Physiological Measurement

ACCEPTED MANUSCRIPT

Ranking of the most reliable beat morphology and heart rate variability
features for detection of atrial fibrillation in short single lead ECG
To cite this article before publication: Ivaylo Christov et al 2018 Physiol. Meas. in press https://doi.org/10.1088/1361-6579/aad9f0

Manuscript version: Accepted Manuscript

Accepted Manuscript is “the version of the article accepted for publication including all changes made as a result of the peer review process,
and which may also include the addition to the article by IOP Publishing of a header, an article ID, a cover sheet and/or an ‘Accepted
Manuscript’ watermark, but excluding any other editing, typesetting or other changes made by IOP Publishing and/or its licensors”

This Accepted Manuscript is © 2018 Institute of Physics and Engineering in Medicine.

During the embargo period (the 12 month period from the publication of the Version of Record of this article), the Accepted Manuscript is fully
protected by copyright and cannot be reused or reposted elsewhere.
As the Version of Record of this article is going to be / has been published on a subscription basis, this Accepted Manuscript is available for reuse
under a CC BY-NC-ND 3.0 licence after the 12 month embargo period.

After the embargo period, everyone is permitted to use copy and redistribute this article for non-commercial purposes only, provided that they
adhere to all the terms of the licence https://creativecommons.org/licences/by-nc-nd/3.0

Although reasonable endeavours have been taken to obtain all necessary permissions from third parties to include their copyrighted content
within this article, their full citation and copyright line may not be present in this Accepted Manuscript version. Before using any content from this
article, please refer to the Version of Record on IOPscience once published for full citation and copyright details, as permissions will likely be
required. All third party content is fully copyright protected, unless specifically stated otherwise in the figure caption in the Version of Record.

View the article online for updates and enhancements.

This content was downloaded from IP address 132.174.255.250 on 15/08/2018 at 02:17

Page 1 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3
4

pt
5
6
TITLE PAGE
7
8
9

cri
10
11
12
Ranking of the most reliable beat morphology and heart rate variability features
13
14 for detection of atrial fibrillation in short single lead ECG
15
16

us
17
18
19
20
21
22
23
24
25
26
27
28
1
an
Ivaylo Christov1,Vessela Krasteva1*, Iana Simova2, Tatiana Neycheva1, Ramun Schmid3

Institute of Biophysics and Biomedical Engineering, Bulgarian Academy of Sciences, Acad. G.

Bonchev Str. Bl 105, 1113 Sofia, Bulgaria
dM
29 2
Acibadem City Clinic Cardiovascular Center, 127 Okolovrasten pat Str., 1407 Sofia, Bulgaria
30
31 3
Signal Processing, Schiller AG, Altgasse 68, CH-6341, Baar, Switzerland
32
33
34
35
36
37
38 * Corresponding author.
pte

39 Vessela Krasteva, e-mail: vessika@biomed.bas.bg

40
Address:
41
42 Institute of Biophysics and Biomedical Engineering,
43
Bulgarian Academy of Sciences
44
45 Acad. G. Bonchev str. bl.105, 1113, Sofia, Bulgaria
46 Tel: +3592 9793631
ce

47
48 Fax: +3592 8723787
49
50
51
52
Ac

53
54
55
56 1
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 2 of 29

1
2
3
4 Ranking of the most reliable beat morphology and

pt
5 heart rate variability features for detection of atrial fibrillation
6
7 in short single lead ECG
8
9

cri
10 Ivaylo Christov1, Vessela Krasteva1*, Iana Simova2,
11 Tatiana Neycheva1 and Ramun Schmid3
12
1
13 Institute of Biophysics and Biomedical Engineering, Bulgarian Academy of Sciences, Acad
14 G Bonchev Str. Bl 105, 1113 Sofia, Bulgaria
15 2
Acibadem City Clinic Cardiovascular Center, 127 Okolovrasten pat Str., 1407 Sofia,
16 Bulgaria

us
17 3
Signal Processing, Schiller AG, Altgasse 68, CH-6341, Baar, Switzerland
18
19 *E-mail: vessika@biomed.bas.bg
20
21 Abstract
22
23
24
25
26
27
28
Objective

an
This study participated in the 2017 PhysioNet/CinC Challenge dedicated to classification of
atrial fibrillation (AF), normal sinus rhythm (Normal), other arrhythmia (Other) and strong
noise, using single lead electrocardiogram (ECG) recordings with duration <60s. The aim is
dM
29 to apply a linear threshold-based strategy for arrhythmia classification, ranking the most
30
31
powerful time-domain ECG features that could be easily reproduced on any platform.
32 Approach
33
34 An algorithm for time-domain ECG analysis was designed to extract 44 features with focus
35 on: noise detection; heart rate variability analysis (HRV); beat morphology analysis and
36
37 delineation of P, QRS, T waves on robust average beat; detection of atrial activity by the
38
presence of P-wave in the average beat and atrial fibrillatory f-waves during TQ intervals.
pte

39
40 Linear discriminant analysis (LDA) classifier was optimized on the Challenge training set
41
42 (8528 ECGs) by stepwise selection of a non-redundant feature set until maximization of the
43 Challenge F1 score. Heart rate (HR) was an independent factor for LDA classifier design,
44
45 particular to bradycardia (HR50bpm), normal rhythm (HR=50-100bpm), tachycardia
46
(HR100bpm).
ce

47
48 Main results
49
50 The algorithm obtained official Challenge F1 score: 0.80 (Overall), 0.90 (Normal), 0.81
51
(AF), 0.70 (Other), 0.54 (Noise) on hidden Challenge testset (3658 ECGs). This is equivalent
52
Ac

53 to true positive rate TPR = 90.1% (Normal), 81.5% (AF), 67.7% (Other), 69.5% (Noise) and
54
55 false positive rate FPR = 13.6% (Normal), 2.3% (AF), 7.7% (Other), 1.5% (Noise).
56 2
57
58
59
60
Page 3 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 Significance
4

pt
5 The top-5 features, which together contributed to about 94% of maximal F1 score were
6
ranked: (1) Proportion of RR-intervals differing by >50ms from the preceding RR interval;
7
8 (2) Poincaré plot geometry estimated by the ratio of the minor-to-major semi-axes of the
9

cri
10 fitted ellipse; (3) P-wave presence in the average beat; (4) Mean percentage value of the RR-
11 interval first differences; (5) Mean correlation of all beats against the average beat. The
12
13 global rank of feature extraction methods highlighted HRV that was alone able to provide
14 92.5% of maximal F1 score (0.74 vs. 0.8). The added value of more complex ECG
15
16 morphology analysis was less significant for Normal, AF, Other rhythms (+0.02 to 0.08

us
17
points) than Noise (+0.19 points), however, indispensable in wearable ECG recording
18
19 devices with frequent artifact disturbance.
20
21
22 Keywords: 2017 PhysioNet/CinC Challenge, atrial fibrillation, arrhythmia detection, noise rejection, heart
23
24
25
26
27
28
1. Introduction
an
rate variability, average beat morphology, atrial fibrillatory waves.
dM
29 Atrial fibrillation (AF) is the most common cardiac arrhythmia, and is the major risk factor for death,
30
31 stroke, hospitalization, heart failure and coronary artery disease (Lip et al 2016, Camm et al 2010). It
32
33 affects about 2-3% of the population in Europe (Zoni-Berisso et al 2014). The prevalence of AF
34 increases with age (from about 0.14% of younger <49 years old, to about 14% of older >80 years old)
35
36 and gender (male to female ratio is 1.2:1).
37
38 AF appears as a result of reentry within multiple circuits in the atria and typically becomes visible in
pte

39
40
the electrocardiogram (ECG) as rapid oscillations or fibrillatory waves (f-waves) with different sizes,
41 shapes and timings (Camm et al 2010). Those f-waves are present in overall ECG, but are masked by
42
43 high amplitude QRS and T waves, thus can be observed only in TQ intervals, predominantly in V1,
44
45 and occasionally in the peripheral leads. AF detection is a challenging task because f-waves, when
46
visible, have typically low amplitudes (<200 uV).
ce

47
48 AF detection based on f-waves recognition was applied during the isoelectric TQ intervals (Christov
49
50 et al 2001a, Petrėnas et al 2012). Former studies of our team have been focused on the improvement
51
52 of AF detection by additional analysis of RR interval irregularity (Christov et al 2001b) in
Ac

53 combination with QRS morphology descriptors for discrimination between supraventricular and
54
55 ventricular arrhythmias (Krasteva et al 2006). Another important feature to the AF detection has been
56 3
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 4 of 29

