximum volume wascalculated using the biplane area length method. Allmeasurements were made in $3 consecutive cardiaccycles, and average values were used for final ana-lyses. The pulsed-wave Doppler-derived transmitralvelocity and digital color tissue Doppler-derivedmitral annular velocities were obtained from theapical 4-chamber view. The early diastolic wave ve-locity (E) and late diastolic atrial contraction wavevelocity (A) were measured using a pulsed-waveDoppler recording. Continuous-wave Doppler wasapplied on the tricuspid valve in different windows(apical 4-chamber view, parasternal right ventricularinflow view, and parasternal short-axis view). Thetricuspid regurgitation signal was recorded, and thetricuspid regurgitation maximum velocity wasmeasured as the highest value recorded from allviews. Spectral pulsed-wave tissue Doppler-derivedearly diastolic relaxation velocity (eʹ) were also ob-tained from the septal and lateral mitral annular po-sition. Finally, the E/eʹ and E/A ratios were calculatedas a Doppler echocardiographic estimate of the LVfilling pressure. All measurements were made in $3consecutive cardiac cycles and average values wereused for final analyses. The following echocardio-graphic parameters were defined as normal orabnormal, with the given cutoffs for abnormality: 1)abnormal e0 (septal eʹ velocity <7 cm/s and/or laterale0 velocity <10 cm/s); 2) abnormal E/e0 (averaged fromseptal and lateral wall E/e0 >14); 3) abnormal TR ve-locity (>2.8 m/s); and 4) abnormal left atrial volumeindex (>34 ml/m2) (4).

MACHINE LEARNING–BASED CLASSIFICATION. Ouranalytical, unsupervised approach to machinelearning (ML)–based classification included the first

TABLE 1 Comparison of Clinical Characteristics Across Study Groups Defined by Low e0

Low e0 (n ¼ 133) Normal e0 (n ¼ 55) p Value

Age, yrs <0.0001*

<60 60 (45.1) 44 (80.1)

$60–<70 48 (36.1) 8 (14.5)

$70 25 (18.8) 3 (5.4)

Sex 0.70*

Male 54 (40.6) 24 (43.6)

Female 79 (59.4) 31 (56.4)

Race 0.88*

White 63 (47.4) 25 (45.5)

Black 33 (24.8) 13 (23.6)

Hispanic/Mexican 25 (18.8) 13 (23.6)

Asian 12 (9.0) 4 (7.3)

Ever smoker 0.18*

Yes 60 (45.1) 19 (34.5)

No 73 (54.9) 36 (65.5)

Obesity, BMI $30 kg/m2 61 (45.8) 14 (25.4) 0.0093*

Large body surface (BSA $2.30 m2) 36 (27.0) 12 (21.8) 0.45*

NYHA functional class 0.27*

I 81 (61.4) 40 (72.7)

II 41 (31.1) 14 (25.5)

III 8 (6.1) 1 (1.8)

IV 3 (2.4) 0 (0.0)

Blood pressure, mm Hg

Systolic 125.8 � 19.5 115.2 � 14.6 0.0004†

Diastolic 72.8 � 9.2 69.6 � 9.7 0.0336†

Mean 90.3 � 11.4 84.7 � 10.7 0.0020†

Clinical chemistry‡

Blood glucose, mg/dl (n ¼ 108/47) 102.3 � 32.3 94.6 � 19.2 0.13†

Serum Na, mmol/l (n ¼ 108/47) 140.3 � 2.4 140.0 � 2.2 0.45†

Serum K, mmol/l (n ¼ 109/47) 4.3 � 0.3 4.2 � 0.3 0.40†

Serum cholesterol, mg/dl

Total (n ¼ 120/49) 187.2 � 44.2 186.9 � 41.6 0.97†

HDL (n ¼ 122/49) 53.1 � 14.2 54.9 � 17.8 0.47†

LDL (n ¼ 122/48) 107.6 � 36.3 104.1 � 33.7 0.55†

Serum triglycerides, mg/dl (n ¼ 122/49) 136.0 � 89.4 142.3 � 93.4 0.68†

Serum creatinine, mg/dl (n ¼ 115/50) 0.9 � 0.2 0.9 � 0.2 0.67†

Stenosis confirmed by CT 0.0010

No stenosis 48 (36.1) 36 (65.4)

Mild/moderate stenosis 61 (45.9) 15 (27.3)

Severe stenosis 24 (18.0) 4 (7.3)

Values are n (%) or mean � SD. *Chi-square test. †Student’s t-test. ‡Numbers are available observations from thelow e0/normal e0 groups.

BMI ¼ body mass index; CT ¼ computed tomography; eʹ ¼ early diastolic relaxation velocity;HDL ¼ high-density lipoprotein; LDL ¼ low-density lipoprotein; NYHA ¼ New York Heart Association.

