Separation of Ventricular Tachycardia FromSinus Rhythm Using a Practical, Real-TimeTemplate Matching Computer System

SAUL E. GREENHUT, THOMAS F. DEERING,* BRUCE M. STEINHAUS,JO L. INGRAM,* STEVEN R. CAMP/ and LORENZO A. DiCARLO**

From the Applied Research Division, Telectronics Pacing Systems, Englewood, Colorado, the*Cardiac Electrophysiology Laboratory, Georgia Baptist Medical Center and Piedmont Hospital,Atlanta, Ceorgia, and the * •Cardiac Electrophysiology Laboratory, St. Joseph Mercy Hospital,Ann Arbor, Michigan

GREENHUT, S.E., ET AL.: Separation of Ventricular Tachycardia From Sinus Rhythm Using a Practical,Real-Time Template Matching Computer System. Template matching morphology analysis of the intra-ventricular eiectrogram (rVEGj has been proposed for incJusion in impJanfabie cardioverter de/ibrilJatorsfICDs) to reduce the number of false ventricular tachyarrhythmia detections caused by rate overlapbetween ventricular tachycardia (VT) and sinus tachycardia and/or supraventricular tachycardia. Tem-plate matching techniques have been developed that reduce the computationaJ complexity while prese.rv-ing the perceived important aspects of electrogram ampJitode and baseline independence found in suchcomputationally unsolved methods as correlation waveform analysis (CWA). These methods have beenshown to work as well as CWA for separation of VT, however, they have not been proven in reaJ-time ona sysfem that incorporates many of the constraints of present day ICDs. The present study was undertakenwith two purposes: (1) to determine if reaJ-time IVEG template matching analysis on an ICD sensingemulator was accurate in separating VTfrom sinus rhythm (SR) eiectrograms; and (2) to compare ampli-tude normalized area of difference (NAD) with signature analysis fSIG), a new, computationally lessexpensive technique that normalizes for amplitude variation within the expected physiological levei ofvariability. In this study, IVEGs, obtained from 16 patients who underwent eJectrophysiologicaJ study(EPS) for evaluation of sustained ventricular arrhythmia, were digitized to 250 Hz with 6'bit quantizationafter filtering (16-44 Hz) and differentiation. After an SR template was selected and periodically updated,it was compared to subsequent IVEGs using NAD and SIG. In general, SIG calculates the fraction ofsamples occurring outside template window boundaries. Eleven-heat running medians from beat-by-beatNAD and SIG results were determined. The maximum median during VT was subtracted from the mini-mum median during Sfl with the result equal to the separation margin. With the minimum separationthreshold set to 0 (i.e., no overlap], 0.1, and 0.2, NAD separated 16/16, 14/16, and 9/16 VTs, while SIGseparated 15/16, 14/16, and 13/16 VTs, respectively. While NAD separated more VT episodes on the strictbasis of no overlap, SIG separated more than NAD as the safety margin was further increased. Conclu-sions: (1] template matching morphology techniques can potentially be implemented in ICDs; (2) usinga patient specific threshold, NAD and SIG appear capable of separating VT from Sfl in most patients;and (3) SIG and NAD appear to be similar in accuracy. Thus, SIG may be preferable since it significantlyreduces the computational load. (PACE, Vol. 15, November, Part II 1992)

arrhythmia detection, implantable cardioverter defibriliator, ventricular tachycardia, signaJ processing,computer analysis

Introduction

Ajj c • . c i T 7 r ^ L . n ^ . r ^ n L J Appropriate implantable cardioverter deii-Address for reprints: Saul E. Greenhut, Ph.D., Research and , .,, . nr^^.^ ., . , j j.Development, Telectronics Pacing Systems, 7400 S. Tucson brillator (ICD) therapy IS dependent upon accurateWay, EnglewDod, CO 80112. Fax: (303) 643-2213. detection of ventricular tachyarrhythmias by the

