Accepted Manuscript

Automated real-time method for ventricular heartbeat classification

Silvia Ortı́n, Miguel C. Soriano, Miquel Alfaras, Claudio R. Mirasso

PII: S0169-2607(18)30245-1
DOI: https://doi.org/10.1016/j.cmpb.2018.11.005
Reference: COMM 4815

To appear in: Computer Methods and Programs in Biomedicine

Received date: 21 February 2018

Revised date: 31 August 2018
Accepted date: 19 November 2018

Please cite this article as: Silvia Ortı́n, Miguel C. Soriano, Miquel Alfaras, Claudio R. Mirasso, Auto-
mated real-time method for ventricular heartbeat classification, Computer Methods and Programs in
Biomedicine (2018), doi: https://doi.org/10.1016/j.cmpb.2018.11.005

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights

• A fully automated and real-time heartbeat ventricular arrhythmia classi-

fier based on a single lead designed to be fast and efficient in the training
and evaluation phase.

T
IP
• The arrhythmia classifier can be applied to different ECG leads with ex-
cellent performance. Potential application to wearables with unconven-

CR
tionally placed electrodes.

• Trained and evaluated with different databases, allowing the inter-patient

and inter-database classification.

US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT

Automated real-time method for ventricular heartbeat

classification

Silvia Ortı́na,∗, Miguel C. Sorianoa , Miquel Alfarasa,1 , Claudio R. Mirassoa

T
a Instituto de Fı́sica Interdisciplinar y Sistemas Complejos, IFISC (CSIC-UIB), Campus
Universitat de les Illes Balears, E-07122 Palma de Mallorca, Spain

IP
CR
Abstract

Background and Objective: In this work, we develop a fully automatic and

US
real-time ventricular heartbeat classifier based on a single ECG lead. Single
ECG lead classifiers can be especially useful for wearable technologies that
provide continuous and long-term monitoring of the electrocardiogram. These
AN
wearables usually have a few non-standard leads and the quality of the signals
depends on the user physical activity.
Methods: The proposed method uses an Echo State Network (ESN) to clas-
M

sify ECG signals following the Association for the Advancement of Medical
Instrumentation (AAMI) recommendations with an inter-patient scheme. To
ED

achieve real-time classification, the classifier itself and the feature extraction
approach are fast and computationally efficient. In addition, our approach al-
lows transferring the knowledge from one database to another without additional
PT

training.
Results: The classification performance of the proposed model is validated on
the MIT-BIH arrhythmia and INCART databases. The sensitivity and precision
CE

of the proposed method for MIT-BIH arrhythmia database are 95.3 and 88.8 for
the modified lead II and 90.9 and 89.2 for the V1 lead. The results reported are
AC

∗ Corresponding author. Telephone number: +34 971259629

Email addresses: silvia@ifisc.uib-csic.es (Silvia Ortı́n), miguel@ifisc.uib-csic.es
(Miguel C. Soriano), malfaras@plux.info (Miquel Alfaras), claudio@ifisc.uib-csic.es
(Claudio R. Mirasso)
1 Present address: PLUX-Wireless Biosignals, S.A., Av. 5 de Outubro 70, 1050-059, Lisbon,

Portugal

Preprint submitted to Computer methods and programs in biomedicine November 20, 2018
 ACCEPTED MANUSCRIPT

further compared to the existing methodologies in literature. Our methodology

is a competitive single lead ventricular heartbeat classifier, that is comparable
to state-of-the-art algorithms using multiple leads.
Conclusions: The proposed fully automated, single-lead and real-time heart-

T
beat classifier of ventricular heartbeats reports an improved classification accu-

IP
racy in different leads of the evaluated databases in comparison with other sin-
gle lead heartbeat classifiers. These results open the possibility of applying our

CR
methodology to wearable long-term monitoring devices with an unconventional
placement of the electrodes.
Keywords: biomedical signal processing, ECG heartbeat classification,
reservoir computing, template matching
US
AN
M
ED
PT
CE
AC

3
 ACCEPTED MANUSCRIPT

1. Introduction

Long-term electrocardiogram (ECG) monitoring is the most common nonin-

vasive tool to diagnose patients with arrhythmias. Arrhythmias produce heart-
beats with unusual timing or unusual electrocardiogram morphology that can

T
5 be detected by expert cardiologists. Nevertheless, visual analysis of long-term

IP
ECG recordings is tedious and time-consuming. Thus, the development of com-
puter aided techniques to efficiently process the large amounts of data stored

CR
by long-term ECG monitoring has been an active area of research over the last
decades.
Most of the proposed algorithmic techniques differ in two main aspects:

US
10

the discriminating features extracted from the ECG for classifying the heart-
beats and the classifier algorithm (an extensive review can be found in [1]).
AN
Widespread discriminating features are those extracted from the time domain
[2, 3, 4], features obtained from the frequency domain [5, 6, 7, 8, 9], wavelet
15 transforms [10, 11, 12, 13, 14, 15, 16] or higher order statistics [7, 8, 12, 14].
M

Popular choices for the learning algorithm include Neural Networks [13, 14, 15,
16, 17, 18], Support Vector Machines [3, 4, 6, 9, 10, 14], conditional random
fields [19], linear discriminant analysis [2, 12, 20], decision trees [7, 8, 20] and
ED

deep learning [21, 22, 23, 24].

20 Nowadays there is an increasing development of compact long-term ECG
PT

monitoring devices (often wearables). These devices demand real-time classifi-

cation without expert assistance. However, they usually have one or a limited
number of non-standard leads that can be significantly noisy. This can invalidate
CE

the reliability of many of the methods proposed so far for heartbeat classifica-
25 tion. Most of the heartbeat arrhythmia classifiers use a combination of leads II
and V1 and little work has been done using a single lead for ECG arrhythmia
AC

classification. Moreover, single lead classifiers are usually lead-specific and do

not perform well for other leads [4].
Another well-known problem of heartbeat classification is the huge inter-
30 patient variability of ECG signals. Heartbeats of the same class can exhibit

4
 ACCEPTED MANUSCRIPT

significant differences among patients. Patient-adapted algorithms can deal with

the inter-patient variability and outperform non-patient-adapted methods [11].
However, most of the previously proposed patient-adapted classifiers require
local expert assistance (semi-automatic approaches), which renders unpractical

T
35 the real-time applications [25]. To achieve a fully automatic patient-adapted

IP
heartbeat classifier we use Template Matching (TM) [26]. The idea behind TM
is to compare the heartbeats with a template specific for a certain class. The

CR
TM allows the classifier to be lead-independent and suitable for non-standard
leads.
40 The classifier is an Echo State Network (ESN) with a ring topology [27].