1
2
3 shown to be the absence of P-waves (Dotsinsky 2007, Ladavich and Ghoraani 2015).
4

pt
5 Some authors were paying attention to the noise that accompanies the ECG and its impact on the AF
6
7 detection algorithms. Oster and Clifford (2015) were analyzing the performance of the AF detection
8
9
algorithms as a function of the QRS detection performance, RR interval irregularity, P-wave absence,

cri
10 presence of f-waves, and presence of noise. They were showing a linear decrease of the AF detection
11
12 accuracy with reduction of the signal-to-noise ratio. Christov et al (2001a) was reporting a false
13
14 positive detection of their ‘wave rectification method’ in the presence of electromyographic (EMG)
15
noise, and a false negative detection after EMG filtering.
16

us
17 AF was generally considered an exclusion criterion for analysis of heart rate variability (HRV)
18
19 because presumably, the apparent total irregularity of ventricular rhythm in atrial fibrillation has been
20
21 daunted most investigators to study significant HRV changes, associated with autonomic system
22
23
24
25
26
27
28
an
status. A few studies were indicating that atrial fibrillation patients had a decreased vagal input in the
heart rate regulation (van den Berg et al 1997, Barauskiene et al 2016). HRV was suggested as an
important marker for the development of recurrent AF (Lombardi et al 2001, Akyürek et al 2003).
HRV has also been shown to be a powerful tool for the detection of AF. Park et al (2009) reported an
dM
29 accuracy of 91.4-92.9% by analysis of the Poincaré plot, relying on its irregularly irregular shape for
30
31 AF vs. non-AF patients. Kim et al (2008) improved AF detection accuracy from 93.8% to 95.2%
32
33 using different logistic regression formulas for evening time and daytime, prving that HRV features
34
have certain time-dependent changes. Zhou et al (2015) reported accuracy in the range 87-98% on
35
36 four public MIT-BIH databases, clearly showing larger dynamics of the instantaneous heart rate (HR)
37
38 during AF by means of Shannon energy estimation on the HR symbolic sequence.
pte

39
40 The 2017 PhysioNet/CinC Challenge (Clifford et al 2017) provided the ground for competitive
41 improvement of AF detection algorithms, applicable to the Challenge’s database of single lead, short
42
43 duration (<1 minute) ECG recordings, often accompanied by strong noises. This study participated in
44
45 the Challenge, aiming to explore the feasibility of simple time-domain ECG analysis techniques for
46
noise detection; HRV analysis; beat morphology analysis after robust synthesis of an average beat
ce

47
48 and delineation of P, QRS, T waves; detection of atrial activity by the presence of a P-wave in the
49
50 average beat and atrial fibrillatory waves during TQ intervals. We showed that ranking of the most
51
52 reliable features by a linear discriminant classifier provided a cost-effective solution (maximal
Ac

53 running time <11 % of the Challenge server quota) with a robust performance on both training and
54
55 blinded test sets (equal score of 0.8) that was just 0.03 points lower than the winners in the Challenge.
56 4
57
58
59
60
Page 5 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 2. Challenge ECG database
4

pt
5
6 All available ECG recordings in the 2017 PhysioNet/CinC Challenge database (Clifford et al 2017)
7 were used for training (8528 ECGs), employing third version (V3) of published rhythm annotations
8
9 over short ECG episodes (9-61s) into 4 labeled classes (Table 1). The test set (3658 ECGs) was used

cri
10
11 by the Challenge to score the official participants' entries, considering that both signals and
12
annotations were hidden for the public. All Challenge participants were also blinded about the
13
14 number of subjects and the number of recordings per subject included in both datasets.
15
16

us
The ECG analysis considered one-lead ECG (equivalent to lead I), taking the available full-length of
17
18 the recording (limited to 60s), sampled at 300 Hz and pre-filtered in a non-diagnostic bandwidth (0.5-
19
20 40 Hz). Detailed description of the data acquisition and rhythm annotation process could be found in
21
22 Clifford et al (2017). Here, we could note that the ECG signals were representative to the recording
23
24
25
26
27
28
arms). an
conditions met by wearable ECG devices with hand-held electrode contact (between fingers of both
dM
29
30
Table 1. Data profile for the training/test set, given as number (%) of recordings in four rhythm
31 classes.
32
33 Rhythm class Training set Test set
34
35 Normal sinus rhythm (Normal) 5076 (60 %) 2437 (67 %)
36
37
Atrial fibrillation (AF) 758 (9 %) 286 (8 %)
38 Other arrhythmia (Other) 2415 (28 %) 683 (19 %)
pte

39
40 Too noisy to be classified (Noise) 279 (3 %) 252 (7%)
41
42
43
44 3. Methods
45
46 We present a multi-parametric algorithm for time-domain analysis of single lead ECG recordings
ce

47
48 with short duration (<60s) that has been implemented in Matlab (MathWorks Inc.) and registered as
49
50 an official open-source entry in the 2017 PhysioNet/CinC Challenge (Christov et al 2017). The
51
algorithm flow diagram is outlined in figure 1, formally presented by 3 processing stages:
52
Ac

53 - Pre-processing stage: implements a conventional QRS detector with a specific noise correction
54
55 feedback.
56 5
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 6 of 29

1
2
3 - Feature extraction: implements 3 modules for calculation of >40 features by HRV analysis,
4

pt
5 average beat morphology analysis, analysis of atrial fibrillatory waves.
6
7 - Rhythm classification: implements 3 linear discriminant functions, optimized for the detection of
8
9
four classes (Normal, AF, Other, Noise), considering their specific profiles in 3 heart rate ranges

cri
10 (bradycardia, normal rate, tachycardia).
11
12 Details for each analysis block are further presented in Methods.
13
14 ECG (single-lead)
15 QRS detection
16

us
17 Noise correction
18
19 HRV analysis Average beat morphology analysis Analysis of atrial fibrillatory waves
20
➢ RR-Tachogram (8 features) ➢ Average beat synthesis (3 features) ➢ Merging of isoelectric [Tend; Q]
21
➢ dRR-Tachogram (9 features) ➢ Detection of fiducial points: R, S, Q, J, T-end, T-peak, P-peak intervals in a continuous signal
22
23
24
25
26
27
28
➢ RR-Histogram (1 feature)
➢ Poincaré Plot (3 features)
an
➢ Calculation of intervals: QRS, PQ, QT (3 features)
➢ Calculation of amplitudes: QRS, J, T, P (4 features)
➢ Detection of P-wave (1 feature)
➢ Detection of atypical ventricular morphologies (4 features):
QRS fragmentation, LBBB, Inverted QRS and T, J-shift
➢ DC filtering (1st difference in 20ms)
➢ EMG filtering (moving average 30ms)
➢ Rectification
➢ Calculation of median value (1 feature)
dM
➢ Cross-correlation analysis (3 features)
29
30 ➢ Calculation of curvature (4 features):
maxc(QRS), maxc(QRS/T), maxc(QRS/P), maxc(T/P)
31
32
33 Classifier: Linear Discriminant Analysis (LDA)
34
➢ 1 LDA for Bradycardia (HR≤50bpm): 194 cases (7% Normal, 4% AF, 86% Other, 3% Noise)
35
➢ 1 LDA for Normal HR (HR=50-100bpm): 7710 cases (65% Normal, 7% AF, 25% Other, 3% Noise)
36
37 ➢ 1 LDA for Tachycardia (HR≥100bpm): 624 cases (3% Normal, 37% AF, 53% Other, 7% Noise)
38
pte

39 Figure 1. Flow chart of the presented algorithm for analysis of 2017 PhysioNet/CinC Challenge data.
40
41
42
43
44 3.1. QRS detection and noise correction
45
46 The QRS detector was based on the Pan and Tompkins (1985) algorithm, relying on the energy
ce

47
48 threshold from the ECG slope, amplitude and width. We used the publicly available implementation
49
of the QRS detector in Matlab (qrs_detect2.m) distributed by the Challenge organizers along with the
50
51 sample code (https://physionet.org/challenge/2017/sample2017.zip). We observed two general
52
Ac

53 sources for QRS detection failure (blockade of the QRS detector over >2s), which we were trying to
54
55 compensate by an additional feedback function for noise correction, as follows:
56 6
57
58
59
60
Page 7 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 - Low amplitude ECG. Due to weak signal energy below certain default thresholds, a blockade of the
4

pt
5 QRS detector over the full recording length was observed. Such QRS detector failure was corrected
6
7 by stepwise signal amplification until the detector starts to find QRS complexes on a regular basis.
8
9
- High-amplitude, high-frequency artifacts. Due to the ‘offline’ processing concept of the QRS

cri
10 detector (certain energy thresholds have been set by taking the signal characteristics over the full
11
12 recording length), the presence of strong artifacts in an arbitrary segment of the recording could
13
14 impede the QRS detector in remaining noise free episodes. The detection of unrealistically small
15
number of QRS (<recording duration (s)/2s) was indicated as false QRS due to high-amplitude,
16

us
17 high-frequency artifacts (figure 2a). Such QRS detector failure was corrected by rejecting (zeroing)
18
19 part of the signal in the vicinity (±1s) around all falsely detected QRS (figure 2b). Then the QRS
20
21 detector wasA00341
restarted. If further
Before RR-intervals>2s were still present, they were excluded from the
noise correction
22
23
24
25
26
27
28
RR-interval series in the next analyses.

