Precision of QT Interval Measurement by Advanced Electrocardiographic Equipment

COMPUTER COLUMN

Precision of QT Interval Measurement by AdvancedElectrocardiographic EquipmentKATERINA HNATKOVA, PH.D.,* YI GANG, M.D., PH.D.,*VELISLAV N. BATCHVAROV, M.D.,* and MAREK MALIK, PH.D., M.D., DS.C.*, †From the *Department of Cardiac and Vascular Sciences, St. George’s University of London and †St. Paul’s CardiacElectrophysiology, London, England

The costs of clinical investigations of drug-induced QT interval prolongation are mainly related tomanual processing of electrocardiographic (ECG) recordings. Potentially, however, these costs can bedecreased by automatic ECG measurement. To investigate the improvements in measurement accuracyof the modern ECG equipment, this study investigated QT interval measurement by the “old” and “new”versions of the 12SL ECG algorithm by GE Healthcare (Milwaukee, WI, USA) and compared the results tocarefully validated and reconciled manual measurements. The investigation used two sets (A and B) ofECG recordings that originated from large clinical studies. Sets A and B consisted of 15,194, and 29,86610-second ECG recordings, respectively. All the recordings were obtained with GE Healthcare recordersand were available in digital format compatible with ECG processing software by GE Healthcare. The twosets of recordings differed significantly in ECG quality with set B being substantially more noise polluted.Compared to careful manual QT interval readings in recording set A, the errors of the automatic QTinterval measurement were (mean ± SD) +3.95 ± 5.50 ms, and +0.51 ± 12.41 ms for the “new” and “old”12SL algorithm, respectively. In recording set B, these numbers were +2.41 ± 9.47 ms, and –0.17 ± 14.89ms, respectively (both differences were highly statistically significant, P < 0.000001). In recording set A,95.9% and 76.6% of ECGs were measured automatically within 10 ms of the manual measurement by the“new” and “old” versions of the 12SL algorithm, In recording set B, these numbers were 83.9% and 59.5%.The errors made by the “new” and “old” version of 12SL algorithm were practically independent each ofthe other (correlation coefficients of 0.031 and 0.281 in recording sets A and B, respectively). The studyshows that (a) compared to the “old” version of the 12SL algorithm, the QT interval measurement by the“new” version implemented in the most recent ECG equipment by GE Healthcare is significantly better,and (b) the precision of automatic measurement by the 12SL algorithm is substantially dependent on thequality of processed ECG recordings. The improved accuracy of the “new” 12SL algorithm makes it feasibleto use modern ECG equipment without any manual intervention in selected parts of drug-developmentprogram. (PACE 2006; 29:1277–1284)

12SL algorithm, manual ECG measurement, drug-induced QT changes

IntroductionThe necessity of investigating every new phar-

maceutical compound in terms of its propensityto alter myocardial repolarization and to prolongQT/QTc interval1,2 leads to additional drug devel-opment expenditure. The costs of QT/QTc orien-tated clinical studies are not negligible and a sub-stantial part of these costs relates to electrocardio-gram (ECG) processing and measurement. It is gen-erally believed that all QT/QTc studies require allECG intervals to be measured under visual controland manual verification.2 The need for such an ar-

Address for reprints: Dr. Katerina Hnatkova, Department of Car-diac and Vascular Sciences, St. George’s University of London,Cranmer Terrace, London SW17 0RE, England. Fax: +44-20-8725-0846s; e-mail: [email protected]

Received March 28, 2006; revised June 13, 2006; accepted July14, 2006.

rangement results from a number of publicationsthat previously reported substantial discrepanciesbetween manual and automatic QT interval mea-surement.3–5 At the same time, this very need isthe major contributor to the high costs of QT/QTcorientated clinical studies.

