Multichannel QRS Morphology Clustering

Data Preprocessing for Ultra-High-Frequency ECG Analysis

Filip Plesinger

, Juraj Jurco

, Josef Halamek

, Pavel Leinveber

Tereza Reichlova

and Pavel Jurak

Institute of Scientific Instruments of the Czech Academy of Sciences, Brno, Czech Republic

International Clinical Research Center at St. Anne’s University Hospital, Brno, Czech Republic

Keywords: ECG, SAECG, QRS, Ultra-High-Frequency, Clustering, Multi-thread, Ventricle Dyssynchrony.

Abstract: Ultra-high-frequency ECG (UHF-ECG) in a range of 500–1,000 Hz has been tested as a new information

source for analysis of left-ventricle dyssynchrony and other myocardial abnormalities. The power of UHF

signals is extremely low, for which reason an averaging technique is used to improve signal-to-noise ratio.

Since ventricle dyssynchrony is different for various QRS complex types, the detected QRS complexes must

be clustered into morphology groups prior to averaging. Here, we present a fully-automated method for

clustering. The first goal of the method is to separate previously detected QRS complexes into different

morphology groups. The second goal is to precisely fit the QRS annotation marks to the exact same position

against the QRS shape. The method is based on the Pearson correlation and is optimized for parallel

processing. In our application with UHF-ECG data the number of detected groups was 3.24 ± 3.41 (mean

and standard deviation over 1,030 records). The method can be used in other areas also where the clustering

of repetitive signal formations is needed. For validation purposes, the method was tested on the MIT-BIH

Arrhythmia and INCART databases from Physionet with results of purity of 98.24 % and 99.50 %.

1 INTRODUCTION

The electrocardiogram (Fig. 1A) is one of the most

important sources of knowledge about heart

function. The analyzed frequency band is mostly

limited to 150 Hz. At higher frequencies the most

limiting factor is the signal-to-noise ratio, which can

be surpassed using the signal-averaged ECG

(SAECG), suppressing noise thanks to averaging

large number (over tens or more) of QRS

complexes. Using SAECG, fetal QRSs were

extracted from maternal ECG (Hon and Lee, 1963)

and weak signals in frequency range 40-300 Hz

revealed high-frequency QRS potentials (Goldberger

et al., 1981). Moreover, SAECG technique allowed

further research in the fields of ventricular late

potentials (Simson 1983; Haberl et al. 1988; Jarrett

and Flowers, 1991) and atrial fibrillation (Fukunami

et al., 1991). Several devices for SAECG were

developed in the past as ART 1200 EPX, Corazonix

Predictor or VCM-3000.

But every step on this high-frequency road was

laid into the range below 250-300 Hz, assuming that

there is nothing useful above. It was correct for early

times when bit-depth of common analog/digital

converters allowed 12 bits, dislodging weak

potentials of higher frequencies into quantization

darkness. Detection of Reduced Area Zones (RAZ)

also uses the frequency range of 150-250 Hz

(Abboud et al., 1987). Accessibility of this technique

(later implemented in Hyper-Q devices) gained new

research in myocardial ischemia (Schlegel et al.,

2004), ECG during anesthesia (Spackman et al.,

2005) or in myocardial infarction (Amit et al., 2013).

Even the spread of SAECG and continuously

increased technical level of analog to digital

converters both in sampling frequency and bit depth

(note common 192 kHz and 24 bits in the field of

digital audio recording), the clinical community still

relies on well-proven frequency range of 0-150 Hz

in classic ECG while only fearless specialists use

RAZ analysis in high-frequency range 150-250 Hz

(HF-ECG).

Against those habits, our team has developed and

tested an innovative method for ultra-high-frequency

ECG (UHF-ECG; up to 2 kHz) analysis (Jurak et al.,

2013) that provides information about spatial and

temporal distribution of depolarization phase of

action potentials. Furthermore, it is able to reveal

ventricle dyssynchrony with the common 12-lead

Plesinger, F., Jurco, J., Halamek, J., Leinveber, P., Reichlova, T. and Jurak, P..

