Biomedical Signal Processing and Control 47 (2019) 41–48

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

Heartbeat classification fusing temporal and morphological

information of ECGs via ensemble of classifiers
V. Mondéjar-Guerra a,b,∗ , J. Novo a,b , J. Rouco a,b , M.G. Penedo a,b , M. Ortega a,b
a
Department of Computing, University of A Coruña, A Coruña, Spain
b
CITIC-Research Center of Information and Communication Technologies, University of A Coruña, A Coruña, Spain

a r t i c l e i n f o a b s t r a c t

Article history: A method for the automatic classification of electrocardiograms (ECG) based on the combination of mul-
Received 6 April 2018 tiple Support Vector Machines (SVMs) is presented in this work. The method relies on the time intervals
Received in revised form 12 June 2018 between consequent beats and their morphology for the ECG characterisation. Different descriptors based
Accepted 8 August 2018
on wavelets, local binary patterns (LBP), higher order statistics (HOS) and several amplitude values were
Available online 23 August 2018
employed. Instead of concatenating all these features to feed a single SVM model, we propose to train
specific SVM models for each type of feature. In order to obtain the final prediction, the decisions of
Keywords:
the different models are combined with the product, sum, and majority rules. The designed methodology
Electrocardiogram (ECG)
Heartbeat classification
approaches are tested on the public MIT-BIH arrhythmia database, classifying four kinds of abnormal and
Support vector machine (SVM) normal beats. Our approach based on an ensemble of SVMs offered a satisfactory performance, improving
Combining classifiers the results when compared to a single SVM model using the same features. Additionally, our approach also
Ensemble of classifiers showed better results in comparison with previous machine learning approaches of the state-of-the-art.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction unipolar limb leads (AVR, AVL, and AVF), and six unipolar chest
leads, also called precordial or V leads (V1, V2, V3, V4, V5 and V6).
Disturbances in the heart rate, popularly known as arrhythmias, Each lead is a view of the electrical activity of the heart from a par-
may be life-threatening, requiring immediate care and often an ticular angle across the body. The record contains approximately
intervention with defibrillator [1]. Nevertheless, most of arrhyth- 2.5 s of duration for each lead. Additionally, to accurately assess
mias are harmless; but even then, they may require therapy to the cardiac rhythm, a prolonged recording from one lead is used to
prevent further severe problems [2]. Arrhythmias are often asso- provide a rhythm strip of 10 s. Lead II is the most commonly used for
ciated with other forms of heart disease. According to the World the rhythm strip [3], since it usually gives a good view of the most
Health Organization (WHO), “Cardiovascular diseases (CVDs) are important waves: P, Q, R, S and T (see Fig. 1 ). Each beat of the heart
the number 1 cause of death globally: more people die annually contains a series of deflections away from the baseline on the ECG,
from CVDs than from any other cause. An estimated 17.7 mil- or waves, that reflect the time evolution of electrical activity in the
lion people died from CVDs in 2015, representing 31% of all global heart. P-wave is a small defection caused by atrial depolarisation,
deaths”. Electrocardiograms (ECG) are a noninvasive and inexpen- Q, R, and S waves are usually considered as a single event known
sive technique commonly employed by cardiologist in their clinical as the QRS-complex, which is the largest-amplitude portion of the
practice routine. They are frequently used to detect cardiac rhythm ECG, being caused by ventral depolarisation. T wave is caused by
abnormalities, measuring the electrical activity of the heart over a ventral repolarisation. Finally, in some cases, an additional U wave
period of time. For a routine analysis of the heart’s electrical activ- may follow the T wave.
ity, an ECG recorded from 12 separate leads is typically used. The The different types of arrhythmias can be detected through the
12-lead ECG consists of three bipolar limb leads (I, II, and III), the analysis of the changes that appeared on these waves. The devel-
opment of a fully automatic system that is able to classify the ECG
heartbeats has been a research topic of high interest throughout the
last decades. Fig. 2 shows a diagram of a general automatic system
∗ Corresponding author at: Department of Computing, University of A Coruña, A
for arrhythmia classification. First, the signals that were captured
Coruña, Spain.
E-mail addresses: v.mondejar@udc.es (V. Mondéjar-Guerra),
through the device are preprocessed. This step usually includes the
jnovo@udc.es (J. Novo), jrouco@udc.es (J. Rouco), mgpenedo@udc.es (M.G. Penedo), baseline removal and the cleaning of high-frequency noise. Next, a
mortega@udc.es (M. Ortega).