The previously published observations ofpoor precision of automatic QT interval mea-surement by electrocardiographic equipment havemostly, if not exclusively, been based on older al-gorithms that were developed already some timeago. The manufacturers of ECG machines recog-nize the need of improving these algorithms andthe evolving advances in electronic componentsmake it possible to increase the computationalpower of standard electrocardiographs. The timehas therefore come to reevaluate the performanceof new commercial algorithms for QT interval mea-surement and to investigate whether the claims of

C©2006, The Authors. Journal compilation C©2006, Blackwell Publishing, Inc.

PACE, Vol. 29 November 2006 1277

HNATKOVA, ET AL.

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

00001000100101

Noise product [technical units]

Cum

ulat

ive

freq

uenc

y

Set B

Set A

Figure 1. Noise distribution in the two sets of evaluatedECGs. Each tracing was characterized by the product ofSD and RMSSD noise (in technical units—see the textfor details). The graph shows the cumulative distribu-tion of these noise product values (the higher the value,the greater the noise contents in the tracing). The figureshows that while the two sets did not differ in the qual-ity of the “best” tracings, set B was substantially morenoise polluted than set A. Note the logarithmic scale ofthe vertical axis.

increased precision often made by manufacturersof electrocardiographic equipment are justified.

Having this in mind, this study was designedto compare the performance of the “old” and“new” versions of QT interval measurement al-gorithms incorporated in the 12SL ECG analyticalsoftware by GE Healthcare (Milwaukee, WI, USA).

MethodsEvaluated Sets of ECG Data

For the purposes of this investigation, two dif-ferent sets of digital ECGs were available that werepreviously recorded as a part of two large clinicalstudies. Both clinical studies were conducted inhealthy subjects of both sexes and involved ECGrecordings uninfluenced by any drug treatment aswell as influenced by drugs that are known to pro-long QT interval. Since the objective of this evalu-ation was to investigate the precision of ECG pro-cessing, the character and details of the sourceclinical investigations are of no importance.

In this text, the two series of ECG data willbe denoted as set A and set B. These sets A andB consisted of 15,194, and 29,866 ECG tracings. Inboth sets, the ECGs were obtained using equipmentby GE Healthcare.

These two particular sets of ECGs were se-lected since they differed substantially in the qual-ity of the ECG tracings. Using previously publishedtechnology,6 standard deviation (SD) and rootmean square of successive differences (RMSSD)noise were calculated in all these ECGs and in each

tracing, the overall noise was characterized by theproduct of SD and RMSSD values. The compari-son of the noise in the two sets of data is shown inFigure 1. The differences in the noise contents be-tween the two sets were highly statistically signif-icant (P-value below the resolution of 32-bit arith-metic) with set B being polluted by substantiallyhigher noise and poorer data quality.

Manual QT Interval Measurement

In each ECG of both sets, QT interval was mea-sured manually. The same procedure was appliedto both sets. Individual QRS complexes were rec-ognized by a pattern-matching algorithm7–9 andtheir positions verified visually on a computerscreen. From the QRS positions, representativebeat waveform was constructed by calculatingsample-by-sample median voltage separately in all12 leads after baseline wander removal.10 The rep-resentative waveforms of individual leads weresuperimposed on the same isoelectric axis and dis-played on a computer screen together with com-puter generated suggestions of Q onset and T off-set triggers. These suggestions were calculated us-ing pattern matching and classification algorithms.This on-screen image was used by two medicallyqualified observers with sufficient electrocardio-graphic experience who were operating indepen-dently of each other. Each observer was requiredto decide whether the QRS-T/U image was mea-surable in terms of QT interval and if yes, to repo-sition the Q-onset and T-offset triggers so that theywould best represent the QT interval duration inthe given tracing. After all the images were pro-cessed by both observers, those tracings were iden-tified in which the observers agreed on the QT in-terval measurability, and for which (if measurable)the observers provided QT interval measurementdiffering by ≤10 ms. In these ECGs, the positions ofthe Q-onset and T-offset triggers provided by twoobservers were averaged. Other ECGs, i.e., thosein which the observers did not agree on the QTinterval measurability or for which they returnedQT interval intervals differing by >10 ms, were re-turned to the same observers for a second indepen-dent measurement. Those ECGs for which the ob-servers still did not agree (using the same criteria ofagreement) after the second independent readingwere reconciled by a senior cardiologist who hadthe previous measurements by the initial observersavailable and who provided the final measurabil-ity decision and, where appropriate, the final QTinterval measurement.

