COMPUTER APPLICATION

Systematic Comparisons of ElectrocardiographicMorphology Increase the Precision of QT IntervalMeasurementKATERINA HNATKOVA, PH.D.,* PETER SMETANA, M.D., PH.D.,† ONDREJ TOMAN, M.D.,‡AXEL BAUER, M.D.,§ GEORG SCHMIDT, M.D.,§ and MAREK MALIK, PH.D., M.D.*

From the *St. Paul’s Cardiac Electrophysiology, London, United Kingdom; †Wilhelminenspital der Stadt Wien,Vienna, Austria; ‡University Hospital Brno, Masaryk University, Brno, Czech Republic; and §1. MedizinischeKlinik, Technische Universitat Munchen, Munich, Germany

Decreased intrasubject variability of QTc values is needed to increase the power and reduce the size ofthe so-called thorough QT studies. One source of QTc variability is the lack of systematic measurementswhen electrocardiograms (ECG) with closely matching morphologies are not measured in an exactly corre-sponding way. The inaccuracy can be eliminated by postprocessing of QT measurements by ECG patternmatching. This study tested the effects of pattern matching in ECG measurements in two populationsof healthy subjects (n = 48 + 56) and in a population of patients with advanced Parkinson’s disease(n = 130) in whom both day-time and night-time data were available. Intrasubject QTc variability wasmeasured by intrasubject standard deviations (SD) of QTc values obtained with manual measurementsbefore and after pattern-matching measurement alignments. In each subject, QT values (n = 230–320)in one drug-free long-term ECG recording were evaluated. The pattern-matching adjustment of the QTmeasurement decreased the intrasubject QTc variability from 5.2 ± 1.0 to 4.5 ± 1.0 ms (P < 10−14) from6.4 ± 1.7 to 5.5 ± 1.6 ms (P < 10−10) from 5.6 ± 1.5 to 4.6 ± 1.4 ms (P < 10−34) and from 6.1 ± 1.9 to5.0 ± 1.7 ms (P < 10−33), in the two populations of healthy subjects and in the day-time and night-timerecordings of Parkinson’s disease patients, respectively. Hence, morphological pattern adjustment of QTinterval measurements improves the quality of the QT data with substantial practical implications. Re-ductions in intrasubject QTc variability were reproducibly found in different populations and thus thetechnology might be recommended for every thorough QT/QTc study. Noticeable reductions of necessarystudy size are likely achievable in this way. (PACE 2009; 32:119–130)

thorough QT/QTc study, QTc variability, ECG morphology, pattern matching

IntroductionPresent regulatory guidance requires the so-

called thorough QT/QTc study (TQTS) to be con-ducted for practically every new pharmaceuticalcompound.1 To assess the danger of drug-inducedTorsade de Pointes tachycardia, the regulatorsneed to evaluate the proarrhythmic potential sig-nals for every drug that prolongs the QTc inter-val. This includes evaluation of cases when thisdrug-induced QTc prolongation is rather small.Specifically, TQTS is considered negative only ifthe upper single-sided 95% confidence interval ofplacebo and baseline corrected QTc prolongationon active treatment is below 10 ms.1

Address for reprints: Marek Malik, Ph.D., M.D., St. Paul’s Car-diac Electrophysiology, 16 Verulam Avenue, Purley, Surrey,London CR8 3NQ, United Kingdom. Fax: +44-20-8660-6031;e-mail: marek.malik@btinternet.com

Received August 18, 2008; revised September 25, 2008; ac-cepted October 11, 2008.

The requirement of assessing the drug-induced QTc changes with this high precisionmakes TQTS exceptional in terms of the mea-surement of standard 12-lead electrocardiograms(ECG). Neither clinical practice, e.g., for the diag-nosis of congenital long QT syndrome, nor otherECG investigations, such as risk stratification ofcardiac patients, require ECG measurement to bemade this precisely. It is therefore not surprisingthat TQTS are frequently conducted in large pop-ulations (e.g., group sizes >50) even when con-ducted in healthy subjects.2–4 Large numbers ofstudy subjects not only allow averaging out mea-surement outliers but also decrease the width ofconfidence intervals of QTc changes.

Because of the well-known substantial inter-subject variability of QTc values,5,6 TQTS needto evaluate intrasubject differences between base-lines and on treatment readings or between read-ings on placebo and on active drug. This intra-subject analysis applies irrespective of whetherthe studies are of crossover or parallel design.

C©2009, The Authors. Journal compilation C©2009, Blackwell Publishing, Inc.

PACE, Vol. 32 January 2009 119

HNATKOVA, ET AL.