https://doi.org/10.1016/j.bspc.2018.08.007
1746-8094/© 2018 Elsevier Ltd. All rights reserved.
42 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48

i.e., combining several SVM models each one trained with a dif-
ferent feature. Several feature descriptors based on R-R intervals,
wavelets, HOS, LBP, and several amplitude values were employed.
Besides, the suitability of each single feature is also evaluated in this
work. Our approach is similar to the work of Zhang et al. [22], which
also used an ensemble of SVMs for the automatic arrhythmia classi-
fication. However, they extracted features from the leads II and V1
and posteriorly, they trained a separated model from the features of
each lead. Finally, they combined the decisions of both models with
the product rule. In our approach, a SVM model is created for each
type of feature being all the features extracted from lead II. Addi-
tionally, an extensive experimentation was made evaluating all the
possible combinations of the selected features. Finally, we tested
several combination rules, including the commonly employed sum,
product, and majority rules [31].
Fig. 1. Waves of a lead II ECG.
In the literature, we can distinguish two popular paradigms for
evaluation of arrhythmia classification task, known as intra-patient
heartbeat segmentation algorithm is applied to split the signal at and inter-patient. In the first paradigm, the whole database can be
beat level. This is usually done detecting the QRS-complex. Then, employed to generate and test the classification models without
several descriptors are applied to each beat in order to extract the any restriction. This paradigm presents a main drawback regard-
features to feed a classifier, which finally determines the type of ing the generalisation of the classifier. Since the model can learn
heartbeat. Many algorithms were proposed in the literature for the particularities of the patients during the training, the score
the heartbeat segmentation [4–7], reaching up to near optimal achieved in the evaluation step may not be highly reliable. Ideally,
results in well-known databases like MIT-BIH [8]. In this work, we an automatic arrhythmia classifier must give an accuracy diagnosis
focus on the two last steps, feature extraction and classification. for any patient, even if the system does not contain any previous
Many features were proposed to describe the ECG heartbeats, high- information about it. Therefore, a method with high generalisa-
lighting the use of wavelets [9,10], higher order statistics (HOS) tion is desirable, since a trained database with records from all the
[11,12], and heartbeat intervals, popularly known as R-R inter- possible patients would be unviable. In order to employ a more real-
vals [13,14]. To built the classification model, numerous previous istic scenario, Chazal et al. [2] proposed the inter-patient paradigm.
works reported the feasibility of machine learning algorithms for They performed a division of the MIT-BIH database records in two
this task [15]; including methods such as Linear Discriminant (LD) different sets: one for training and other testing. These sets were
[2], AdaBoost [16], Multilayer Perceptron (MLP) [9,17,18], Genetic carefully designed avoiding the inclusion of any record from the
Algorithm-Back Propagation Neural Network (GA-BPNN) [19], Con- same patient in both sets. We followed the inter-patient paradigm
volutional Neural Networks (CNN) [20], and, especially, Support to evaluate our approach.
Vector Machine (SVM) [17,18,21–24]. In the next section, the used database, the selected features and
An ensemble of classifiers combines the decisions of the indi- the proposed approach for the ECG classification are detailed. The
vidual classifiers that compose it, in order to improve the final employed performance measurements, the experiments and the
prediction. There are many techniques in the literature to create obtained results are explained on Section 3. Finally, Section 4 details
an ensemble of classifiers [25]. Some methods train each classifier the conclusions extracted from this work.
with a different subset of the training examples like Bagging [26],
or AdaBoost [27]. Dietterich and Bakiri [28] deal with a problem
2. Material and methods
that requires a large number of classes, partitioning the number of
outputs in different sets, generating an ensemble of classifier. Other
The well-known Massachusetts Institute of Technology-Beth
works train each classifier in a different subset of the input features.
Israel Hospital (MIT-BIH) arrhythmia database [8], from Physionet
Duin and Tax [29] performed a large experimentation of this alter-
[32], was employed to train and test our classification models,
native and concluded that the combination of classifiers trained on
allowing in turn the comparison of our results with those from the
different feature sets was very useful, especially when the single
state-of-the-art methods.
classifiers offered a good performance. Waske and Benediktsson
[30] employed an ensemble of SVMs in a multi-source land cover
classification problem using a balanced dataset. Their ensemble of 2.1. MIT-BIH arrhythmia database
SVMs, training each SVM with a different data source, significantly
improved the results in comparison to a single SVM that was trained This database contains 48 ECG records of about 30 min, sampled
with the whole data sources. at 360 Hz with 11-bit resolution from 47 different patients. Each
The main goal of this work is to evaluate the benefits of using record comprises two signals, the first one is, for all the records,
an ensemble of SVMs for the arrhythmia classification problem, the modified-lead II (MLII), whereas the second one corresponds

