Automated T Wave End Detection Methods

Comparison of Four Different Methods for T Wave End Detection

Jonathan Moeyersons

1,4

, Griet Goovaerts

1,4

, Suzy Huijghebaert, Bert Vandenberk

2,3

Rik Willems

2,3

and Sabine Van Huffel

1,4

KU Leuven, Department of Electrical Engineering (ESAT), STADIUS, Kasteelpark Arenberg 10, 3001 Leuven, Belgium

KU Leuven, Department of Cardiovascular Sciences, Herestraat 49, 3000 Leuven, Belgium

University Hospitals Leuven, Cardiology, Herestraat 49, 3000 Leuven, Belgium

Imec, 3001 Leuven, belgium

Keywords: T Wave End Detection, Electrocardiogram (ECG).

Abstract: T wave end detection is essential for electrocardiogram (ECG) processing and analysis. Several methods have

been proposed and tested, but an objective comparison is lacking. In this paper, four different (semi-)

automated methods are compared with the manually annotated T wave ends of the PhysioNet QT database.

The first method is a semi-automatic method, based on a template matching algorithm. The second method

uses the tangent of the steepest point of the descending limb of the T wave. The third and fourth method

perform a maximum area search of, respectively, a trapezium and the area under the curve. In order to evaluate

the accuracy and repeatability of the proposed algorithms, the mean and standard deviation (sd) of the

detection errors were computed. This was performed for leads I and II separately, after selection of the best

annotated T wave end per beat and after selection of the best lead. We demonstrated that the trapezium method

is the least repeatable of all methods tested (sd=29.7ms), whilst the integral method scores best in terms of

accuracy (mean=2.2ms). These findings were strengthened by the analysis of the generated Bland-Altman

plots, where the smallest bias was observed for the integral method (-1.89ms).

1 INTRODUCTION

The QT interval is an indirect measurement of the

time of the depolarization and repolarization of the

ventricular cells. Prolongation of this interval is

associated with the occurrence of lethal ventricular

arrhythmias in patients with the congenital long QT

syndrome (Schwartz and Wolf 1978)(Goldenberg et

al. 2008), in patients taking QT-prolonging non-

antiarrhythmic medication (De Ponti et al., 2002) and

even in the general population (Goldenberg et al.

2006). Therefore, accurate measurement of the QT

interval is of major importance.

Manual detection of the T wave end requires a

time consuming effort of the clinician. Unfortunately,

the great morphological variation in ECG signals

makes it hard to design an automated and widely

applicable algorithm (Manriquez and Zhang, 2007).

Whereas the QRS onset is easily detected, because of

its sharpness, it can be quite challenging to determine

the end of the T wave, since it gradually merges with

the baseline (Couderc and Zareba, 2005).

Furthermore, the presence of U waves might cause

additional difficulties. Large U waves, fused with the

T wave, should be included in the measurement, in

contrast to small and/or separate U waves which

should not be included (Vohra, 2007). During

exercise, the problems with T wave end detection are

even more distinct, since at fast heart rates, the T

wave might fuse with the following P wave (Chauhan

et al., 2002). These facts make it difficult to

automatically detect the end of the T wave.

Despite all these challenges, several algorithms

have been developed using different methodologies.

Since these different algorithms can differ in QT

interval measurement by 10 to 20ms it is important

that a correct detection method is selected (Panicker,

Karnad, Joshi, et al., 2009). This paper is the first to

compare four different (semi-)automated methods on

the same manually appointed T wave ends of the

PhysioNet QT database.

Moeyersons J., Goovaerts G., Huijghebaert S., Vandenberk B., Willems R. and Van Huffel S.

Automated T Wave End Detection Methods - Comparison of Four Different Methods for T Wave End Detection.

DOI: 10.5220/0006171700920098

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 92-98

ISBN: 978-989-758-212-7

The first method is a semi-automatic method,

based on a template matching algorithm in which the

user manually selects, hence the semi, the beginning

and the end of the QT interval on a template beat

(Berger et al., 1997). The second method determines

the end of the T wave using the tangent of the steepest

point of the descending limb of the T wave. The third

and fourth method perform a maximum area search

of, respectively, a trapezium (Vázquez-Seisdedos et

al., 2011) and the area under the curve (Qinghua

Zhang et al., 2006). In the following, we will briefly

summarize the four methods mentioned.