2146 November, Part II 1992 PACE, Vol. 15

SEPARATION OF VT FROM SR

ICD. Presently, rate based detection algorithms donot accurately differentiate pathological from non-pathological tachycardias well in all patients andpresent morphology algorithms, such as the proba-bility density function,^ have not proven effective^and are therefore infrequently used. Improved dif-ferentiation algorithms are necessary in order toprevent erroneous classification, which may resultin failure to deliver treatment during ventriculartachycardia (VT), or spurious shocks during a su-praventricuiar mechanism with possible physicaldiscomfort, battery drainage, or proarrhythmia.^'*

In order to increase the specificity of VT de-tection, many methods have been proposed as anadjunct to rate based detection,^ including time-domain morphological analysis of the intraven-tricular electrogram (IVEG).^'^"" Previous studieshave shown the success of template matching mor-phology methods in differentiating VT from sinusrhythm (SR).^-^" '̂̂ ' Correlation waveform analy-sis (CWA) has been considered the standard tem-plate matching method applied to the analysis ofelectrocardiological signals.^'^ Because of thecomputational limitations of state-of-the-art ICDs,more computationally efficient techniques, mostof which are amplitude and baseline normaliza-tion variations of the difference in area betweenelectrograms, have been developed and tested,^'^^The constraints in implementing these algorithmsinto present day ICDs have not been considered.

The purpose of the present study was to evalu-ate the effectiveness of template matching tech-niques in a real-time environment using process-ing limitations that simulate important aspects ofcurrent level ICD technology. A secondary pur-pose of the study was to determine the effective-ness of a new template matching technique, signa-ture analyis (SIG), which significantly reduces thenumber of beat-to-beat computations required ofother template matching algorithms that also nor-malize for amplitude variations. Since IVEC am-plitude has been shown to vary from baseline dur-ing exercise,^^ Valsalva maneuvers,^^ ventricularvolume,^^ and ischemia,^^ a template matchingmorphology measure that is independent of ampli-tude has been perceived as potentially more desir-able for increased specificity. All template match-ing measures that are independent of electrogramamplitude and have been shown to work well indifferentiating VT from SR, incorporate at least

several multiplications beat-to-beat. SIG, whichnormalizes for amplitude within the expectedphysiological range and requires no costly multi-plications heat-to-beat, was developed and testedas part of the current study. The results from SIGwere compared to NAD, a previously describedtechnique, which was analyzed simultaneously.

Methods

Electrophysiology Study and Patient Population

Sixteen patients (12 male, 4 female) with in-duced sustained monomorphic VT during a stan-dard electrophysiological study (EPS), defined asVT lasting more than 30 seconds or terminated be-cause of hemodynamic compromise, were stud-ied. The patient population ranged in age from 43to 86 years (mean = 68 years). Thirteen patientshad coronary artery disease and one had hypertro-phic cardiomyopathy. Two patients had right bun-dle branch block as their underlying SR and fourhad a nonspecific intraventricular conductiondelay. Seven of the 16 patients were on an antiar-rhythmic drug at tbe time of EPS.

Ventricular electrograms were analyzed inreal-time from the distal bipolar pair (separation= 1 cm) of a 6 French quadripolar catheter locatedat the right ventricular apex during EPS for theinduction and termination of VT. An atrial elec-trode catheter and surface electrocardiogram veri-fied the presence of VT.

External ICD Sensing Emulator

The ICD sensing emulator (ISE) consists of apace/sense amplifier that filters, amplifies, anddigitizes IVEGs. The digitized signal is processedby a Z-80 coprocessor board within an IBM PC(IBM Corp., Armonk, NY, USA). The analysis soft-ware is downloaded to coprocessor memory priorto processing. The IBM PC itself is used only as akeyboard, graphical, and data storage interface.

Electrograms are filtered (hardware 16-44 Hz,1-pole high pass, 2-pole low pass), differentiated,then digitized at 500 Hz. The resulting first differ-ences are 6 bits in length giving a dynamic rangeof ± 31 quantization levels. Consecutive samplesare scanned prior to analysis, resulting in an effec-tive sampling rate of 250 Hz and ± 62 possiblediscrete amplitude levels.