J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8 Sengupta et al.A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0 ECG Wavelets for Assessing Myocardial Relaxation

1653

step of choosing the most appropriate combination ofclassifier and sample-splitting methods. We used theR package chromosomal microarray (R Foundation forStatistical Computing, Vienna, Austria) (19), whichpermitted a direct comparison of 8 different classi-fiers (component-wise boosting, diagonal discrimi-nant analysis, partial least squares-lineardiscriminant analysis, shrinkage discriminant anal-ysis, feed-forward neural networks, probabilisticneural networks, random forest, and support vectormachine) and 3 methods of sample-splitting (10-foldcross-validation, Monte Carlo cross-validation, andbootstrapping) with a fixed random seed. For each ofthese combinations, we estimated the misclassifica-tion rates, Brier scores, and average estimated prob-ability of low e0. Based on these parameters, we chosea classifier and splitting method that best capturedthe data structure.

We used the random forest ensemble classifier(20) with a Monte Carlo cross-validation procedure toclassify the study subjects. The random forest clas-sifier was performed with the following specifica-tions: input number of variables, 370; type ofoutcome, dichotomous classification; number oftrees, 500; available number of variables for splittingat each node, 19; and minimum node size, 5. Errorrates were estimated from the out-of-bag trees.Variable importance was measured using apermutation-based metric (21) that captured themean improvement in prediction attributable to agiven variable. The novel procedure described byJanitza et al. (22) was used to test the significance ofvariable importance. We performed a total of 20 it-erations of the model, with each splitting the sampleinto a training (67%) and a testing (33%) subset. Themultiple iterations were designed to accumulate atleast 2 occurrences of each study participant in thetesting subset. Predicted probabilities of low e0,variable importance, as well as out-of-bag error rateswere aggregated and averaged over the 20 iterationsof random forests.

STATISTICAL ANALYSIS. Between-group compari-sons were conducted using the Pearson’s chi-squaretest (for goodness of fit) or Fisher exact test (for cat-egorical variables) and Student’s t-test (for contin-uous variables) after testing for normal distributionusing the Kolmogorov-Smirnov test. Predictive accu-racy and screening performance was assessed byestimating the area under the receiver-operatingcharacteristic curve. Differences between potentialmoderator effects of clinical and ECG covariates onthe predictive accuracy were evaluated usingCochrane’s Q test for heterogeneity of the area under

the curve across categories of a moderating variable.Incremental value of the random forest-based classi-fier over other clinical predictors of low e0 wasdetermined using the incremental discriminationimprovement and continuous version of the netreclassification index (23). We used Stata 14.0 (StataCorp, College Station, Texas) for all statistical ana-lyses. The integrated discrimination index and netreclassification index were estimated using the Statapackage IDI (Michael Lunt, University of Manchester,

TABLE 2 Association of Low e0 With Other Echo Parameters

n* Low e0 Normal e0 p Value

Ejection fraction, % 133/55 62.6 � 6.8 62.7 � 4.4 0.90

LV internal diameter

End systole, cm 132/55 2.8 � 0.5 2.9 � 0.4 0.38

End diastole, cm 133/55 4.5 � 0.6 4.5 � 0.6 0.68

Interventricular septal thickness at enddiastole, cm

133/55 1.0 � 0.2 0.9 � 0.2 0.0001

LV posterior wall thickness at enddiastole, cm

133/55 1.0 � 0.2 0.8 � 0.1 <0.0001

LV mass index, g/m2 133/55 82.7 � 24.1 74.7 � 27.9 0.05

LVH classification 133/55 0.0013

Normal 54 (40.6) 38 (69.1)

Concentric remodeling 58 (43.6) 10 (18.2)

Eccentric hypertrophy 8 (6.0) 5 (9.1)

Concentric hypertrophy 13 (9.8) 2 (3.6)

Left atrium volume index 129/52 35.5 � 13.8 33.4 � 12.7 0.36

Mitral valve E velocity, cm/s 131/55 0.7 � 0.2 0.7 � 0.1 0.10

Mitral value A velocity, cm/s 131/55 0.7 � 0.2 0.6 � 0.1 <0.0001

E/A ratio 130/55 0.9 � 0.3 1.3 � 0.4 <0.0001

Average e0, cm/s 133/55 6.9 � 1.4 10.7 � 1.5 Not tested

Average E/e0 131/55 10.7 � 3.9 7.2 � 1.8 <0.0001

Mitral valve deceleration time, ms 131/55 250.4 � 69.4 227.6 � 46.4 0.0269

Moderate-to-severe regurgitation

Mitral valve 133/55 5 (3.7) 5 (9.0) 0.15†

Aortic valve 132/55 2 (1.5) 1 (1.8) 1.00†

Pulmonary valve 132/55 — — —

Tricuspid valve 132/55 1 (0.7) 1 (1.8) 0.50†

Number of DD defining variables 130/55

0 2 (1.5) 32 (58.2)

1 59 (45.4) 22 (40.0)

2 48 (36.9) 1 (1.8)

3 17 (13.1) 0 (0.0)

4 4 (3.1) 0 (0.0)

Overall test of significant difference <0.0001‡

Test of linear trend <0.0001

AS, aortic valve peak velocity >3 m/s 132/55 7 � 5.3 0 � 0.0 0.10

Right ventricular systolic pressure,mm Hg

93/37 26.5 � 13.7 23.9 � 5.9 0.27

Values are mean � SD or n (%). All p values are from Student’s t-test, unless otherwise indicated. *Number ofavailable observations from the low e0/normal e0 groups. †Fisher exact test. ‡Chi-square test.