2 METHODS

2.1 Method 1: Semi-automatic (SEMI)

T Wave End Detection

The first step of this algorithm is the storage of all

beats, 0.35s before and 0.75s after each R peak.

Hereof the trimmed mean is calculated, excluding the

upper and lower 10% percent of the data, to create a

template representing the average beat. In order to

calculate the QT interval, the distinct features of the

template have to be detected. The template’s R peak

is known, since we segmented the signal based

hereon. The time location of the R peak is denoted 



Hereafter, the user manually selects the beginning

and the end of the QT interval. This segment is

denoted (n), where  is the sample number. Thus,



(

)

=

(



)





  = 



∶ 



(1)

where () is the ECG signal and 



and 



are

the manually selected beginning and end points of the

QT interval. The duration of the interval is samples.

For the purpose of matching all other beats to the

template, only the region of the template from =





+



 = 



is used, with 



equal to 50ms.

Per beat, an error function 



() is defined:





(



)

∑

[















(







)

−(



+)]²

(2)

where  is the time-stretching factor and 



is the

R peak under investigation. The result is the sum of

squared differences between the template T wave and

the stretched or compressed T wave for beat . A

progressive search in the interval [0.91.1] is

conducted in order to find the value of  that

minimizes 



(



)

. The best value of  is denoted 



and

the QT interval of the 

beat is defined as





=







(3)

with, as stated before,  equal to the duration of

the template QT interval. The 

T wave end is

defined by the sum of the 

QT interval and the

location of the according R peak, minus the length of

the template’s QR interval. In summary, the

algorithm finds the QT interval for each beat such that

the T wave shape best matches the template T wave

under the time-stretch model.

The working principle of this method is

demonstrated graphically in Figure 1. The top panel

presents the template beat. Here, the manually

selected beginning and end of the QT interval are

indicated. The region highlighted in bold is the

segment used to compute the error function. In the

next panel, several time-compressed versions of a

new beat’s T wave are superimposed on the template.

The area of difference between the template T wave

and the uncompressed T wave of the beat is then

calculated and the optimal value of  is determined.

Figure 1: Operator selects beginning and end points of the

QT interval from the template (top panel). For each of the

other beats in the epoch, multiple time-compressed or time-

stretched versions of the QT interval are generated (second

panel). These will be used for comparison with the template

QT interval to derive the optimal time-stretching factor .

2.2 Method 2: Tangent Method (TAN)

The first step of this algorithm is again the

segmentation of all beats, 0.35s before and 0.75s after

each R peak. Every beat will be processed and

analysed separately. The second step is an additional

cleaning of each beat using a cubic Savitzky-Golay

filter in order to remove high frequency noise.

Afterwards, the isoelectric baseline of the beat is

Automated T Wave End Detection Methods - Comparison of Four Different Methods for T Wave End Detection

aligned with the zero line by subtracting the median

of the first 320ms of the studied beat.

In the third step, a search window is defined in

which the T wave peak is selected. The left bound is

set at 60ms after the R peak, in order to exclude the

QRS complex, but to include the whole T wave. The

right bound is set in the interval between the

suspected end of the T wave and the next R peak. In

this search window the derivative is calculated,

followed by a detection of all sign changes. This

operation results in the location of all peaks and

valleys in the selection window. Subsequently, the T

peak is defined as the peak or valley with the maximal

absolute amplitude in the selection window, but

located maximally 850ms from the start of the beat.

Figure 2: Tangent Method. The red line represents the

tangent of the steepest point of the descending limb of the

T wave and the red dot represents the T wave end.

The fourth step is the selection of the T wave end.

First the steepest point of the descending or ascending

limb of the T wave is defined. Secondly, the tangent

through this point is calculated and the cross point

between this line and the zero baseline is searched.

