Realtime Multichannel System for Beat to Beat QT Interval Variability
Vito Starc, MD, PhD* and Todd T. Schlegel, MD**
* Institute of Physiology, School of Medicine, University of Ljubljana, Ljubljana, Slovenia
** NASA Johnson Space Center, Human Adaptation and Countermeasures Office, Houston,
Short title: Realtime Multichannel QTV
Vito Starc, MD
Ljubljana University School of Medicine
Institute of Physiology
Tel: +386 1 543 7500
Fax: +386 1 543 7501
Grant No.: PO510381
Ministry of Education, Science and Sport
The measurement of beattobeat QT interval variability (QTV) shows clinical promise for
identifying several types of cardiac pathology. However, until now, there has been no device
capable of displaying, in real time on a beattobeat basis, changes in QTV in all 12
conventional leads in a continuously monitored patient. While several software programs
have been designed to analyze QTV, heretofore, such programs have all involved only a few
channels (at most) and/or have required laborious user interaction or offline calculations and
postprocessing, limiting their clinical utility. This paper describes a PCbased ECG software
program that in real time, acquires, analyzes and displays QTV and also PQ interval
variability (PQV) in each of the eight independent channels that constitute the 12lead
conventional ECG. The system also processes certain related signals that are derived from
singular value decomposition and that help to reduce the overall effects of noise on the real
time QTV and PQV results.
The measurement of beattobeat QT interval variability (QTV) shows clinical
promise for identifying several types of cardiac pathology, including coronary artery
disease(13), acute myocardial ischemia(4) and infarction(5), left ventricular hypertrophy(6)
and various types of cardiomyopathy(79). Importantly, in patients referred for
electrophysiologic studies and in animals who receive drugs that prolong the QT interval,
increased repolarization lability as reflected by increased beattobeat QTV is strongly
predictive of future arrhythmic events(10, 11). As such, several researchers have recently
developed software programs to quantify QTV. There are, however, a number of limitations
associated with these existing software programs. First, most if not all require that the
operator actively perform certain functions, for example that he or she manually choose the
onset and offset of an initial QT interval template. This step of human intervention is both
time and labor consuming, and affects reproducibility(12). Second, the majority of these
existing programs also allow for QTV analyses in only one single channel at a time (i.e.,
usually in limb lead II or limb lead I), and not in multiple channels simultaneously. This can
lead to spuriously high QTV values, particularly when the T wave in the channel being
studied is small, noisy or otherwise poorly defined. Third, to our knowledge, no existing
programs for QTV analyses perform in real time, only offline, with the exception of one real
time program that is focused on dynamic QT vs. RR plots(13) and which in any case has the
other limitations noted above. A lack of realtime capability necessarily entails that any acute
or subacute change in QTV, as might occur for example during myocardial ischemia(4) (or
the treatment thereof), cannot be readily observed or acted upon. All of these same limitations
also apply to existing software designed to analyze PQ interval variability (PQV)(1416).
We describe herein a PCbased software program that allows for the realtime
monitoring of multiple indices of QTV and PQV, simultaneously in each of the 8
independent channels that constitute the standard 12lead ECG. The program also contains
utilities for automatic formation of templates and detection of intervals such that
dependencies on the operator are completely eliminated. In addition, several selfadjusting
procedures control the software’s associated signal averaging techniques and thus provide
quality control for the consistent determination of the QT and other intervals. These iterative
averaging techniques ensure more precise determinations of QTV and PQV by preventing the
analysis of noise and by eliminating important artifacts such as jitter in the detected QRS
wave fiducial points. As a result, several indices of variability are provided instantaneously
and on a beattobeat basis, allowing for the online monitoring of multiple cardiac electric
processes. The program’s analysis of ECG waveforms consists of first identifying individual
beats and then particular waves, followed by the formation of template beats, construction of
time series for the RR, QT and PQ intervals, and finally the statistical description of the
variability of the last “n” recorded beats for each interval type. The performance of the
software – especially with respect to the stability and precision of key measured indices – is
also briefly illustrated herein through a study that involved 12lead ECG recordings during
both normal breathing and during deep, slow breathing in 19 healthy individuals.
The realtime analysis of QTV and PQV is performed on a Windows 2000 or XP
based PC using a “client” software program (“PQT”, developed in Slovenia) that
communicates with a “server” software program (CardioSoft, Houston, TX). The
communication between the client and server software is established via a Windowsbased
named pipe. A commercial 12lead PCbased ECG device (e.g., CardioSoft, Houston, TX or
Cardiax, IMED, Budapest, Hungary) provides eight channels of incoming raw ECG data via
the named pipe to the client at a sampling rate of 1 ms.