AS ¼ aortic stenosis; DD ¼ diastolic dysfunction; LV ¼ left ventricular; LVH ¼ left ventricular hypertrophy.

Sengupta et al. J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8

ECG Wavelets for Assessing Myocardial Relaxation A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0

1654

Manchester, United Kingdom). Statistical significancewas tested at global type I error rate of 0.05.

RESULTS

Clinical characteristics of the 188 study subjects areshown in Table 1. The prevalence of low e0 was 70% inthe study sample, and the risk of low e0 was signifi-cantly higher in individuals age $60 years and inthose who were obese, hypertensive, or had a CT-confirmed stenosis (Table 1). Several echocardio-graphic features coexisted with low e0 status,including LV wall thickness, concentric remodeling or

hypertrophy, as well as a number of variables that areused clinically for defining the presence of DD(Table 2).

PREDICTIVE PERFORMANCE OF THE SIGNAL-PROCESSED

SURFACE ECG–BASED RANDOM FOREST CLASSIFIER. Ofthe 8 ML classifiers used to identify low e0 based on370 features from the signal-processed surfaceelectrocardiography (spECG), consistency and theleast error-prone performance was provided by acombination of random forest classification andMonte Carlo cross-validation splitting strategy(Online Table 1). The area under the curve for pre-diction of low e0 using this random forest-basedclassifier was 91% (95% confidence interval [CI]:86% to 95%), with a sensitivity and specificity of80% and 84%, respectively (Figure 2). The Janitzatest indicated that a total of 257 features (of the 370detailed in the Online Supplementary Data File)were significantly important. The top 25 features areshown in Figure 3A. The variable that was most vital inbuilding the classifier was ICON, which represents aderived energy measure that captures overall cate-gories of the energy kinetics of the myocardium.Detailed analyses revealed that the ICON variablealone had a significantly independent and additiveuse in the prediction of low e0 (Online Table 2).To validate the distribution of ICON was not age- andpopulation-specific, we analyzed the distributionof ICON repolarization energy index in a young(45 � 8.7 years, 49% male) cohort from West Virginiawith normal echocardiograms. The ICON energy gradedistribution (Grade 1: normal, Grade 2: intermediate,and Grade 3: low) of the young cohort with normaleʹ closely mirrored the subject from New York withnormal eʹ (Online Table 3). The error rates of therandom forest model from multiple iterations areshown as mean (dark red curve) and confidencebands (pink curves) in Figure 3B. It is noticeablethat the error rates converged after nearly 150 trees inall iterations.

Coronary artery stenosis and the presence of 2 ormore variables of DD significantly coexistedwith low e0

(Tables 1 and 2). The spECG-based random forestclassifier could also detect 23 of the 28 (82%) caseswith significant CAD (>50% stenosis) and 54 of70 (77%) cases with $2 abnormal echocardiographicfeatures of DD. Therefore, when a low e0 was pre-dicted, the post-test probability of stenosis rose to64% from a pre-test probability of 55% (likelihoodratio: 1.42; 95% CI: 1.11 to 1.82). Similarly, thepost-test probability of DD increased to 48% from apre-test probability of 38% (likelihood ratio: 1.50;95% CI: 1.21 to 1.87).

FIGURE 2 Role of the spECG Classifier in Prediction of Relaxation Abnormality

1.00

0.75

0.50

0.25

0.00

0.00 0.25 0.50

Area under the ROC curve0.9057 (95% CI: 0.8634 - 0.9480)

0.75 1.001 – Specificity

Predictive Performance of the RF ModelIllustrative Example 1

Sign

al p

roce

ssed

ECG

Tiss

ue D

oppl

erEC

G

Illustrative Example 2 BA

Sens

itivi

ty

Sensitivity:Specificity:LR+LR-

0.8045 (0.7268 - 0.8681)0.8364 (0.7120 - 0.9223): 4.92 (2.69 - 8.99): 0.23 (0.16 - 0.34)

(A) Two illustrative examples of energy characteristics obtained for 2 patients: 1 with low e0 (right) and 1 with a normal e0 (left). The spECG recordings were syn-

chronized with the tissue Doppler (top track) and conventional ECG (second track from top). Energy dynamics are as follows: blue represents low energy, and red

represents high energy. Note the normal e0 associated with a corresponding strong T-wave energy spectrum in example 1. In contrast, note the attenuated e0 and

reduced T-wave energy spectrum in example 2. (B) ROC for the prediction of low e0 using the ascribed probabilities of the spECG-based random forest classifier. The

optimal cutoff is indicated with an orange circle and its approximation to the upper left corner is shown using a dashed line. At the optimal cutoff, the diagnostic

performance was quantified and is shown in the box next to the cutoff. eʹ¼ early diastolic relaxation velocity; LRþ ¼ positive likelihood ratio; LR- ¼ negative likelihood

ratio; RF ¼ random forest; ROC ¼ receiver-operating characteristic; spECG ¼ signal-processed surface electrocardiography. Other abbreviation as in Figure 1.