Automatic QT Interval Measurement

All the ECGs of both sets were processedwith the new research version of ECG researchworkstation (Magellan package, GE Healthcare).

1278 November 2006 PACE, Vol. 29

PRECISION OF AUTOMATIC QT MEASUREMENT

The processing was repeated twice. During thefirst processing, the workstation package was pro-grammed to provide the so-called “matrix” QT in-terval measurement that corresponded to the pre-vious version of 12SL algorithm (called here “old”12SL). During the second processing, the work-station package was programmed to provide the“full 12SL” QT interval measurement that corre-sponded to the most recent version of the 12SL al-gorithm (called here “new” 12SL) as presently im-plemented in the most recent electrocardiographicrecorders by GE Healthcare.

Both “old” and “new” algorithms form a T-wave detection function based on vector magni-tude signal from eight independent lead inputs.Specifically, the “old” system forms a “superlead”of the T wave, which is the sum of the absolute val-ues of leads I, II, and V1–V6. An initial estimate ofthe T offset is made based on the point where thedownslope (after the peak of the T wave) becomesless than one quarter of the maximum downslope.Subsequent refining part of the algorithm consid-ers the sum of slopes of all leads. If the sum inthe area immediately preceding the initial start-ing point is greater than a proprietary numericalvalue, T offset is extended until the individualslopes level out. The “new” T offset algorithm alsoforms a “superlead” of the T wave but also createsa “superlead” of the T-wave slope. The algorithmfirst identifies the final peak or valley (negative Twave) of the T wave. Subsequently, it repositionsthe T offset following the contribution of the “su-per slope” window to the overall T-wave segmentuntil the ratio of the contribution to the originalsignal drops under a proprietary value or the “su-per slope” changes its direction (nadir or peak ofT-U, T-P).

The analysis of the study was made indepen-dently of GE Healthcare. Specifically, the equip-ment and software manufacturer had neither any

'new

' 12S

L

'old

' 12S

L

'new

' 12S

L

'old

' 12S

L

Figure 2. Examples of cases that were found nonmeasurable during manual processing and that were processedby both 12SL algorithms. In both panels, the representative waveforms of P-QRS-T/U complex of all 12 leads aresuperimposed together with vertical lines that show measurement triggers (black fine line corresponds to Q waveonset, the bold grey lines show T-wave offset determined by the 12SL algorithm).

access to the ECG data nor any influence onthe conduct of clinical studies from which thererecordings originated.

Measurement Comparison and Statistics

For both sets, the automatic measurements ofQT interval by the “old” and “new” 12SL algo-rithms were compared with the manual measure-ment. Errors of automatic measurement by both al-gorithms were evaluated using the Bland-Altmanstatistics.11 Absolute values of the errors were fur-ther compared using Wilcoxon matched pair test.The errors of “old” and “new” 12SL measurementswere compared using the Kolmogorov-Smirnovtest and by evaluating scatter diagrams of the er-rors of both 12SL algorithms measuring the sameECGs.

ResultsManual Measurement and NonmeasurableRecordings

In recording sets A and B, 58 tracings (0.38%)and 689 tracings (2.31%), respectively, were foundnonmeasurable during manual processing. Theseincluded ECGs that were too noise polluted andECGs that showed T-wave patterns too flat and/orboth flat and fusing with U wave patterns so that T-wave offset could not have been determined witha sufficient precision. In all other ECGs, QT inter-val was determined manually using the describedmeasurement strategy.

All the ECGs that were found manually non-measurable were measured by both automatic al-gorithms (there is no measurable/nonmeasurabledistinction made in the 12SL algorithm). These in-cluded not only cases of extreme noise pollutionbut also cases in which the T wave were visuallyjudged too flat and/or nondistinguishable from Uwaves (Fig. 2).