Two processes run in parallel within the PQT client: the acquisition of data and ECG
wave analysis. Both processes run in a timeshared mode controlled by two separate threads
on the Windows 2000 or XPbased PC. The first thread reads the named pipe to which raw
data are provided by the server program, which results in the writing of eight channels of
ECG data into a circular buffer that can contain up to the last “M” minutes of 1 mssampled
data. Besides the raw data, the commercial 12lead PCbased ECG devices also provide the
positions of the instantaneous QRSwave fiducial points. The PQT client program uses this
information as well as information from its own QRSwave detector to perform assessments
of the amount of QRSwave jitter.
Beat selection for averaging
When analysis begins, templates for the overall ECG signal in each channel are first
formed. These initial global templates are formed somewhat differently from later templates.
To construct the initial global templates, the first 20 beats are collected, the probability
function(17) of the RR interval is created for those beats, and its maximum determined.
Finally, those ten beats that are closest to the determined maximum are selected, and a
template ECG wave for the same beats is then obtained by superimposing the ECG signal
based on the fiducial point of the QRS wave. When later ECG templates are constructed,
separate templates for the QRS complex and for the T and P waves are formed as described
Automated formation of separate templates for the QRS, T and P waves
After the initial templates have been constructed, breaking points called PQbreak
(between the P wave and the QRS complex), QTbreak (between the QRS complex and the T
wave), Tend (the final point of the T wave) and Pini (the initial point of the P wave) are used to
construct templates for the three principal waveforms QRS, T and P. Thus, the QRS complex
will ultimately be confined to the time interval between PQbreak and QTbreak, the T wave to the
time interval between QTbreak and Tend, and the P wave to the time interval between Pini and
PQbreak. The algorithms that define the breaking points and thus the templates themselves are
a. QRS template
The characteristics of the template QRS complex are determined by searching for
local minima and maxima as well as by the minimal and maximal time derivative values
(slopes) between them (Figure 1). The initial slope is defined as the first deflection before the
fiducial point that is larger than 1/10 of the maximal slope of the given QRS complex. The
crossing of the initial slope with the preceding nearly horizontal segment determines the
initial point of the QRS complex, Qini, for each measured ECG lead. Next, the program
selects the first segment of the QRS complex that is larger than 15% of the total QRS signal
amplitude, and determines the first and the last point of that segment, QRS1 and QRS2,
respectively (Figure 1). This segment is used to determine the ∆QRS, the deviation of the
initial part of the QRS complex of the incoming beat from the same region in the template.
This is necessary to correct any QRS shift (jitter) or an improperly detected fiducial point for
the incoming signal.
The time instant of the Qini points obtained from all eight measured signals are then
treated statistically to find the mean and the standard deviation (SD), and a point called Qini0
is defined as the earliest of the eight Qini that occurred within one SD from the mean. The
final point of the QRS complex, QRSend, is determined similarly to Qini, and then the point
QRSend0 is determined similarly to Qini0. It is Qini0 and QRSend0 that in turn determine the
breaking points (PQbreak and QTbreak)between waves. Specifically, in the present
implementation of the program, PQbreak is considered to occur 20 ms before Qini0, and QTbreak
30 msec after QRSend0.
When PQ and QT variability are determined, the borders for the QRS template (Qini0
and QRSend0) as well as the QRS breaking points (PQbreak and QTbreak) are treated as the same
for all leads. However, new borders and breaking points are computed every time the
program (through automatic preset) or the user (through manual interaction) instigates a new
template in order to optimally refresh the monitoring going forward.
b. T wave template
The onset and offset of the Twave template are constructed in the following manner.
First, a biparabolic curve is constructed consisting of two interconnected parabolic segments,
both with an amplitude one half of the total biparabolic curve height, the first part nearer to
the QRS complex and the second (distal) part with its apex convex toward the zero line
(Figure 2). The first and the last point of the biparabolic curve are designated as T1 and T2,
respectively. Both parabolic segments are further divided into three parts, with the apex
occurring in the outer third of each parabola. The apex of the distal parabola is later used to
determine the final point of the T wave, Tend. In addition, in order to more closely replicate
the T wave, the horizontal line between points Tend and T2 replaces the most distal aspect
arising from Tend. The biparabolic curve is shifted toward the QRS complex, while varying its
height and width until finding the best fit to the initial template T wave. Because for
waveforms in both the positive and negative directions, several matches might have occurred
while approaching from the distal part towards the QRS complex, the one with the largest
overall amplitude is selected as the most representative one. The point Tend is then determined
by the position of the distal apex point. The interval QT0 = Tend Qini0 is then the absolute
duration of the QT interval in that particular ECG lead. In addition to determining Tend, the
program uses both of the extremities of the representative biparabolic curve, T1 and T2, to
determine the interval for matching the incoming Tend to the Tend of the existing template.
However, unlike the case for the QRS template borders and breaking points, Tend is
subsequently treated as unique for each lead.
c. P wave template
A procedure similar to that used to form the Twave template is also used to
determine the appropriate borders for the P wave template in all ECG leads. In addition, once
determined, Pini, like Tend, is subsequently treated as unique for each lead.