US
ESNs are Recurrent Neural Networks where only the weights of the output layer
are trained [28]. This makes them attractive for handling data in real-time and
simplifies enormously their hardware implementation [29, 30].
AN
In this work, we present a fully automated, single-lead and real-time heart-
45 beat classifier that can be useful for long-term ECG monitoring devices. Our
model discriminates between heartbeats with supraventricular origin (normal
M

morphology) and ventricular origin (abnormal morphology). The rest of the

manuscript is organized as follows. In section 2 we give a detailed description of
ED

the proposed classifier. The results and comparison with other state-of-the-art
50 approaches are presented in Section 3 and conclusions are provided in Section
4.
PT

2. Materials and Methods

CE

The proposed scheme for the automatic heartbeat classifier is presented in

Figure 1. The processing stage includes filtering the ECG recording, heartbeat
detection and segmentation. Each heartbeat is then characterized by a set of
AC

55

morphological and dynamical features that constitute the input to the classifier.
Finally, the model is evaluated over two databases different to those used to train
the model.

5
 ACCEPTED MANUSCRIPT

Interpolation
NO 250 Hz Heartbeat Normalized
ECG Signal
ECG detection & segmented
Filtering
Acquisition segmentation heartbeat
Sampling
= 250Hz?

YES

T
Feature

IP
Extraction
F11 F12 F13 ...
.
F21 F22 F23 ...
F31 F32 F33 ...

CR
........
1. TRAINING 2. TESTING

Features Features Features Features

from AHA from MIT-SV from MIT-AR from INCART
database

Lead 1 Lead 2
database

Lead 1 Lead 2 US database

Lead 1 Lead 2 Lead 1

database

Lead 12
AN
Training ESN ESN ESN ESN
class class class class class class
SVB' VB' SVB' VB' SVB'
ESN parameters VB'
M

Figure 1: Schematic of the heartbeat classifier.

ED

2.1. ECG Databases

60 The heartbeat classifier is trained and evaluated on different databases (see

Figure 1). The ECG recordings of the American Heart Association ventricu-
PT

lar arrhythmia ECG database (AHA) and MIT-BIH Supraventricular Arrhyth-

mia database (MIT-SV) [31, 32] are used for training whereas the classifier is
CE

evaluated on the MIT-BIH Arrhythmia (MIT-AR) [32, 33] and INCART [32]
65 databases. Information about these databases is summarized in Table 1. The
inter-database approach guarantees the generalization capability of the classifier
AC

to other patients (inter-patient approach) and databases.

The original database beat annotation labels are converted to the five types
recommended by AAMI [34]: N (Non-ectopic beat), SVEB (Supraventricular
70 ectopic beat), VEB (Ventricular ectopic beat), F (Fusion beat) and Q (Unclas-

6
 ACCEPTED MANUSCRIPT

Table 1: Databases used for training and testing the heartbeat classifier
Training Test
Database AHA MIT-SV MIT-AR INCART
No. ECG 155 78 48 75

T
No. Patients - - 47 32

IP
No. Leads 2 2 2 12
Sampling (Hz) 250 128 360 257

CR
No. SVB’ 317612 174317 92754 155624
No. VB’ 32403 9953 7803 20227

US
sified and paced beats). Our goal is a binary class model to distinguish between
beats of normal and abnormal morphology:
AN
• SVB’ class (label 0): beats with normal morphology and supraventricular
origin (N+SVEB).

75 • VB’ class (label 1): beats with abnormal morphology and ventricular ori-
M

gin (VEB+F).

The annotated ECG recordings are 30 minutes long. Recordings with paced
ED

beats are omitted in agreement with the AAMI recommended practice. Table
1 shows the number of annotated heartbeats analyzed in each database.
The MIT-AR database has two leads: the modified lead II (MLII) and mostly
PT

80

V1 (some patients have V2 or V5 leads instead of V1). The MIT-SV ECG

recordings were chosen to supplement the MIT-AR database, making it likely
CE

that they present the same origin. There is no information about the origin
of the AHA leads. The INCART database has the 12 standard leads obtained
85 from an ECG recorder with 10 electrodes.
AC

2.2. ECG Signal Preprocessing and Heartbeat Segmentation

ECG recordings are processed with a common sampling rate of 250 Hz. The
databases are linearly interpolated to 250 Hz using the PhysioToolkit software
[32].

7
 ACCEPTED MANUSCRIPT

90 Each lead of the ECG recording is filtered in a bandwidth range of 0.5-

35 Hz to remove unwanted artifacts, high-frequency noise and to correct the
baseline wander. We use a second order Butterworth high-pass filter with a
cut-off frequency of 0.5 Hz and a finite impulse response low-pass filter of 12th

T
order with a cut-off frequency of 35 Hz. The power spectra of the ECG does

IP
95 not contain significant information beyond 35 Hz [35] and frequencies below 0.5
Hz are not relevant for arrhythmia classification [36].

CR
The ECG recording is segmented into heartbeats by extracting 0.52 seconds
of signal surrounding each heartbeat annotation (130 samples at 250 Hz). Af-
ter assessing suitable window durations within the normal interval limits, we
100

US
found in the data that a window size of 0.52 s captures most of P and T waves
information while avoiding misleading information by neighboring heartbeats.
Each segmented heartbeat includes all the samples collected from 0.2 seconds
AN
before to 0.32 seconds after the fiducial point that determines the position of
the heartbeat. The segmented heartbeats are individually normalized to zero
105 mean and standard deviation one.
M

The position of the heartbeat can be determined using an automatic heart-

beat detection algorithm [37]. In this work, we use the temporal heartbeat
ED

locations provided by the databases. We have adjusted the original positions

of the training databases AHA and MIT-SV to the local extremum of the QRS
110 complex as in the MIT-AR database.
PT

The specific fiducial points used to determine the position of the heartbeat
are irrelevant for the processing as long as they are all at the same location with
CE

respect to the corresponding heartbeat waveform. Therefore, for the evaluation

of the INCART database, we use the point of minimum derivative within a small
115 window around the annotated fiducial point to detect the heartbeat instead of
AC

the local extremum of the QRS complex used in the training databases and the
MIT-AR. In this way, we test the robustness of the method to changes in the
fiducial point.