J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8 Sengupta et al.A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0 ECG Wavelets for Assessing Myocardial Relaxation

1655

PUTATIVE MODERATION OF THE PREDICTIVE

PERFORMANCE OF THE spECG-BASED RANDOM

FOREST CLASSIFIER. We investigated whetherimportant clinical and ECG covariates are likely toinfluence the predictive performance of the spECG-based random forest classifier. Our results indicatedthat individuals $60 years of age, those who wereobese, and those with hypertension had a betterpredictive performance compared with their respec-tive counterparts (Table 3). Similarly, individuals witha “borderline” or “abnormal” rating on the GlasgowInterpretive Analysis were associated with a margin-ally better prediction of low e0 by the random forestclassifier. However, this gradient across the ECGdiagnosis criteria was not statistically heterogeneous(Table 3). No other variables demonstrated a statisti-cally significant moderator-type association with thepredictive performance of the random forest classi-fier, including CT-confirmed stenosis.

INCREMENTAL VALUE OF THE spECG-BASED

RANDOM FOREST CLASSIFIER TO DETECT LOW E 0.

We next determined whether addition of the random

forest classifier-based prediction of low e0 had incre-mental value above and beyond the three clinical fea-tures associated with a high risk of low e0, age $60years, obesity, and hypertension. We found a sharpimprovement in prediction due to the classifier suchthat it improved the area under the curve by 19%(p < 0.001), integrated discrimination index by 0.42(p<0.001), and correctly reclassifiedover 80%of lowe0

and normal individuals (p < 0.001) (Table 4, Model 1).Furthermore, upon addition of Glasgow riskcategorization for surface ECG to this model, thespECG-based random forest classifier continued tosignificantly improve the area under the curve (by16%), integrated discrimination index (by 0.40), andreclassification (correctly reclassifying over 80% ofindividuals) (Table 4, Model 2).

DISCUSSION

The high prevalence of LVDD in the general commu-nity, the inability of a physical examination to reli-ably detect LV dysfunction, and the limited utility of

FIGURE 3 Features of the spECG-Based Random Forest Classifier

2 2.5 3 3.5 4

Variable ImportancePermutation Variable Importance Measure

4.5 5 5.5 6

icon

irem2

ilo1 value

avrp51 value

v1p21 value

viemlv/vilmlv

v4p21 value

v6p21 value

avrp21 value

v3dtm

v1rlm

v1rem

iiirlm2

ip51 value

iip53 value

ip21 value

avfrlm

avfp53 value

v1drate

v1rem2

idrate

iram

iip51 value

irlm

irem

0.45

0.41

0.37

0.33

0.29

0.251 50045040035030025020015010050

Tree Number

Out-of-Bag Error Rate

A

B

OOB

Erro

r Rat

e

(A) Variable importance for the top 25 variables. The importance measure

was based on permutation and contribution of the variable to the

accuracy of the model. Details of the importance estimates are provided in

Online Table 2. (B) Out-of-bag (OOB) error rates for the RF model.

Other abbreviation as in Figure 2.

Sengupta et al. J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8

ECG Wavelets for Assessing Myocardial Relaxation A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0

1656

a standard ECG mandate physician discernment toidentify those individuals who should undergo anechocardiogram. Although ECG is a widely usedtechnical procedure for the evaluation of cardiovas-cular function, there is no single or explicit ECGpattern that is predictive of the presence of LVDD. Inthis investigation, we combined measurements fromwavelet-transformed ECG with ML techniques toextract time-frequency indexes and features of thesignal-processed ECG. This was done with the over-arching aim of developing an algorithm for auto-mated diagnosis of abnormal myocardial relaxation.The ML method aided in circumventing the extremelytedious task of manually recognizing features fromthe wavelet-transformed ECG signal, where patternsconvey significant information regarding develop-ment of DD (Central Illustration). The spECG demon-strates a robust prediction of myocardial relaxationabnormalities as seen on echocardiography. More-over, the prediction of abnormal relaxation alsoallowed ready recognition of subjects with moreadvanced stages of DD and concurrent CAD withsignificantly more incremental value compared withclinical variables and surface ECG.

With the advent of computerized techniques, therehas been a resurgence of interest in correlating sur-face ECG with features of LVDD. There has also been agrowing interest in different linear and nonlineartechniques for identifying subjects with relaxationabnormalities and for determining risks for heartfailure. The present study extends these observationsand provides strong evidence that use of ML tech-niques can help isolate discernible features from ECGwavelets for robust estimation of LV relaxation ab-normalities. Specifically, the wavelet transformtechnique in this investigation utilized the waveletenergy in the peak of QRS to scale the appearance ofthe energy in the T-wave display. This relative dis-tribution of energy colors improved the signal-to-noise ratio and magnified the repolarization energysignals, which allowed for extraction of featuresassociated with reduced early diastolic myocardialrelaxation velocity.