PACE. Vol. 15 November, Part II 1992 2147

GREENHUT, ET AL.

Template Matching Algorithms

The template matching algorithmic approachfor the detection of VT from IVEGs uses the follow-ing general procedure:

1. Select and store an SRIVEG depolarizationtemplate.

2. As IVEGs are sensed, align the templatewith the sensed IVEG.

3. Galculate a measure of similarity betweenthe template and sensed IVEG that determines therhythm classification of the sensed depolarization,

4. Update the template periodically to adjustto moderate, time-dependent morphologicalchanges.

Procedures used for selecting, aligning, and updat-ing templates for this study are described in detailelsewhere.^^ Template acquisition is done semiau-tomatically under operator control, templates are80 msec in duration, and alignment is by a bestfit method. The two template matching algorithmscalculated were normalized area of difference(NAD) and SIG.

NAD

NAD is a template matching approach previ-ously described,^^ which measures the differencein area between two waveforms when normalizingfor the difference in amplitude between the wave-forms. Baseline normalization over an IVEG depo-larization was accomplished by high pass filteringand differentiation. Although NAD is less compu-tationally demanding than CWA, NAD performedas well or better than GWA in previousNAD has been described as:

NAD = 1 -N

k = -Sk

(1)

where ti are the template points, Si are the detectedIVEG (signal) points, and N is the number of pointsin the template. A perfect match between templateand signal results in an NAD of 1. In the proposedimplementation, once an SR template and VTthreshold have been selected, NAD requires N -I-1 multiplications per IVEG, which is half as manyas GWA, but is still 21 multiplications per depolar-ization for an 80-msec template at 250 Hz.

SIG

SIG is a technique developed by the authorsin an attempt to increase the sensitivity of VT de-tection, maintain amplitude independence withinthe physiological range, and eliminate powercostly multiplications from beat-to-beat analysis.SIG was so named because it creates a boundarywindow enclosing all template points that form a"signature" of the waveform to be compared toincoming IVEGs. The number of points occurringoutside the signature template boundaries arecounted. The greater the fraction of points occur-ring outside the boundaries, the more the mis-match between template and signal. The equationfor SIG is;

SIG ='cnt

(2)

where Tent is the number of template points whoseabsolute value is greater than a minimum am-plitude.

In the present study, this minimum amplitudewas selected as 5 quantization levels. Tent elim-inates points close to the baseline that contain lit-tle discriminatory information, is set after partialtemplate acquisition, and is adjusted at each tem-plate updating. Sout is the number of signal pointsthat occur outside template boundaries. SIG == 1when all signal points are within template bou::id-aries. SIG values < 0 are possible since Soui canbe a maximum of the numher of template points,while Tent is usually less than the total number oftemplate points.

After the VT threshold and template are se-lected, Equation 2 may be rewritten as:

> (3)

where 7 = (1 - Q]T^nu 9 is the SIG VT detectionthreshold, and -y is calculated only after templateacquisition and updating. If Sout is greater than 7the rhythm would be called VT, otherwise SR.Only a compare and increment, if required, arenecessary at each template sample to evaluateEquation 3. Additional computations (adds and bitshifts) are required as part of the SIG normaliza-tion procedure described in Equation 5.

Template signature boundaries are crealedafter template acquisition and are adjusted after

2148 November, Part II 1992 PACE, Vol. 15

SEPARATION OF VT FROM SR

each template update. The window is set at ± 8ifrom each sample point where

5, = max minimum-window, :̂ i = 1. , . N

(4)and ti is each template point and mini'mum_win-dow equals the minimum amplitude for Tent givenahove (i.e., 5). The value for minimum.ivindowand

can be varied and have the effect of increasing ordecreasing the sensitivity and specificity of VT de-tection. The template signature boundaries (Equa-tion 4) are calculated only after template acquisi-tion and updating.

SIG includes a separate amplitude normaliza-tion procedure. The amplitude normalizationcompares the sum of the signal points to the sumof the template points:

If < 0.75 X then Si*—2Si,i = 1

N N

If 2 | s i | > 1 . 5 X 2 | t , | then

Otherwise s^*—s,.