Finally, since this cross point does not exactly align

with the T end, the point 20ms after the cross point is

selected to be the T end. The working principle of this

method is demonstrated graphically in Figure 2.

2.3 Method 3: Trapezium’s Area

(TRA) Method

The TRA method assumes that T peaks are previously

detected, following the described steps. First, a search

window is defined, which encloses the whole T wave.

Second, the search window is narrowed. The left and

right bound are replaced by the samples with the

maximal and minimal slope in the search window.

Finally, the first point with an absolute slope smaller

than 0.1 is selected and a maximum search of the

absolute values around this point is conducted. The

maximal is referred to as the T peak.

This T wave end detection method is based on the

calculation of consecutive areas of a rectangular

trapezium defined by three fixed and one variable

vertex.

The first fixed point is located 100ms past the T

peak in order to ensure the inclusion of the T wave

end (



, 



). The second fixed point is defined by

subtracting each value of a search window in between

the T peak and the first fixed point. A maximum

search is performed and the steepest point of the

descending limb of the T wave is selected (



, 



The third fixed point is the cross point between a

vertical line through the first and a horizontal line

through the second fixed point (



, 



). The variable

point starts at the second fixed point and follows the

graph, until it reaches the first fixed point (



, 



This is demonstrated graphically in Figure 3. The T

wave end is defined as the point where the area  of

the trapezium is maximal.  is calculated by the

following formula:



=0.5∗(



−



)∗(2



−



−



)

(4)

This means that the area  will be zero when the

variable point equals the second fixed point and

maximum when it is located at the end of the T wave.

Figure 3: Determination of the T wave end by the

computation of the area A of several trapezes. T end denotes

the maximum area.

2.4 Method 4: Integration (INT)

Operation

Based on the R peak detection for each beat, an

interval, [







], is delimited so that the T wave end

is inside this interval, with no overlap with the other

wave forms. The proposed algorithm mainly consists

of the computation of an indicator 



which reaches

its maximal value when the T end is detected. It is

computed through an integration operation in a

sliding window, with the window size smaller than

the length of the whole T wave. 



is computed by

the following formula:

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing





∑





−





(5)

where  is the sliding window size in discrete

time, 



is the signal value at time point  and ̅



is the

mean value of the signal in a small window around .

For each instance  inside the earlier defined interval,

the value of 



is computed and the T wave end is

located at the value of  maximizing 



3 EVALUATION WITH

MANUALLY ANNOTATED ECG

SIGNALS

The performance of the presented algorithms is

evaluated on the PhysioNet QT database. We

compared the performance of the different detection

methods with the manually annotated T wave ends in

this database.

3.1 The PhysioNet QT Database

The PhysioNet QT database has been designed to

serve as a reference for the validation and comparison

of T wave end detecting algorithms (Laguna et al.

1997). It contains 105 records of 15min two-lead

ECG signals and a total of 3944 T wave end

annotations. The annotations were performed

manually by two expert cardiologists. 3542 T wave

ends were annotated by the first cardiologist and 402

were annotated by the second cardiologist, in 11

recordings. At least 30 beats per record were

annotated, except for 2 records in which no T wave

ends were annotated. The signals are sampled at

250Hz.

Since each detection method uses different types

of filtering, their performance could depend on the

characteristics of the filters. To homogenize this

dependence, the pre-processing was generalized. All

signals were filtered with a zero-phase bandpass filter

between 1 and 40Hz, to correct for baseline drift and

high frequency noise.

3.2 Performance Comparison

In order to evaluate the accuracy and repeatability of

the proposed algorithms, the mean and standard

deviation (sd) of the detection errors, that is the time

difference in ms, between the manually and

automatically detected T wave ends, were computed

for the four different methods in the two ECG leads.

The mean and standard deviation of the errors

were computed as follows. First, the four algorithms

were applied to each of the two leads of an ECG

record in the QT database. Each T wave end is

annotated once by the cardiologists and located twice

by the four algorithms, once per ECG lead. Second,

the detection error is computed. Each manually

annotated T wave end was compared with the

corresponding four automatically annotated T wave

ends. For each lead, the mean detection error and

standard deviation per algorithm was computed.