HNATKOVA, ET AL.

'new

' 12S

L

'old

' 12

SL

man

ual

'new

' 12S

L

'old

' 12S

L

man

ual

'new

' 12S

L

'old

' 12S

L

man

ual

'new

' 12S

L

'old

' 12S

L

man

ual

Figure 3. Example of “extreme” errors in which the “new” 12SL algorithm differed by more than 100 ms from manualQT interval measurement. In all panels, the representative waveforms of P-QRS-T/U complex of all 12 leads aresuperimposed together with vertical lines that show measurement triggers (black fine lines corresponds to Q wave onsetand to manual measurement of T-wave offset, the bold grey lines show T-wave offset determined by the 12SL algorithm).The two top panels show examples of the very rare cases in which the 12SL algorithm worked unsatisfactorily inrecordings that were relatively noise free. The two bottom panels show examples in which the errors of automaticprocessing were caused by extreme noise in one of the ECG leads (fine grey lines in the superimposed images of therepresentative waveform) that already influenced QRS detection by the 12SL algorithm.

Extreme Errors of Automatic Measurement

“Extreme” errors were defined as cases inwhich automatic measurement differed from themanual measurement by >100 ms. In recordingset A, there were four (0.026%) and one (0.006%)such cases with the “new” and “old” 12SL algo-rithms, respectively. In recording set B, these num-bers were 21 (0.070%) and 20 (0.067%), respec-tively.

In a vast majority of cases, these extreme er-rors occurred in tracings that had at least one leadhighly noise polluted (e.g., in cases where it ap-peared that an electrode must have been discon-nected) and in which the 12SL algorithm madenoise-related errors in determining the positionof individual QRS complexes. Cases in which ex-treme errors occurred in ECGs that were relativelyclean and noise free were very rare (below 0.02%in both data sets). Examples of extreme errors areshown in Figure 3.

Statistical Evaluation of Measurement Errors

In recording set A, the errors of the auto-matic QT interval measurement were (mean ± SD)+3.95 ± 5.50 ms and +0.51 ± 12.41 ms for the“new” and “old” 12SL algorithm, respectively.In recording set B, these numbers were +2.41± 9.47 ms and −0.17 ± 14.89 ms, respectively.Corresponding Bland-Altman plots are shown inFigure 4.

In both sets, the absolute errors of the “new”12SL algorithm were highly statistically signifi-cantly lower than the errors of the “old” 12SL al-gorithm (Z statistics of 51.83 and 73.25 for sets Aand B, respectively, P <10−6 for both).

Cumulative distributions of absolute errors(i.e., differences from true manual QT intervalmeasurement regardless of their sign) are shownin Figure 5. The exact percentages of tracingsthat were within measurement errors of 5, 10,and 15 ms are shown in Table I. In both ECG



ECG Set A

-100

-80

-60

-40

-20

0

20

40

60

80

100

275 300 325 350 375 400 425 450 475 500 525

Manually measured QT interval [ms]

Err

or o

f 'ol

d' 1

2SL

QT

mea

sure

men

t [m

s]

ECG Set A

-100

-80

-60

-40

-20

0

20

40

60

80

100

275 300 325 350 375 400 425 450 475 500 525


Err

or o

f 'ne

w' 1

2SL

QT

mea

sure

men

t [m

s]

ECG Set B

-100

-80

-60

-40

-20

0

20

40

60

80

100

275 300 325 350 375 400 425 450 475 500 525


Err

or o

f 'ol

d' 1

2SL

QT

mea

sure

men

t [m

s]

ECG Set B

-100

-80

-60

-40

-20

0

20

40

60

80

100

275 300 325 350 375 400 425 450 475 500 525


Err

or o

f 'ne

w' 1

2SL

QT

mea

sure

men

t [m

s]

Figure 4. Bland-Altman scatter diagrams showing errors by automatic QT interval measurement versus true manuallymeasured QT interval values. The top and bottom panels correspond to ECG sets A and B, respectively; the left and rightpanels show errors by the “old” and “new” 12SL algorithm, respectively. In each panel, individual marks correspondto separate ECG tracings, the grey bold lines show average error of the automatic measurement, and the dashed boldblack lines show a band of ±1 SD of the measurement error. Note the obvious differences between the “old” and “new”12SL algorithm. Note also that the “extreme” errors exceeding 100 ms are not shown in the scatter diagrams.