Cellular work, animal models, and human cohortshave demonstrated that electrocardiographic repo-larization changes in the T-wave are accompanied byechocardiographic signs of DD (24–28). The link be-tween electrical repolarization and diastolic me-chanics may be mediated by changes in calciumhandling. Indeed, myocardial ischemia, hypertrophy,aging, and risk factors such as hypertension, diabetes,or smoking are associated with sarcoplasmic

TABLE 4 Evaluation of Moderator Effects on the Diagnostic Performance of MyoVista to

Detect Low e0

Moderator Variablesand Categories n AUC (95% CI) Q (df)* phet Value

Overall 188 0.9057 (0.8634–0.9480) —

Age, yrs 11.64 (1) 0.0006

<60 104 0.8803 (0.8182–0.9424)

$60 84 0.9913 (0.9769–1.0000)

Sex 3.62 (1) 0.06

Male 78 0.8519 (0.7694–0.9353)

Female 110 0.9428 (0.9002–0.9855)

Race 3.37 (3) 0.33

White 88 0.9048 (0.8413–0.9681)

Black 46 0.9510 (0.8952–1.0000)

Hispanic/Mexican 38 0.8862 (0.7771–0.9953)

Asian 16 0.7083 (0.3811–1.0000)

Ever smoker 0.69 (1) 0.40

Yes 79 0.9289 (0.8680–0.9899)

No 109 0.8931 (0.8342–0.9519)

Obese (BMI $30 kg/m2) 4.57 (1) 0.0326

Yes 75 0.9625 (0.9224–1.0000)

No 113 0.8852 (0.8269–0.9435)

BSA ($2.30 m2) 0.09 (1) 0.76

Top quartile 48 0.8912 (0.8001–0.9823)

First to third quartile 140 0.9067 (0.8574–0.9561)

NYHA functional class 0.30 (1) 0.58

I 121 0.9071 (0.8558–0.9584)

II 55 0.9321 (0.8588–1.0000)

Hypertension NA —

Yes 33 1.0000

No 155 0.8986 (0.8517–0.9455)

Traditional ECG classification 3.65 (2) 0.16

Normal 108 0.8829 (0.8219–0.9440)

Borderline 55 0.9451 (0.8795–1.0000)

Abnormal 25 0.9697 (0.8975–1.0000)

CT-confirmed stenosis 0.26 (2) 0.87

No stenosis 84 0.9080 (0.8490–0.9670)

Mild/moderate stenosis 76 0.8809 (0.7938–0.9679)

Severe stenosis 28 0.9063 (0.7160–1.0000)

*Cochrane’s Q statistic of heterogeneity.

AUC ¼ area under the curve; BMI ¼ body mass index; BSA ¼ body surface area; CI ¼ confidence interval;CT ¼ computed tomography; df ¼ degrees of freedom; ECG ¼ electrocardiogram; eʹ ¼ early diastolic velocity;NYHA ¼ New York Heart Association functional classification.

TABLE 3 Improved Prediction of Low e0 by the RF-Based

Classifier Above and Beyond Other Clinical Correlates

Measure of Improvement Estimate Significance

Model 1: Covariates—Age Categories, Obesity, Hypertension

Improvement in area under the ROC curve 0.1933 <0.0001

Improvement in LR chi-square test (1 df) 90.44 <0.0001

Akaike information criterion 94.81 <0.0001

Integrated discrimination index 0.4189 <0.0001

Net reclassification improvement <0.0001

Proportion of individuals with low e0 withhigher probabilities

84.21%

Proportion of normal individuals withlower probabilities

81.82%

NRI 1.3206

Model 2: Covariates—Age Categories, Obesity,Hypertension, and Conventional ECG Classification

Improvement in area under the ROC curve 0.1605 <0.0001

Improvement in LR chi-square test (1 df) 88.60 <0.0001

Akaike information criterion 94.81 <0.0001

Integrated discrimination index 0.4008 <0.0001

Net reclassification improvement <0.0001

Proportion of individuals with low e0 whohad higher probabilities

84.96%

Proportion of normal individuals who hadlower probabilities

81.82%

NRI 1.3356

df ¼ degrees of freedom; ECG ¼ electrocardiogram; LR ¼ likelihood ratio;NRI ¼ Net Reclassification Index; ROC ¼ receiver-operating characteristic.