Finally, the overall mean and standard deviation for

all ECG records were computed.

In Table 1 the results of the validation of the four

proposed algorithms are presented. In row 1 and 2 the

overall mean and sd of lead I and II are presented. The

mean value expresses how close the algorithms are to

the manually annotated markers (accuracy), and the

sd value provides information about the stability

(repeatability) of the detection criteria. The mean

values have to be interpreted with caution, since over-

and underestimation of the manually annotated T

wave ends cancel each other out. This might result in

a lower overall mean value.

Table 1: Comparison of the overall mean and standard

deviation (sd) of the differences, in ms, between the

automatic and manually annotated T wave ends for all

methods in both leads separately.

SEMI TAN TRA INT

Lead mean sd mean sd mean sd mean sd

I -6.2 15.4 -18.0 17.8 25.7 37.2 10.1 21.4

II -2.1 15.9 -5.4 18.0 19.7 36.1 14.9 29.1

The results of Table 1 show that, in terms of the

overall mean error and standard deviation, the TRA

method is outperformed by the other three algorithms,

when evaluated on the PhysioNet QT database.

Although obtained on the same database, this method

showed worse results, compared to the original paper

(Vázquez-Seisdedos et al., 2011). A first explanation

might be the exclusion of some beats in the original

paper due to poor quality of T wave end detection.

Eliminating cardiac cycles of poor quality generally

improves the results. Therefore, evaluating the

detection errors without removing the bad cardiac

cycles puts the proposed algorithms in the least

favourable evaluation condition, which might explain

the difference. The second explanation is the choice

of lead. In this paper, the algorithm is applied on both

leads separately, whilst in the original paper the a

posteriori best result among two computed positions

was chosen for error evaluation (best beat per record

(BB)). In clinical practice, the human operator could

choose the best lead for each patient individually

(best lead per record (BL)), but it is less reasonable to

Automated T Wave End Detection Methods - Comparison of Four Different Methods for T Wave End Detection

switch leads per cardiac cycle (Zhang et al., 2005). In

order to take this difference in lead selection into

account, the BB and BL values were also calculated.

The BB values were computed according to the

method adopted first in (Martínez et al., 2004) and

later in (Zhang et al. 2005; Vázquez-Seisdedos et al.,

2011). This method defines the T wave end per beat

by selecting the lead in which the detection error,

between the automatically and manually annotated T

wave end, is minimal. The BL method selects the

ECG lead which contains the most T wave ends,

appointed by the previously described method. If an

equal amount of T wave ends were appointed in both

leads, the first lead was selected. From the viewpoint

of a human operator, this is a more realistic

procedure. The results of both methods can be found

in Table 2.

Table 2: Comparison of the overall mean and standard

deviation (sd) of the differences, in ms, between the

automatic and manually annotated T wave ends for all

methods with the supplementary BB and BL protocol.

SEMI TAN TRA INT

Lead mean sd mean sd mean sd mean sd

BB -4.9 15.1 -8.0 16.3 11.8 29.7 2.2 20.0

BL -6.2 17.6 -7.6 19.0 14.3 37.5 3.9 22.9

When applying the BB protocol, it was observed

that the overall sd, obtained by each of the methods,

was lower compared to the sd obtained for lead I and

II. This was expected, since the lead in which the

detection error is minimal was selected per beat. This

protocol is most in accordance with the annotation

method of the cardiologists, since they made their

annotation by examining both leads and based their

decision on the best lead (Martínez et al., 2004).

In clinical practice, the best lead can be selected

after the ECG recorder is set up. Hence, the BL results

are clinically the most relevant, concerning everyday

T wave end detection. We demonstrated that the TRA

method is the least repeatable of all methods tested

(sd=37.5ms), whilst the SEMI method is the most

repeatable one (sd=17.6ms). The integral method

scores best in terms of accuracy (mean=3.9ms).