3
an 5
4

3
A00341 after noise correction
dM
ECG (mV)

29 2 2
30
1 1
31
32 0 0
33
-1 -1
34
35 -2 -2
36 0 5 10 15 20 25 30 0 5 10 15 20 25 30
37 Time (s) Time (s)
(a) (b)
38
Figure 2. Illustration of the noise correction technique for high-amplitude, high-frequency artefacts
pte

39
40
in file A00341 (annotation ‘Normal’ rhythm): (a) only the artifact is detected (‘*’ marks); (b) after
41
42 rejection of the artifact, the major part of QRS complexes are detected (‘o’ marks).
43
44
45
46
3.2. HRV analysis
ce

47
48
49 We further calculated time-domain features of short-term HRV. The standard frequency domain
50
51 HRV measurements (ESC/NASPE Task Force 1996) were not considered due to potentially
52 inaccurate spectrum calculation from the limited RR-interval series available within the short
Ac

53
54 Challenge ECG recordings (<60s).
55
56 7
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 8 of 29

1
2
3 - RR-Tachogram. 7 features were measured for the series of all RR-intervals as: mean value
4

pt
5 (RRmean), median value (RRmedian), standard deviation (RRstd), mean deviation (RRmeand),
6
7 proportions of the standard and mean deviation from the mean value (RRstd%, RRmeand%), and
8
9
ratio of mean-to-median value (RRrat). The mean HR was further reported as a reciprocal

cri
10 surrogate of RRmean: HRmean=60/RRmean (bpm).
11
12 - dRR-Tachogram. 9 features were measured for the series of all RR-interval first differences as:
13
14 mean value (dRRmean), standard deviation (dRRstd), median deviation (dRRmedian) and their
15
percentage to RRmean (dRRmean%, dRRmedian%, dRRstd%), the proportion of RR intervals
16

us
17 differing by >50ms from the preceding RR interval (PNN50), the square root of the mean squared
18
19 differences of successive RR intervals (RMSSD) and its proportion to meanRR (RMSSD%).
20
21 - RR-Histogram. HRV Triangular Index was used to count the total number of RR intervals divided
22
23
24
25
26
27
28
Hz) has been implemented.
an
by the number of RR intervals in the modal bin. This measure is dependent on the length of the bin,
therefore, the common standard for discrete scale resolution equal to 7.8125 ms (sampling rate 128

- Poincaré Plot [RRn, RRn-1]. 3 features were used for quantification of the Poincaré plot geometry:
dM
29 ▪ Short-to-long term HRV was calculated as the ratio SD1/SD2, where SD1 and SD2 are the
30
31 minor and the major semi-axes of the fitted ellipse (Golińska 2013):
32
33 2 dRRstd 2
34 SD1  dRRstd , SD 2  2 RRstd 2
 .
35
2 2
36 ▪ Variability in the temporal structure was calculated by the descriptor Complex Correlation
37
38 Measure (CCM), which quantifies the point-to-point (dynamic) variation of the Poincaré plot
pte

39
40 (Karmakar et al 2011):
41 N 2
1
42
43
CCM   Ai ,
SD1 SD2 ( N  2) i 1
44
45 where Ai represents the area of the i-th triangle, formed by 3 consecutive points with Poincaré
46
plot coordinates [RRi, RRi-1], [RRi+1, RRi], [RRi+2, RRi+1] from the RR-interval series with
ce

47
48
length N.
49
50 ▪ The Pearson's correlation coefficient (corRR) represents the goodness of linear fitting of all
51
52 points in the Poincaré plot [RRn, RRn-1]:
Ac

53
54
55
56 8
57
58
59
60
Page 9 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 N
4  RR RR

pt
i i 1
5 i2
6
corRR  .
N N
7  RR  RR
2 2
i i 1
8 i2 i2
9

cri
10 The presented example shows noticeably different profiles of all HRV plots (RR-tachogram, dRR-
11
12 Tachogram, RR-Histogram and Poincaré plot), between Normal rhythm (figure 3a) and AF (figure
13 3b). Therefore, we suggested that HRV analysis would be a powerful feature extraction technique for
14
15 quantification of the intrinsic AF heart rate irregularity with large deviation from the normal.
16

us
File: A00254, Annotation: 'Normal' rhythm File: A00225, Annotation: 'AF' rhythm
17 0.3 1

18 0.2 0.8

19 0.1 0.6
ECG (mV)

ECG (mV)
20 0 0.4

21 -0.1 0.2

22 -0.2 0

23
24
25
26
27
28
1000

900
-0.3

-0.4
0 5

RR-TACHOGRAM
10 15
Time (s)

>125
100
75
20

dRR-TACHOGRAM
25

an 30

1000

900

800
-0.2

-0.4
0 5

RR-TACHOGRAM
10 15
Time (s)

>125
100
75
20

dRR-TACHOGRAM
25 30
dM
800 50 50

29 25 700 25
ms

700
ms

ms
ms

0 0
30 600 -25
600 -25

31 500 -50 500 -50

-75
-75
32 400
-100
400
-100

33 300
0 10 20 30
<-125
#RR 0 10 20 30 #RR
300
0 10 20 30 40 50 60
<-125
0 10 20 30 40 50 60

34 RR-HISTOGRAM POINCARE PLOT RR-HISTOGRAM #RR #RR

POINCARE PLOT
20 1000 10 1000
35 18
900 9
900
16
36
8
14 800 7 800
RR(n-1) [ms]
RR(n-1) [ms]

37 12 6
% beats

% beats

700 700

38
10 5
600 600
8 4
pte

39 6 500 3 500

40 4
2
400
2
1
400

41 0
<-40 -40 -30 -20 -10 0 +10 +20 +30 +40>+40
300
0 100 200 300 400 500 600 700 800 9001000
0
<-40 -40 -30 -20 -10 0 +10 +20 +30 +40>+40
300
300 400 500 600 700 800 900 1000
42 % dev of average RR RR(n) [ms] % dev of average RR RR(n) [ms]

43 (a) (b)
44 Figure 3. HRV analysis plots for: (a) ‘Normal’ rhythm (file A00254), (b) AF rhythm (file A00225).
45
46 The ECG signal and the QRS detection marks (‘o’) are shown in the top trace. The RR-interval series
ce

47
48 are shown in the RR-Tachogram (red lines highlight the RRmean value and RRmeanRRstd range).
49
50 The RR-interval first differences are shown in the dRR-Tachogram (red lines highlight the PNN50
51
range 50ms). The histogram of the RR-interval series normalized to RRmean is shown in the RR-
52
Ac

53 Histogram (red lines highlight the RRstd/RRmean range). The RR-interval series [RR(n), RR(n-1)]
54
55 are shown in the Poincaré Plot.
56 9
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 10 of 29

1
2
3 3.3. Average beat morphology analysis
4

pt
5
6 3.3.1. Average beat synthesis. The beat morphology was evaluated in a window (-300ms;
7 0.6*meanRR) around the QRS fiducial point (R or S peak, with the highest positive or negative
8
9 amplitude, indicated by ‘o’ marks in figures 2-4). The peak-to-peak amplitude range within the

cri
10
11 analysis windows of all QRS fiducial points in the recording was reported as mean value (meanAmp)
12
and standard deviation (stdAmp). We validated only the beats with the most sustained amplitudes
13
14 (those within the amplitude range meanAmpstdAmp) for synthesis of a robust average beat. Cross-
15
16 correlation was used to align the beats within analysis windows before beat averaging. The rejected

us
17
18 outliers with non-sustained amplitudes were suspected as artifacts or abnormal beats (figure 4, ‘x’
19
20
beats). The beat synthesis module was designed to return the average beat (figure 5) and 3 additional
21 features, further used in the rhythm classification process:
22
23
24
25
26
27
28
- the absolute amplitude deviation stdAmp (mV),
- the normalized amplitude deviation stdAmp/meanAmp (%) an
- the amplitude rejection ratio rejAmp (%), representing the number of beats with non-sustained
amplitudes to the total number of beats.
dM
29
File: A00003, Annotation: 'Normal' rhythm File: A00137, Annotation: 'AF' rhythm
30 0.5 2

31
32 1
0
33
ECG (mV)

ECG (mV)

34 0

35 -0.5
36 -1

37
38 -1
0 10 20 30 40 50 60
-2
0 5 10 15 20 25 30
Time (s) Time (s)
pte

39
40 (a) (b)
File: A00008, Annotation: 'Other' rhythm
41 3
File: A00056, Annotation: 'Noise'

42 1
2
43
44
ECG (mV)
ECG (mV)

0.5 1
45
46 0
0
ce

47 -1
48 -0.5
49 0 10 20 30 40 50 60 -20 5 10 15 20 25 30
50 Time (s) Time (s)
51 (c) (d)
52
Ac

53 Figure 4. Examples for validation of sustained beats (‘o’ mark) and rejection of the outliers (‘x’
54
55 marks) out of the range meanAmpstdAmp (enclosed within the blue lines) for 4 rhythm classes: (a)
56 10
57
58
59
60
Page 11 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 ‘Normal’ rhythm (file A00003), (b) AF rhythm (file A00137), (c) ‘Other’ rhythm (file A00008), (d)
4

pt
5 ‘Noise’ (file A00056).
6
7 File: A00003 File: A00137 File: A00008 File: A00056
8 0.1
Annotation: 'Normal' rhythm Annotation: 'AF' rhythm Annotation: 'Other' rhythm Annotation: 'Noise'
0.8 0.5
9 S R R R

cri
Q J T-end 0.8
10 0
0.7 0.4

11 0.6 0.7 0.3

-0.1 P-peak T-peak
12 0.6
average beat (mV)

average beat (mV)

average beat (mV)

average beat (mV)