Multichannel QRS Morphology Clustering - Data Preprocessing for Ultra-High-Frequency ECG Analysis.

In Proceedings of the 3rd International Congress on Cardiovascular Technologies (CARDIOTECHNIX 2015), pages 11-19

ISBN: 978-989-758-160-1

ECG that cardiology specialists are used to. As the

preliminary results show, dyssynchrony can be

derived from UHF signal envelopes in the QRS

complex region in leads V1 and V6. A high-dynamic

acquisition system must be used simultaneously with

new processing methods to acquire UHF-ECG.

We have used input data with a sampling rate of

5 kHz and 24-bit dynamic range which allows us to

analyze UHF envelopes in a frequency range of

500–1,000 Hz (Fig. 1B), being high above currently

accepted frequency range. The power in this range is

very weak; the QRS amplitude in the UHF envelope

is approximately 80 dB lower than the low-

frequency (up to 150 Hz) QRS complex amplitude

(compare Figures 1A and B).

Figure 1: A – ECG signal with P-wave, QRS complex and

T-wave; B – signal envelope in the 500–1,000 Hz band;

C – averaged envelope, smoothed (40 samples, rectangular

window).

Signal averaging (mentioned as SAECG before)

is applied to increase the signal-to-noise ratio using

QRS annotation marks as reference points. QRS

complexes must be clustered into groups specified

by QRS morphology due to the possible presence of

more beat types within a single ECG record.

Furthermore, QRS annotation marks must point

to exactly the same sample (Rompelman and Ros,

1986a; Rompelman and Ros, 1986b) inside the QRS

shape (Fig. 1A). Envelopes in the range 500–

1,000 Hz (Fig. 1C) can then be averaged using tens

to hundreds of beats depending on the signal quality.

The electrical activity of the myocardium, expressed

by UHF-ECG envelopes, carries specific temporal

and spatial information that can be further analyzed

(Jurak et al., 2013).

The clustering process is, therefore, an essential

step in UHF-ECG analysis as it is needed to assure

that UHF envelopes belonging to different QRS

types will not be mixed together during averaging. A

similar need has been described in a study (Amit et

al., 2013) before averaging the signal to obtain RAZ.

Current clustering methods (Castro and Paulo,

2014; Lagerholm and Peterson, 2000; Cuesta-Frau et

al., 2003; Chang et al., 2005) aim to assign

previously detected QRS to known beat types. This

goal is not sufficient for our objective due to the fact

that a specific beat type may have different

morphologies (leading to different UHF-ECG

envelopes) which have to be distinguished. Also,

existing approaches do not correct the positions of

QRS annotation marks which is an important step to

maintain detail in averaged envelopes.

We are, therefore, proposing a new clustering

method that can be used in UHF-ECG analysis and

works with multiple leads without human

intervention, allowing full automation.

2 METHOD

2.1 Method Inputs and Preprocessing

The mounted signals from ECG V–leads (V1 to V6)

and a list of QRS annotations are inputs for the

method. QRS annotations are acquired by a robust

multi-lead detection method (Plesinger et al., 2014)

and cleared of pacemaker activity. Thus, only signal

preprocessing steps are mounting and pacemaker

activity removal, where areas influenced by

pacemaker activity are replaced by the linear

interpolation.

2.2 Processing

Processing is presented in a flowchart (Fig. 2).

2.2.1 Primary Clustering

At the beginning (Fig. 2A), the first QRS annotation

is declared the first member of the first morphology

group. Next, the segment around the first unassigned

QRS annotation is compared by Pearson correlation

to each of the QRS annotations already assigned.