QT interval variability algorithm
In the PQT program, the P, QRS and T templates are constructed by signal averaging
rather than by using a single beat from the measured ECG signal. In addition, matching of
any new incoming complex to the templates in any region of interest is thereafter achieved by
shifting of the ECG signal(1, 2, 16), not by “stretching”(8). It should be understood that the
PQT program is principally concerned with quantifying the variability of intervals rather than
their absolute values. Thus, if for any reason the algorithm chooses a false beginning or end
point for a given template interval, then all interval values will be biased proportionately, but
the beattobeat variability in the computed intervals will be unaffected. The specific steps
performed in analyzing QT interval variability are described in detail below.
1) The template beat φ(n), where n is the sample number, is constructed from the
selected beats using a signal averaging technique. Only those beats with shape similar to the
template are selected for averaging. Since the program automatically determines the borders
of each wave component (P, QRS, and T templates, as described above) and therefore the
time window for matching of waves, its remaining task is to shift the particular incoming
wave component with respect to the template until obtaining an acceptable match. The
matching algorithm is based on the least square deviation of the incoming wave versus the
2) The matching of waves is performed in two substeps. First, a broader time interval
containing the complete wave component is used to reach the best fit. Specifically, the points
QTbreak and T2 are used for the T wave (Figure 3a), PQbreak and QTbreak for the QRS complex,
and P1 and PQbreak for the P wave. In this substep, the amplitude of the incoming wave
component is normalized with respect to the template, i.e., it is adjusted to reach the same
surface area under the curve in the same time interval as that of the template (Figure 3a). This
substep is important for providing a parameter called the “norm”, an index of the quality of
matching described in greater detail below. In the second substep, shifting of the normalized
wave in time is performed to achieve the best fit in a smaller time window. Specifically, the
points T1 and T2 define the best fit for the end of the T wave (Figure 3b), the points QRS1 and
QRS2 the best fit for the beginning the QRS complex, and the points P1 and P2 the best fit for
the beginning of the P wave.
3) Each wave of any incoming beat x(n) is shifted to or from the trigger point to
achieve the best alignment with the template in the appropriate time window. For this
purpose, an error function of time shifting εi(ak) is defined as the sum of the squared
differences between the template wave (P, QRS or T) and the appropriate shifted version of
the incoming wave of beat i:
aTj x Taj ε=+−+ +
where ak is the timeshifting parameter of ECG lead k, Tk is the triggering point of the
template, Ti is the triggering point of the beat i, and n1 and n2 are the beginning and the end of
the matching windows, e.g., QTbreak and T2 in the first matching, and T1 and T2 in the second
matching, respectively, in the case of the T interval. The parameter ak is changed for that beat
to find that particular ak, called âk, which occurs at the minimal εi(ak) until the step of ak is <
0.1 ms. To achieve shifting of the P, QRS and T waves by a noninteger value of a,
interpolation of the waves between the measured sampled values is necessary.
4) The value of εi(ak) after the first matching represents the quality of matching, or the
norm. The norm is mathematically a dimensionless quantity that represents the average
deviation of the incoming wave from the template, divided by the amplitude (A) of the wave
in that interval; i.e., norm = √ εi(ak)/( n2 n1+1)/A. The lower the norm, the better the fit, and
when the incoming wave and the template are perfectly matched, the norm = 0.
During construction of the template, the norm in each particular ECG lead is treated
statistically to obtain the mean value and SD in that lead. The match between the incoming
wave and the template is considered acceptable if the norm of any beat is within one SD of
the mean norm in the corresponding ECG lead. This procedure helps in providing an
acceptable number of beats for averaging even in case of noisy signals. For the QRS and T
waves, an acceptable norm is usually less than 0.15.
5) Instead of using a single bin for the signal averaging of any particular ECG
waveform, the PQT program uses seven bins that cover the expected variability interval of
each waveform type. Hence, for QT variability, the first bin is used for averaging of the
shortest QT intervals, and the successive bins for averaging of successively longer QT
intervals. The variability interval is defined from the variability record during previous
construction of the template, and is initially set to 15 ms. Thus, for example, if the T wave
variability interval in lead II was 14 ms during template construction, then each bin for
averaging will be 2 ms wide. Before adding the wave into the corresponding bin for signal
averaging, the wave is appropriately shifted to adjust for jitter in that particular bin. This
creation of several bins for signal averaging helps to prevent an overreliance on initial
values. At least seven templates are therefore constructed for each waveform in each ECG
lead, and the bin with the highest number of accepted beats is the one ultimately selected to
provide the template. The norm is therefore not the only criterion controlling the averaging
procedure, and the final percentage of beats that ultimately constitute a template is typically
between 35 and 50%. Despite the potential for a reduction in the number of beats, the quality
of the template is nevertheless improved due to selection of the most similar beats.