0.5 0.2
13 -0.2 T-peak
0.4
0.5
0.1
14 0.4
-0.3 0.3 0
15 0.3 Q
J
T-peak
16 0.2
0.2
-0.1 T-end

us
-0.4 T-peak P-peak
17 0.1 0.1 -0.2
Q S
18 -0.5
R 0
J
0 -0.3
19 -0.6 -0.1
S T-end -0.1
Q J T-end
-0.4
S
0 0.2 0.4 0.6 0 0.2 0.4 0.6 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6
20 Time (s) Time (s) Time (s) Time (s)
21 (a) (b) (c) (d)
22
23
24
25
26
27
28
Figure 5. Average beat synthesis for the rhythms in figure 4: (a) Normal rhythm (file A00003), (b)
AF rhythm (file A00137), (c) Other rhythm (file A00008), (d) Noise (file A00056). The detected
fiducial points (R, S, Q, J, T-end, T-peak) are shown for all average beats (‘o’ marks). The P-peak
an
dM
29 fiducial points are marked only in (a, c) cases, where ‘P-wave presence’ was detected.
30
31
32
33 3.3.2. Detection of fiducial points (R, S, Q, J, T-end, T-peak, P-peak). The two peaks (R, S) were
34
35 distinguished as the most positive and negative extremities in a window of 140ms around the QRS
36
37
fiducial point. The detection of Q, J, T-end, T-peak was adopted from our previous studies (Christov
38 and Simova 2006, Daskalov and Christov 1999a, Daskalov and Christov 1999b). P-peak was
pte

39
40 searched as the maximal amplitude deflection in the interval (Q-300ms; Q). P-peak was validated if
41
42 found in a physiologically reasonable interval (Q-230ms; Q-30ms). Examples of detected fiducial
43
points on the average beat are shown in figure 5.
44
45
46 3.3.3. Calculation of intervals (QRS, PQ, QT). The durations of intervals between the fiducial
ce

47 points were measured as: QRSwidth=J-Q, PQ=Q-Ppeak, QT=Tend-Q.

48
49
50 3.3.4. Calculation of amplitudes (QRS, J, T-peak, P-peak). The peak-to-peak QRS amplitude
51 (QRSp-p) was reported. The amplitudes of the J point, T-peak, P-peak were calculated against the
52
Ac

53 offset of the isoelectric Q point.

54
55
56 11
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 12 of 29

1
2
3 3.3.5. Detection of P-wave. The algorithm for P-wave detection was following several empirical
4

pt
5 criteria. The violation of any criterion was interpreted as ‘P-wave absence’; otherwise the respect to
6
7 all criteria was interpreted as ‘P-wave presence’:
8
9
- P-peak was validated, i.e. found in interval (Q-230ms to Q-30ms), according to 3.3.2;

cri
10 - P-peak sign = R-peak sign;
11
12 - P-peak slope (+10ms) > thr1, where thr1 = 0.5V;
13
14 - P-peak slope (-10ms) > thr1;
15
16 - P-peak slope (+20ms) > 4*thr1, where 4*thr1 = 2V;

us
17 - P-peak slope (-20ms) > 4*thr1;
18
19 - P-peak amplitude > thr2, where thr2 = 16% of QRSp-p amplitude.
20
21 The P-peak slope was calculated as the first absolute difference between the amplitudes of the P-peak
22
23
24
25
26
27
28
an
and a distant sample taken before (-delay) or after (+delay) the P-peak. The thresholds (thr1, thr2)
were optimized for the training Challenge dataset.
An illustration of the P-wave detection performance is presented in figure 5, where P-waves are
detected to be ‘present’ for ‘Normal’ and ‘Other’ average beats, while ‘absent’ for ‘AF’ and ‘Noise’
dM
29
average beats.
30
31
32 3.3.6. Detection of atypical ventricular morphologies. We introduced formal criteria for detection
33
of 4 types of atypical beats, pertinent to ‘Other’ arrhythmia class:
34
35 - QRS fragmentation (figure 6a): inversion of the slope to the left of the R-peak (R-fragm) that could
36
37 not be physiologically accepted as Q point due to:
38
(i) short interval between R-fragm and R-peak (<80ms)
pte

39
40
41
(ii) small amplitude drop at the point of the slope inversion (R-peak - R-fragm<30%QRSp-p).
42 The QRS fragmentation was detected as ‘present’ if both criteria were satisfied.
43
44 - Inverted QRS and T-wave (figure 6b): Opposite sign of T wave and maximal QRS peak.
45
46 - Left bundle branch block (LBBB) (figure 6c): a specific case of inverted QRS and T-wave. It was
ce

47
48 detected by two additional criteria: wide QRS>140ms and high-amplitude T-peak>1/3*QRSp-p.
49
50 - J-shift (figure 6d): Offset of J elevation or depression in respect to the Q point. The J-shift value
51
52 was reported in absolute units J-shift (mV) and normalized units J-shift (%), the latter normalized
Ac

53 to the QRSp-p amplitude.

54
55
56 12
57
58
59
60
Page 13 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 File: A00096 File: A00176 File: A00055 File: A00162
4 T-peak 0.5

pt
0.4
5 0.2
T-peak 1
6 0.3
R-fragm
0.8
average beat (mV)

7 0.1
0.6
8 0.2
0
9 0 0.4

cri
0.1
10 QRS>140ms
0.2
11 0 -0.1
0
12
-0.1 -0.2
13 -0.2
S
-0.5
S J
14 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8 0.2 0.4 0.6 0.8
-0.4
0.2 0.4 0.6 0.8
15 Time (s) Time (s) Time (s) Time (s)
16 (a) (b) (c) (d)

us
17 Figure 6. Four examples of average beats with atypical ventricular morphologies: (a) QRS
18
19 fragmentation, (b) Inverted QRS and T-wave, (c) LBBB, (d) J-depression.
20
21
22
23
24
25
26
27
28
an
3.3.7. Cross-correlation analysis. The morphologies of all beats were compared against the average
beat by maximal cross-correlation (corBeat). Statistics of corBeat as mean value, 25%, 50%
percentiles of all beats was used for quantification of the beat waveform variation. Smaller corBeat
dM
values were associated with larger waveform variations, usually induced by provisional atypical
29
30 morphologies of ventricular extrasystoles or artifacts.
31
32 3.3.8. Calculation of curvature. The curvature was defined as c=1/r, where r is the radius of the
33
34 circle, which best fits signal data x(t)-mean(x(t)) in the least mean square (LMS) sense, denoted in
35
36 Gander et al (1996) as ‘algebraic fit’. The solution of this problem was further described in details.
37
Let us first consider the algebraic representation of the circle in the time domain plain of signal x(t):
38
pte

39 (1) a(t 2  x 2 )  bt  cx  d  0
40
41 The LMS fit of a circle to number of N data samples (N>3) was solved by the linear system of
42
43 equations Xu  r , where the coefficients u  (a, b, c, d) T were calculated to minimize r , considering
44
45 the data vector X:
46
 t12  x12 1
ce

47 t1 x1
48  
49 (2) X   ..... ... ... ... .
50  
51 t 2  x2 tN xN 1
 N N 
52
Ac

53 To obtain a non-trivial solution, we imposed a constraint on u  1 , i.e. a 2  b 2  c2  d 2  1 .

54
55
56 13
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 14 of 29

1
2
3 The Matlab implementation of this solution used the embedded singular value decomposition
4

pt
5 function: [U,S,V] = svd(X,0), where the coefficients [a,b,c,d]=V(1:4,1) were calculated by choosing
6
7 the eigenvector from V with the largest eigenvalue.
8
9 Further, the coordinates of the center of the circle (t c , x c ) were calculated by considering a  0 and

cri
10
11 transforming eq. (1) to:
12
13 b 2 c b2  c2 d
14 (3) (t c  )  (x c  ) 2  - ,
2a 2a 4a 2 a
15
16

us
17 b c b2  c2 d
18 so that t c   , xc   and the radius of the circle is r  - .
2a 2a 4a 2 a
19
20 In our Challenge application, we defined N=5 points for curvature calculation over the average beat,
21
22 and further found the maximal curvatures during P, QRS, T waves, and their ratios.
23
24
25
26
27
28
an
Some examples are shown in figure 7. The curvature ratios maxc(QRS/P)=maxc(QRS)/maxc(P),
maxc(QRS/T)=maxc(QRS)/maxc(T) and maxc(T/P)=maxc(T)/maxc(P) quantify the relative activity
of different waves, used to distinguish abnormalities, e.g. small/missing P-waves (figure 7b for AF),
dM
29 extremely large P, T-waves (figure 7d for ‘Noise’), taking as a reference the examples of normal
30
31 activity (figure 7a,c for Normal and Other rhythm).
32
File: A02587, Annotation: 'AF'
33 0.4
File: A02088, Annotation: 'Normal'
0.4
File: A05720, Annotation: 'Other' File: A01585, Annotation: 'Noise'
maxc(QRS)
34 maxc(QRS)
0.8
maxc(QRS/P)=9 0.1
0.3 maxc(QRS/T)=7
35 0.3
0.7 maxc(T/P)=1.2
36 0.2
0.2
0.6 0.05 maxc(P)
maxc(QRS)
37 0.1
maxc(P) maxc(T) 0.5 maxc(T)
38 0.1
maxc(P) maxc(T) 0 0.4 0
pte