ECG Set A

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

0 5 10 15 20 25 30 35 40 45 50

Absolute error of automatic QT measurement [ms]

Cum

ulat

ive

freq

uenc

y

new 12SLold 12SL

ECG Set B

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

0 5 10 15 20 25 30 35 40 45 50

Absolute error of automatic QT measurement [ms]

Cum

ulat

ive

freq

uenc

y

new 12SLold 12SL

Figure 5. Cumulative distributions of absolute errors of automatic QT interval measurement. The step-lines in thegraphs show in what percentage of the ECG tracings was the absolute error of automatic QT interval measurementbelow or equal to the threshold given on the horizontal axis. In each panel, the bold black graph corresponds to the“new” 12SL measurement while the bold grey graph corresponds to the “old” 12SL measurement. The left panel showsresults obtained with the ECG set A, the right panel shows results obtained with ECG set B.


HNATKOVA, ET AL.

Table I.

Percentages of ECG Tracings with Successful AutomaticQT Interval Measurement

ECG Set A ECG Set BAbsoluteMeasurement “New” “Old” “New” “Old”Error 12SL 12SL 12SL 12SL

≤5 ms 73.7 47.8 54.4 33.5≤10 ms 95.9 76.6 83.9 59.5≤15 ms 99.3 91.7 94.0 77.3

The table shows percentages of ECG tracings in which the errorof automatic QT interval measurement was below the giventhreshold (e.g., in ECGs set A, while the “new” 12SLmeasurement algorithm was within 10 ms of the manualmeasurement in 95.9% of ECGs, this was the case with only76.6% of the ECGs with the “old” 12SL measurement algorithm).

sets, the differences in the error distributions werehighly statistically significant (P < 0.00001 in bothcases).

Finally, Figure 6 shows that the “new” 12SLmeasurement algorithm is truly new, i.e., that theerrors made by this algorithm are not only smallerbut also practically independent of the errors madeby the “old” 12SL algorithm. The correlation coef-ficients between the errors of the “new” and “old”12SL algorithms were 0.031 in ECG set A, and0.281 in ECG set B.

DiscussionThe study shows clearly (a) that the QT in-

terval measurement by the “new” 12SL algorithm

ECG set A

-100

-80

-60

-40

-20

0

20

40

60

80

100

-100 -80 -60 -40 -20 0 20 40 60 80 100

Error of 'new' 12SL QT measurement [ms]

Err

or o

f 'ol

d' 1

2SL

QT

mea

sure

men

t [m

s]

ECG set B

-100

-80

-60

-40

-20

0

20

40

60

80

100

-100 -80 -60 -40 -20 0 20 40 60 80 100

Error of 'new' 12SL QT measurement [ms]

Err

or o

f 'ol

d' 1

2SL

QT

mea

sure

men

t [m

s]

Figure 6. Scatter diagrams between the errors of QT interval measurement made by the “new”and “old” 12SL algorithms. The left panel corresponds to ECG set A, the right panel to ECGs ofset B. Note that the “extreme” errors exceeding 100 ms are not shown in the scatter diagrams.

implemented in the most recent electrocardio-graphic equipment by GE Healthcare is highlysignificantly better that the QT interval measure-ment by the previous “old” version of this al-gorithm, and (b) that the precision of automaticmeasurement by the 12SL algorithm is substan-tially dependent on the quality of processed ECGrecordings.

The first observation provides justification ofthe claims by the manufacturer that their new ap-proach to QT interval measurement substantiallyimproves previous versions of ECG analysis.