(5)

where Si are the signal samples and are ti the cur-rent template points. Although equation 5 in-cludes multiplications and divisions, the valuesfor these operations were chosen so that the opera-tions could he accomplished using a minimumnumber of bit shifts and adds on a microprocessor.For example, the multiply by 0.75 requires a total

of 3 right shifts and 1 add of 2 ft l I ti I and themultiply by 2 is 1 left shift of Si. In addition, opera-tions involving tj are performed only after tem-plate acquisition and updating. The amplitudenormalization procedure (Equation 5) is calcu-lated prior to Equation 2 or 3.

The template window boundaries allow formodest changes in amplitude without beat-by-beatnormalization. The amplitude adjustment of Equa-tion 5 only occurs when incoming IVEGs are <0.75 or > 1.5 of the template amplitude. The proce-dure ideally maintains the template points withinthe SIG window for amplitude changes rangingfrom 0.25 to 3 times the template size, which isthought to fully account for electrophysiologicalvariability in the waveform.

Statistics

Eleven-heat running medians of NAD and SIGwere calculated over passages of SR and VT. The

PATIENT

U OD

n ji I) II 14 13 I*

Figure 1. Results for NAD and SIG 11 -beat runningmedians/or separation of VT from SR. Patient num-bers are listed along the horizontai axis, on the leftfor NAD and the right for SIG results. Ranges of the11-beat running Sfi medians are indicated by soiidblack bars with the mean indicated hy an unfilledsquare. Below and slightly to the right of the SRranges for each patient are the 11-beat VT medianranges funfilJed bars) and means (bJack squares).

NAD SIO

PACE. Vol. 15 November, Part II 1992 2149

GREENHUT, ET AL.

NAD

Ienplate 2713^NAD 0,745SIG 0,846

Rate 112 UEGM Interv 529NSNUN21120

Figure 2. Patient 2 ICD sensing emulator screen during sinus rhythm (SR). The ventricular rateis 112 beats/min. The lower right window is the ventricular eJectrogram (before differentiation)with sensing markers indicated (S). The window at the lower left displays the SR templatewaveform fderivative) after each update. The small window to the right displays each sensedeJectrogram (derivative)/or more carefuJ visual observation. The two windows above show thetime course of NAD and SIG over the most recant 234 sensed depolarizations as a wrap-aroundbuffer. The most recent NAD and SIG values, 0.745 and 0.846, respectively, are displayed at thebottom of the screen. The number of depolarizations analyzed (211) and the number of templateupdates (20) are indicated at the lower right. Occasional decreases in NAD and SIG values arebecause of changes in IVEG morphology resulting from premature ventricular depolarizations.The minimum 11-beat running median was 0.72 for NAD and 0.77 for SIG during SR forpatient 2.

2150 November, Part II 1992 PACE, Vol. 15

SEPARATION OF VT FROM SR

first three beats following the pacing inducedonset of VT were not included in the analysisbecause of electrode polarization artifact. Them i n i m u m m e d i a n d u r i n g s i n u s r h y t h m(min_mediansR) and the maximum median dur-ing VT {max_medianvT) for each patient werecompared for NAD and SIG. NAD and SIG SRA^Tseparations (a) were defined as:

= min_mediansR — max_medianVT- (6)

Results

The NAD and SIG median data during SR andVT for all patients is summarized in Figure 1. The

mean, minimum, and maximum median for SRand VT are shown for each patient. The pairedaverage separation of SIG medians {0.37 ± 0.21[mean ± standard deviation]) was not signifi-cantly greater than for NAD medians (0.32 ± 0.21).The mean SR duration analyzed for all patientswas 155 seconds while the mean VT duration was20 seconds. An example of the ISE screen duringSR {Fig. 2) and VT {Fig. 3) for patient 2 is included.Premature ventricular depolarizations during SRdecreased the SR minimum median, therefore re-ducing the separation margin.