This is performed simultaneously (as a multi-thread

process) in leads V1 to V6 within an area of ±120

ms around the QRS annotation and six correlation

coefficients are obtained. If the lowest correlation

coefficient C

min

is higher than correlation threshold

, the unassigned QRS is linked to the group of

correlating QRS annotation. If the tested QRS

annotation cannot be assigned to any of the existing

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

morphology groups, a new group is created and the

unassigned QRS is linked to it as its first member.

This loop is repeated until none of the QRS

annotations remains unassigned.

Figure 2: Method flowchart. A – primary clustering, B –

correlation threshold reduction, C – annotation jitter

reduction, D – computation of averaged shapes, E – shift-

test, F – connecting small groups into a joined group, G –

group reordering, H – group count check. C

min

is the

lowest correlation from leads V1 to V6, C

is the current

corr. threshold (0.98 at the beginning of the process), C

tmin

the minimal permitted correlation threshold (0.75), C

the

decrement of the correlation threshold, N

the maximum

number of groups permitted after primary clustering, N

gmax

the maximum number of groups.

2.2.2 Annotation Jitter Reduction

QRS annotations inside each morphology group

must be exactly aligned with one another. The

highest correlation between each QRS segment

(leads V1 to V6) and the first member of each group

is found by shifting the QRS annotation to the left

and right (Fig. 1C).

2.2.3 Group Average Multi-lead Shapes

A group average multi-lead shape (GAMS) is

created (Fig. 2D and Fig. 3C) for each morphology

group by averaging corresponding samples from all

QRS complexes from the specific group. Each

GAMS contains six averaged shapes (leads V1-V6).

2.2.4 Group Shift-test

It is possible that two or more groups contain a

similar type of QRS morphology, merely shifted to

the left or right. To merge such groups together, the

central part (a width of 120 ms) of each GAMS is

correlated with the GAMS of other groups when

shifted to the left and right (Fig. 2E). If the

correlation maximum is higher than threshold C

the groups are joined together and the program

jumps back to the jitter reduction (Fig. 2C).

2.2.5 Joined Group

Groups containing less than three QRS annotations

are assigned to the “Joined Group” (Fig. 2F and Fig.

3 – the last column). This usually contains

misidentified QRS annotations and artefacts. A large

Joined Group can be produced if the source is too

noisy.

2.2.6 Order Groups by Size

Next, the shape groups are arranged by the number

of related QRS annotations except the Joined Group

(Fig. 2G). The shape group with the largest number

of QRS is named Group 1, the second Group 2 etc.

2.2.7 Check Groups Count

If the number of groups is still too high (>50) and C

is higher than C

tmin

, correlation threshold C

decreased and the computation is restarted (Fig. 2H).

2.3 Method Outputs

The method produces a modified QRS annotations

list, where each of the QRS annotations retains the

Multichannel QRS Morphology Clustering - Data Preprocessing for Ultra-High-Frequency ECG Analysis

information identifying the morphology group to

which a specific QRS belongs. Information about

the location of each QRS is updated when the offset

correction and group shift-test tasks (Fig. 2C and

2E) are completed. Statistical properties providing

information about the correlation between GAMS

and each member of a specific group are also saved

for statistical processing.

3 RESULTS

3.1 Application

The source data for our method are records 8–15

minutes long in a resting supine position; sampling

rate 5 kHz and bit resolution 26 bits (25 kHz and 24

bits before down-sampling). This dataset (UHF-

ECG) has been recorded at the International Clinical

Research Center at St. Anne’s University Hospital,

Brno, Czech Republic using a recording device from

the company M&I, Prague. A total number of 1,030

recordings have been made (262 ischemic heart

disease, 36 hypertrophic cardiomyopathy, 302

dilated cardiomyopathy, 261 heart transplant and

169 healthy subjects).