39 maxc(QRS)
(mV)
(mV)

(mV)

(mV)

0 -0.1 0.3
40
-0.2 0.2
41 -0.1
maxc(T)
-0.05
0.1
42 -0.3
-0.2 0
43 -0.4 maxc(P) -0.1
maxc(QRS/P)=31 -0.1 maxc(QRS/P)=0.9
44 -0.3 maxc(QRS/P)=14
-0.5 maxc(QRS/T)=7 maxc(QRS/T)=2
maxc(QRS/T)=8 -0.2
45 maxc(T/P)=1.7 maxc(T/P)=4 maxc(T/P)=0.4
-0.4
46 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
ce

Time (s) Time (s) Time (s) Time (s)

47
48 (a) (b) (c) (d)
49 Figure 7. Examples for average beat curvatures for 4 rhythm classes: (a) Normal rhythm (file
50
51 A02088), (b) AF rhythm (file A02587), (c) Other rhythm (file A05720), (d) Noise (file A01585). The
52
Ac

53 black curve represents the average beat; the blue curve represents its curvature.
54
55
56 14
57
58
59
60
Page 15 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 3.4. Analysis of atrial fibrillatory f-waves (AFF)
4

pt
5 The analysis of f-waves was based on our previous study (Christov et al 2001a, Christov et al 2001b),
6
7 which rejected all [Q; T-end] intervals, where f-waves have been masked by QRS and high T
8
9
amplitudes. All isoelectric intervals [T-end; Q] were merged in a continuous signal, which was

cri
10 subjected to DC filtering (1st difference in 20ms), rectification, and EMG filtering (moving average
11
12 over 30ms). The resultant TQ signal is shown in figure 8. Its median value (thick line in figure 8)
13
14 quantifies the isoelectric level of the ECG signal, shown to be more prominent while influenced by
15
the AFF activity during AF (figure 8a) than other arrhythmia and noise (figure 8b). The potential
16

us
17 sources of errors are: noisy environments causing high isoelectric levels during non-AF rhythms; or
18
19 insensitive ECG lead to atrial depolarization, leading to only little AFF activity during AF.
20
21 0.04
File: A00137, Annotation: 'AF' rhythm
0.04
File: A00056, Annotation: 'Noise'

22
23
24
0.03

an 0.03
TQ (mV)

TQ (mV)

0.02 0.02
25
26 0.01 0.01

27 0 AFF=11.7 uV 0
28 AFF=6.3 uV
dM
29 -0.01
0 2 4 6 8 10 12
-0.01
0 2 4 6 8 10 12 14 16
30 Time (s) Time (s)

31 (a) (b)
32 Figure 8. Merged TQ signals for the rhythms in figure 4: (a) AF rhythm (file A00137), (b) Noise (file
33
34 A00056). The median signal amplitude (thick line) is highlighted to quantify the AFF activity.
35
36
37
38 3.5. Linear Discriminant Analysis (LDA)
pte

39 We designed a four-class LDA classifier, considering rhythms: ‘Normal’, ‘AF’, ‘Other’, ‘Noise’
40
41 (defined in Table 1) and taking an input vector with 44 features (21 HRV, 22 average beat
42
43 morphology, 1 AFF), summarized in the feature extraction process in figure 1.
44
45
The classifier performance was estimated by the global Challenge F1 score (Clifford at al 2017):
46 3

 F1
ce

47 Rx
48 2TPRx
F1  x 1
, where F1Rx  is the specific F1 score for rhythm classes
49 3 TPRx  FN Rx  FPRx
50
51 Rx={‘Normal’, ‘AF’, ‘Other’}, calculated from the number of true positives (TP), false negatives
52
Ac

53 (FN) and false positives (FP) in each class.

54
55
We also reported the standard metrics of classifiers’ performance, including true positive rate
56 15
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 16 of 29

1
2
3 (TPRRx) and false positive rate (FPRRx) for rhythm class Rx, being comparable to general rhythm
4

pt
5 classification studies, non-participants in the PhysioNet/CinC Challenge 2017:
6
7 TPRx FPRx
TPRRx  100 * , (%) , FPR Rx  100 * , (%) .
8 TPRx  FN Rx FPRx  TN Rx
9

cri
10 Non-redundant LDA models were trained by stepwise feature selection and adjustment the prior
11
12 4

13
probabilities of 4 rhythm classes (  PRx  1 , where Rx={‘Normal’, ‘AF’, ‘Other’, ‘Noise’}), until
x 1
14
15 maximization of the global Challenge F1 score. Our top-scored model on training was further
16

us
17 evaluated by the Challenge organizers on a hidden test set for unbiased estimation of its accuracy.
18
19
20 4. Results
21
22
23
24
25
26
27
28
an
Using training data, we observed non-uniform distributions of 4 rhythm classes within different HR
ranges (figure 9), considering the mean HR value over the full recording length:
- Bradycardia (HR<50 bpm): 194 cases (2.3 % of total training database) with prevalent proportion
of Other rhythms (86%).
dM
29
- Normal HR (HR=50-100 bpm): 7710 cases (90.4 % of total training database) with prevalent
30
31 proportion of Normal (65%) and Other (25%) rhythms.
32
33 - Tachycardia (HR>100 bpm): 624 cases (7.3 % of total training database) with prevalent proportion
34
35 of Other rhythms (53%) and AF (37%).
36
37
38
pte

39
40
41
42
43
44
45
46
ce

47
48
49
50
51
52
Ac

53
Figure 9. Histogram of the mean HR categorized for 4 rhythm classes. The sum of all cases
54 (normalized proportions) within a HR range equals 100%.
55
56 16
57
58
59
60
Page 17 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3
4

pt
5
6 We considered relevant to design an LDA classifier, particular to HR range because LDA would best
7 fit to the prior probabilities of different rhythms appearing with such HR. Therefore, 3 LDA
8
9 classifiers, applicable to Bradycardia, Normal HR and Tachycardia were further optimized by

cri
10
11 stepwise feature selection for maximization of the F1 score. The contribution of the first 15 features
12
to the LDA performance is shown in table 2. We noted that the 1st ranked feature provided F1 score
13
14 in the range 0.56-0.61, which was about 60-76% of the maximal F1 score achieved by the respective
15
16 LDA classifier. The combination of the top-3 features in normal HR, top-9 features in Bradycardia

us
17
18 and top-5 features in Tachycardia contributed up to 90% of the maximal F1 score.
19 Figure 10 illustrates the statistical distributions of the top-5 features in LDA (Normal HR) that
20
21 highlight the potential of PNN50, SD1/SD2, dRRmean% to distinguish Normal and Other rhythms;
22
23
24
25
26
27
28
an
PNN50, dRRmean%, P-wave presence to distinguish AF; and corBeat(mean) to distinguish Noise.

Table 2. Evaluation of the top-15 features for 3 LDA classifiers. The features are listed in order of
their forward stepwise selection in each model. The features are ranked by global Challenge F1-score
dM
29
30 and its proportion from the classifier’s maximal F1 score.
31
32 LDA Normal HR (HR=50-100 bpm) LDA Bradycardia (HR50 bpm) LDA Tachycardia (HR100 bpm)
33 7710 cases 194 cases 624 cases
34 Stepwise F1 % Stepwise F1 % Stepwise F1 %
35 selection score max selection score max selection score max
36 of features (F1) of features (F1) of features (F1)
37 1. PNN50 (%) 0.579 76.3 RRmean (ms) 0.556 60.2 PNN50 (%) 0.608 76.2
38 2. SD1/SD2 0.654 86.2 RMSSD (%) 0.606 65.6 corBeat(25%) 0.693 86.8
pte

39 3. P-wave presence (y/n) 0.682 89.9 PQ (ms) 0.64 69.3 QRSwidth (ms) 0.702 88.0
40 4. dRRmean (%) 0.702 92.5 J-shift (mV) 0.693 75.0 T-peak (mV) 0.706 88.5
41 5. corBeat(mean) 0.715 94.2 maxc(QRS), (mV) 0.72 77.9 QT (ms) 0.725 90.9
42 6. corRR (%) 0.721 95.0 maxc(T/P) 0.763 82.6 QRSp-p (mV) 0.731 91.6
43 7. J-shift (%) 0.728 95.9 maxc(QRS/T) 0.783 84.7 RRmedian (ms) 0.75 94.0
44 8. QRSp-p (mV) 0.73 96.2 AFF (mV) 0.812 87.9 dRRmean (ms) 0.758 95.0
45 9. maxc(QRS/P) 0.734 96.7 Noise correction (y/n) 0.829 89.7 RRmeand (%) 0.763 95.6
46 10. AFF (mV) 0.739 97.4 maxc(QRS/P) 0.849 91.9 AFF (mV) 0.769 96.4
ce