The number of patients separated at increas-ing minimum separation thresholds is shown inFigure 4. Zero minimum separation represents nooverlap of VT and SR. At a zero minimum separa-

s s s s s s s

lenplate 5713NAD 0,407SIG 0.538

Bate 178 UEGH Intepv 333MSNUN42529

Figure 3. Corresponding to Figure 2 (patient 2), ICD sensing emulator screen during ventriculartachycardia. The ventricular rate is 178 beats/min (see Fig. 2 for description]. The sudden de-crease in NAD and SIG values shown in the top graphs occurs when VT is induced. Note thedifference in morphology between the SR template and VT derivative waveforms as shown. Thethird depolarization within the continuous IVEG window is probably a fusion beat that causesNAD and SIG values to increase briefly as indicated near the end of the NAD and SIG windows.The maximum 11-beat running median was 0.42 for NAD and 0.46 for SIG during VT, giving0.30 and 0.31 SR/VT separation margins for NAD and SIG, respectively.

PACE, Vol. 15 November. Part II 1992 2151

GREENHUT, ET AL.

16-

14-

10-

8-

6-1

4-

24

>•....->

1

* ^

1- ••^

\

i--"̂

t—1

\

NS

< »\

1 1

:

i \

: 1 !

t

i

K *

r- <

1 1 1 1 1 1 1 1 1 1 I 1 1 M H (-I0.1 0.2 0.3 0.4 0.5 0.6

Minimum Separation Threshold0.7 0.8

NAD SIG

Figure 4. Shows the number of pa-tients whose maximum m e d i a nduring VT and minimum medianduring SR are separated at increas-ing minimum separation thresh-olds. As the minimum acceptableseparation threshold increases, thenumber of patients whose SR andVT can be separated using A7AD orSIG decreases.

tion threshold NAD separated all 16 patients andSIG separated 15 of 16 (94%). However, as the min-imum separation threshold was increased, theability to discriminate SR from VT is reduced, al-though more slowly for SIG.

Discussion

Preliminary data obtained during this studysuggests that template matching morphologymethods for the separation of VT from SR can beeffective when utilizing a patient specific thresh-old within the processing constraints inherent instate-of-the-art ICDs, including reduced samplingrate, reduced amplitude resolution, reduced band-width, and real-time operation on a moderatelysimple microprocessor. Since SIG appears prelim-inary to be as effective as NAD in differentiatingVT from SR, SIG might be preferable to NAD be-cause of the significantly reduced processing re-quirements of the former.

A number of limitations are noted with thispreliminary study. The correlation between sig-nals obtained with a temporary catheter comparedto those of a permanent catheter are unknown; thestability of the signals over time has not yet beentested; and the band-pass filter used in this study

was not previously independently tested for itsdiscriminatory capability. Other publications sug-gest that the discriminatory capability of templatematching morphology measures can be improvedwith use of high pass filters, e.g., a 10-Hz, 2-polehigh pass filter, or degraded with other high passfilters, e.g., 20-Hz, 2-pole filter.^^ In addition,physiological and/or clinical parameters mayfunction to significantly alter the baseline tem-plate or may cause electrograms to be falsely de-tected as VT. Such factors might include ambientsympathetic tone and parasympathetic t one , aswell as other neural factors; heart rate; ventricularloading characteristics; postural changes; is-chemia; and antiarrhythmic agents.'^^^^'^^ T h e 11-beat running median was also not compared toother methods of analysis that may have resultedin a different discriminatory power.

Another limitation of the study is the analysisof only one VT configuration in each patient. Itis not known whether other VT morphologies inindividual patients, if present, would have sepa-rated from SR as well. Also, it has previously beenshown that occurrence of paroxysmal bund lebranch block of supraventricular origin is a proba-ble source of arrhythmia misclassification usingtime-domain morphology methods.'^

2152 November, Part II 1992 PACE, Vol . 15

SEPARATION OF VT FROM SR

Despite a number of potentially apparent limi-tations, preliminary data would suggest that tem-plate matching morphology methods might havesignificant impact in allowing for the separationof VT from SR. The use of such methods, shouldthen be demonstrated as efficacious in future

generation ICDs, will help to prevent inappro-priate shocks without decreasing the ability todeliver effective therapy in a timely fashion. Ad-ditional clinical research is required to confirmthe validity of these techniques in patient popula-tions.