Using the presented method, all the available

records were clustered into morphology groups with

an average number of groups per record of 3.24 ±

3.41. The average percentage of QRS assigned to the

“Joined Group” (the group of QRS which did not

correlate well enough with any of the other groups)

was 0.99 % ± 2.09 and the average percentage of

QRS in Group 1 was 95.42 % ± 9.85. The median

correlation between each member of the majority

group (Group 1) and the corresponding averaged

shape was 0.997 ± 0.005 (over 1030 records).

Overall results for UHF-ECG dataset are shown in

Table 1.

Table 1: UHF-ECG dataset results. G1 - amount of QRS

in largest group, JG - amount of QRS in Joined Group, Ng

– number of detected groups.

Med. Mean ± SD Min Max

G1 [%] 99.51 95.37 ± 9.94 24.54 100.0

JG [%] 0.25 0.99 ± 2.11 0.00 27.65

Ng 2.00 3.27 ± 3.45 1.00 39.00

The application is presented in Fig. 3 on a

subject with dilated cardiomyopathy. QRS

complexes were detected from a 12-lead UHF-ECG

record (Fig. 3A) and clustered using leads V1 to V6

into five morphology groups (Fig. 3B). The majority

group – Group 1 – (Fig. 3B, the first column and

Fig. 4A) contains 711 QRS. The second group (Fig.

4B) contains 235 QRS. Group 3 and Group 4 (not

displayed) contains only 27 and 4 QRS, for which

reason it is not usable for averaging.

Figure 3: Method application: A – raw ECG data, B –

clustering method result (clusters in columns, leads in

rows) with median correlation inside each cell,

C – averaged QRS shapes for two largest groups, D –

averaged envelopes in range 500–1,000 Hz for two largest

QRS groups. UHF-ECG subject 0766.

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

Figure 4: Categorized QRS complexes, ECG V1 lead in

240 ms window. A – Group 1 QRS complexes, B – Group

2 QRS complexes. Vertical line in the middle shows QRS

annotation mark (i.e. trigger). UHF-ECG subject 0766.

The “Joined Group” (Fig. 3B, the last red

column) contains 16 QRS which were not attached

to any of the existing groups due to noise or

artefacts. Each cell in the grid (Fig. 3B) contains a

240-millisecond-long window and each drawn QRS

is normalized to fill the predefined height of the cell

(i.e. auto-scaled). Once QRS complexes are

clustered into groups, it is possible to compute and

average amplitude envelopes in a range of 500–

1,000 Hz using FFT and Hilbert transform. The

averaged envelopes for Group 1 and Group 2 are

shown in Fig. 3D. In comparison with the greatly

magnified averaged QRS shape of Group 1 (Fig.

3C), we can see large electrical dyssynchrony

revealed between leads V1 and V6 (Fig. 3D).

3.2 Method Validation using Physionet

Annotated Databases

The MIT-BIH Arrhythmia (Moody and Mark, 2001)

and INCART databases from Physionet (Goldberger

et al., 2000) were used with 47 and 75 records,

respectively, for validation and comparison

purposes. The data quality does not allow the use of

these databases for UHF-ECG analysis (the

sampling rate and bit resolution are insufficient), but

they are carefully annotated by specialists and

existing QRS annotations can be used to evaluate the

purity of the clustered groups.

Specificity and sensitivity values or confusion

matrixes (usual ways to validate clustering method)

could not be computed because the presented

method does not aim to cluster QRS by any known

pathological morphology. Instead, purity (P) values

for both databases were computed to show the level

of contamination of clustered groups by different

QRS types (defined in Physionet beat annotations).

P was computed over all subjects and groups (except

for Joined Groups) as:

  100  100 

∑











 





(1)

where N

qrs

is the number of all QRS complexes

within the specific group, N

major

is the maximal

occurrence of any beat type (specified by Physionet

annotations) in the specific group, and N

total

is the

sum of all QRS over all subjects and groups (except

for Joined Groups). The overall purity of MIT-BIH

was 96.69 % and 98.38 % for the INCART database.