Figure 1. Example of representative superimposed median beats of all 12-leads of two separate10-second ECGs recorded in a healthy subject approximately 20 minutes apart (FDAEcg graphicssystem, AMPS, New York, NY). In each panel, four different T wave offset measurements areshown (full and open arrows below and above the ECG tracings). All these four measurements(shown in steps of 5 ms) as well as measurements in between are plausible and depend on ECGinterpretation strategy. If the same interpretation and measurement strategy is used for bothtracings (thus if the measurements are systematic) corresponding measurement positions areselected.

The required final precision makes it thereforeimportant that, among others, the ECG readingsare systematic, i.e., that the same or closelysimilar ECG patterns are measured in the sameway.

To explain this in more detail, Figure 1 showsrepresentative beats of two 10-second ECGs thatwere recorded in the same healthy subject approx-imately 20 minutes apart. While the QT intervalsare clearly different, both the Q-onset and T-offsetmorphological patters, i.e., the shape of the ECGtracings around the beginning of the QRS complexand the end of the T wave, are almost identical inboth ECGs. In each of the two ECGs, four differ-ent measurements of T offset are shown withina range of 15 ms. Since the definition of T waveoffset depends on different possibilities of ECG in-terpretation, none of these measurement possibil-ities can be declared more correct than the others.However, it is obvious that the QTc values wouldbe made more variable if different interpretationswere used for these two ECGs. Because the TQTSevaluation depends on differences between QTcvalues, it is irrelevant which of the different ECGinterpretations shown in Figure 1 is used as longas it is used systematically.

Hence, ensuring systematicity of QT measure-ments means securing that the very same interpre-tation of ECG patterns and morphologies is usedwhen detecting Q onset and T offset. (By patterns

and morphologies, we understand the shapes andcomposites of the recorded tracings that allow theECG interpretation.)

Not surprisingly, the systematicity of ECGmeasurements can be more easily guaranteed withcomputer rather than manual readings. With cor-rectly designed algorithms, the results based onidentical input data are always the same. Unfortu-nately, in spite of the recent substantial advances,computerized measurements of ECG patterns areknown to be far from 100% reliable.7 Even themost advanced algorithms occasionally lead tovery substantial errors that can overcome the ad-vantage of systematicity and invalidate a TQTSentirely. Therefore, although not strictly requiredby regulatory guidance, the majority of TQTSpresently use manual ECG measurements, mostfrequently performed with electronic calipers oncomputer screen. Human visual evaluation of ECGpatterns is naturally less systematic than algo-rithmic evaluation and thus, even in cases whenthe human observer only corrects initial computerreadings, the systematicity of the measurement isfrequently lost. (For example, using solely visualinterpretation, it is practically impossible to guar-antee that the two ECGs shown in Figure 1 aremeasured in a fully systematic way.)

At the same time, computerized pattern recog-nition and classification of ECG morphologies canbe made entirely independent of the measurement

120 January 2009 PACE, Vol. 32

SYSTEMATIC QT INTERVAL MEASUREMENT

procedures providing that the initial measure-ments are free of gross errors. This leads to thepossibility of setting up a postprocessing systemaligning the ECG measurements in the same sub-ject and repositioning the measurement triggers sothat that mutually corresponding ECG patterns aremeasured in the same and systematic way.

This text reports on the development of a post-processing system of this kind and on its testingin three independent sets of data that were ob-tained in studies selected from previously con-ducted TQTS.

MethodsPattern Matching and Alignment System

Mathematical and algorithmic principles ofthe system that was set up for ECG measure-ment alignment are described in the Appendix.In brief, the system calculates a “distance” be-tween different ECG morphologies and using it,sorts all measured ECGs according to the closenessof their morphologies. For every measured ECG, aweighted average is subsequently calculated of thedifferences between measurement triggers in thegiven ECG and in all other ECGs, using semire-ciprocal values of the distances as weights. Fromthe weighted average of the measurement differ-ences, new positions of measurement triggers arecalculated.

Populations

The system was tested with ECG data ofthree separate populations. The first population(Healthy subjects A) consisted of 48 healthy sub-jects (mean age 31.1 ± 8.7, range 18–48 years,25 women). The second population (Healthy sub-jects B) consisted of 56 healthy subjects (mean age34.6 ± 9.4, range 19–60 years, 24 women). Thethird population (Parkinson’s disease) consistedof 130 patients with advanced Parkinson’s dis-ease (mean age 63.3 ± 10.1, range 29–88 years,41 women). All the populations were taken fromstudies that were conducted at regulatory regis-tered clinical facilities. In every case, the sourcestudy was approved by the Institutional EthicsCommittee or Review Board and each study sub-ject gave informed written consent.

The populations were selected such that ofthe available data, the population of Healthy sub-jects A had near the best quality of ECG record-ings while the population of Healthy subjects Bhad, among previous studies of healthy subjects,near the most noise-polluted ECG recordings (dueto electrical background noise in the environmentand electrode noise). The population of Parkin-son disease was included because the underly-ing pathology is well known to affect ECG qual-

ity. This spectrum allowed assessing how muchthe pattern alignment system is dependent onthe quality and noise contents of the measuredECGs.