6) It is expected that QT intervals differ among different leads. Hence, the QTk,i
interval of ith incoming beat and of kth ECG lead differs from the template QTk,0 by an
, ,0, k iki k i
QTQT QRST = + ∆+ ∆
where ∆Tk,i is the interval obtained by shifting the T1 – T2 segment of the ith T wave to
match the template, and ∆QRSi by shifting the QRS1 – QRS2 segment of the QRS complex,
In detail, the sum of time displacement of the T (∆Tk,i) and QRS waves (∆QRSi) forms
the QT interval change for the ith beat and the kth ECG lead (ΔQTk,i), and is calculated
according to the equation:
ΔQTk,i = ∆Tk,i + ∆QRSi
ΔTk,i = âk,i(T)∆t, and ∆QRSi = âi(QRS)∆t
and where âk,i(T) and âi(QRS) represent the timeshifting factors of the T and QRS wave
matching windows and∆t the digitalization interval (1 ms). Because Qini0 (and QRSend0) is the
same for all leads, this same procedure also applies to ∆QRSi. The latter is selected among all
eight measured ECG channels as that ∆QRSk,i which is placed to the time instant of the
highest density of the ∆QRSk,i distribution, using a technique(17) similar to that described for
selection of the initial 20 beats with respect to the RR interval. We have found that such
∆QRSi differ less than 0.3 ms from a corresponding ∆QRSchange obtained from the derived
vector ECG using the method of Edenbrandt and Pahlm(18).
Finally, the QT interval for the ith beat is calculated according to the equation:
QTk,i = QTk,0 + ∆QTk,i
where QTk,0 is the duration of the template QT interval in the kth lead.
The algorithm thus provides the QT interval for each beat such that the T wave and
QRS complex best match the template T wave and QRS complex under the timeshift model.
PQ interval variability algorithm
The sum of time displacement of the P (∆Pk,i) and QRS waves (∆QRSi) constitutes the PQ
interval change for the ith beat and of kth ECG lead (ΔPQk,i) and is calculated according to
ΔPQk,i = ΔPk,i + ∆QRSi
ΔPk,i = âk,i (P)∆t, and ∆QRSi = âi(QRS)∆t
and where âk,i(P) and âi(QRS) represent timeshifting factors of the P and QRS wave matching
windows and∆t the digitalization interval (1 ms).
Finally, the PQ interval for the ith beat is calculated according to the equation:
PQk,i = PQk,0 + ∆PQk,i
where PQk,0 is the duration of the template PQ interval.
Time series analysis
The PQT program uses the QTV algorithm as described above to generate, in real
time, the time series of the QT interval along with that of the RR interval. Time series are
analyzed according to the recommendations of the Task Force of the European Society of
Cardiology and the North American Society of Pacing and Clinical Electrophysiology(19)
using specific indices such as the standard deviation of normaltonormal RR and QT
intervals (SDNN RR and SDNN QT, respectively), the root mean square of the successive
interval difference (rMSSD RR and RMSSD QT), etc. In addition, realtime changes in
related indices that attempt to “correct” QTV (or PQV) for simultaneous variability in RR
intervals (see for example (1, 2, 16)) can also be enabled if appropriate, as can (principally
for academic reasons(20)) a beattobeat QT dispersion function (21) for realtime display.
The Multilead QT interval signal
Because the standard ECG provides several leads in which the QT signal varies in a
similar manner with time, the PQT program constructs a “multilead” QT signal, defined as an
average of the four most similar signals from the eight independently measured ECG
channels. The multilead QT interval is specifically derived in the following way: In each of
the eight measured signals, an average QT interval value for the last n beats is calculated,
QTav, which is in turn subtracted from each of the individual QT intervals within the n beats
to obtain the deviation from the mean, QTQTav. In case of the kth lead, the result is denoted
as dQTk,i = QTk,i QTav,k. For each of any two mutually considered signals, kth and lth, a
covariance of the median beat for the nbeats interval is computed as:
Cov(QTk, QTl) = ∑
(dQTk,i – dQTl,i) 2 /(n1).
We thus obtain 8*7/2 = 28 separate covariance values from the eight ECG channels.
Subsequently, the program accepts those half of the channels that overall have the
most similar variations and rejects those half that have the least similar variations. Suppose
for example that a particular ECG channel appeared in j mutual covariances in the accepted
set. The program sums all of those covariances, divides the sum by the number j, and then
takes the square root of the result to get a complex covariance value. Finally, both for this
channel and for the other seven independent ECG channels, it calculates the time average
over the last nbeat covariance values and sorts the ECG channels by those values to select
the four channels with the most similar QT signals. The time averaging is necessary to retain
some stability in case of occasional signal artifact in the accepted ECG channels. The four
accepted ECG channels are then used for the construction of the multilead QT interval signal
simply by taking the average of their corresponding QT interval values for each ECG beat.