47 11. QRSwidth (ms) 0.741 97.6 dRRstd (ms) 0.849 91.9 maxc(T/P) 0.773 96.9
48 12. CCM (%) 0.745 98.2 RMSSD (ms) 0.849 91.9 HRmean (bpm) 0.774 97.0
49 13. PQ (ms) 0.746 98.3 dRRmean (%) 0.87 94.2 RRmeand (ms) 0.784 98.2
50 14. J-shift (mV) 0.747 98.4 stdAmp/meanAmp (%) 0.883 95.6 maxc(QRS/P) 0.784 98.2
51 15. P-peak (mV) 0.749 98.7 P-wave presence (y/n) 0.865 93.6 LBBB (y/n) 0.784 98.2
52
Ac

53
54
55
56 17
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 18 of 29

1
2
3
4

pt
5
6
7
8
9

cri
10
11
12
13
14
15
16

us
17
18
19
20
21
22
23
24
25
26
27
28
an
Figure 10. Statistical distributions of the top-5 features, entered in the LDA model for classification
dM
29
30 of 90.4% of the training database, including 7710 cases with normal HR range (50-100 bpm).
31
32
33 Table 3. Performance of our official Challenge entry on the training database: Confusion matrix
34
35
(number of cases, % from total rhythm) and performance metrics (F1 score, TPR and FPR) for 4
36 rhythm classes. The highlighted entries of the confusion matrix correspond to TP, rhythm
37
38 classification (Total) corresponds to TP+FN, rhythm annotation (Total) corresponds to TP+FP.
pte

39
40 Confusion matrix
41 Performance metrics
Rhythm classification
42
43 F1 TPR FPR
Normal AF Other Noise Total
44 score (%) (%)
45 Normal 4572 32 422 50 5076 0.890 90.1 13.6
46
Rhythm annotation

(90%) (1%) (8%) (1%)

ce

47
AF 12 618 115 13 758 0.790 81.5 2.3
48
49
(1.5%) (81.5%) (15%) (2%)
50 Other 572 144 1635 64 2415 0.707 67.7 7.7
51 (23.5%) (6%) (68%) (2.5%)
52 Noise 37 12 36 194 279 0.647 69.5 1.5
Ac

53 (13%) (4%) (13%) (70%)

54
55
Total 5193 806 2208 321 8528 - - -
56 18
57
58
59
60
Page 19 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3
4

pt
5
6 Table 4. F1 score of the official Challenge entry.
7
8 Rhythm class Training set Test set (hidden)
9

cri
10 Normal 0.890 0.899
11 AF 0.790 0.809
12
13 Other 0.707 0.697
14
15 Overall 0.80 0.80
16

us
17
18
19 Detailed analysis of the training classification performance, gathering the decision of LDA
20 (Bradycardia) for 194 cases, LDA (Normal HR) for 7710 cases and LDA (Tachycardia) for 624 cases
21
22
23
24
25
26
27
28
an
is shown in table 3. Table 4 presents the final evaluation of our official phase entry, published online
by the PhysioNet/Computing in Cardiology Challenge (2017). The overall F1 score on the hidden test
database was 0.80, identifying the algorithm on 18th place, just 0.03 points lower than the winners in
the Challenge with score 0.83 (see table 5). The maximal running time of the algorithm was reported
dM
29 to be relatively low, taking <11% of the Challenge server quota.
30
31 In a further step, we ranked the most powerful feature extraction method for classification of atrial
32
fibrillation, i.e. based only on RR-interval measurements or beat morphology measurements. Figure
33
34 11 shows the performances of both feature extraction methods when independently applied on
35
36 different rhythms in the training database. In comparison to the official Challenge entry (using all
37
38 features), we observed certain deficiencies of the reduced feature set:
pte

39 - F1 drop was minimal for HRV (-0.02 to -0.06 points) and maximal for morphological features (-
40
41 0.15 to -0.22 points) in detection of specific rhythms (Normal, AF, Other).
42
43 - F1 drop was minimal for morphological features (-0.04 points) and maximal for HRV (-0.19
44
45
points) in Noise detection.
46
ce

47
48
49
50
51
52
Ac

53
54
55
56 19
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 20 of 29

1
2
3
4

pt
5
6
7
8
9

cri
10
11
12
13
14
15
16

us
17
18
19
20
21
22
23
24
25
26
27
28
an
Figure 11. Training F1 score of two basic feature extraction methods (Beat morphology and HRV)
compared to the official Challenge entry (using all features). F1 drop values: (all features vs. HRV)
and (all features vs. beat morphology) are indicated above the respective bars, evaluated for different
dM
29
rhythms in the training database.
30
31
32
33 5. Discussion
34
35 This study was an official participant in the PhysioNet/CinC Challenge 2017 dedicated to AF
36 classification from short single lead ECG recordings. It showed that a simple linear discriminant
37
38 model, including a few common HRV and average beat morphology features, was able to accurately
pte

39
40 classify AF with overall score F1=0.8 (Table 4), and specific AF score: F1 AF=0.81, TPRAF=81.5%,
41
42 FPR=2.3% (Table 3). The score was equal on both datasets, including training (8528 ECGs) and
43 hidden test (3658 ECGs). For comparison, the hidden test set scores of the top-9 Challenge
44
45 participants are presented in table 5. In summary, their methods were based on complex neural
46
networks either for feature extraction or classification. They computed a large constellation of ECG
ce

47
48
features – part of them involved complex transformations; some were presented by not formal
49
50 definitions. Many algorithms were rather Matlab based research models, without practical
51
52 perspectives for their real-life implementation in conventional point of care ECG diagnostic devices.
Ac

53
54 In contrast, the latter was the target of our cost-effective algorithm design, including clear and
55
56 20
57
58
59
60
Page 21 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 formally presented time-domain features based on standard mathematical computations that could be
4

pt
5 easily reproduced on any platform.
6
7
8 Table 5. Top-9 studies published in the PhysioNet/Computing in Cardiology Challenge 2017 official
9 score, evaluated on the hidden test set. The methods for feature extraction and classification are

cri
10 summarized for each study, using the available information and specific terminology in the
11 referenced publication.
12
13 F1 score
Rank Authors Feature extraction Classification
14 Normal AF Other Overall
15 1 Teijeiro et 79 features: Abductive ECG Recurrent Neural 0.90 0.85 0.74 0.83
16 al (2017) interpretation, morphology Networks,

us
17 features, HRV, signal quality; Tree Gradient
18 global and sequence (per beat) Boosting
19
classifications, classification (XGBoost),
20
stacking LDA
21
22
1 Datta et 150 features: morphology, AF, Multi-layer 0.92 0.82 0.75 0.83
23
24
25
26
27
28
1
al (2017)

Zabihi et
al (2017)
HRV, frequency, statistical,

Spectrogram based noise

removal
491 features: time domain,
frequency domain, time-
an
abnormalities, noise detection;
cascaded binary
classifier,
Adaptive boosting
(AdaBoost)
LDA, quadratic
discriminant
0.91 0.84 0.73 0.83
dM
29 frequency domain, analysis QDA,
30 morphology, phase space Random forest
31 reconstruction
32 1 Hong et Expert features from statistical, Deep Neural 0.91 0.81 0.75 0.83
33 al (2017) signal processing and medical Networks (DNNs)
34 area, features from the most
35 representative wave, features
36
extracted by DNNs
37
5 Zihlmann Logarithmic ECG spectrogram, Deep convolutional 0.91 0.82 0.73 0.82
38
et al stack of convolutional layers neural networks
pte

39
40 (2017) for feature extraction, (CNN)
41 aggregation of features across Deep convolutional
42 time by averaging, recurrent neural
43 linear classifier networks (CRNN)
44 5 Bin et al 30 features: AF, morphology, Decision tree 0.91 0.82 0.73 0.82
45 (2017) RR intervals, similarity index ensemble
46 between beats (AdaBoost)
ce

47 5 Xiong et Feature extraction by CNN 16-layer CNN 0.90 0.82 0.73 0.82
48 al (2017) convolution layer, based on
49 sliding 15x1 filter with
50 trainable weights;
51 pooling layers for feature
52 reduction; fully connected
Ac

53
layer for rhythm class
54
probability calculation
55
56 21
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 22 of 29