Electrocardiographic Recordings

The ECG recordings of all three populationswere obtained from TQTS that investigated theeffects of different drugs, including moxifloxacinthat was consistently used as the positive con-trol. In this manuscript, however, the evaluationof the adjustment algorithm is based only on ECGdata obtained from baseline recordings that werenot influenced by any treatment. Specifically, forthe purposes of testing the algorithm, we reporton data that were obtained during one baselineday-time recording made in each subject of all3 populations. For the Parkinson disease popula-tion, we also tested the algorithm in one night-timerecording made in each subject of the population(night-time recordings were not available from thetwo studies in healthy subjects). Because of theknown day/night differences in QT interval du-ration, we report on the day-time and night-timerecordings in the Parkinson Disease populationseparately.

All the recordings were initially obtainedas long-term 12-lead ECGs. From each long-termrecording, several 10-second ECG segments wereselected for accurate measurement. In the studiesof all three populations, this selection process wasbased on the same principles and organized in twosteps. First, ECG segments were selected withinprespecified data-point windows defined by theprotocol of the original study. In each case, theECG selected within these time points were pre-ceded by stable heart rate allowing < ±2 beats/minheart rate fluctuations for at least 90 seconds beforethe selected 10-second segment. Of all possible se-lections satisfying these conditions, ECGs with thelowest noise contents8 were selected.

The second selection step was made to char-acterize individualized QT/RR relationship forthe purposes of designing subject-specific heartrate correction formulae. Individualized QT/RRpattern characterization that is needed for thesubject-specific heart rate correction must includesufficient heart rate range with comprehensivesampling between the slowest and fastest ratesfound in the source recording. This is not achiev-able using only ECG data obtained in prespeci-fied study data-points since at such moments, thesubjects are usually at supine rest and their heartrates do not differ sufficiently. Therefore, in eachbaseline long-term recording, all 10-second ECGswere found that were preceded by at least 120 sec-onds of stable heart rate, again permitting vari-ability up to ±2 beats/min. These ECG segments

PACE, Vol. 32 January 2009 121

HNATKOVA, ET AL.

were sorted according to the preceding stable rateand divided, in each baseline long-term recording,into a number of bins as equidistant as possible inthe underlying heart rate. In each bin, the ECGwith the lowest noise content8 was subsequentlyselected for accurate measurement. In the record-ings of Healthy subjects A and Healthy subjectsB, 100 and 200 such bins were used, respectively(the exact number was dictated by the protocol ofthe given source TQTS) while in each day-timeand each night-time recording of Parkinson dis-ease subjects, 150 bins were used.

Altogether, for the purposes of the algorithmtest reported in this text, 230, 320, 305, and 265ECG segments of 10-second duration were con-sidered in each individual of Healthy subjects A,Healthy subjects B, and in day-time and night-timerecording of each patient of the Parkinson dis-ease population, respectively. (These exact num-bers were again derived from the protocols of thesource TQTS.)

Similar ECG extractions were also made fromother baseline recordings of the source TQTSas well as from on-treatment recordings. Alto-gether, per study protocol, 1620, 920, and 2800ECG segments were measured in each individ-ual of Healthy subjects A, Healthy subjects B, andParkinson disease, respectively. The algorithm ofmorphological pattern adjustment was applied toall these measured ECG segments in each subjectbut the evaluation of the algorithm reported hereconsidered only the ECGs extracted from one sin-gle baseline recording, as described in the preced-ing paragraph.

Electrocardiographic Measurements

In all three source TQTS, the same measure-ment strategy was used.

Each selected 10-second ECG was filtered andits baseline wander removed. (Modification ofpreviously published filtering and baseline wan-der removal methods was used in the studies ofhealthy subjects9,10; study-specific filtering andbaseline removal techniques were developed forthe processing of the ECGs in the Parkinson pa-tient’s study.) From these signals, median repre-sentative beats were constructed and QT intervalwas measured using an in-house developed tech-nology based on an advanced version of patternclassification algorithm.

Manual Measurements

In each ECG, the Q-onset and T-offset triggerswere subsequently visually checked and, wherenecessary, manually corrected by two indepen-dently working cardiologists. These visual checksand measurement corrections were performed ona computer screen in the median beats of all

12-leads images superimposed on the same iso-electric axis in a 1 kHz resolution. Results of thesemeasurement checks by the two independent car-diologists were compared and those ECGs identi-fied in which the cardiologists differed in the QTinterval measurement by >6 ms. These ECGs werereturned to the same pair of cardiologists for re-peated verifications. Otherwise, the positions ofthe Q-onset and T-offset triggers were averaged.After the remeasurement of the ECGs that orig-inally led to discrepancies, the results providedby the two cardiologists were again compared andcases where they still disagreed were reconciledby a senior cardiologist. All the ECGs of the samepatient were processed by the same personnel.This procedure led to quality assured manual mea-surement of the selected 10-second ECG segments.Nonmeasurable ECG samples were identified atthis step.