The time series of the QT interval thus obtained represents an idealized time series derived
from the four channels containing the most similar information, thereby reducing the effects
of noise. In addition, for each beat the program calculates the SD of the four accepted QT
signals from their mean, a quantity called “QTdev”. The QTdev is not only a measure of the
quality of the derived multilead QT interval signal, but also of the spatial variability in the
selected ECG channels. When the beattobeat spatial variability (QTdev) is less than the
simultaneous beattobeat temporal variability for the multilead QT interval (also measured
as SD), then the temporal variability is more likely a consequence of genuine (as opposed to
Selecting only four ECG channels for determination of the multilead QT interval
signal tends to eliminate those ECG channels that produce spurious QTV values due to poor
delineation of the T and/or QRS waveforms. In relatively noisefree recordings wherein the
waveform peaks and troughs are well defined, a high QTdev might theoretically be a
consequence of true regional differences (heterogeneity) in repolarization, but this hypothesis
has not yet been investigated.
QTRR cross correlation
A cross correlation is also determined between the multilead QT interval and the RR
interval for the last n beats. This is calculated as a covariance between the RR and QT
interval signals, cov(QT, RR), divided by the square root of each variance:
QTRRxc = cov(QT, RR)/√ QΤv √RRv.
It is expected that QTRRxc will approach 1.0 if QT and RR intervals vary in parallel,
and –1.0 if they vary in the opposite way, with low absolute QTRRxc values occurring if
variations of QT and RR signals are poorly correlated, as is common in disease(8, 22).
Although the program calculates QTRRxc between the RR interval and the QT in each of the
eight measured ECG channels, the QTRRxc obtained from the multilead QT interval signal
provides a singularly representative measure for practical use.
QT and PQ variability based on decomposed signals
The PQT program also calculates ECG waves along the principal axes, obtained by
singular value decomposition (SVD)(20, 23, 24) of the incoming eight channels of raw data.
In this technique, the ECG is reconstructed after SVD in an orthogonal eight lead system, the
first 3 orthogonal components of the T wave representing the traditional 3dimensional T
wave vector or dipolar signal contents. The abundancy of each component is given by its
maximal value along its principal axis (eigenvalue). However, because the overall electrode
placement in standard ECG is predominantly oriented along the ventricular axis, the
eigenvalue of the main orthogonal component obtained by SVD is usually several times
larger than that of any individual standard lead. The program therefore employs a modified
SVD algorithm to construct the complete ECG signal, including the T wave and QRS
complex along the principal axes. Specifically, it attenuates each of the measured ECG
signals by that amount required to reach the isotropic distribution with respect to the signal’s
orientation in space. As an approximation, the program employs a method similar to that of
Edenbrandt and Pahlm(18), who reconstructed the Frank vector ECG (x, y and z leads) from
the standard leads. However, it uses the multiplication factors ax, ay, and az of the x, y and z
component of the corresponding lead to calculate the main diagonal, dk, instead. For the kth
lead, dk = √( ax
2 + ay
2 + az
2 ). Next, the ECG signal of the kth lead is multiplied by dk and
applied in the SVD algorithm. As a result, the principal axes of such an approach nearly
coincide with those obtained by the SVD algorithm using the three derived orthogonal
leads(18). Since the eigenvalues of the orthogonal ECG components fall rapidly from the first
through the eighth orthogonal components, the SVD method represents the least square
solution to an overdetermined set of linear equations provided by the set of eight measured
ECG signals. It is therefore a sound approach to eliminate any kind of superimposed noise.
Because for each of the ECG wave components (P, QRS and T), the third through
eighth eigenvalues are usually only a small fraction of the first two eigenvalues, the program
uses only the first (or alternatively the first and second) orthogonal signals along the principal
axes (SVD1 and SVD2) to measure what we have termed the QTSVD and PQSVD intervals. In
addition, the program constructs a diagonal value from the three largest orthogonal
components (in the Frank lead system this would correspond to the radius vector amplitude,
or RVA) with time course represented by R0(t) = √( SVD1(t) 2 + SVD2(t) 2 + SVD3(t) 2 ). The
RVA signal can only be positive. Nevertheless, it is the only signal we derive that contains
the complete threedimensional information regarding the beginning and the end of the
various intervals. Since the resulting signals show by far the lowest variability in waveform
shape when compared to their initial templates (see below), they are used to determine the
(socalled) QTSVD and PQSVD variability. The program also determines the values of the so
called Twave residuum (TWR)(20, 25, 26) both instantaneously for each beat, and in a
slidingwindow fashion for the last n signalaveraged beats, independently of the TWR of the
llustration of software performance: a study of healthy individuals
In order to illustrate the performance of the software, standard supine 12lead ECGs
were recorded for 5 min in each of 19 healthy young male subjects (age range 2023 years)
during both normal breathing and during deep, slow breathing. The deep, slow (~0.1 Hz)
breathing was performed to help maximize the RR interval excursions(27) (and thus also the
QT interval excursions) so that the stability of software's determinations of the QTc interval
could be assessed. In the same study, we were also interested in determining any regional
variation in the quality of the signals as evaluated from the corresponding norm values – i.e.,
for the P, QRS and T waves in the eight standard channels and in two other derived signals:
the first orthogonal SVD signal (SVD1) and radius vector amplitude (RVA) signal (defined
above). Another goal was to compare the spatial variation of the 4 components of the
multilead QT interval signal with that same signal's temporal variation. Finally, we also
desired to study the relative QRS wave jitter in each signal in order to determine the most
appropriate signal selection for Qini, specifically by measuring the SDNN of the QRS interval
within each channel or derived signal. The Institutional Review Board at the University of
Ljubljana approved the study, and all subjects, who were awake and in sinus rhythm without
any sign of heart disease or conduction disturbance, gave informed consent.