J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8 Sengupta et al.A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0 ECG Wavelets for Assessing Myocardial Relaxation

1657

reticulum calcium uptake inhibition (25). Delayeduptake of calcium is pathophysiologically associatedwith a prolongation of action potential and QT inter-val (26). A prolonged QT interval has been furthercorrelated with progressive reduction of tissueDoppler-derived relaxation velocities in several clin-ical studies (27–29). In addition to the QT interval,animal and cellular experiments have also investi-gated a specific relationship between transmuralheterogeneity of relaxation and the time interval be-tween peak and end of the T-wave (30). A recentstudy indicated an inverse, linear correlation be-tween the time interval between peak and end of theT-wave observed on surface ECG and tissue Doppler-derived early diastolic LV longitudinal myocardialrelaxation velocities (31). Another study used unsu-pervised machine learning to divide heart failure withpreserved ejection fraction into 3 phenotypes withdifferent risk profiles. The T-wave peak to end dura-tion was associated with higher B-type natriureticpeptide level, lower eʹ velocity, and a high-riskphenotype classification (32). In all of these studies,however, the correlations were modest, and the ECGindexes were not robust as diagnostic tests, mostlikely due to simplification by using only 1 ECG

feature. We therefore selected a multifeature extrac-tion method using machine learning to assess thecapability of spECG to predict abnormal myocardialrelaxation in patients with risk factors who wereotherwise scheduled for CT coronary angiogram. ThespECG not only revealed an ability to predict cardiacrelaxation abnormalities, but 23 of the 28 cases (82%)with significant CAD (defined as >50% stenosis of theepicardial coronary arteries) were identified usingspECG. A technique similar to spECG may, therefore,be a useful tool in daily clinical cardiology practices

CENTRAL ILLUSTRATION Continuous Wavelet-BasedTime-Frequency Analysis of an ECG Signal for PredictingMechanical Relaxation Abnormalities

Sengupta, P.P. et al. J Am Coll Cardiol. 2018;71(15):1650–60.

The conventional electrocardiogram (ECG) (A) and signal-processed surface electro-

cardiography (spECG) (B) are aligned with tissue Doppler echocardiography-derived

longitudinal mitral annular velocity waveforms from left ventricular septum (C) during

ejection (sʹ), early (eʹ) and late diastolic relaxation (aʹ). The left- and right-side of each

panel shows the findings from subjects with normal and impaired left ventricular (LV)

relaxation, respectively. Energy dynamics in spECG are as follows: blue represents low

energy, and red represents high energy. Note the relationship of normal repolarization

energies during the T-wave (B, dotted box) associated with normal e0. The dotted

arrows (C) show the timing of the end of the T-wave in relationship to the tissue

Doppler waveforms. In contrast, note the reduced T-wave energy spectrum associated

with attenuated eʹ in the subject with impaired LV relaxation. Features from spECG were

extracted using a random forest classifier to build a diagnostic spECG algorithm for

predicting LV mechanical relaxation abnormalities.

Sengupta et al. J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8

ECG Wavelets for Assessing Myocardial Relaxation A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0

1658

where screening strategies to identify subclinicalLVDD and CAD risks are often clustered together.

STUDY LIMITATIONS. Some limitations of this studyshould be considered before the results can begeneralized. First, this was a feasibility study with alimited number of patients. We determined the posthoc statistical power of this dataset to detectobserved differences in the area under the curveconsequent to inclusion of the random forest clas-sifier into the discriminating model using the pROCR package (33). Analyses revealed that the esti-mated post hoc statistical power of Models 1 and 2were 0.9788 and 0.9562, respectively (Table 3).

These data combined with the knowledge that theRF-based classifier somewhat obviates the need formultiple comparisons, because it demonstrates thatthese estimates had adequate statistical power toaddress the study question. Second, the trade-offbetween a simpler predictor (e.g., based on theICON variable only as shown in Online Table 3) or amore comprehensive model such as the one used inthis study remains to be assessed in future studies.Third, the sensitivity and specificity were 80% and84%, respectively, and the reason for false positiveor negative results may be related to a relativelyfixed definition of abnormal eʹ without clinical ad-justments for age- and sex-related variations in eʹ.For example, a value of 7 cm/s for septal eʹ may beextremely low velocity for defining abnormalrelaxation in a younger population. Future studieswould need to address the distribution of spECG incomparison with age- and sex-related ranges in eʹand diastolic dysfunction parameters. Fourth,wavelet transform analysis over time has beenapplied to a variety of biomedical signals withpredictable accuracy; however, the assessment ofLVDD is a new application, and the reproducibilityover time during serial visits would require evalu-ation in future investigations. Fifth, the use ofspECG was evaluated for detecting abnormal LVrelaxation, which refers to early stages of diastolicdysfunction where many patients may not have yetadvanced features, including left atrial enlargementand elevation of left atrial and LV filling pressures.Future studies would be required to understandthe value of spECG in predicting progressivegrades of LVDD. Similarly, it would be valuable toinvestigate the role of spECG in assessing LVstrain abnormalities, which are more sensitivemarkers of cardiac dysfunction. Finally, etiology-specific differences in diagnostic performance ofspECG were not addressed in the present investi-gation and would require future in-depthexploration.