8
 ACCEPTED MANUSCRIPT

2.3. Feature Extraction

120 Our aim is to design a fast and real-time heartbeat classifier to distinguish
between SVB’ and VB’ classes using information from a single ECG lead. The
TM together with some statistical features extracted from the time domain

T
(chosen to be robust against noise and as lead-independent as possible) are used

IP
as discriminant features of our automatic and real-time classifier. The selected
125 features that characterize the heartbeat should not imply a high computational

CR
cost to support the real-time implementation. They should also be independent
of the ECG lead to allow non-standard lead placements. We also avoid the
calculation of fiducial points that can be compromised in presence of noise.

130 US
It is well known that Normal heartbeats (N) are characterized by regular
cardiac cycles (RR intervals), the presence of P wave and narrow QRS. In con-
trast, VB heartbeats are characterized by shorter RR intervals, the absence of
AN
P wave and wider QRS. As the P wave is not present in some leads, we do not
consider the presence or absence of P wave as a discriminant feature for our
classifier.
M

135 The RR interval is defined as the time interval between successive heartbeats.
The features related to the RR intervals used by our classifier are:
ED

1. Previous RR interval (the time difference between the beat and the pre-
vious beat).
2. Next RR interval.
PT

140 3. Average over the past 500 RR intervals. For the first 500 heartbeats of the
ECG recording, this average is calculated using only the available previous
CE

RR intervals.
4. Standard Deviation of Successive Differences (SDSD) between 5 adjacent
RR intervals. SDSD is a standard feature to classify arrhythmic ECG
AC

145 segments [38].

The area under the heartbeat waveform and its derivative are also used as
discriminant features:

5. Average of the absolute value of the segmented beat: mean(|Beat|).

9
 ACCEPTED MANUSCRIPT

6. Average of the absolute value of the derivative of the segmented beat:

150 mean(|Beat0|).

where the derivative is calculated with the Central-Difference Formula (f 0(x) ≈

(f (x + h) − f (x − h))/(2h)). We expect that VB’ beats exhibit a broad area

T
because their associated QRS complexes are wider. VB’ beats also tend to rise

IP
and fall slower than SVB’ signals. Hence, the area under the derivative should
155 be smaller.

CR
The previous features are fast and easy to compute and lead-independent
but they do not account for the inter-patient variability. To that end, we use
the template matching as a simple and effortless patient-adaptable approach.

160
US
It is based on evaluating a similarity measure between the input signal and
a template or reference [39]. In this work, the reference is an estimation of
AN
the average SVB’(N+SVEB) heartbeat of each patient. The calculation of the
reference beat is based on the observation that the abnormal samples are un-
derrepresented and the most common heartbeat is the normal beat (N). The
M

reference is updated at runtime to address changes in the ECG morphology due

165 to physical activity or noise. In Figure 2, we show an example of a reference
heartbeat.
ED

The reference beat is calculated by taking an average over the 50% of 500
consecutively detected beats. The 250 beats used for calculating the reference
are those with the lowest ratio between features 5 and 6. This ratio should be
PT

170 larger for the VEB class heartbeats as they usually are slower evolving signals
(slow rising and falling) with bigger areas than the SVB’ class heartbeats. 500
CE

beats represent around 5-8 minutes in the normal resting state. By averaging
only the 50% of the 500 beats with more probability of belonging to the SVB’
class, we try to guarantee that the majority of the beats used to calculate the
AC

175 reference beat indeed belong to the SVB’ class. This is crucial in cases with
bigemia or trigemia to prevent a bad estimation of the average SVB’ heartbeat
waveform.
During the initial section of the patient’s ECG recording, the 50% of the

10
 ACCEPTED MANUSCRIPT

T
IP
CR
US
Figure 2: Examples of segmented heartbeats of the patient 2106 of the AHA database. Red
AN
(blue) heartbeat belongs to the VB’ (SVB’) class. In black, the reference beat used for these
heartbeats.

first 500 beats are averaged to calculate the initial template. After this initial
M

180 warm-up stage, the template is updated and calculated at runtime in agreement
with the real-time nature of the system. A new template is calculated each 15
ED

beats and replaces the old one. 15 beats is approximately the number of beats
in the typical 10 s ECG strip used by doctors for 80 beats per minute (bpm). To
work in real time, the current heartbeat is compared with the currently available
PT

185 template.
Once the reference beat is calculated, the Pearson’s correlation coefficient is
CE

used to measure the similarity between the current heartbeat and the reference.
The set of correlations to evaluate the similarity between the heartbeat and the
template is:
AC

190 7. C(Ref, Beat): Correlation coefficient between the reference beat and the
heartbeat.
8. C(Ref 0 , Beat0 ): Correlation coefficient between the derivatives of the ref-
erence beat and the heartbeat.

11
 ACCEPTED MANUSCRIPT

9. C(Ref 2 , Beat2 ): Correlation coefficient between the squares of the refer-

195 ence beat and the heartbeat.
10. C((Ref 0 )2 , (Beat0 )2 ): Correlation coefficient between the squares of the
derivatives of the reference beat and the heartbeat.

T
The use of a reference beat for each patient can reduce classification errors

IP
due to the inter-patient variability of the ECG, reducing the dependence with
200 respect to the particular patient. Moreover, as the reference is updated in

CR
real time, it can address for the intra-patient variations in the ECG long-term
continuous monitoring when the subject experiences changes in heart conditions
or physical states.

205
US
Finally, we consider two more features related to the reference beat:

11. mean(|Ref |): Average of the absolute value of the reference beat.
AN
12. mean(|Ref 0 |): Average of the absolute value of the derivative of the ref-
erence beat.
M

At the end of the feature selection stage, each heartbeat is represented as a

12th-dimensional feature vector that is the input for the classification algorithm.
ED

210 2.4. Classification algorithm: ESN

The classification algorithm is an Echo State Network (ESN) [28]. ESNs

are part of the Reservoir Computing (RC) paradigm, a new machine learning
PT

approach that has been successfully applied to many tasks [40]. The main idea
behind RC lies in the random projection of the input onto a high dimensional
CE

215 nonlinear system (the reservoir). The reservoir is fixed and only the connections
between the reservoir and the output layer are tuned, thus the training is fast
and involves a low computational cost. Another advantage of ESNs is that they
AC

are not prone to overfitting. Since the internal connections between the neurons
and the input are not trained, the free parameters of the system are limited to
220 the output weights.
In a standard ESN, the reservoir is a recurrent neural network where the
input weights, biases and connection weights between the internal neurons are