The second observation is not surprising andwas expected.12 In the presence of substantialnoise and artefact, automatic ECG processing al-gorithms implemented in the recording equipmentclearly cannot perform as precisely as when pro-cessing clean data with properly shaped and welldefined ECG patterns. GE Healthcare has recentlyintroduced a method for the measurement andtrending of signal quality, which has previouslybeen correlated with the quality of computerizedECG interpretation.13 However, this capability wasnot evaluated in this study.

Both observations as well as the details re-ported in this study have direct practical im-plications for the investigations of drug-inducedQT/QTc changes. Considering the precision re-quired to conduct QT/QTc oriented study effec-tively,14 the occasional errors of automatic QTmeasurement make it questionable whether such“thorough” studies2 can presently be supportedby fully automatic ECG measurement without anymanual intervention. Likewise, although very rare,large errors in automatic QT interval measurementexist and thus the fully automatic measurementsare unlikely suitable for registries of drug-induced



changes in which every singular large error mighthave serious implications.

Nevertheless, before the truly “thorough”QT/QTc studies are conducted in drug devel-opment program, “not-so-thorough” studies arefrequently conducted with the aim of approxi-mating drug-induced QT/QTc changes and elim-inating compounds with large repolarization ef-fects with which the rather expensive “thorough”QT/QTc study would only lead to waste of drug-development resources. The precision of such“not-so-thorough” studies clearly does not needto be as tight as that of the ultimate “thorough”studies since only the compounds that have a pro-found effect on cardiac repolarization need to beidentified. The results of the present study sug-gest that in such a scenario, fully automatic ECGmeasurement may have a valuable role especiallyif good quality ECG recordings can be guaran-teed, e.g., when using in-house or other clinicalunits under direct influence by drug developers.Although, we were unable to investigate repeata-bility of QT measurement errors in serial ECGs,our anecdotal experience suggests that gross mea-surement errors are not reproduced within se-rial ECG tracings when “new” 12SL algorithmis used (which was not the case with the “old”12SL algorithm). The “not-so-thorough” QT/QTcstudies should therefore benefit substantially fromthe analysis of a large multitude of ECG tracings.While with manual ECG measurement such an ap-proach might not be economical, once the auto-matic measurement can be obtained practically forfree for each single ECG, analysis of huge numbersof recordings (e.g., many thousands per subject)becomes a viable option especially if their acqui-sition may also be made inexpensive, e.g., whenusing continuous monitors and 12-lead Holtertechnology.

While reporting occasional inaccuracies of QTinterval measurement by modern electrocardio-graphs, technological restrictions of manufactur-ing the electrocardiographic equipment need to beremembered. Modern digital electrocardiographsare produced to serve clinical practice rather thandrug-development investigations. When observ-ing abnormal QT/QTc interval in clinical practice,ECG recording is repeated and/or remeasured be-fore any clinical decisions are made. In a vast ma-jority of cases, the precision within less than ±20ms as achieved by the “new” 12SL algorithm ingood quality ECG tracings (see top left panel of Fig.4) is fully adequate for clinical practice. Our in-house experience suggests that automatic QT inter-val measurement can be further substantially im-proved, e.g., with pattern classification algorithms,if all necessary computational power is available.

However, it would be entirely unrealistic to ex-pect each modern electrocardiograph to have asmuch computational power and data capacity as atop range personal computer, and to spend consid-erable time on the processing of each single ECGrecording. In clinical applications, which are thetrue domain of the equipment, such a technologi-cal advantage would just be wasted.