Pattern Aligned Measurement

Subsequently, the pattern matching and align-ment algorithm was applied and suggested mod-ifications in the manual QT interval measure-ments which were again reviewed on the computerscreen. In this way, the systematically adjusted QTinterval measurements were obtained.

Automatic Measurements

In order to compare these QT interval mea-surements with widely available full automaticmeasurements, all of the 10-second ECG segmentsused in this investigation were also measured bythe most recent version of the 12SL software pack-age (GE Healthcare, Milwaukee, WI, USA).7 Auto-matic QT interval measurements were obtained inthis way.

Hence, three different QT interval values wereobtained for each measured ECG segment, i.e., the“automatic,” “manual,” and “pattern-matching”measurements. (Note that the “manual” measure-ment involved initial computerized measurementand thus belongs to the category that is occasion-ally also called “semiautomatic.”)

Heart Rate Correction

With all three QT interval measurements, theQT/RR patterns were obtained in each baselinerecording considered in this report. In each case,the QT intervals values were related to the RR in-terval representative of stable heart rate precedingthe measured ECG.

Individually Optimized Heart Rate Correction (QTcIvalues)

The resulting QT/RR patterns were convertedinto individualized heart rate correction formulaewhich included curvature optimization.6,11 The

122 January 2009 PACE, Vol. 32

SYSTEMATIC QT INTERVAL MEASUREMENT

QTcI values were obtained in this way (i.e., 3 dif-ferent sets of QTcI values were obtained in eachpopulation, corresponding to the 3 different setsof QT interval values).

Individualized Fixed Formula Heart Rate Correction(QTcIf values)

In addition, the parabolic log/log formula wasoptimized for each subject and each set of sourceQT measurements. That is, for each subject andeach of the “automatic,” “manual,” and “pattern-matching” QT values, an individual-specific pa-rameter α (the correction exponent) was found thatcorresponded to the individualized correction for-mula in the fixed mathematical form of QTc =QT/RRα.12 The QTcIf values were obtained in thisway.

Fridericia Correction (QTcF values)

Finally, Fridericia formula13 was also appliedand for each set of the “automatic,” “manual,” and“pattern-matching” QT values, the QTcF valueswere obtained using this formula.

Hence, in summary, nine different sets ofheart rate corrected QTc values were obtained, thatis, three heart rate correction possibilities (QTcI,QTcIf, and QTcF) for each of the three QT inter-val measurements (automatic, manual, and patternmatching). The RR interval values used in thesenine different QTc sets were always the same (rep-resenting the stable rate preceding the measuredECG—as described earlier).

Statistics

The goal of the adjustment algorithm is to in-crease the systematicity of intrasubject QT mea-surements. Therefore, the effect of the algorithmmay be measured by studying the intrasubjectstandard deviations (SD) of QTc values.

Specifically, in each subject of each popu-lation (considering the day-time and night-timerecordings in Parkinson’s disease separately), SDof the QTc values were calculated for all nine setsof QTc values. The algorithm of morphologicalpattern adjustment was considered successful ifit led to statistically significant reduction of in-trasubject SDs from manually measured QTc val-ues to pattern-matching QTc values. The reductionwas tested separately for QTcI, QTcIf, and QTcFusing paired two-tail t-test.

Where appropriate, further comparisonswithin subjects were made using paired two-tailt-test while intersubject comparisons were madeusing two-sample two-tail t-test assuming differ-ent variances. Data are presented as mean ± SD;statistical significance was assumed if P < 0.05.

ResultsAfter excluding nonmeasurable ECGs, the

tests of the algorithm of morphological pattern ad-justment were made involving 11,031 (99.92% perprotocol), 17,455 (97.41%), 38,767 (97.77%), and33,707 (97.84%) ECG segments in Healthy subjectsA, Healthy subjects B, and day-time and night-timerecordings of Parkinson’s disease subjects, respec-tively.

Figure 2 shows the distribution of the changesmade by the adjustment algorithm (that is, the dis-tribution of differences between the manual andpattern matching QT values). In an overwhelmingmajority, the adjustment changes were small onlybetween –5 and 5 ms.