12
 ACCEPTED MANUSCRIPT

Input Layer Reservoir: r Output Layer

T
IP
CR
Win W Wout

(a)

Input Layer US
Reservoir: r Output Layer
AN
M
ED

Win W Wout

(b)
PT

Figure 3: Schematic representation of a (a) standard ESN with a random internal nodes
connectivity and (b) ring ESN. Random weights are depicted with discontinuous arrows,
whereas training or predefined weights are depicted with continuous arrows.
CE

randomly set (see Figure 3 (a)). ESNs with random input weights (Win ) but
simple and deterministic connection (W) topologies perform as well as ESNs
AC

225 with random W [41]. An advantage of ESNs with non-random connection

topologies is that they can be easily implemented in hardware allowing for low
power consumption and high-speed processing [29].
In this work, we use a cycle ESN where the internal neurons are connected in

13
 ACCEPTED MANUSCRIPT

a ring (non-random connection topology) [41] (see Figure 3 (b)). The response
230 of the cycle ESN with N internal neurons of the ESN to a d-dimensional input
x(n) at the nth time step is defined by:

T
r(n) = F (γWin x(n) + βWr(n − 1)), (1)

IP
where Win ∈ RN ×d is a random matrix drawn from a uniform distribution
over [1, −1], F is the ESN activation function and γ and β are the input and

CR
the connection scaling parameters, respectively. The connection matrix W ∈
235 RN ×N is deterministic and has only non-zero elements in the lower sub-diagonal
Wi−1,i = 1 and at the upper-right corner W1,n = 1. The activation function is

US
the classic sigmoid function shifted to become symmetric around zero.
The response of the ESN to the input, r(n), is used to calculate the expected
AN
outputs, o(n), according to:

o(n) = Wout r(n), (2)

M

where Wout ∈ Rl×N are the output weights of the ESN and l is the number
of output nodes. The output weights are the only ones that are optimized to
minimize the error between the train outputs and their corresponding target
ED

240

outputs, using a simple linear regression solved by normal equations. The num-
ber of output nodes is l = 1, as the target is a binary scalar (SVB’=0 or VB’=1).
PT

The continuous output given by equation 2 is converted into a binary one by

a decision threshold of 0.5. The input dimensionality is given by the number
245 of features that characterize the heartbeat (d = 12) described in the previous
CE

subsection.
The ESN retains background information from previous inputs without the
need to represent this information explicitly in the model. The parameter β
AC

modulates the weight of the background information on the classification. For

250 β = 0, the ESN is no longer a recurrent network but just a single layer feedfor-
ward neural network with random input weights and no memory of the previous
input [42].

14
 ACCEPTED MANUSCRIPT

3. Results

The metrics recommended by the AAMI for evaluating the classification are:
255 Sensitivity (Se), Positive predictive value or Precision (+P ), False positive rate
or specificity (F P R) and overall accuracy (Acc). The F1 score is the harmonic

T
mean of Se and +P , F 1 = 2(Se) · (+P )/((Se) + (+P )).

IP
3.1. Training of the heartbeats classifier

CR
The hyper-parameters of the ring ESN are the network size (N ), the input
260 connection matrix W in and the scaling parameters γ and β. These hyper-
parameters together with the output weights are optimized in the training phase
and are fixed in the test phase.
US
In the training phase, the parameters of the ring ESN, N , γ and β, are esti-
mated according to a 5-fold cross-validation (CV) procedure over the heartbeats
AN
265 of the AHA and MIT-SV databases. In the 5-fold CV, the training dataset is
split into 5 folds. Each fold is used once as validation set while the rest of the
folds are used for training. The final results are the combination of the pre-
M

dicted outputs from the validation folds. The folds are randomly made, while
preserving the percentage of samples for each class.
ED

270 Heartbeats from all the available leads of the training databases are consid-
ered for the 5-fold CV, i.e., heartbeats of the same ECG but corresponding to
a different lead are considered as independent data (see Figure 1). To avoid an
PT

undesired dependence on the input weights, we average over 10 different input

random connection matrices Win . The criterion to choose the optimum set of
CE

275 ESN parameters in this imbalanced scenario is the one that maximizes the F1
score over the training set.
We first fix the number of neurons to a low value (N = 100) and search the
AC

γ and β pair with the highest F1 score on the CV. The result of the CV as a
function of β and γ for N = 100 is plotted in Figure 4. The best performance
280 is obtained for low values of γ > 0 and β > 0 and the optimum pair is γ = 0.1
and β = 0.2. The worst case scenario is β = 0 (no memory or information of
previous heartbeats).

15
 ACCEPTED MANUSCRIPT

0.9 0.9

0.8
0.895
0.7

0.6 0.89

T
0.5
0.885
β

IP
0.4

0.3 0.88

CR
0.2
0.875
0.1

0 0.87
0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

US γ

Figure 4: F1 score of the 5-fold CV as a function of β and γ for N = 100 averaged over 10
AN
different random input matrices.

The number of neurons is also optimized. Figure 5 shows the F1 score of

the 5-fold CV as a function of the N with β = 0.2 and γ = 0.1. The F1 score
M

285 increases with N until N ≈ 1500 and then saturates.

Accordingly, we set β = 0.2, γ = 0.1 and N = 1500 and randomly gener-
ED

ate several input random matrices (ten in this particular case). The one that
performs better in the training set is chosen to be used in the final classifier.
PT

3.2. Evaluation over the MIT-AR and INCART databases

290 The MIT-AR and the INCART databases are used for the evaluation (see
Figure 1). The final performance of the ESN is evaluated over databases dif-
CE

ferent from those used in the training. The origin of the leads used for the
training (the AHA and the MIT-SV databases) are not reported in the liter-
AC

ature, although it is likely that the MIT-SV and MIT-AR databases have the
295 same leads. The leads of the testing databases are MLII and mostly V1 in
the case of the MIT-AR database and the 12 standard leads for the INCART
database. Tables 2 and 3 show the performance over each lead of the MIT-AR
and the INCART databases, respectively. The ESN parameters are the ones

16
 ACCEPTED MANUSCRIPT

0.98

0.96

0.94

0.92

T
0.9

IP
0.88 +P
0.86 F1

CR
0.84 Se

0.82

0.8
0 1000
US
2000
N (neurons)
3000 4000

Figure 5: The F1 score of the 5-fold CV as a function of the number of neurons of the ring
AN
ESN (N ) with β = 0.2 and γ = 0.1. The error bars are the standard deviation over 10 different
random input matrices.
M

Table 2: Performance of the heartbeat classifiers on both leads of the MIT-AR database.
The ESN parameters are β = 0.2, γ = 0.1 and N = 1500. The input random matrix is the
one with the best performance on the training set.
ED

Lead Se (%) +P (%) Acc (%) FPR (%)

A (MLII) 88.7 89.2 98.3 99.1
B (Mostly V1) 87.1 89.8 98.2 99.2
PT

that yield the maximum F1 score during the train phase.