As shown in Table 1, the multilead QT interval and the QT intervals of the SVD1 and
RVA signals exhibited no statistically significant changes during normal breathing compared
to deep, slow breathing, although there was a trend toward shortening of the QT intervals
during the deep, slow breathing. Since this shortening was paralleled by shortening of the RR
interval (from 909.7 ± 147 ms to 834.5 ± 139 ms), we corrected the multilead QT interval
according to the formula QTc = QT/RR 0.314 (28). The resulting QTc interval changed from
457.3 ± 18.9 ms to 461.8 ± 21.9 ms, with a relative change of 1.9 ± 1.4%. In other words,
even ignoring any inaccuracy associated with one of the bettervalidated generic QTc
correction formulas (28), the software determined QTc across the two different conditions
with a precision below 5 ms, or 2% .
As shown in Figure 4 (left side), for all three waveform types (P, QRS and T), the
lowest norm values (best performance) during normal breathing generally occurred in the two
derived signals rather than in any of the standard leads. Whereas the norm values in standard
lead II were amongst the lowest for the P and QRS waves, they were amongst the highest for
the T wave, suggesting that lead II is probably not the most ideal lead for QTV analyses. In
general, the P wave norm values were approximately double those for the QRS and T waves
across the various leads. Deep, slow breathing adversely affected the norms in the majority
of leads (Figure 4, right side). Nonetheless, one or the other of the derived signals still
remained the best overall performer for all three waveform types, regardless of the breathing
The spatial variability (QTdev) associated with the four constituents of the multilead
QT signal was 1.79 ± 0.98 ms during normal breathing versus 2.42 ± 1.3 ms during deep,
slow breathing, while the simultaneous temporal variability of the signal as measured by
SDNN was 3.14 ± 2.0 ms versus 5.47 ± 4.60 ms, respectively. Thus, given the ratios of
spatial to temporal variability, much of the temporal variability was likely genuine and not
due to noise.
Values for PQ variability, QRS jitter, and QT variability during normal breathing (all
as measured by SDNN) are shown in Figure 5 (left side). During deep, slow breathing, both
variability and jitter increased (right side of Figure 5). As with the norm, the mean variability
and jitter values were also generally lowest for one or the other of the derived signals.
The QRS jitter value (by SDNN) in lead II during normal breathing was 1.07 ± 0.71 ms. This
compares to an actual PQ and QT variability (also by SDNN) of 4.31 ± 1.88 ms and 3.30 ±
1.55 ms, respectively, in the same lead during normal breathing. Thus, QRS jitter itself could
theoretically account for up to a quarter of the PQV and up to a third of the QTV in lead II
based calculations. The potential contribution of QRS jitter to PQV and QTV appeared to be
even higher during deep, slow breathing.
Finally, Figure 6 shows an example of RR interval variability (panel a) and the
corresponding QTV (panel b) in a healthy young male performing deep, slow (0.1 Hz)
breathing. As expected, the QT interval mostly followed the RR interval, which resulted in a
high absolute value for QTRRxc. The quality of matching (norm) in this case was also better
for the T wave than for the QRS wave (panel c), while the QRS signal showed considerable
jitter with respect to the trigger (panels c and d). As illustrated by this case, the results of
QTV using our method (shifting) do not differ in any significant way from those using the
method of Berger et al (stretching) when the latter method is also corrected for QRS trigger
point jitter. When such is the case, both methods result in nearly identical QT interval signals
and norms for the T wave (panels b and c). Very pronounced though was the exaggerative
effect that QRS jitter had on QTV. Thus, in any case where the Berger et al. (or our) method
relies strictly on a QRS fiducial point uncorrected for jitter, it results in a distorted
(exaggerated) QTV, as shown in panel b.
In conclusion, our institutions have collaboratively developed an advanced ECG
software program that in conjunction with commercially available PCbased ECG hardware
and software, acquires, analyzes and displays the QT and PQ variability within each of the
independent ECG channels in real time on a beattobeat basis. It also performs the same
functions utilizing certain derived signals that reduce the effects of both noise and
physiological jitter and thus improve the reliability of the overall results. Although additional
clinical validation is necessary, the automated, multichannel, realtime and noisereduction
aspects of the software will hopefully aid in bringing what has heretofore been regarded as a
highly promising research technique into eventual clinical usage.