Purity can also be evaluated separately for each

subject/record and Tables 2 and 3 show purity,

group sizes and group count statistics for the MIT-

BIH and INCART databases.

Table 2: Results for MIT-BIH arrhythmia database

records. P – purity, G1 - amount of QRS in the largest

group, JG - amount of QRS in Joined Group, Ng – number

of detected groups.

Med. Mean ± SD Min. Max.

P [%] 99.84 96.87 ± 8.26 51.15 100

G1 [%] 79.41 79.23 ± 16.39 28.68 99.41

JG [%] 4.13 9.09 ± 9.63 0.20 31.03

Ng 18.00 23.77 ± 17.89 2.00 70.00

Table 3: Results for INCART database records.

Med. Mean ± SD Min. Max.

P [%] 99.95 98.47 ± 4.98 69.21 100

G1 [%] 73.92 75.09 ± 17.01 35.24 99.59

JG [%] 3.82 7.87 ± 8.68 0.00 33.48

Ng 13.50 19.29 ± 15.38 2.00 71.00

The overall purity of MIT-BIH was 96.69 % and

98.38 % for the INCART database, meaning that the

resultant groups are slightly contaminated by

different beat types. Further insight into the

contamination issue showed it is possibly caused by

two reasons:

Multichannel QRS Morphology Clustering - Data Preprocessing for Ultra-High-Frequency ECG Analysis

First, our method is not able to distinguish

between premature and escape beats having the

same shape as another QRS type (i.e. atrial

premature beats – A – and normal sinus beats – N –

in record 202 in MIT-BIH). Joining these beats

together (which is acceptable for our goals) will

increase average purity to 98.24 % (MIT-BIH) and

99.50 % (INCART).

Second, due to noise, individual shapes are not

able to correlate enough with existing morphology

groups and a large number of groups can be created.

If the number of groups is too high, the correlation

threshold can be decreased which may lead to the

unwanted linking of different beat types into one

group (as in the case of records 208 and 213 from

MIT-BIH).

Another issue can occur when the bottom limit to

correlation threshold (0.75) is met, but it is not

possible to create larger groups (N

QRS

>3). In this

way, the majority of QRS are moved to the Joined

Group (as in records 2, 3 and 58 from INCART).

3.2.1 Results Comparison

Specific results for validation records are presented

by Table 4 (MIT-BIH), showing comparison values

for groups detected by our method and an existing

method (Castro and Paulo, 2014).

Figure 5: Joined Group sizes for two Physionet databases

and UHF-ECG dataset. Medians 4.13 (MIT-BIH), 3.82

(INCART) and 0.26 (UHF-ECG).

The number of QRS complexes in Joined Groups

(Fig. 5) indicates how many QRS complexes are not

(hypothetically) suitable for the following UHF-

ECG analysis. This value should be as low as

possible. The number of groups generated by

clustering (Fig. 6) is lowest for the UHF-ECG

database, though it strongly depends on subject-

specific pathology as well as signal quality.

Table 4: Results for MIT-BIH database records. P –

purity, G1 – amount of QRS in Group 1, JG – amount of

QRS in Joined Group, N

– number of groups generated

by our method, N

– number of groups generated by

compared method (Castro and Paulo, 2014).