The performance of the classifier is very similar for both available leads of
CE

300

the MIT-AR database, with Se and the +P around 88%. In the case of the
INCART database, there are leads that give better VB’ classification accuracy
AC

than others. However, for most of the leads, both Se and +P are above 90%.
The leads that give the best test results for the INCART database are V1, V6
305 and lead II, respectively. Leads V1 and II are used in the training phase and
they are also the most commonly used for arrhythmia monitoring due to their
ability in distinguishing VEB and SVEB heartbeats from normal heartbeats.

17
 ACCEPTED MANUSCRIPT

T
IP
CR
Table 3: Performance of the heartbeat classifiers on each lead of the INCART database. The
ESN parameters are β = 0.2, γ = 0.1 and N = 1500. The input random matrix is the one
with the best performance on the training set.
Lead
I
II
Se (%)
76.8
92.6
US
+P (%)
92.2
97.4
Acc (%)
96.6
98.9
FPR (%)
99.6
99.7
AN
III 91.9 94.4 98.6 99.3
AVR 91.9 96.2 98.7 99.5
AVL 88.7 87.8 97.3 98.4
M

AVF 91.8 96.7 98.7 99.6

V1 93.7 98.0 99.1 99.8
ED

V2 85.8 96.5 98.0 99.6

V3 85.4 95.4 97.8 99.5
V4 90.6 96.9 98.6 99.6
PT

V5 91.7 97.5 98.8 99.7

V6 92.6 97.8 98.9 99.7
CE
AC

18
 ACCEPTED MANUSCRIPT

Leads V6, V5, AVR and AVF also give a very good performance even if they
are not explicitly included at all in the training phase. Therefore, we think
310 that the final performance of our classifier is directly related to the intrinsic
informative properties of the lead to detect ventricular heartbeats regardless

T
of the lead inclusion in the training sets. In fact, the worst case scenario for

IP
the INCART database is lead I, where the Se of the classifier drops to 76.8%,
following by lead AVL. These leads (I and AVL) look at the heart activity from

CR
315 view angles very different from the best ones (V1 and II), so they probably have
lower intrinsic properties for detecting arrhythmias and are prone to give worse
results.

320
US
The results of Tables 2 and 3 have been obtained with a standard threshold
value of 0.5, the average between the target values (0 for class SVB’ and 1 for
VB’). However, since the training data is highly imbalanced, a more heuristic
AN
threshold might fit better. We have swept the threshold value in the training
set keeping the rest of the parameters fixed. In this case, the best F1 score
is obtained for a threshold of 0.4. Applying this new threshold value to the
M

test set, we have a better Se but at the cost of a worse +P for the MIT-AR
325 and the INCART databases. These results agree with the idea that changing
ED

the decision threshold simply makes a trade off between the number of true
positive and the number of true negative predictions. The empirically obtained
threshold of 0.4 leads to a higher false alarm rate that does not compensate for
PT

the obtained higher sensitivity. In the case of the MIT-AR database, the results
330 with the 0.4 decision threshold are a Se = 91.1(91.4) and a +P = 84.6(82.3) for
CE

the lead II (V1).

Our approach is a low computational cost algorithm that can be suitable
for real-time implementations. We have performed the experiments in MAT-
AC

LAB(R2010b) on a PC (configuration: 256GB RAM memory and 32 cores).

335 Average testing runtime of 30 minutes of ECG with our classifier is 2.2 s (1.7
s for the feature extraction and 0.5 s for the ESN classifier). The preprocess-
ing and segmentation are beyond the scope of this work and were performed
beforehand and excluded from runtime assessment. However, this should not

19
 ACCEPTED MANUSCRIPT

be a problem for the real-time implementation, since there are already methods
340 that can preprocess and segment the ECG in real-time [43]. Once the hyper-
parameters of the model are fixed it only takes 10 min to train the full model
with the AHA and MIT-AR databases (i.e. around 1 million heartbeats since

T
each lead is considered independent data). The short training time allows to

IP
easily re-train the model if new data are available. The classification method
345 in the testing phase only involves N nonlinear transformations and, thanks to

CR
the sparse connectivity in the ESN, a reduced number of multiplications and
additions. In the training phase, we use a simple least squares regression to
optimize the ESN output weights. Least squares regression is solved using the

350 US
normal equations with M training examples and N features with M > N , with
a resulting total time complexity of O(N2 M).
The method proposed here allows for a real-time application thanks to the
AN
low computational demands associated to our methodology. In our implemen-
tation, the only practical issue is that the classifier has to wait for the first
500 heartbeats of each patient to compute the reference. Subsequently, the al-
M

355 gorithm gives the results for these first 500 heartbeats very fast (note that it
takes only 2.2 seconds to classify 30 minutes of the ECG). After these initial 500
ED

heartbeats, the system starts to work in real-time: a new reference is calculated

every 15 heartbeats and once it is ready, it substitutes the current reference.
The system always uses the latest available reference to obtain the correspond-
PT

360 ing heartbeat features. In a clinical environment, however, the best practice
would be to just ignore the first 500 heartbeats and start to classify the rest of
CE

the ECG from the heartbeat 501.