1. Vrtovec B, Starc V, Starc R. Beattobeat QT interval variability in coronary patients.
J Electrocardiol. 2000;33(2):11925.
Vrtovec B, Sinkovec M, Starc V, Radovancevic B, Schlegel TT. Coronary artery
disease alters ventricular repolarization dynamics in type 2 diabetes. Pacing Clin
Electrophysiol. 2005;28 Suppl 1:S17881.
Johansson M, Gao SA, Friberg P, et al. Elevated temporal QT variability index in
patients with chronic renal failure. Clin Sci (Lond). 2004;107(6):5838.
Murabayashi T, Fetics B, Kass D, Nevo E, Gramatikov B, Berger RD. Beattobeat
QT interval variability associated with acute myocardial ischemia. J Electrocardiol.
Nahshoni E, Strasberg B, Adler E, Imbar S, Sulkes J, Weizman A. Complexity of the
dynamic QT variability and RR variability in patients with acute anterior wall
myocardial infarction: a novel technique using a nonlinear method. J Electrocardiol.
Piccirillo G, Germano G, Quaglione R, et al. QTinterval variability and autonomic
control in hypertensive subjects with left ventricular hypertrophy. Clin Sci (Lond).
Atiga WL, Fananapazir L, McAreavey D, Calkins H, Berger RD. Temporal
repolarization lability in hypertrophic cardiomyopathy caused by betamyosin heavy
chain gene mutations. Circulation. 2000;101(11):123742.
Berger RD, Kasper EK, Baughman KL, Marban E, Calkins H, Tomaselli GF. Beatto
beat QT interval variability: novel evidence for repolarization lability in ischemic and
nonischemic dilated cardiomyopathy. Circulation. 1997;96(5):155765.
Berger RD. QT variability. J Electrocardiol. 2003;36 Suppl:837.
Atiga WL, Calkins H, Lawrence JH, Tomaselli GF, Smith JM, Berger RD. Beatto
beat repolarization lability identifies patients at risk for sudden cardiac death. J
Cardiovasc Electrophysiol. 1998;9(9):899908.
Thomsen MB, Verduyn SC, Stengl M, et al. Increased shortterm variability of
repolarization predicts dsotalolinduced torsades de pointes in dogs. Circulation.
Gao SA, Johansson M, Hammaren A, Nordberg M, Friberg P. Reproducibility of
methods for assessing baroreflex sensitivity and temporal QT variability in endstage
renal disease and healthy subjects. Clin Auton Res. 2005;15(1):218.
GovreenSegal D, Radai MM, Sivan Y, Abboud S. Realtime PCbased system for
dynamic beattobeat QTRR analysis. Comput Biomed Res. 1999;32(4):33654.
Forester J, Bo H, Sleigh JW, Henderson JD. Variability of RR, P wavetoR wave,
and R wavetoT wave intervals. Am J Physiol. 1997;273(6 Pt 2):H285760.
Shouldice R, Heneghan C, Nolan P, Nolan PG, McNicholas W. Modulating effect of
respiration on atrioventricular conduction time assessed using PR interval variation.
Med Biol Eng Comput. 2002;40(6):60917.
Arnol M, Starc V, Starc R. Atrioventricular conduction variability in coronary
patients. J Electrocardiol. 2003;36(4):3119.
Press WH. Numerical recipes in C : the art of scientific computing. 2nd ed.
Cambridge ; New York: Cambridge University Press; 1992:610615.
Edenbrandt L, Pahlm O. Vectorcardiogram synthesized from a 12lead ECG:
superiority of the inverse Dower matrix. J Electrocardiol. 1988;21(4):3617.
19. Heart rate variability. Standards of measurement, physiological interpretation, and
clinical use. Task Force of the European Society of Cardiology and the North
American Society of Pacing and Electrophysiology. Eur Heart J. 1996;17(3):35481.
Malik M, Acar B, Gang Y, Yap YG, Hnatkova K, Camm AJ. QT dispersion does not
represent electrocardiographic interlead heterogeneity of ventricular repolarization. J
Cardiovasc Electrophysiol. 2000;11(8):83543.
Figueredo EJ, Ohnishi Y, Yoshida A, Yokoyama M. Usefulness of beattobeat QT
dispersion fluctuation for identifying patients with coronary heart disease at risk for
ventricular arrhythmias. Am J Cardiol. 2001;88(11):12359.
Frljak S, Avbelj V, Trobec R, Meglic B, Ujiie T, Gersak B. Beattobeat QT interval
variability before and after cardiac surgery. Comput Biol Med. 2003;33(3):26776.
Press WH. Numerical recipes in C : the art of scientific computing. 2nd ed.
Cambridge ; New York: Cambridge University Press; 1992:5970.
Acar B, Yi G, Hnatkova K, Malik M. Spatial, temporal and wavefront direction
characteristics of 12lead Twave morphology. Med Biol Eng Comput.