CONCLUSIONS

ECG remains one of the widely used diagnosticscreening test procedures in cardiology. AlthoughLVDD occurs early in cardiovascular disease, the useof ECG as a screening technique for detecting LVDDremains unrecognized. The present investigationused signal-processing techniques to magnify smallchanges on the surface ECG frequency spectrumassociated with development of myocardialrelaxation abnormalities and showed incremental

PERSPECTIVES

COMPETENCY IN PATIENT CARE AND PROCEDURAL

SKILLS: Early repolarization abnormalities identified by spECG

are associated with abnormal myocardial relaxation detected by

tissue Doppler echocardiography. These findings are incremental

to clinical symptoms and signs and information obtained from

the surface ECG.

TRANSLATIONAL OUTLOOK: Population-based studies are

needed to evaluate the diagnostic use of spECG for detection of

the earliest stages of diastolic dysfunction in various types of

cardiac disease and monitoring the myocardial response to

therapy.

J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8 Sengupta et al.A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0 ECG Wavelets for Assessing Myocardial Relaxation

1659

value of spECG over conventional ECG for robustprediction of myocardial relaxation abnormalities.These data suggest a potential role for spECG as anovel screening mechanism for LVDD that can beutilized in population-based studies to identify pa-tients that would benefit from echocardiographicevaluations.

ACKNOWLEDGMENTS The authors thank Drs. AlaaOmar, Allen Weiss, and Ahmad Mahmoud for theirhelp in compiling the echocardiography database.

ADDRESS FOR CORRESPONDENCE: Dr. Partho P.Sengupta, Heart and Vascular Institute, West VirginiaUniversity, 1 Medical Center Drive, Morgantown,West Virginia 26506-8059. E-mail: Partho.sengupta@wvumedicine.org.

RE F E RENCE S

1. Wan SH, Vogel MW, Chen HH. Pre-clinical dia-stolic dysfunction. J Am Coll Cardiol 2014;63:407–16.

2. Kane GC, Karon BL, Mahoney DW, et al.Progression of left ventricular diastolic dysfunc-tion and risk of heart failure. JAMA 2011;306:856–63.

3. Hunt SA, Abraham WT, Chin MH, et al. ACC/AHA2005 guideline update for the diagnosis andmanagement of chronic heart failure in the adult: areport of the American College of Cardiology/American Heart Association Task Force on PracticeGuidelines (Writing Committee to Update the2001 Guidelines for the Evaluation and Manage-ment of Heart Failure). J Am Coll Cardiol 2005;46:e1–82.

4. Nagueh SF, Smiseth OA, Appleton CP, et al.Recommendations for the evaluation of left ven-tricular diastolic function by echocardiography: anupdate from the American Society of Echocardi-ography and the European Association of Cardio-vascular Imaging. Eur Heart J Cardiovasc Imaging2016;17:1321–60.

5. Manning WJ, Patel AR, Picard MH, et al. ACCF/ASE/AHA/ASNC/HFSA/HRS/SCAI/SCCM/SCCT/SCMR 2011 appropriate use criteria for echo-cardiography: a report of the American Col-lege of Cardiology Foundation Appropriate UseCriteria Task Force, American Society of Echo-cardiography, American Heart Association,American Society of Nuclear Cardiology, HeartFailure Society of America, Heart Rhythm Soci-ety, Society for Cardiovascular Angiography andInterventions, Society of Critical Care Medicine,Society of Cardiovascular Computed Tomogra-phy, Society for Cardiovascular Magnetic Reso-nance American College of Chest Physicians.J Am Coll Cardiol 2011;57:1126–66.

6. Brado J, Dechant MJ, Menza M, et al. Phase-contrast magnet resonance imaging revealsregional, transmural, and base-to-apex dispersionof mechanical dysfunction in patients with long QTsyndrome. Heart Rhythm 2017;14:1388–97.

7. Sengupta PP, Khandheria BK, Korinek J, et al.Apex-to-base dispersion in regional timing of leftventricular shortening and lengthening. J Am CollCardiol 2006;47:163–72.

8. Crowe JA, Gibson NM, Woolfson MS,Somekh MG. Wavelet transform as a potential toolfor ECG analysis and compression. J Biomed Eng1992;14:268–72.

9. Shekatkar SM, Kotriwar Y, Harikrishnan KP,Ambika G. Detecting abnormality in heart dy-namics from multifractal analysis of ECG signals.Sci Rep 2017;7:15127.

10. Addison PS. Wavelet transforms and the ECG:a review. Physiol Meas 2005;26:R155–99.

11. Li Q, Rajagopalan C, Clifford GD. A machinelearning approach to multi-level ECG signal qual-ity classification. Comput Methods Programs Bio-med 2014;117:435–47.

12. Sudarshan VK, Acharya UR, Oh SL, et al.Automated diagnosis of congestive heart failureusing dual tree complex wavelet transform andstatistical features extracted from 2s of ECG sig-nals. Comput Biol Med 2017;83:48–58.