References

10.

Langer A, Heilman MS, Mower MM, et al. Consid-erations in the development of the automatic im-plantable defibrillator. Medical Instrument 1976;10:163-167.Lin D, DiCarlo LA, fenkins JM. Identification ofventricular tachycardia using intracavitary ven-tricular electrograms: Analysis of time and fre-quency domain patterns. PACE 1988; 11:1592-1606.Cohen TJ, Chien WW, Lurie KC, et al. Implantablecardioverter defibrillator proarrhythmia: Case re-port and review of the literature. PACE 1991; 14:1326-1329.Johnson NJ, Marchlinski FE. Arrhythmias inducedby device antitachycardia therapy due to diagnos-tic nonspecificity. J Am Coll Cardiol 1991; 18:1418-1425.Lang DJ, Bach SM. Algorithms for fibrillation andtachyarrhythmia detection. J Electrocardiol;23(Suppl.):46-50.Davies DW, Wainwright RJ, Tooley MA, et al. De-tection of pathological tachycardia by analysis ofelectrogram morphology. PACE 1986; 9:200-208.Langberg JL, Cibb WJ, Auslander DM, et al. Identi-fication of ventricular tachycardia with use of themorphology of the endocardial electrogram. Circu-lation 1988; 77:1363-1369.Steinhaus BM, Wells RT, Greenhut SE. et al. Detec-tion of ventricular tachycardia using scanning cor-relation analysis. PACE 1990; 13:1930-1936.Throne RD, Jenkins JM, DiCarlo LA. The bin areamethod: A computationally efficient technique foranalysis of ventricular and atrial intracardiac elec-trograms. PACE 1990; 13:1286-1297.Paul VE, Farrell T, Gill J, et aL Automatic recogni-tion of ventricular arrhythmias using temporalelectrogram analysis. PACE 1991; 14:1265-1273.

11. Throne RD, Jenkins JM, DiGarlo LA. A comparisonof four new time-domain techniques for discrimi-nating monomorphic ventricular tachycardia fromsinus rhythm using ventricular waveform mor-phology. IEEE Trans Biomed Eng 1991; 38:561-570.

12. Bricker JT, Ward KA, Zinner A, et al. Decrease incanine endocardial and epicardial electrogramvoltages with exercise: Implications for pacemakersensing. PACE 1988; 11:460-464.

13. Rosenqvist M, Lagergren H, Strandberg H, et aLValsalva-induced variations in the intracardiacsignaL PACE 1985; 8:856-861.

14. Lekven J, Chatterjee K, Tyberg JV, et al. Pro-nounced dependence of ventricular endocardialQRS potentials on ventricular volume. Br Heart J1978; 40:891-901.

15. Bonoris PE, Greenberg PS, Christison GW, et al.Evaluation of R wave amplitude changes versusST-segment depression in stress testing. Circula-tion 1978; 57:904.

16. Greenhut SE, Steinhaus BM. Template matchingtechniques for electrophysiologic signals. BiomedSci Instrument 1992; 28:37-42.

17. Jenkins JM, DiCarlo LA, Chiang G-MJ. Impact offiltering upon ventricular tachycardia identifica-tion by correlation waveform analysis. PAGE 1991;14:1809-1813.

18. Einelli CJ, DiCarlo LA, Jenkins JM, et al. Effectsof increased heart rate and sympathetic tone onintraventricular electrogram morphology. Am ]Cardiol 1991; 68:1321-1328.

19. Throne RD, DiGarlo LA, Jenkins JM, et al. Paroxys-mal bundle branch block of supraventricular ori-gin: A possible source of misdiagnosis in detectingventricular tachycardia using time domain analy-ses of intraventricular electrograms. PAGE 1990;13:453-468.

PAGE, Vol. 15 November, Part II 1992 2153