Record P [%] G1 [%] JG [%] N

100 98.58 99.34 0.66 2 4

101 99.86 68.47 23.43 52 4

102 99.86 92.36 1.28 17 10

103 100.00 85.94 7.20 29 10

104 98.77 56.98 15.84 54 16

105 100.00 68.23 26.87 28 10

106 100.00 44.89 30.93 30 27

107 100.00 96.40 0.94 10 11

108 99.70 60.64 25.47 18 22

109 100.00 97.08 1.70 7 13

111 100.00 90.58 6.54 17 8

112 99.96 82.63 8.00 39 4

113 100.00 96.99 1.06 10 5

114 99.58 66.58 11.66 70 8

115 100.00 97.75 1.54 8 11

116 99.96 92.91 2.99 13 10

117 99.93 99.41 0.59 2 4

118 95.84 93.81 3.86 13 3

119 100.00 77.45 0.20 3 6

121 99.94 95.60 4.13 4 5

122 100.00 82.96 9.37 55 1

123 100.00 99.41 0.26 4 3

124 98.10 53.24 2.66 16 14

200 96.99 28.68 31.03 48 20

201 97.95 84.11 3.16 11 15

202 98.62 96.77 1.45 9 9

203 99.50 62.08 26.31 43 33

205 99.92 79.07 6.81 49 14

207 94.92 79.41 6.94 22 61

208 51.15 67.11 29.48 32 28

209 90.34 71.11 27.29 20 10

210 100.00 79.70 11.32 39 27

212 100.00 66.05 14.56 38 5

213 81.83 98.40 1.32 5 17

214 99.95 87.75 2.61 13 21

215 99.96 76.06 17.45 38 16

217 98.82 70.15 4.08 30 28

219 99.63 96.75 0.56 7 14

220 95.62 98.88 0.83 4 2

221 100.00 61.27 18.34 34 14

222 83.68 73.66 12.12 64 8

223 96.57 79.31 3.72 21 23

228 99.84 65.76 9.40 25 14

230 100.00 58.07 1.68 8 3

231 99.87 79.31 1.08 12 5

232 77.76 92.47 5.79 15 4

233 99.67 72.04 2.60 29 24

A comparison of the results from the UHF-ECG

dataset and MIT-BIH and INCART databases in

Figure 5 shows that data quality greatly influences

the number of unsuccessfully clustered beats.

MIT-BIH INCART UHF-ECG

Percents

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

Figure 6: Comparison of number of group count results for

MIT-BIH (median 18), INCART (median 13.5), UHF-

ECG dataset (median 2) using the presented method and

MIT-BIH results (median 10) acquired from (Castro &

Paulo 2014) as the last bar - N

4 DISCUSSION

The proposed method is able to categorize QRS

(Fig. 7) in high sampled data (5,000 Hz) and great

bit-depth (26 bits), allowing to see ultra-high

frequency potentials in individual morphology

groups and able to reveal ventricle dyssynchrony as

in Fig. 3. Thanks to the direct comparison among all

of registered QRS this method is able to catch

continuous changes in QRS morphology (which we

encountered while detecting QRS from isolated

hearts).

This is in contrast to building of morphology

template (Breithardt et al., 1991). On the other hand,

due to this behavior the processing time increases to

uncomfortable lengths while processing long (hours)

UHF-ECG recordings.

Figure 7 presents method results, showing

temporal distribution of detected morphology groups

in part of 15-minute record. Such technique may be

used even for standard 6-12 leads ECG.

The legitimacy of using frequencies above usual

300 Hz is shown in Figure 8. Even the currently

used HF frequency range of 150-250 Hz (Fig. 8B)

shows dyssynchrony of specific QRS, it is obvious

that frequency range of 500-1,000 Hz (Fig. 8C)

provides more precise image of ventricles electrical

activation with higher temporal resolution. On the

other hand, range of 1,000-2,000 Hz (Fig. 8D)

brings less evident activity due to significantly lower

signal-to-noise ratio.

Figure 7: Temporal distribution of QRS morphology

groups during the measurement of heart-transplant subject

(UHF-ECG subject 0086). A – RR intervals (beat-to-beat)

excluding beats from Joined Group, B – morphology

groups temporal distribution, where Group 1 (major

occurrence) is a normal QRS, rarely interrupted with

ventricle beats from Group 2. Unrecognized or false

positive QRS complexes are in Joined Group.