Zabel M, Malik M, Hnatkova K, et al. Analysis of Twave morphology from the 12
lead electrocardiogram for prediction of longterm prognosis in male US veterans.
Okin PM, Malik M, Hnatkova K, et al. Repolarization abnormality for prediction of
allcause and cardiovascular mortality in American Indians: the Strong Heart Study. J
Cardiovasc Electrophysiol. 2005;16:17.
Bernardi L, Sleight P, Bandinelli G, et al. Effect of rosary prayer and yoga mantras on
autonomic cardiovascular rhythms: comparative study. Bmj. 2001;323(7327):14469.
Malik M. Problems of heart rate correction in assessment of druginduced QT interval
prolongation. J Cardiovasc Electrophysiol. 2001;12(4):41120.
Table 1. Changes in intervals after breathing change
RR interval QTcML
442.8±31.1 442.8 ± 30.51 442.3 ± 30.6 909.6 ± 147.6 457.27 ± 18.9
434.7±26.5 437.7 ± 30.7 435.3 ± 25.2 834.5 ± 139.4 461.78 ± 21.9
Values are means ± standard deviations in ms. N = 19 subjects. QTML, QTSVD1 and QTRVA:
QT interval values based on the multilead QT interval, SVD1and RVA signals, respectively.
QTcML: the corrected multilead QT interval value. No statistically significant changes in QT
intervals were noted by ttest.
Figure 1. The borders of the template QRS complex are determined by searching for local
minima and maxima as well as by the minimal and maximal slopes. The crossing of the
initial slope with the preceding nearly horizontal segment determines the initial point of the
QRS complex, Qini, of each measured ECG lead, and the last crossing the point QRSend,
necessary for determination of the breaking points between waves, PQbreak and QTbreak, which
occur 20 ms before Qini0 and 30 msec after QRSend0, respectively. The points QRS1, and QRS2
determine the smaller time window for the final matching of the QRS wave.
Figure 2. To determine the Twave onset and offset, a biparabolic curve is constructed using
two interconnected parabolic segments, both with an amplitude one half of the total
biparabolic curve height. The segments are connected at the median point of the biparabolic
curve. In the distal parabolic segment a horizontal line replaces the most distal ‘hook’
between Tend and T2. The initial biparabolic curve (thick solid dark gray curve) is shifted
toward the QRS complex, while varying its height and width until finding the best fit to the
template T (thick solid black curve). The position of the biparabolic curve then determines
the border points on the template wave (dashed curve), for defining the smaller time window
for later matching (see Figure 3b). These border points are T1 and T2 for the timeshift
parameter, and Tend for determining the end point of the T wave.
Figure 3. a) Initial shifting and normalizing: The incoming wave (taller dashed curve,
black) is shifted toward the template wave (solid dark gray curve) for matching.
Simultaneously, it is normalized to have the same amplitude in the broader time window
from QTbreak to T2 (lower dashed curve, black). The squared difference between the
normalized wave and the template represents the norm for the T wave. b) Final shifting: The
distal segment (lower right dashed curve, black) of the normalized incoming wave (dashed
dark gray curve under the template T wave) is shifted further toward the template wave (solid
dark gray curve) for matching in the smaller time window from T1 to T2. The required total
shift from the incoming wave to the final position of the distal segments represents the time
Figure 4. Norm values for the P, QRS and T waves during normal breathing (left plots) and
during deep, slow breathing (right plots) in a group of 19 healthy males. Norm values for the
P wave were higher than those for the QRS and T waves, and for all wave types, the
performance of one or the other of the derived signals (i.e., of SVD1, the first orthogonal
component from singular value decomposition, or of RVA, the radius vector amplitude) was
generally superior to that of the best of the eight individual standard channels.
Figure 5. Variability of the PQ, QRS and QT intervals during normal breathing (left plots)
and during deep, slow breathing (right plots) in the same group of 19 healthy males. The
SDNN (standard deviation of normal intervals) of all three interval types generally increased
during deep, slow breathing compared to normal breathing, with the increased variability of
the QRS interval reflecting increased “jitter”. As with the norm values, the mean variability
and jitter values were also generally lowest for one or the other of the derived signals.
Figure 6. QT and other interval variability during deep, slow (0.1 Hz) breathing in a healthy
male. Panel a (uppermost) shows the RR interval changes during a 100 s interval. Panel b
shows the corresponding changes of the QT intervals in lead V3. Within panel b, the lower,
QRS jittercorrected QT trends were obtained by the “shifting” method (described herein)
and by the “stretching” method of Berger et al(8), respectively. The two trends are
effectively superimposed, demonstrating the practical equivalence of the two techniques. The
upper trend in the panel b represents variability of the T wave only, without considering QRS
wave variability (“jitter”). Panel c demonstrates that the norms for the T wave also do not
differ between the “shifting” and “stretching” methods, although in this case the Twave
norm was considerably lower than that for the QRS. Panel d demonstrates that QRS jitter can
be significant, particularly during deep breathing.
Page 33 Download full-text