1
2
3 9 Kropf et 380 features: beat classification Random forest 0.90 0.83 0.72 0.81
4 al (2017) by cross-correlation, RR-

pt
5
intervals, morphology features,
6
atrial f-waves, QRS quality
7
8 9 Plesinger 120 features: frequency CNN and bagged 0.91 0.80 0.74 0.81
9 et al envelograms (1-5 Hz, 5-10 Hz, tree ensemble

cri
10 (2017) … , 35-40 Hz), general signal
11 features, RR-intervals,
12 statistical description of
13 average beat, statistical
14 features from CNN, noise
15 detection
16

us
17 18 This study 44 features: HRV, average beat LDA 0.90 0.81 0.70 0.80
18 morphology, atrial f-waves
19
20
21
22 The F1 score drop by 0.03 points with respect to the winners in the Challenge was mainly due to a
23
24
25
26
27
28
an
proportional performance limitation for AF and Other arrhythmia. This was evident in the confusion
matrix (Table 3) with largest errors: 23.5% of Other arrhythmia were classified as Normal rhythm,
15% of AF were classified as Other arrhythmia. We suggest that more features, quantifying the
dM
29
morphological diversity of various rhythm disorders included in Other arrhythmia, would be a
30 potential strategy for improving the performance. For example, an extended set of morphological and
31
32 time-frequency beat descriptors via Matching Pursuits decomposition were shown in our previous
33
34 study (Christov et al 2006) to be a powerful approach for detection of the most common heartbeat
35
types (normal beats, premature ventricular contractions (PVCs), left and right bundle branch blocks,
36
37 paced beats). These refined criteria were not included in our Challenge algorithm because they
38
pte

39 required suitable training dataset with annotation of heartbeats by their type, while the particular AF
40
41 detection task was related to annotation of long episodes. We suggest that training on additional data
42 with significant PVCs could be a good starting point for refined classification of Other arrhythmia,
43
44 recognizing this very common and mostly harmless rhythm abnormality.
45
46 On the Challenge training data, our feature selection process was separately performed for
ce

47
48 bradycardia, normal HR and tachycardia (table 2), taking into consideration that:
49
50 - The prior probabilities of 4 rhythm classes were non-uniformly distributed from slow to rapid
51
52
rhythms (figure 9);
Ac

53 - The morphology of the beats has been found to vary depending on the heart rate, so that HR
54
55 increment was significantly associated with shortening of RR, PR, QRS, QT intervals, decrement
56 22
57
58
59
60
Page 23 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 of R-wave amplitude and more negative S-wave in lateral leads, slightly more negative Q wave,
4

pt
5 increment of P-wave and T-wave amplitudes, J-point and ST-level changes (Clifford 2006).
6
7 - HRV was significantly associated with average heart rate that is not only a physiological
8
9
phenomenon but also due to non-linear (mathematical) relationship between RR intervals and HR

cri
10 (Sacha and Pluta 2007).
11
12 Thus, following the strategy for heart-rate specific design of LDA models, we found different
13
14 features which best separate the rhythm classes for slow, normal and rapid rhythms (table 2):
15
- Normal HR (HR=50-100 bpm). The top-5 features, which together contribute to about 94% of
16

us
17 maximal F1 score were 3 HRV and 2 beat morphological features, ranked in the order:
18
19 ▪ PNN50 (76.3% of max F1 score). According to figure 10, PNN50 exhibited the largest
20
21 separation between different rhythm classes with median values: 6.8% (Normal rhythm), 28%
22
23
24
25
26
27
28
an
(Other rhythm), 53.3% (Noise) and 78.8% (AF). Thus the proportion of RR-intervals differing
by >50ms from the preceding RR interval was found to best distinguish the intrinsic rhythm
irregularity in AF (about 2.8 to 12 times superior than Other and Normal rhythms,
respectively).
dM
29 ▪ SD1/SD2 (adds +9.9% of max F1 score). According to figure 10, short-to-long term HRV was
30
31 minimal for Normal rhythms (median value 0.305) and maximal for Other rhythms (0.633),
32
33 thus found to be the most powerful feature for separation of those two classes. The
34
phenomenon of lower SD1/SD2 for AF (0.5) was due to the combination of both large short-
35
36 term and large long-term HRV, while Other rhythms manifested a combination of large short-
37
38 term and low long-term HRV.
pte

39
40 ▪ P-wave presence (adds +3.7% of max F1 score). According to figure 10, the absence of P-
41 waves in the average beat was an outstanding feature for recognition of AF. We have
42
43 considered reliable the developed algorithm for P-wave detection, which found P-waves in
44
45 only 5% of AF, being about 9, 11 and 15 times rare than Noise (46%), Other rhythms (56%)
46
and Normal rhythms (75%), respectively.
ce

47
48 ▪ dRRmean% (adds +2.6% of max F1 score). According to figure 10, the distribution of
49
50 dRRmean% was similar to PNN50, revealing the trend for maximal RR-interval first
51
52 differences in AF and Noise, being 2.4 to 8 times larger than Other and Normal rhythms,
Ac

53 respectively.
54
55 ▪ corBeat(mean) (adds +1.7% of max F1 score). According to figure 10, the mean correlation
56 23
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 24 of 29

1
2
3 of all beats against the average beat was an outstanding feature for Noise detection. The
4

pt
5 largest waveform variations during noise were quantified by significant drop of inter-beat
6
7 cross correlation (median value of 0.76), which was about 0.22 points lower than all other
8
9
noise-free rhythms (0.98).

cri
10 - Bradycardia (HR<50 bpm): The top-5 features, which together contributed to about 78% of
11
12 maximal F1 score were 2 HRV and 3 beat morphology features, ranked in the order: RRmean
13
14 (60.2% of max F1 score), RMSSD% (+5.5% of max F1), PQ interval (+3.7% of max F1), J-shift
15
(+5.7% of max F1), maxc(QRS) (+2.9% of max F1).
16

us
17 - Tachycardia (HR>100 bpm): The top-5 features, which together contributed to about 91% of
18
19 maximal F1 score were 1 HRV and 4 beat morphology features, ranked in the order: PNN50
20
21 (76.2% of max F1 score), corBeat(25%) (+10.6% of max F1), QRS width (+1.2% of max F1), T-
22
23
24
25
26
27
28
an
peak amplitude (+0.5% of max F1), QT-interval (+2.4% of max F1).
Although the presence of f-waves was found to be a key feature for AF detection in our previous
studies (Christov et al 2001a, Christov et al 2001b), the equivalent feature AFF in this study had a
relatively low impact on the overall accuracy (+0.5% of max F1), ranked 10th by the LDA classifier
dM
29 (table 2). We suggested that the use of single-lead ECG via the fingers of both arms (lead I) was
30
31 slightly sensitive to the main f-wave component, typically observed in leads V1, II and aVF. In our
32
33 opinion, the hand-held two-electrode technique for self-monitoring of AF at home could not rely on
34
the standard criteria for analysis of the P-wave activity via f-waves.
35
36 After having computed a lot of features, the global rank of the feature extraction methods clearly
37
38 indicated that the most important ones were based on RR interval changes and, secondary, beat
pte

39
40 morphology (figure 11). HRV features alone were able to provide a simple and accurate rhythm
41 classification (0.06 points below maximal F1 score). The added value of the more complex ECG
42
43 waveform analysis was negligible for Normal rhythms (0.02 points), slightly increasing for AF and
44
45 Other rhythms (0.06 to 0.08 points) and most significant for the detection of Noise (0.19 points). This
46
might open the discussion about the design of a cost-effective device for AF detection based only on
ce

47
48 RR-interval analysis if the recording conditions might be rigorously controlled and not interrupted by
49
50 artifacts. The “best” wearable arrhythmia detection device should be, however, robust to noise and
51
52 maximally accurate, therefore, incorporating both HRV and morphology features. Among the latter,
Ac

53 we could distinguish the P-wave absence in the average beat sensitive to AF, as well as the inter-beat
54
55 correlation coefficient sensitive to ECG morphological instability. Such instability might warn the
56 24
57
58
59
60
Page 25 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 user to double-check the hand-held electrode contact and repeat the measurement, thus avoiding
4

pt
5 arrhythmia classification errors on diagnostically unacceptable episodes with artifacts that commonly
6
7 disturb the manually controllable signal acquisition by non-professionals.
8
9

cri
10
11
12
13
14
15
16

us
17
18
19
20
21
22
23
24
25
26
27
28
an
dM
29
30
31
32
33
34
35
36
37
38
pte

39
40
41
42
43
44
45
46
ce

47
48
49
50
51
52
Ac

53
54
55
56 25
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 26 of 29

1
2
3 References
4

pt
5
6 Akyürek O, Diker E, Güldal M and Oral D 2003 Predictive value of heart rate variability for the
7 recurrence of chronic atrial fibrillation after electrical cardioversion. Clin Cardiol. 26(4):196-200,
8
9 doi: 10.1002/clc.4960260411

cri
10
11 Barauskiene V, Rumbinaite E, Karuzas A, Martinkute E and Puodziukynas A 2016 Importance of
12
13 heart rate variability in patients with atrial fibrillation. J Cardiol Clin Res 4(6): 1080.
14
15 Bin G, Shao M, Bin G, Huang J, Zheng D and Wu S 2017 Detection of atrial fibrillation using
16

us
17 decision tree ensemble. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.342-204
18
19 Camm AJ, Kirchhof P, Lip GY, Schotten U, Savelieva I, Ernst S et al. 2010 Guidelines for the
20
21 management of atrial fibrillation: the Task Force for the Management of Atrial Fibrillation of the
22
23
24
25
26
27
28
an
European Society of Cardiology (ESC), Eur. Heart J. 31 2369−429