Fig. 2. Diagram of general steps of a full automatic arrhythmia classification system.

 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48 43

Table 1 tion step, therefore the QRS annotations included in MIT-BIH were
MIT-BIH labelling and the standard AAMI classes.
employed.
AAMI MIT-BIH For each beat, a window of size 180 was centred around its R-
Normal (N) N, L, R peak and, then, all the features were computed inside that region.
Supraventricular ectopic beat (SVEB) e, j, A, a, J, S Fig. 3(a) shows the mean values of all the beats from the MIT-BIH
Ventricular ectopic beat (VEB) V, E group by the four AAMI classes, whereas Fig. 3(b–f) show the mean
Fusion (F) F values obtained over each feature descriptor. The following features
Unknown beat (Q) /, f, Q
were employed since they showed a good performance on similar
previous works.

to V1, V2, V4, or V5, depending on the record. Therefore, only the 2.3.1. Wavelets
MLII is provided by all the records. The database contains approx- The wavelet transforms present the capability of allowing infor-
imately 110,000 beats, all of them were independently annotated mation extraction from both frequency and time domains, which
by two or more expert cardiologists and the disagreements were make them suitable for the ECG description. The use of wavelet
resolved. Following the Association for the Advancement of Medi- transforms were successfully proved by different authors on ECG
cal Instrumentation AAMI recommended practice [33], the MIT-BIH classification [9,10]. Here, we used the Daubechies wavelet func-
heartbeat types are grouped into five heartbeat classes as shown in tion (db1) with 3 levels of decomposition, making a 23-dimensional
Table 1 . As recommended by the AAMI, the records with paced descriptor.
beats were not considered, namely 102, 104, 107, and 217. The
database is highly imbalanced, as near a 90% of the beats belong 2.3.2. HOS
to the class N whereas the remaining 3%, 6%, and 1% of the beats The use of higher order statistics (HOS), i.e., cumulants of the
belong to classes SVEB, VEB, and F. We adopted the decision of second, third, and fourth order were proposed as a better alter-
ignoring the Q AAMI class like other authors [22,9], since it is prac- native for the morphological ECG description in [11,12]. Here, a
tically non-existent. Only 15 samples belong to class F. In order to 10-dimensional feature was created dividing each beat into 5 inter-
make a fair comparison between our results and those from other vals, computing the kurtosis and skewness value over each one.
previous works, we used the popular inter-patient scheme divi-
sion proposed by Chazal et al. [2], which divided the database in 2.3.3. 1D-LBP
two datasets. Each dataset contains data from 22 records with a A 1D variant of the well-known descriptor, the 2D-Local binary
similar proportion of beat types: patterns (LBP), was previously proposed for feature extraction of
raw Electroencephalogram (EEG) signals in [35]. The 1D version
• Training (DS1): 101, 106, 108, 109, 112, 114, 115, 116, 118, 119, maintains the idea of the original 2D version. For each data point
122, 124, 201, 203, 205, 207, 208, 209, 215, 220, 223, 230. in a beat, a binary code is produced by the comparison of its value
• Testing (DS2): 100, 103, 105, 111, 113, 117, 121, 123, 200, 202, with the value of their neighbours. Then, a histogram that contains
210, 212, 213, 213, 219, 221, 222, 228, 231, 232, 233, 234. the frequency of each binary pattern is built. Here, we used the 8
neighbour uniform LBP code, making a 59-dimensional descriptor.
The first dataset was employed for training whereas the second
one was used to evaluate the performance of the model. None of 2.3.4. Our morphological descriptor
the patients was repeated in both datasets. We proposed a morphological descriptor that relies on several