Figure 3 shows scatter diagrams of individualintrasubject SDs of QTc values without pattern-matching (i.e., automatic and manual measure-ments) plotted against the SDs of pattern-matchedQTc values. Systematic reduction of SD of QTcvalues is obvious in the figure (note that the scatterdiagrams are above the dotted and dashed lines ofidentity). This was statistically confirmed. Specif-ically, considering only the QTcI values, the al-gorithm of morphological pattern adjustment re-duced the SDs of QTcI values from 5.23 ± 1.04 to4.50 ± 1.04 ms, P = 5.1 × 10−15; from 6.38 ± 1.69to 5.50 ± 1.65 ms, P = 1.5 × 10−11; from 5.55 ± 1.52to 4.55 ± 1.45 ms, P = 2.7 × 10−35; and from6.09 ± 1.88 to 5.04 ± 1.73 ms, P = 1.4 × 10−34 inthe data of Healthy subjects A, Healthy subjects B,and day-time and night-time data of Parkinson’sdisease patients, respectively. We have also no-ticed systematic and highly significant reductionsof the SDs of QTc values from automatic to manualmeasurement. Specifically, the SDs of QTcI valuesbased on manual measurement that we have justlisted were smaller than the corresponding valuesof SDs based on automatic measurement whichwere 6.21 ± 1.26, P = 8.3 × 10−8; 8.00 ± 2.71,P = 1.4 × 10−8; 7.43 ± 3.50, P = 4.6 × 10−12;and 7.70 ± 3.89, P = 8.3 × 10−8 (the P-values re-fer to the comparison with SDs based on manualmeasurement). Similar differences with similarlystrong statistical significances were found with theSD of QTcIf and QTcF values (Fig. 4).

Figure 5 shows examples of QT/RR distribu-tions taken from the individual populations. Theeffect of the pattern matching and alignment al-gorithm is obvious. After the application of thealgorithm, the QT/RR plots become much morecompact with substantially lower variability of QTintervals measured at close RR intervals.

In spite of the tight intrasubject SD of QTcvalues, the intersubject differences were marked.Figure 6 shows the distribution of the individ-ual QTcI intervals (based on pattern-matching

PACE, Vol. 32 January 2009 123

HNATKOVA, ET AL.

Figure 2. Distributions of the “corrections” of QT interval measurement made by the patternmatching and alignment algorithm in the four sets of data used in the algorithm tests. Note thatin all cases the majority of the corrections were very small. Specifically, corrections exceeding ±5ms were made only in 7.3%, 12.4%, 14.0%, and 17.7% of all ECGs in Healthy subjects A, Healthysubjects B, and day-time and night-time recordings of Parkinson’s disease patients, respectively.

measurement) found in different subjects of thepopulations. Among others, the figure shows thatin all populations, there were significant differ-ences in QTcI between men and women. Consid-ering the QTcI values based on pattern matchingQT measurement, the differences in intrasubjectmean QTcI values between women and men were:428.1 ± 16.3 versus 408.4 ± 9.4 ms, P = 7.1 × 10−6;427.7 ± 16.3 versus 406.2 ± 22.7 ms, P = 0.0001;426.3 ± 18.3 versus 416.3 ± 15.9 ms, P = 0.004;and 440.3 ± 18.3 versus 425.1 ± 15.7, P = 0.00002in the data of Healthy subjects A, Healthy subjectsB, and day-time and night-time data of Parkin-son’s disease patients, respectively. In the data ofParkinson’s disease patients, the intrasubject dif-ferences between QTcI at day and at night werealso statistically significant, both in women (P =2.5 × 10−12) and in men (P = 1.2 × 10−16).

The intrasubject SD of QTcI values (pattern-matching QT measurement) tended to be higherin women compared to men although the differ-ence did not reach statistical significance system-atically: 4.80 ± 0.97 versus 4.18 ± 1.04 ms, P =0.036; 5.62 ± 1.93 versus 5.41 ± 1.43 ms, P = 0.66;4.87 ± 1.25 versus 4.4 ± 1.51 ms, P = 0.065; and5.46 ± 1.44 versus 4.85 ± 1.82 ms, P = 0.042; inthe data of Healthy subjects A, Healthy subjects B,and day-time and night-time data of Parkinson’sdisease patients, respectively.

DiscussionThe investigation shows that highly statis-

tically significant improvement of the intrasub-

ject stability and thus of the quality of QTc datacan be obtained with post hoc morphologicalpattern adjustment of quality controlled manualmeasurements. Since we observed success of theadjustment algorithm irrespective of source ECGquality (including ECG data obtained in patientswith advanced Parkinson’s disease) the adjust-ment algorithm seems to be robust and stableindependently of technical and biological noise,providing that nonmeasurable ECG samples areproperly excluded. The study also confirms thatthe quality of the QTc data, as judged by the in-trasubject SDs, is poorer in fully automatic mea-surements compared to carefully reconciled andquality controlled manual readings.7 While the au-tomatic algorithmic reading of the QT interval is100% reproducible when applied to identical sig-nal repeatedly, tiny signal changes from one ECGto another may cause differences in the algorith-mic results even when the true QT interval has notchanged.