3.3. Comparison with other methods in the literature

AC

It is not straightforward to make a fair comparison between heartbeat clas-

365 sifiers that use different databases and training and test sets. To be as fair
as possible we only compare our method with fully automatic heartbeat classi-
fiers tested over the MIT-AR database (by far the most used one) and whose
training set does not share subjects with the testing set. Classifiers with intra-

20
 ACCEPTED MANUSCRIPT

patient classification give better results than those that follow the inter-patient
370 approach [8]. In Table 4 we show the VB performance, i.e. the classification of
V heartbeats versus non-V heartbeats, of different algorithms. In order to illus-
trate that our method is a competitive alternative to classify heartbeats with

T
ventricular origin (VB’), Table 4 also includes the classification performance of

IP
the VB’ class (V+F) versus the SVB’ class (N+S). The VB’ performance is
375 usually not reported in the literature and these results have been obtained from

CR
the confusion matrices, when available.
Our approach outperforms most of the other single lead classifiers and gives
state-of-the-art results when it is compared with methods that combine infor-

380 US
mation from more than one lead to classify. Moreover, our approach is the only
one that works well for both leads of the MIT-AR database. Although some of
the presented methods give better results, many require a high computational
AN
cost that invalidates their use for classification in real time (e.g. [44, 46]). In
most cases, the computational cost of these methods, either during the training
or the test phases, are not reported.
M

385 Regarding the INCART database, a reported VB performance for a fully

automatic classifier yields a Se = 88 % and a +P = 96 %, which was obtained
ED

using a combination of the leads II and V1 [25]. Our VB results for lead II and
V1 of the INCART database (see Table 3) are clearly better, with Se > 92 %
and similar +P using only one lead.
PT

390 Our method is a competitive alternative to classify heartbeats with ven-

tricular origin in comparison with other existing methods with the additional
CE

advantage of using a single lead and being computed in real-time.

4. Conclusions
AC

In this work, we have presented a fully automated, single-lead and real-

395 time classifier of heartbeats with ventricular origin. The proposed classifier is
a competitive single lead ventricular heartbeat classifier whose performance is
comparable to the state-of-the-art approaches that use more than one lead to

21
 ACCEPTED MANUSCRIPT

Table 4: VB and VB’ performance of the heartbeat classifiers on the MIT-AR database. Only
the best fully automatic work result is reported. The VB performance has been calculated in

T
our algorithm without taking into account the ’F’ class heartbeats.

IP
VB (V) VB’ (V+F)
Work Classes Leads Se +P Se +P

CR
Chazal [2] 5 MLII+V1 77.7 81.9 84.1 42.1
Ye (2012) [10] 5 MLII+V1 81.5 63.1 81.7 40.3
Zhang [3] 4 MLII+V1 85.5 92.8 91.8 58.3
Mar [12]
Garcia [6]
Llamedo [25]
4
3
5
MLII+V1
MLII+V1
MLII+V1
US 86.8
87.3
89.0
75.9
59.4
87.0
86.9
-
82.3
61.3
-
78.2
AN
Ye (2016) [11] 5 MLII+V1 91.8 98.0 84.4 96.7
Tejeiro [44] 5 MLII+V1 92.8 92.2 91.4 92.7
Ghorbani [8] 5 MLII+V1 96 77.6 87.2 75.4
M

Krasteva [20] 2 MLII+V1 96.7 99.2 94.1 99.3

Wu [21] 4 MLII 80.5 81.4 83.2 85.3

ED

Lannoy [19] 4 MLII 85.1 - - -

Rahhal [24] 5 MLII 91.0 79.5 - -
Raj [9] 5 MLII 87.5 65.4 86.0 43.5
PT

Sultan [7] 5 MLII 95.4 94.1 89.6 96.4

Herry [45] 4 MLII 77.5 79.1 82.3 37.8
CE

V1 79.6 62.7 - -
Huang [4] 3 MLII 93.9 90.9 - -
V1 78.1 43.8 - -
AC

This work 2 MLII 95.3 88.8 88.7 89.2

V1 90.9 89.2 87.1 89.8

22
 ACCEPTED MANUSCRIPT

classify. Our classifier exhibits a good performance for different leads, support-
ing its possible application to unconventional placement electrodes as it is quite
400 independent of the origin of the lead. The classification method can generalize
to unseen subjects and most of the unseen leads for the identification of ven-

T
tricular heartbeats, but it still struggles to do so for some leads (such as I and

IP
AVL).
The proposed method is designed to overcome some of the typical drawbacks

CR
405 of previous heartbeat classifiers, such as the inability to operate in real-time.
The classifier is designed to be fast and computationally efficient not only in
the evaluation phase but also in the training stage. This allows the real-time

410
US
implementation of the method as well as the easy incorporation of new training
data to improve the system’s performance. Moreover, the proposed method
generalizes across different ECG databases and therefore accounts for the inter-
AN
patient classification, which is demonstrated by using different databases for
training and testing.
The proposed algorithm only distinguishes between heartbeats with and
M

without ventricular origin. Future work will focus on the extension of these re-
415 sults to the five heartbeats types recommended by the AAMI. This would require
ED

including new features to differentiate normal and supraventricular heartbeats

from the class SVB’.
PT

5. Highlights

A fully automated and real-time heartbeat ventricular arrhythmia classifier

CE

420 based on a single lead designed to be fast and efficient in the training and
evaluation phase.
The arrhythmia classifier can be applied to different ECG leads with ex-
AC

cellent performance. Potential application to wearables with unconventionally

placed electrodes.
425 Trained and evaluated with different databases, allowing the inter-patient
and inter-database classification.

23
 ACCEPTED MANUSCRIPT

Acknowledgment

The authors would like to thank Ingo Fischer, Xavier Ibàñez Català and
Agustı́n Maciá for valuable scientific discussions. This work is partially sup-
ported by the Spanish Ministerio de Economa y Competitividad (MINECO)

T
430

and Fondo Europeo de Desarrollo Regional (FEDER) through project TEC2016-

IP
80063-C3-3-R. Silvia Ortı́n was supported by the Conselleria d’Innovació, Re-
cerca i Turisme del Govern de les Illes Balears and the European Social Fund.

CR
Miguel Cornelles Soriano was supported by the Spanish Ministerio de Economa,
435 Industria y Competitividad through a Ramon y Cajal Fellowship (RYC-2015-
18140).

References
US
AN
References

[1] E. J. da S. Luz, W. R. Schwartz, G. Cámara-Chávez, D. Menotti, ECG-

M

440 based heartbeat classification for arrhythmia detection: A survey, Com-

puter Methods and Programs in Biomedicine 127 (2016) 144–164. doi:
http://dx.doi.org/10.1016/j.cmpb.2015.12.008.
ED

[2] P. de Chazal, M. O’Dwyer, R. B. Reilly, Automatic classification of

heartbeats using ECG morphology and heartbeat interval features, IEEE
PT

445 Transactions on Biomedical Engineering 51 (7) (2004) 1196–1206. doi:

10.1109/TBME.2004.827359.
CE

[3] Z. Zhang, J. Dong, X. Luo, K. S. Choi, X. Wu, Heartbeat classification

using disease-specific feature selection, Computers in Biology and Medicine
46 (1) (2014) 79–89. doi:10.1016/j.compbiomed.2013.11.019.
AC

450 [4] H. Huang, J. Liu, Q. Zhu, R. Wang, G. Hu, A new hierarchical method
for inter-patient heartbeat classification using random projections and RR
intervals, BioMedical Engineering Online 13 (1) (2014) 90. doi:10.1186/
1475-925X-13-90.