13. Clark EN, Sejersten M, Clemmensen P,Macfarlane PW. Automated electrocardiograminterpretation programs versus cardiologists’triage decision making based on teletransmitteddata in patients with suspected acute coronarysyndrome. Am J Cardiol 2010;106:1696–702.

14. Macfarlane PW, Norrie J, Committee WE. Thevalue of the electrocardiogram in risk assessmentin primary prevention: experience from the Westof Scotland Coronary Prevention Study.J Electrocardiol 2007;40:101–9.

15. Fayn J, Rubel P, Macfarlane PW. Can the les-sons learned from the assessment of automatedelectrocardiogram analysis in the Common Stan-dards for quantitative Electrocardiography studybenefit measurement of delayed contrast-enhanced magnetic resonance images?J Electrocardiol 2007;40:246–50.

16. Engin M, Cidam O, Engin EZ. Wavelet trans-formation based watermarking technique for

human electrocardiogram (ECG). J Med Syst 2005;29:589–94.

17. Minhas FU, Arif M. Robust electrocardio-gram (ECG) beat classification using discretewavelet transform. Physiol Meas 2008;29:555–70.

18. Murugappan M, Murugappan S, Zheng BS.Frequency band analysis of electrocardiogram(ECG) signals for human emotional state classifi-cation using discrete wavelet transform (DWT).J Phys Ther Sci 2013;25:753–9.

19. Slawski M, Boulesteix A, Bernau C. CMA: syn-thesis of microarray-based classification. Availabeat: https://www.bioconductor.org/packages/release/bioc/html/CMA.html. Accessed December2017.

20. Breiman L. Random forests. Machine Learning2001;45:5–32.

21. Janitza S, Celik E, Boulesteix A-L.A Computationally Fast Variable ImportanceTest for Random Forest for High DimensionalData. Munich, Germany: University of Munich,2015.

22. Janitza S, Strobl C, Boulesteix AL. An AUC-based permutation variable importance measurefor random forests. BMC Bioinformatics 2013;14:119.

23. Pencina MJ, D’Agostino RB Sr., Demler OV.Novel metrics for evaluating improvement indiscrimination: net reclassification and integrateddiscrimination improvement for normal variablesand nested models. Stat Med 2012;31:101–13.

24. Prenner SB, Shah SJ, Goldberger JJ, Sauer AJ.Repolarization heterogeneity: beyond the QT in-terval. J Am Heart Assoc 2016;5:e003607.

25. Periasamy M, Janssen PM. Molecular basis ofdiastolic dysfunction. Heart Fail Clin 2008;4:13–21.

26. Vyas H, O’Leary PW, Earing MG, Cetta F,Ackerman MJ. Mechanical dysfunction in extremeQT prolongation. J Am Soc Echocardiogr 2008;21:511.e15–7.

Sengupta et al. J A C C V O L . 7 1 , N O . 1 5 , 2 0 1 8

ECG Wavelets for Assessing Myocardial Relaxation A P R I L 1 7 , 2 0 1 8 : 1 6 5 0 – 6 0

1660

27. Nador F, Beria G, De Ferrari GM, Stramba-Badiale M, Locati EH, Lotto A, Schwartz PJ. Un-suspected echocardiographic abnormality in thelong QT syndrome. Diagnostic, prognostic, andpathogenetic implications. Circulation 1991;84:1530–42.

28. Savoye C, Klug D, Denjoy I, Ennezat PV,Le Tourneau T, Guicheney P, Kacet S. TissueDoppler echocardiography in patients with longQT syndrome. Eur J Echocardiogr 2003;4:209–13.

29. Haugaa KH, Edvardsen T, Leren TP, Gran JM,Smiseth OA, Amlie JP. Left ventricular mechan-ical dispersion by tissue Doppler imaging:a novel approach for identifying high-risk

individuals with long QT syndrome. Eur Heart J2009;30:330–7.

30. Antzelevitch C, Sicouri S, Di Diego JM, et al.Does Tpeak-Tend provide an index of transmuraldispersion of repolarization? Heart Rhythm 2007;4:1114–6.

31. Sauer A, Wilcox JE, Andrei AC, Passman R,Goldberger JJ, Shah SJ. Diastolic electromechan-ical coupling: association of the ECG T-peak to T-end interval with echocardiographic markers ofdiastolic dysfunction. Circ Arrhythm Electrophysiol2012;5:537–43.

32. Oskouie SK, Prenner SB, Shah SJ, Sauer AJ.Differences in repolarization heterogeneitymmong heart failure with preserved ejection

fraction phenotypic Subgroups. Am J Cardiol 2017;120:601–6.

33. Robin X, Turck N, Hainard A, et al. pROC: anopen-source package for R and Sþ to analyze andcompare ROC curves. BMC Bioinformatics 2011;12:77.

KEY WORDS continuous wavelettransform, diastolic dysfunction, signal-processed ECG, tissue Doppler imaging

APPENDIX For an online supplementary datafile and supplemental tables, please see theonline version of this paper.

MV-JNRPT-001(A)