The observed reduction in intrasubject SDmay seem numerically small. However, since thepower sample calculations of TQTS depend on in-trasubject SD of QTc values quadratically,14,15 theimplications of these reductions are of substan-tial importance (e.g., increase of intrasubject SDby a quarter leads to the increase of study size byaround 60%). Hence the application of the mor-phological pattern adjustment to improve manu-ally measured QT values makes apparent practi-cal sense. Even in carefully selected good qualityECGs (population of Healthy subjects A) the effect

124 January 2009 PACE, Vol. 32

SYSTEMATIC QT INTERVAL MEASUREMENT

Figure 3. Each panel shows a scatter diagram of intrasubject SD of QTc values obtained without pattern matching(vertical axes) plotted against the SD of QTc values obtained after pattern matching (horizontal axes). The open markscorrespond to automatic measurement, the shaded marks to the manual measurement—both these are shown on thevertical axes. The dashed lines show the axes of identity. The four columns correspond to the investigated four setsof data, the top, middle, and bottom rows show standard deviations of QTcI, QTcIf, and QTcF values (see the text forexact definitions).

was comparable to the other populations. Substan-tial power increases and/or size reduction of TQTSmay thus be systematically achieved by the patternadjustment technology.

At the same time, it seems obvious that thetask of making the measurements systematic can-not be replaced by increased accuracy of man-ual judgments. The manual measurements of theTQTS that provided data for these tests were sub-ject to excessive quality control that was moremeticulous than in other conventional TQTS.Moreover, the adjustments made by the morpho-logical pattern-matching algorithm were, in mostinstances, completely beyond the precision ofcareful visual recognition (Fig. 7). This is becausean accurate identification of T-offset in surfaceECG is, to some extent, a convention, and to re-produce this convention from one ECG to an-other with nearly absolute systematicity is be-

yond human visual and manual means (as we al-ready discussed in Introduction, Fig. 1). Thus thepattern-matching algorithm improves the system-aticity of the measurements even when manualmeasurements are undergoing extensive qualitycontrol.

To some extent, it may be surprising thatwith these data, the differences between manualand automatic QT measurement were not evengreater. However, this was largely made possibleby the selection of ECG segments for measure-ment. Even in situations when the source long-term ECG recording was of poor quality, the bestpossible ECG segments were used based on sys-tematic noise contents evaluation.8 This also in-fluenced the intrasubject SDs of QTc values thatwould be difficult to obtain otherwise. Similar suc-cess of the automatic readings cannot be expectedin data that were not carefully preselected since

PACE, Vol. 32 January 2009 125

HNATKOVA, ET AL.

Figure 4. Statistical summaries of the intrasubject SD of QTc values shown in Figure 3. The fourparts of each panel correspond to the four investigated data sets; the open, shaded, and filledbars correspond to automatic, manual, and pattern-matching QT measurement; the top, middle,and bottom panels show SD of QTcI, QTcIf, and QTcF values (see the text for exact definitions).

126 January 2009 PACE, Vol. 32

SYSTEMATIC QT INTERVAL MEASUREMENT

Figure 5. Examples of QT/RR distributions in four subjects selected from the four different data sets. Each rowcorresponds to one subject and shows QT/RR distributions based on automatic QT measurements (left panels withopen marks), manual QT measurements (middle panels with light shaded marks), and pattern-matched and alignedQT measurements (right panels with dark shaded marks). Note that the QT/RR patterns are getting progressivelymore compact from left to right. (The examples taken from day-time and night-time recordings in Parkinson’s diseasepatients show two different subjects.)

the accuracy of automatic readings is substantiallyinfluenced by ECG quality.7 To a substantial ex-tent, the observed intrasubject SDs of QTc valuesalso depended on the strict stability of heart ratepreceding the ECG measurements. We know fromindependent observations that without this selec-tion criterion (as well as when using different RRinterval measurement such as those coming onlyfrom the measured 10-second ECG segment), theQTc variability would have been much larger.16

As far as we are aware, no other attempts havepreviously been made to adjust manual (or semi-automatic) measurements of QT interval with mor-phological classifications of ECG patterns. There-fore, we can only compare our secondary resultswith published literature. In this sense, our find-

ings agree well with previous reports. The meanvalues of QTc intervals found in individual sub-jects, as shown in Figure 5, are consistent withprevious population reports.17 Similarly, the ob-served differences between women and men andbetween day-time and night-time QTc interval val-ues agree with existing literature.18 We also repro-duced the known fact that irrespective of measure-ment quality, Fridericia correction leads to muchless stable QTc values than the individualized cor-rections. Because of all these consistencies withprevious publications, we can conclude that ourdata are sufficiently robust and broadly represen-tative and that similar findings should also be ex-pected in data obtained from other populationsand other TQTS.

PACE, Vol. 32 January 2009 127

HNATKOVA, ET AL.