It might be noted that the mean and sd calculation

was simplified. One value was computed per record

and the overall mean and sd were computed as the

average of these values. This method does not take

the number of annotated T wave ends in each record

into account. Therefore, we opted to generate Bland-

Altman plots. These allow a direct comparison

between all manual annotations and the T wave end

selections of the four algorithms. Only the BB values

were taken into account, since this protocol is most in

accordance with the annotation method of the

cardiologists. Based on the Bland-Altman plots of the

respective QT intervals, Q being manually annotated

by the cardiologists, an evaluation of the agreement

of the methods was performed. The results of the

evaluation are depicted in Figure 4.

The comparison of the TRA method shows the

largest limits of agreement (-109.10/87.27ms). These

results strengthen the previous findings. In

accordance, the best agreement was determined for

the SEMI method (-75.69/85.01ms), although the

agreement of the INT method was only slightly worse

(-84.48/80.70ms). The obtained biases are in the

range of the ones earlier reported (Panicker, Karnad,

Natekar, et al. 2009; Vázquez-Seisdedos et al., 2011).

The results of the SEMI method could be

explained by the small influence of baseline wander

and U waves on the detection of the T wave end.

Because of their low amplitude, U waves have

significantly less influence on the sum of the squared

differences compared to the T waves. However, it

should be noted that the method is very operator

dependent. This is highlighted by the cluster forming

of the difference points in the Bland-Altman plot. All

QT intervals computed per record will be biased in

accordance to the difference in end point selection of

the template T wave end, compared to the manually

annotated T wave end. This results in a relatively

unaffected QT variability, but alters the QT lengths.

This operator dependency should be taken into

account when using this method in QT interval

analysis.

Besides the agreement intervals, the biggest

difference between the algorithms could be observed

for the cloud on the right. This cloud contains the

longest QT intervals, including biphasic T waves and

fusions with the U wave. Both the TAN and TRA

method were outperformed by the INT method for the

detection of the actual ends of these QT intervals.

Probably, this is due to the fact that the TAN and TRA

method rely on the detection of the T wave peak,

making it harder to detect more complex biphasic T

waves or fused T and U waves.

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing

Figure 4: Bland-Altman plots of the four T wave end

detection algorithms, compared to the manual annotated

beats. The dotted black line indicates the average bias, or

the average of the differences. The dotted red lines

represent the 95% limits of agreement.

3.3 Limitations

We annotated the T wave ends on the template beats

for the SEMI method ourselves. Better results could

have been obtained by annotation of the template by

a cardiologist. Also, the tangent and trapezium

method heavily rely on an accurate T peak detection.

This could not be fully guaranteed and might be an

additional cause for the large limits of agreement of

the TRA method. Finally, although the PhysioNet QT

database provides a large database of annotated beats,

it should be noted that it is not known which lead was

annotated. In this paper the BB approach is further

investigated, but it should be noted that this approach

cannot be applied in a clinical setting.

4 CONCLUSIONS

This paper is the first to compare four different

(semi-)automated methods on the same manually

appointed T wave ends of the PhysioNet QT database.

We demonstrated that, in terms of overall mean error

and standard deviation, the TRA method is

outperformed by the other algorithms. The SEMI and

INT methods perform approximately equivalent, but

the SEMI method is very operator dependent.

Therefore, the INT method is the preferred method.

As presented, an important difference remains

between automatically and manually annotated T

wave ends. This is probably due to the previously

mentioned morphological variation, U wave fusion

and omnipresent noise, which also impede manual

annotation.

ACKNOWLEDGEMENTS

RW is supported as a clinical researcher by the Fund

for Scientific Research Flanders (FWO).