24
 ACCEPTED MANUSCRIPT

[5] Z. Zidelmal, A. Amirou, D. Ould-Abdeslam, J. Merckle, ECG beat classi-

455 fication using a cost sensitive classifier, Computer Methods and Programs
in Biomedicine 111 (3) (2013) 570–577.

[6] G. Garcia, G. Moreira, D. Menotti, E. Luz, Inter-Patient ECG Heartbeat

T
Classification with Temporal VCG Optimized by PSO, Scientific Reports

IP
7 (1). doi:10.1038/s41598-017-09837-3.

CR
460 [7] S. Sultan Qurraie, R. Ghorbani Afkhami, ECG arrhythmia classification
using time frequency distribution techniques, Biomedical Engineering Let-
ters 7 (4) (2017) 325–332. doi:10.1007/s13534-017-0043-2.

US
[8] R. Ghorbani Afkhami, G. Azarnia, M. A. Tinati, Cardiac arrhythmia clas-
sification using statistical and mixture modeling features of ECG signals,
AN
465 Pattern Recognition Letters 70 (2016) 45–51. doi:10.1016/j.patrec.
2015.11.018.

[9] S. Raj, K. C. Ray, O. Shankar, Cardiac arrhythmia beat classification us-

M

ing DOST and PSO tuned SVM, Computer Methods and Programs in
Biomedicine 136 (2016) 163–177. doi:10.1016/j.cmpb.2016.08.016.
ED

470 [10] C. Ye, B. V. K. V. Kumar, M. T. Coimbra, Heartbeat Classification Using

Morphological and Dynamic Features of ECG Signals, IEEE Transactions
on Biomedical Engineering 59 (10) (2012) 2930–2941. doi:10.1109/TBME.
PT

2012.2213253.

[11] C. Ye, B. V. K. Kumar, M. T. Coimbra, An Automatic Subject-Adaptable

CE

475 Heartbeat Classifier Based on Multiview Learning, IEEE Journal of

Biomedical and Health Informatics 20 (6) (2016) 1485–1492. doi:10.1109/
AC

JBHI.2015.2468224.

[12] T. Mar, S. Zaunseder, J. P. Martı́nez, M. Llamedo, R. Poll, Optimization

of ECG Classification by Means of Feature Selection, IEEE Transactions
480 on Biomedical Engineering 58 (8) (2011) 2168–2177. doi:10.1109/TBME.
2011.2113395.

25
 ACCEPTED MANUSCRIPT

[13] Z. Dokur, T. Ölmez, ECG beat classification by a novel hybrid neural

network, Computer Methods and Programs in Biomedicine 66 (2-3) (2001)
167–181. doi:10.1016/S0169-2607(00)00133-4.

[14] F. A. Elhaj, N. Salim, A. R. Harris, T. T. Swee, T. Ahmed, Arrhyth-

T
485

mia recognition and classification using combined linear and nonlinear fea-

IP
tures of ECG signals, Computer Methods and Programs in Biomedicine
127 (2016) 52–63. doi:10.1016/j.cmpb.2015.12.024.

CR
[15] R. J. Martis, U. R. Acharya, L. C. Min, ECG beat classification using PCA,
490 LDA, ICA and Discrete Wavelet Transform, Biomedical Signal Processing

US
and Control 8 (5) (2013) 437–448. doi:10.1016/j.bspc.2013.01.005.

[16] O. T. Inan, L. Giovangrandi, G. T. Kovacs, Robust neural-network-based

AN
classification of premature ventricular contractions using wavelet transform
and timing interval features, IEEE Transactions on Biomedical Engineering
495 53 (12) (2006) 2507–2515. doi:10.1109/TBME.2006.880879.
M

[17] M. Javadi, R. Ebrahimpour, A. Sajedin, S. Faridi, S. Zakernejad, Improv-

ing ECG Classification Accuracy Using an Ensemble of Neural Network
ED

Modules, PLoS ONE 6(10) (2011) e24386. doi:10.1371/journal.pone.

0024386.

500 [18] S. Kiranyaz, T. Ince, M. Gabbouj, Real-Time Patient-Specific ECG Clas-

PT

sification by 1-D Convolutional Neural Networks, IEEE Transactions on

Biomedical Engineering 63 (3) (2016) 664–675. doi:10.1109/TBME.2015.
CE

2468589.

[19] G. De Lannoy, D. François, J. Delbeke, M. Verleysen, Weighted condi-

AC

505 tional random fields for supervised interpatient heartbeat classification,

IEEE Transactions on Biomedical Engineering 59 (1) (2012) 241–247.
doi:10.1109/TBME.2011.2171037.

[20] V. Krasteva, I. Jekova, R. Leber, R. Schmid, R. Abacherli, Superiority

of Classification Tree versus Cluster, Fuzzy and Discriminant Models in a

26
 ACCEPTED MANUSCRIPT

510 Heartbeat Classification System, PLoS One 13 (10(10)) (2015) e0140123.

doi:10.1371/journal.pone.0140123.

[21] Z. Wu, X. Ding, G. Zhang, A Novel Method for Classification of ECG

Arrhythmias Using Deep Belief Networks, International Journal of

T
Computational Intelligence and Applications 15 (04) (2016) 1650021.

IP
515 doi:10.1142/S1469026816500218.
URL http://www.worldscientific.com/doi/abs/10.1142/

CR
S1469026816500218

[22] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, M. Adam, A. Ger-

520
US
tych, R. S. Tan, A deep convolutional neural network model to classify
heartbeats, Computers in Biology and Medicine 89 (2017) 389–396.
doi:https://doi.org/10.1016/j.compbiomed.2017.08.022.
AN
URL http://www.sciencedirect.com/science/article/pii/
S0010482517302810

[23] S. Kiranyaz, T. Ince, M. Gabbouj, Real-Time Patient-Specific ECG Clas-

M

525 sification by 1-D Convolutional Neural Networks, IEEE Transactions on

Biomedical Engineering 63 (3) (2016) 664–675. doi:10.1109/TBME.2015.
ED

2468589.