Figure 6. Distribution of individual QTc values. The top two panels correspond to Healthysubjects A and Healthy subjects B and show, for each subject of the population, the mean QTcvalue and the intrasubject SD. The bottom panel corresponds to Parkinson’s disease patientsand shows a scatter diagram between day-time (vertical axis) and night-time (horizontal axis)intrasubject mean QTc values and their SD. (In all panels, QTcI values based on pattern-matchingmeasurement are shown—see the text for details.) Round shaded marks correspond to femalesubjects, open diamond marks to male subjects. Note that while the intrasubject SD (and thusthe intrasubject variability) of QTc values are narrow, the intersubject spread of the individualQTc values is substantial.

128 January 2009 PACE, Vol. 32

SYSTEMATIC QT INTERVAL MEASUREMENT

Figure 7. Examples of repositioning of the QT intervalmeasurements by the pattern-matching and alignmentalgorithm. Representative median beats of all 12 leadssuperimposed on the same isoelectric axis are shown(the two cases belong to different subjects). In both trac-ings, the dashed vertical lines show the original man-ual measurements, the bold vertical lines the reposi-tioned measurements. Note that in both cases, the ini-tial manual measurements were accurate and withoutproblems. However, they differed from measurementsof other morphologically close ECGs in the same sub-jects. Therefore, the adjustment algorithm repositionedthe measurement of the top and bottom tracings by −4ms and +7 ms, respectively.

Several limitations of the study also need to beacknowledged. Because of the technical difficultyof obtaining the Q-onset and T-offset triggers in therequired 1-kHz precision from the automatic mea-surements, we were unable to apply the morpho-logical pattern adjustment also to the automaticQT measurements. While it might be possible tospeculate that even such measurements would besubstantially improved by morphological pattern

adjustment, we are unable to confirm this by di-rect observations. Also, the adjustment algorithmis based on the assumption that all the initial mea-surements are “fairly correct.” Using data withvery substantial outliers (as is always the case withfully automatic measurement) might spoil the suc-cess of the pattern adjustment greatly. The intra-subject SDs of QTc values that we reported wererelated to repeated observations within one long-term recording. Although it is well recognized thatthe QTc profiles are well reproducible within eachsubject, some additional increase of the SD valuesmight be expected when studying ECG data ob-tained over substantially prolonged periods. Nev-ertheless, our experience suggests that such anadditional increase is truly minimal.19 In theseverifications of the adjustment algorithm, it wasapplied to datasets of ECG recording that con-tained a substantial number of ECG samples. Ourpresent experience suggests that such larger sam-ples are indeed needed for the success of the algo-rithm since an almost continuous spectrum of ECGmorphologies is needed for proper balance of theadjustment. (In TQTS, such an almost continuousspectrum of morphologies is recorded even whenthe investigated drug changes T wave morphologyas TQTS recordings are, as a rule, made at differ-ent plasma and tissue levels of the compound inquestion.) Nevertheless, we have not studied theminimum number of ECGs required by the algo-rithm and thus cannot comment on it. Additionalreduction of intrasubject QTc variability is pos-sible when combining individualized heart ratecorrection with individualized QT/RR hysteresiscorrection16 but we have not used it in this inves-tigation because the heart rate history data werenot systematically available for the QT intervalmeasurements.

Thus, in spite of these limitations, this in-vestigation permits one to conclude that morpho-logical pattern adjustment of QT interval mea-surements improves the quality of the QT datasignificantly and with substantial practical impli-cations. Noticeable reductions in intrasubject QTcvariability were reproduced in different popula-tions with different ECG quality. Therefore, thetechnology might be recommended for every thor-ough QT/QTc study in order to make it more ro-bust and more powerful. Substantial reductionsof necessary study size are likely achievable withthis measurement improvement.

References1. E14 Clinical evaluation of QT/QTc interval prolongation and proar-

rhythmic potential for non-antiarrhythmic drugs. Guidance to in-dustry. Fed Regist 2005; 70:61134–61135.

2. Sarapa N, Nickens DJ, Raber SR, Reynolds RR, Amantea MA. Riton-avir 100 mg does not cause QTc prolongation in healthy subjects:

A possible role as CYP3A inhibitor in thorough QTc studies. ClinPharmacol Ther 2008; 83:153–159.

3. Kubitza D, Mueck W, Becka M. Randomized, double-blind,crossover study to investigate the effect of rivaroxaban on QT-interval prolongation. Drug Safety 2008; 31:67–77.

PACE, Vol. 32 January 2009 129

HNATKOVA, ET AL.

4. Hulhoven R, Rosillon D, Letiexhe M, Meeus MA, Daoust A, StockisA. Levocetirizine does not prolong the QT/QTc interval in healthysubjects: Results from a thorough QT study. Eur J Clin Pharmacol2007; 63:1011–1017.

5. Malik M, Farbom P, Batchvarov V, Hnatkova K, Camm AJ. Relationbetween QT and RR intervals is highly individual among healthysubjects: Implications for heart rate correction of the QT interval.Heart 2002; 87:220–228.