SV: BOF KU Leuven: CoE #: PFV/10/002 (OPTEC),

SPARKLE #: IDO-10-0358, The effect of perinatal

stress on the later outcome in preterm babies #:

C24/15/036; FWO: project #: G.0869.12N (Tumor

imaging), G.0A5513N (Deep brain stimulation);

IWT: project #: TBM 110697-NeoGuard, SWT

150466-OSA+; iMinds Medical Information

Technologies: SBO2016; Belgian Federal Science

Policy Office: IUAP P7/19/ (DYSCO, ‘Dynamical

systems, control and optimization’, 2012-2017);

Belgian Foreign Affairs-Development Cooperation:

VLIR UOS programs (2013-2019); EU: European

Union's Seventh Framework Programme (FP7/2007-

Automated T Wave End Detection Methods - Comparison of Four Different Methods for T Wave End Detection

2013): EU MC ITN TRANSACT 2012, #316679,

ERASMUS EQR: Community service engineer ,

#539642-LLP-1-2013; The research leading to these

results has received funding from the European

Research Council under the European Union's

Seventh Framework Programme (FP7/2007-2013) /

ERC Advanced Grant: BIOTENSORS (n° 339804).

This paper reflects only the authors' views and the

Union is not liable for any use that may be made of

the contained information. (Tumor imaging)

G.0A5513N (Deep brain stimulation);

REFERENCES

Berger, R.D. et al., 1997. Beat-to-Beat QT Interval

Variability. Circulation, 96(5).

Chauhan, V.S. et al., 2002. Sex differences in QTc interval

and QT dispersion: Dynamics during exercise and

recovery in healthy subjects. American Heart Journal,

144(5), pp.858–864.

Couderc, J.-P. & Zareba, W., 2005. Assessment of

Ventricular Repolarization From Body-Surface ECGs

in Humans. In Cardiac Safety of Noncardiac Drugs.

Totowa, NJ: Humana Press, pp. 107–129.

Goldenberg, I., Moss, A.J. & Zareba, W., 2006. QT

interval: How to measure it and what is “normal.”

Journal of Cardiovascular Electrophysiology, 17(3),

pp.333–336.

Goldenberg, I., Zareba, W. & Moss, A.J., 2008. Long QT

Syndrome. Current Problems in Cardiology, 33(11),

pp.629–694.

Laguna, P. et al., 1997. A database for evaluation of

algorithms for measurement of QT and\nother

waveform intervals in the ECG. Computers in

Cardiology 1997, 24, pp.673–676.

Manriquez, A.I. & Zhang, Q., 2007. An algorithm for QRS

onset and offset detection in single lead

electrocardiogram records. Annual International

Conference of the IEEE Engineering in Medicine and

Biology - Proceedings, (2), pp.541–544.

Martínez, J.P. et al., 2004. A wavelet-based ECG

delineator: evaluation on standard databases. IEEE

transactions on bio-medical engineering, 51(4),

pp.570–81.

Panicker, G.K., Karnad, D.R., Natekar, M., et al., 2009.

Intra- and interreader variability in QT interval

measurement by tangent and threshold methods in a

central electrocardiogram laboratory. Journal of

Electrocardiology, 42(4), pp.348–352.

Panicker, G.K., Karnad, D.R., Joshi, R., et al., 2009. Z-

score for benchmarking reader competence in a central

ECG laboratory. Annals of Noninvasive

Electrocardiology, 14(1), pp.19–25.

De Ponti, F. et al., 2002. Safety of non-antiarrhythmic drugs

that prolong the QT interval or induce torsade de

pointes: an overview. Drug safety, 25(4), pp.263–86.

Qinghua Zhang et al., 2006. An Algorithm for Robust and

Efficient Location of T-Wave Ends in

Electrocardiograms. IEEE Transactions on Biomedical

Engineering, 53(12), pp.2544–2552.

Schwartz, P.J. & Wolf, S., 1978. QT interval prolongation

as predictor of sudden death in patients with myocardial

infarction. Circulation, 57(6), pp.1074–1077.

Vázquez-Seisdedos, C.R. et al., 2011. New approach for T-

wave end detection on electrocardiogram: performance in

noisy conditions. Biomedical engineering online, 10, p.77.

Vohra, J., 2007. The Long QT Syndrome. Heart Lung and

Circulation, 16(SUPPL. 3).

Zhang, Q. et al., 2005. Robust and efficient location of T

wave ends in Electrocardiogram. Computing in

Cardiology (CinC), 32, pp.711–714.

BIOSIGNALS 2017 - 10th International Conference on Bio-inspired Systems and Signal Processing