[24] M. M. Rahhal, Y. Bazi, H. Alhichri, N. Alajlan, F. Melgani, R. R. Yager,

Deep learning approach for active classification of electrocardiogram sig-
PT

530 nals, Information Sciences 345 (2016) 340–354. doi:10.1016/j.ins.2016.

01.082.
CE

[25] M. Llamedo, J. P. Martinez, An Automatic Patient-Adapted ECG Heart-

beat Classifier Allowing Expert Assistance, IEEE Transactions on Biomed-
AC

ical Engineering 59 (8) (2012) 2312–2320. doi:10.1109/TBME.2012.

535 2202662.

[26] V. Krasteva, I. Jekova, QRS template matching for recognition of ventric-

ular ectopic beats, Annals of Biomedical Engineering 35 (12) (207) 2065–
2076.

27
 ACCEPTED MANUSCRIPT

[27] A. Rodan, P. Tino, Minimum complexity echo state network, IEEE Trans-
540 actions on Neural Networks 22 (1) (2011) 131–144. doi:10.1109/TNN.
2010.2089641.

[28] H. Jaeger, H. Haas, Harnessing nonlinearity: predicting chaotic systems

T
and saving energy in wireless communication, Science 304 (2004) 78–80.

IP
[29] D. Brunner, M. C. Soriano, C. Mirasso, I. Fischer, Parallel photonic infor-

CR
545 mation processing at gigabyte per second data rates using transient states,
Nature Communications 4 (2013) 1364. doi:10.1038/ncomms2368.

[30] M. C. Soriano, S. Ortı́n, D. Brunner, L. Larger, C. R. Mirasso, I. Fischer,

US
L. Pesquera, Optoelectronic reservoir computing: tackling noise-induced
performance degradation, Opt. Express 21 (1) (2013) 12–20. doi:10.1364/
OE.21.000012.
AN
550

[31] S. D. Greenwald, Improved detection and classification of arrhythmias

in noise-corrupted electrocardiograms using contextual information, Ph.D.
M

thesis, Harvard-MIT Division of Health Sciences and Technology (1990).

[32] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ivanov,

ED

555 R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, H. E. Stanley, Phys-

ioBank, PhysioToolkit, and PhysioNet: Components of a New Research Re-
source for Complex Physiologic Signals, Circulation 101 (23) (2000) e215–
PT

e220.

[33] G. B. Moody, R. G. Mark, The impact of the MIT-BIH Arrhythmia

CE

560 Database, IEEE Eng. in Med. and Biol. 20 (3) (2001) 45–50.

[34] ANSI/AAMI, Testing and reporting performance results of cardiac rhythm

AC

and ST segment measurement algorithms, American National Standards

Institute, Inc. (ANSI), Association for the Advancement of Medical Instru-
mentation (AAMI), ANSI/AAMI/ISO EC57, 1998-(R)2008.

565 [35] J. Sahambi, S. Tandon, R. Bhatt, Using wavelet transforms for ECG
characterization. An on-line digital signal processing system, IEEE En-

28
 ACCEPTED MANUSCRIPT

gineering in Medicine and Biology Magazine 16 (1) (1997) 77–83. doi:

10.1109/51.566158.
URL http://ieeexplore.ieee.org/document/566158/

[36] N. V. Thakor, J. G. Webster, W. J. Tompkins, Estimation of QRS Complex

T
570

Power Spectra for Design of a QRS Filter, IEEE Transactions on Biomed-

IP
ical Engineering BME-31 (11) (1984) 702–706. doi:10.1109/TBME.1984.
325393.

CR
URL http://ieeexplore.ieee.org/document/4121752/

575 [37] M. Elgendi, B. Eskofier, S. Dokos, D. Abbott, Revisiting QRS detection

US
methodologies for portable, wearable, battery-operated, and wireless ECG
systems, PLoS ONEdoi:10.1371/journal.pone.0084018.
AN
[38] Y. J. Kim, J. Heo, K. S. Park, S. Kim, Proposition of novel clas-
sification approach and features for improved real-time arrhythmia
580 monitoring , Computers in Biology and Medicine 75 (2016) 190–202.
M

doi:http://dx.doi.org/10.1016/j.compbiomed.2016.06.009.
URL http://www.sciencedirect.com/science/article/pii/
S0010482516301482
ED

[39] A. K. Jain, R. P. W. Duin, J. Mao, Statistical pattern recognition: a

585 review, IEEE Transactions on Pattern Analysis and Machine Intelligence
PT

22 (1) (2000) 4–37. doi:10.1109/34.824819.

[40] M. Lukoševičius, H. Jaeger, Survey: Reservoir Computing Approaches to

CE

Recurrent Neural Network Training, Comput. Sci. Rev. 3 (3) (2009) 127–
149. doi:10.1016/j.cosrev.2009.03.005.
AC

590 [41] A. Rodan, P. Tino, Minimum complexity echo state network, IEEE Trans.
Neural Netw. 22 (2011) 131–144.

[42] S. Ortı́n, M. C. Soriano, L. Pesquera, D. Brunner, D. San-Martı́n, I. Fis-

cher, J. M. Gutiérrez, A Unified Framework for Reservoir Computing and

29
 ACCEPTED MANUSCRIPT

Extreme Learning Machines based on a Single Time-delayed Neuron, Sci-

595 entific Reports 5 (2015) 14945. doi:10.1038/srep14945.

[43] S.-W. Chen, H.-C. Chen, H.-L. Chan, A real-time QRS detection method
based on moving-averaging incorporating with wavelet denoising, Com-

T
puter Methods and Programs in Biomedicine 82 (3) (2006) 187–195.

IP
doi:10.1016/J.CMPB.2005.11.012.
600 URL https://www.sciencedirect.com/science/article/pii/

CR
S0169260705002592

[44] T. Teijeiro, P. Felix, J. Presedo, D. Castro, Heartbeat classification us-

605
US
ing abstract features from the abductive interpretation of the ECG, IEEE
Journal of Biomedical and Health Informatics 22 (2) (2018) 409–420.
doi:10.1109/JBHI.2016.2631247.
AN
[45] C. L. Herry, M. Frasch, A. J. E. Seely, H.-t. Wu, Heart beat classification
from single-lead ECG using the synchrosqueezing transform, Physiol. Meas.
M

38 (2) (2017) 171–187.

[46] S. Sultan Qurraie, R. Ghorbani Afkhami, ECG arrhythmia classification

ED

610 using time frequency distribution techniques, Biomedical Engineering Let-

ters 7 (4) (2017) 325–332. doi:10.1007/s13534-017-0043-2.
PT
CE
AC

30