Christov I, Bortolan G and Daskalov I 2001a Automatic detection of atrial fibrillation and flutter by
wave rectification method. J. Med. Eng. Tech. 25 217−21
dM
29 Christov I, Bortolan G and Daskalov I 2001b Sequential analysis for automatic detection of atrial
30
31
fibrillation and flutter. Comput. Cardiol. 28 293−6
32
33 Christov I and Simova I 2006 Fully Automated Method for QT Interval Measurement in ECG.
34
Comput. Cardiol. 33 321−34
35
36
37 Christov I, Gómez-Herrero G, Krasteva V, Jekova I, Gotchev A and Egiazarian K 2006 Comparative
38
study of morphological and time-frequency ECG descriptors for heartbeat classification. Medical
pte

39
40 Engineering & Physics, 28, 876−87, doi:10.1016/j.medengphy.2005.12.010
41
42 Christov I, Krasteva V, Simova I, Neycheva T and Schmid R 2017 Multi-parametric Analysis for
43
44 Atrial Fibrillation Classification in ECG. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.175-
45
46 021
ce

47
48 Clifford GD 2006 ECG Statistics, Noise, Artifacts, and Missing Data – Chapter 3. In: Advanced
49
50 methods and tools for ECG data analysis, Artech House, Inc. Norwood, MA, USA, ISBN-13:
51
978-1580539661
52
Ac

53
54 Clifford GD, Liu CY, Moody B, Lehman L, Silva I, Li Q, Johnson AEW and Mark RG 2017 AF
55
56 26
57
58
59
60
Page 27 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 classification from a short single lead ECG recording: The Physionet Computing in Cardiology
4

pt
5 Challenge 2017. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.065-469
6
7 Daskalov IK and Christov II 1999a Electrocardiogram signal preprocessing for automatic detection
8
9 of QRS boundaries. Med. Eng. Phys. 21 37-44

cri
10
11 Daskalov IK and Christov II 1999b Automatic detection of the electrocardiogram T-wave end. Med.
12
13 Biol. Eng. Comput. 37 348-53
14
15 Datta S, Puri C, Mukherjee A, Banerjee R, Choudhury AD, Singh R, Ukil A, Bandyopadhyay S, Pal
16

us
17 A and Khandelwal S 2017 Identifying normal, AF and other abnormal ECG rhythms using a
18
19 cascaded binary classifier. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.173-154
20
21 Dotsinsky I 2007 Atrial wave detection algorithm for discovery of some rhythm abnormalities,
22
23
24
25
26
27
28
Physiol. Meas. 28 595−610

an
Gander W, Golub GH and Strebel R 1996 Least-Squares Fitting of Circles and Ellipses. Bulletin of
the Belgian Mathematical Society. 3Suppl 63−84,
dM
https://www.emis.de/journals/BBMS/Bulletin/sup962/gander.pdf
29
30
31
Golińska A 2013 Poincaré Plots in analysis of selected biomedical signals. Studies in Logic,
32 Grammar and Rhetoric 35 117−127
33
34
Hong S, Wu M, Zhou Y, Wang Q, Shang J, Li H and Xie J 2017 ENCASE: an ENsemble ClASsifiEr
35
36 for ECG classification using expert features and deep neural networks. Comput. Cardiol. 44 1–4,
37
38 doi: 10.22489/CinC.2017.178-245
pte

39
40 Karmakar C, Khandoker A, Voss A and Palaniswami M 2011 Sensitivity of temporal heart rate
41
42 variability in Poincaré plot to changes in parasympathetic nervous system activity. Biomed. Eng.
43
44 OnLine 10 17
45
46 Kim D, Seo Y, Youn CH 2008 Detection of atrial fibrillation episodes using multiple heart rate
ce

47
48
variability features in different time periods. Conf Proc IEEE Eng Med Biol Soc. 2008:5482-5,
49 doi: 10.1109/IEMBS.2008.4650455
50
51
Krasteva V, Jekova I and Christov I 2006 Automatic detection of premature atrial contractions in the
52
Ac

53 electrocardiogram. Electrotechnica & Electronica E+E 9-10/2006 49-55

54
55
56 27
57
58
59
60
 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2 Page 28 of 29

1
2
3 Kropf M, Hayn D and Schreier G 2017 ECG Classification based on time and frequency domain
4

pt
5 features using random forests. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.168-168
6
7 Ladavich S and Ghoraani B 2015 Rate-independent detection of atrial fibrillation by statistical
8
9 modeling of atrial activity. Biomed. Signal Process, Control 18 274−81

cri
10
11 Lip GYH, Fauchier L, Freedman SB, Van Gelder I, Natale A, Gianni C et al. 2016 Atrial fibrillation,
12
13 Nat. Rev. Dis. Primers 2 16016
14
15 Lombardi F, Colombo A, Basilico B, Ravaglia R, Garbin M, Vergani D, Battezzati PM and Fiorentini
16

us
17 C 2001 Heart rate variability and early recurrence of atrial fibrillation after electrical
18
19 cardioversion. J Am Coll Cardiol. 37(1) 157-62, doi: 10.1016/S0735-1097(00)01039-1
20
21 Oster J and Clifford GD 2015 Impact of the presence of noise on RR interval-based atrial fibrillation
22
23
24
25
26
27
28
detection. J. Electrocard. 48 947−51

230−6
an
Pan J and Tompkins WJ 1985 A real-time QRS detection algorithm. IEEE Trans. Biomed. Eng. 32
dM
29 Park J, Lee S and Jeon M 2009 Atrial fibrillation detection by heart rate variability in Poincare plot.
30
31
Biomed Eng Online 8: 38, doi: 10.1186/1475-925X-8-38
32
33 Petrėnas A, Marozas V, Sörnmo L and Lukosevicius A 2012 An echo state neural network for QRST
34
cancellation during atrial fibrillation. IEEE Trans. Biomed. Eng. 59 2950−7
35
36
37 PhysioNet/Computing in Cardiology Challenge 2017 A more detailed breakdown of performance of
38
each official algorithm. URL:
pte

39
40 https://physionet.org/challenge/2017/results_all_F1_scores_for_each_classification_type.csv,
41
42 Published online 10 October 2017.
43
44 Plesinger F, Nejedly P, Viscor I, Halamek J and Jurak P 2017 Automatic detection of atrial
45
46 fibrillation and other arrhythmias in Holter ECG recordings using PQRS morphology and rhythm
ce

47
48
features. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.364-057
49
50 Sacha J and Pluta W 2007 Alterations of an average heart rate change heart rate variability due to
51
mathematical reasons. Internat. J. Cardiol. 128 444–447, doi:10.1016/j.ijcard.2007.06.047
52
Ac

53
54 Teijeiro T, Garcia CA, Castro D and Félix P 2017 arrhythmia classification from the abductive
55
56 28
57
58
59
60
Page 29 of 29 AUTHOR SUBMITTED MANUSCRIPT - PMEA-102425.R2

1
2
3 interpretation of short single-lead ECG records. Comput. Cardiol. 44 1–4, doi:
4

pt
5 10.22489/CinC.2017.166-054
6
7 The Task Force of ESC and NASPE. 1996 Heart rate variability. Standards of measurement,
8
9 physiological interpretation, and clinical use. Eur. Heart J. 17 354−81

cri
10
11 van den Berg MP, Haaksma J, Brouwer J, Tieleman RG, Mulder G and Crijns HJGM 1997 Heart rate
12
13 variability in patients with atrial fibrillation is related to vagal tone. Circulation 96 1209-16, doi:
14
15 10.1161/01.CIR.96.4.1209
16

us
17 Xiong Z, Stiles M and Zhao J 2017 Robust ECG signal classification for the detection of atrial
18
19 fibrillation using novel neural networks. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.066-
20 138
21
22
23
24
25
26
27
28
1–4, doi: 10.22489/CinC.2017.069-336
an
Zabihi M, Rad AB, Katsaggelos AK, Kiranyaz S, Narkilahti S and Gabbouj M 2017 Detection of
atrial fibrillation in ECG hand-held devices using a random forest classifier Comput. Cardiol. 44
dM
Zhou X, Ding H, Wu W and Zhang Y 2015 A Real-Time Atrial Fibrillation Detection Algorithm
29
30 Based on the Instantaneous State of Heart Rate. PLoS One 10(9):e0136544, doi:
31
32 10.1371/journal.pone.0136544
33
34 Zihlmann M, Perekrestenko D, Tschannen M 2017 Convolutional recurrent neural networks for
35
36 electrocardiogram classification. Comput. Cardiol. 44 1–4, doi: 10.22489/CinC.2017.070-060
37
38 Zoni-Berisso M, Lercari F, Carazza T and Domenicucci S 2014 Epidemiology of atrial fibrillation:
pte

39
40 European perspective. Clin. Epidem. 6 213−20
41
42
43
44
45
46
ce

47
48
49
50
51
52
Ac

53
54
55
56 29
57
58
59
60