6. Batchvarov VN, Ghuran A, Smetana P, Hnatkova K, Harries M,Dilaveris P, Camm AJ, et al. QT-RR relationship in healthy sub-jects exhibits substantial inter-subject variability and high intra-subject stability. Am J Physiol Heart Circ Physiol 2002; 282:H2356–H2363.

7. Hnatkova K, Gang Y, Batchvarov VN, Malik M. Precision of QT in-terval measurement by advanced electrocardiographic equipment.Pacing Clin Electrophysiol 2006; 29:1277–1284.

8. Batchvarov V, Hnatkova K, Malik M. Assessment of noise in digitalelectrocardiograms. Pacing Clin Electrophysiol 2002; 25:499–503.

9. Daskalov IK, Dotsinsky IA, Christov II. Developments in ECG acqui-sition, preprocessing, parameter measurement, and recording. IEEEEng Med Biol 1998; 17:50–58.

10. Meyer CR, Keiser HN. Electrocardiogram baseline noise estimationand removal using cubic splines and state-space computation tech-niques. Comput Biomed Res 1977; 10:459–470.

11. Pueyo E, Smetana P, Caminal P, de Luna AB, Malik M, Laguna P.Characterization of QT interval adaptation to RR interval changesand its use as a risk-stratifier of arrhythmic mortality in amiodarone-

treated survivors of acute myocardial infarction. IEEE Trans BiomedEng 2004; 51:1511–1520.

12. Malik M. Problems of heart rate correction in assessment of drug-induced QT interval prolongation. J Cardiovasc Electrophysiol2001; 12:411–420.

13. Fridericia LS. Die Systolendauer im Elekrokardiogramm bei nor-malen Menschen und bei Herzkranken. Acta Med Scand 1920;53:469–486.

14. Zhang J, Machado SG. Statistical issues including design and sam-ple size calculation in thorough QT/QTc studies. J Biopharm Stat2008; 18:451–467.

15. Zhang L, Dmitrienko A, Luta G. Sample size calculations in thor-ough QT studies. J Biopharm Stat 2008; 18:468–482.

16. Malik M, Hnatkova K, Smetana P, Schmidt A. Accurately measuredand heart-rate corrected QTc intervals do not show any day-timevariability. (abstract) Eur Heart J 2007; 28:624.

17. Rautaharju PM, Zhang ZM. Linearly scaled, rate-invariant normallimits for QT interval: Eight decades of incorrect application ofpower functions. J Cardiovasc Electrophysiol 2002; 13:1211–1218.

18. Extramiana F, Maison-Blanche P, Badilini F, Pinoteau J, DeseoT, Coumel P. Circadian modulation of QT rate dependence inhealthy volunteers: Gender and age differences. J Electrocardiol1999; 32:33–43.

19. Malik M, Andreas J-O, Hnatkova K, Hoeckendorff J, Cawello W,Middle M, Horstmann R, et al. Thorough QT/QTc study in patientswith advanced Parkinson’s disease—Cardiac safety of rotigotine.Clin Pharmacol Therap, in press.

AppendixMathematically, a 12-lead ECG is a real num-

ber function e : � → �12 assigning a vector of 12voltage values to each time instant. The similar-ity of ECG patterns defines “distances” betweendifferent morphologies that might need to be con-sidered within different regions of the ECG sig-nal. Let δI define a metric space on all ECG-typefunctions defined within a time interval I ⊂ �.That is, the following is valid for every e1, e2, e3defined within I : δI (e1, e3) ≤ δI (e1, e2) + δI (e2, e3),δI (e1, e2) = δI (e2, e1), and δI (e1, e1) = 0. (In the eval-uation reported here, the similarity metric was im-plemented using distances between 1 and correla-tion coefficients calculated in a single-dimensionrepresentation.)

Let S be a set of all measured ECGs in onegiven subject of the study. Let a measurementpoint (e.g., Q-onset or T-offset) in an ECG e ∈ Sbe manually placed at te. Further let fixed param-eters α and β define a “meaningful surrounding”of the measurement point Te = 〈te − α, te + β〉 (e.g.,in the evaluation reported here, T wave offset was

pattern matched within a region starting 100 msbefore and ending 50 ms after the T-offset mea-surement point while for QRS onset, these limitswere from –20 to +30 ms).

Further, for e, f ∈ S where e = f , let �ef be a

set of all [�,<] → [�,<] homorphisms Tf → Te.By searching �e

f the algorithm finds homorphismhe

f such that

�ef = δTe(e, f ◦ he

f ) = ming∈�e

f

δTe(e, f ◦ g).

Let denote �ef = te − he

f (t f ). Then, the patternmatching and adjustment algorithm changes themeasurement point te in e to

te −∑k∈S

�ek

�ek

/∑k∈S

1�e

k.

In the evaluation reported here, the reposi-tioned measurement points were rounded to thenearest millisecond.

130 January 2009 PACE, Vol. 32