MIT-BIH INCART UHF-ECG Ngc

Number of groups

Multichannel QRS Morphology Clustering - Data Preprocessing for Ultra-High-Frequency ECG Analysis

Figure 8: Comparison of averaged QRS shape (N=1350)

in different frequency ranges for leads V1 (orange) and V6

(black). Only QRSs from Group 1 were taken into

account. A – averaged QRS complex, B – amplitude

envelopes in 150-250 Hz, C – amplitude envelopes in 500-

1,000 Hz, D – amplitude envelopes in 1,000-2,000 Hz

(UHF-ECG subject 0086).

Two Physionet databases (MIT-BIH and

INCART) were processed to validate the presented

clustering method and the purity of the clustered

QRS groups was 98.24 % (MIT-BIH) and 99.50 %

(INCART), respectively. This purity shows the level

of correspondence between annotated QRS types

and clustered groups. The median correlation within

Group 1 for the UHF-ECG dataset was 0.997,

showing extremely high overall morphology

stability inside the majority group. In comparison

with an existing clustering method (Castro and

Paulo, 2014) in Fig. 6 and Table 4, our approach

produces a larger number of groups (the median

from the N

column is 18, while the median from

is 10).

Although the method is designed to work with

specific UHF-ECG data, it can be used in any other

area where the clustering of repetitive signal

formations according to shape is needed (as shown

in figure 9 with P-wave example).

5 CONCLUSIONS

A multichannel clustering method has been

presented as an essential part of UHF-ECG analysis.

The method clusters a list of previously detected

QRS complexes into groups by morphology and

corrects QRS annotation mark positions inside each

group to point to exactly the same location of QRS

shape. This functionality allows averaging of UHF-

ECG envelopes with regard to specific QRS types

and, thanks to annotation jitter reduction, the

averaged UHF envelopes retain the highest possible

amount of detail. Correctly averaged UHF-ECG

envelopes are tested by a related study to reveal

information on heart ventricle dyssynchrony.

Figure 9: Presented clustering method used for clustering

of P-wave. 538 P-waves total, 501 were clustered into

Group 1 and 37 into Joined Group. UHF-ECG subject

0616 with dilated cardiomyopathy.

It is also evident that our method works

significantly better with the UHF-ECG dataset (for

which the method was originally designed) than with

low-resolution data from the MIT-BIH and INCART

databases.

CARDIOTECHNIX 2015 - International Congress on Cardiovascular Technologies

The presented method is part of the software

UHF Solver which is used for autonomous

processing of UHF-ECG data to obtain information

about heart ventricle dyssynchrony. Also, the

method has been implemented as a plugin for

SignalPlant, free signal-processing and visualization

software.

ACKNOWLEDGEMENTS

This research was supported by project no.

P102/12/2034 from the Grant Agency of the Czech

Republic and by MEYS CR (LO1212),

its infrastructure by MEYS CR and EC

(CZ.1.05/2.1.00/01.0017) and by ASCR (RVO:

68081731).

REFERENCES

Abboud, S. et al., 1987. Detection of transient myocardial

ischemia by computer analysis of standard and signal-

averaged high-frequency electrocardiograms in

patients undergoing percutaneous transluminal

coronary angioplasty. Circulation, 76(3), pp.585–596.

Amit, G. et al., 2013. High-frequency QRS analysis in

patients with acute myocardial infarction: a

preliminary study. Annals of noninvasive

electrocardiology : the official journal of the

International Society for Holter and Noninvasive

Electrocardiology, Inc, 18(2), pp.149–156.

Breithardt, G. et al., 1991. Standards for analysis of

ventricular late potentials using high-resolution or

signal-averaged electrocardiography. A statement by a

Task Force Committee of the European Society of

Cardiology, the American Heart Association, and the

American College of Ca. Circulation, 83(4), pp.1481–

1488.

Castro, D. & Paulo, F., 2014. A method for context-based

adaptive QRS clustering in real-time. Biomedical and

Health Informatics, IEEE Journal of, PP(99), pp.1–12.