Limitations of our investigation also need tobe acknowledged. The advances in the automaticQT interval measurement that we report concernonly the equipment by one manufacturer. Al-though GE Healthcare is one of the leading man-ufacturers of electrocardiographic equipment andthus suitably placed to develop robust and ac-curate new technologies, it is possible to expectthat the experience and know-how of other seriousmanufacturers would result in similar improve-ments. However, we are not aware of any similarlarge validation study for automated QT intervalrecognition using other equipment and commer-cial software. Furthermore, the high accuracy ofour manual measurements was achieved by a verylabor intensive approach. Such accuracy wouldbe unlikely achieved with standard manual pro-cessing in large commercial research organizationsthat deal constantly with large volume of ECG dataand do not give as much attention to detail andrepeated quality checks as was the case in oursetting. Finally, we have intentionally restrictedthis study to technical comparison of differentmeasurement approaches. The implications of au-tomatic measurement techniques for an accuratedetection of drug-induced QT/QTc changes wereoutside the scope of this methodological study.Nevertheless, we have not seen any difference inthe improvement from the “old” to the “new” 12SLalgorithm among ECGs that were not influenced byany treatment and ECGs that were obtained whilethe recorded subject were under the influence ofQT prolonging pharmaceuticals (details not pre-sented here).

In spite of all these limitations, the study per-mits to conclude that recent advances in digitalECG processing within the recording equipmentmade substantial impact on the precision of QTinterval measurement. The accuracy of the “new”12SL algorithm is not only much greater than theaccuracy previously observed with QT intervalmeasurement by the “old” 12SL algorithms, it alsomakes it feasible to use the modern equipmentwithout any manual intervention in carefully se-lected parts of drug-development program.

Acknowledgment: The authors are thankful to Dr. JoelXue for providing the descriptive details on the implementa-tion of the proprietary 12SL algorithm.


HNATKOVA, ET AL.

References1. Fenichel RR, Malik M, Antzelevitch C, Sanguinetti M, Roden DM,

Priori SG, Ruskin JN, et al. Drug-induced torsades de pointes andimplications for drug development. J Cardiovasc Electrophysiol2004; 15:475–495.

2. Guidance for industry: E14 Clinical evaluation of QT/QTc intervalprolongation and proarrhythmic potential for non-antiarrhythmicdrugs. www.fda.gov/ohrms/dockets/98fr/2004d-0377-gdl0002.pdf(accessed 20 March 2006)

3. Malik M. Errors and misconceptions in ECG measurement used forthe detection of drug induced QT interval prolongation. J Electro-cardiol 2004; 37(Suppl.):25–33.

4. Savelieva I, Gang Y, Guo X, Hnatkova K, Malik M. Agreementand reproducibility of automatic versus manual measurementof QT interval and QT dispersion. Am J Cardiol 1998; 81:471–477.

5. McLaughlin NB, Campbell RW, Murray A. Comparison of auto-matic QT measurement techniques in the normal 12 lead electro-cardiogram. Br Heart J 1995; 74:84–89.

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

7. Malik M, Kautzner J, Hnatkova K, Camm AJ. Identification ofelectrocardiographic patterns. Pacing Clin Electrophysiol 1996;19:245–251.

8. Daskalov IK, Christov II. Electrocardiogram signal preprocessingfor automatic detection of QRS boundaries. Med Eng Phys 1999;21:37–44.

9. Kohler BU, Henning C, Orglmeister R. The principles of softwareQRS detection—Reviewing and comparing algorithms for detectingthis important ECG waveform. IEEE Eng Med Biol Mag 2002; 21:42–57.

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

11. Bland JM, Altman DG. Statistical methods for assessing agreementbetween two methods of clinical measurement. Lancet 1986;1:307–310.

12. Willems JL, Zywietz CHR, Arnaud P, van Bemmel JH, Degani R,Macfarlane PW. Influence of noise on wave boundary recogni-tion by ECG measurement programs—Recommendations for pre-processing. Comput Biomed Res 1987; 20:543–562.

13. Farrell RM, Young BJ. Effect of lead quality on computerized ECGinterpretation. Comput Cardiol 2004; 31:173–176.

14. Malik M, Hnatkova K, Batchvarov V, Gang Y, Smetana P, CammAJ. Sample size, power calculations, and their implications for thecost of thorough studies of drug induced QT interval prolongation.Pacing Clin Electrophysiol 2004; 27:1659–1669.


Precision of QT Interval Measurement by Advanced Electrocardiographic Equipment

Documents

Transcript of Precision of QT Interval Measurement by Advanced Electrocardiographic Equipment