Cuesta-Frau, D., Pérez-Cortés, J. C. & Andreu-García, G.,

2003. Clustering of electrocardiograph signals in

computer-aided Holter analysis. Computer Methods

and Programs in Biomedicine, 72(3), pp.179–196.

Fukunami, M. et al., 1991. Detection of patients at risk for

paroxysmal atrial fibrillation during sinus rhythm by P

wave-triggered signal-averaged electrocardiogram.

Circulation, 83(1), pp.162–169.

Goldberger, A. L. et al., 1981. Effect of myocardial

infarction on high-frequency QRS potentials.

Circulation, 64(1), pp.34–42.

Goldberger, A. L. et al., 2000. PhysioBank, PhysioToolkit,

and PhysioNet: Components of a New Research

Resource for Complex Physiologic Signals.

Circulation, 101(23), pp.215–220.

Haberl, R. et al., 1988. Comparison of frequency and time

domain analysis of the signal-averaged

electrocardiogram in patients with ventricular

tachycardia and coronary artery disease: methodologic

validation and clinical relevance. Journal of the

American College of Cardiology, 12(1), pp.150–158.

Hon, E. H. & Lee, S. T., 1963. Noise reduction in fetal

electrocardiography. American Journal of Obstetrics

& Gynecology, 87(8), pp.1086–1096. Available at:

http://dx.doi.org/10.1016/0002-9378(63)90104-2.

Chang, K.-C. et al., 2005. a Comparison of Similarity

Measures for Clustering of Qrs Complexes.

Biomedical Engineering: Applications, Basis and

Communications, 17(06), pp.324–331.

Jarrett, J. R. & Flowers, N. C., 1991. Electrophysiology ,

Pacing , and Arrhythmia Signal- Averaged

Electrocardiography : History , Techniques , and

Clinical Applications. Clin. Cardiol., 14, pp.984–994.

Jurak, P. et al., 2013. Ultra-high-frequency ECG

Measurement Institute of Scientific Instruments , AS ,

Brno , Czech Republic. Computing in Cardiology

Conference (CinC), 2013, 40, pp.783–786.

Lagerholm, M. & Peterson, G., 2000. Clustering ECG

complexes using hermite functions and self-organizing

maps. IEEE Transactions on Biomedical Engineering,

47(7), pp.838–848.

Moody, G. B. & Mark, R. G., 2001. The impact of the

MIT-BIH arrhythmia database. IEEE Engineering in

Medicine and Biology Magazine, 20(3), pp.45–50.

Plesinger, F. et al., 2014. Robust Multichannel QRS

Detection. Computing in Cardiology Conference

(CinC), 2014, 41, pp.557–560.

Rompelman, O. & Ros, H. H., 1986a. Coherent averaging

technique: a tutorial review. Part 1: Noise reduction

and the equivalent filter. Journal of biomedical

engineering, 8(1), pp.24–29.

Rompelman, O. & Ros, H. H., 1986b. Coherent averaging

technique: a tutorial review. Part 2: Trigger jitter,

overlapping responses and non-periodic stimulation.

Journal of biomedical engineering, 8(1), pp.30–35.

Schlegel, T. T. et al., 2004. Real-time 12-lead high-

frequency QRS electrocardiography for enhanced

detection of myocardial ischemia and coronary artery

disease. Mayo Clinic proceedings. Mayo Clinic, 79(3),

pp.339–350.

Simson, M. B., 1983. Clinical application of signal

averaging. Cardiol Clin, 1(1), pp.109–119. Available

at: http://www.ncbi.nlm.nih.gov/pubmed/6399988.

Spackman, T. N., Abel, M. D. & Schlegel, T. T., 2005.

Twelve-lead high-frequency QRS electrocardiography

during anesthesia in healthy subjects. Anesthesia and

Analgesia, 100(4), pp.1043–1047.

Multichannel QRS Morphology Clustering - Data Preprocessing for Ultra-High-Frequency ECG Analysis