In this work, only the lead II was considered to describe the mor- amplitude values from the beats. Instead of directly employing
phology of the signal. This decision was motivated by the following several amplitude values like other previous works [2,22], our
facts: it is the only lead that is present for all the records from the descriptor relies on the Euclidean distance (sample, amplitude)
MIT-BIH arrhythmia database; it is also the most commonly used between the R-peak and four points of the beat. The selection of
lead by the experts to analyse the ECG signals; and finally, Chazal this points depends on the amplitude values over the following
proved that using only one lead is sufficient for the arrhythmia intervals:
classification task [34].
• max(beat[0, 40]).
• min(beat[75, 85]).
2.2. Signal preprocessing • min(beat[95, 105]).
• max(beat[150, 180]).
Before computing the features from the ECG signals, a pre-
processing step was applied. Most of the previous works of the
where beat is the 180-dimensional array, centered on the R-peak
literature [2,9,22] usually performed the baseline removal (see Fig.
that contains the amplitude values.
2) followed by a high-frequency noise filtering at this step. In this
case, we have just performed the baseline removal. To compute
2.3.5. R-R Intervals
the baseline of the signal, two consecutive median filters of 200-
A descriptor based on these intervals is certainly the most
ms and 600-ms were applied. Then, this baseline was subtracted
employed feature for the classification of ECGs in the literature
from the original signal, producing the baseline corrected ECG sig-
[15]. Besides the morphological features, R-R intervals computed
nal. We made the decision of not performing any high-frequency
from the time between consequent beats were employed. The next
noise filtering in order to preserve the signal as raw as possible for
intervals were extracted:
the feature extraction step.
• Pre-RR: indicating the distance between the actual heartbeat and
2.3. Selected features the previous one.
• Post-RR: indicating the distance between the actual heartbeat
In practice, a QRS detection algorithm like the proposed by Pan and the next one.
and Tompkins [4] would be required in order to segment the full • Local-RR: containing the average of the 10 previous Pre-RR val-
signal into beats. However, this work is focused on the classifica- ues.
44 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48

Fig. 3. Average beats from the MIT-BIH database grouped by the four AAMI class (N, SVEB, VEB, and F) (a) 180 window centered on R-peak from the raw lead II signal. (b)
Wavelets of family ‘db1’ level 3 decomposition. (c) 5 HOS intervals: skewness and kurtosis. (d) Histogram U-LBP 1D of 8 bits. (e) Our morphological descriptor. (f) 4 R-R
intervals followed by its 4 normalised values. (Best seen in color).

• Global-RR: containing the average of the Pre-RR values produced Table 2

Features employed for ECG description.
in the last 20 min.
Feature name Description Size

R-R R-R intervals + normalized R-R intervals 8

In addition, the normalised version of these intervals [13], i.e., Wavelets Family db1 with 3 level of decomposition 23
dividing them by its mean value within the same ECG record, were HOS Skewness and Kurtosis with 5 intervals 10
also employed, making a total of eight R-R features. LBP Uniform-LBP 1D 8 bits 59
Our Morph. Our morphological descriptor 4

2.4. Classifier models Given N classes and L models, the final decision of a new obser-
vation x is computed using the pairwise a posteriori probability
Due to their good performance showed on previous ECG classi- P(ym |fl (x)) that follows a sigmoid function:
fication works [21,22], we employed SVMs as the classifiers for all
our experiments. 1
P(ym |fl (x)) = , (1)
1 + exp(−ym fl (x))

where M = 2, since each model of the OAO is in fact a binary clas-

2.4.1. Support vector machine sifier, y1 =+1 and y2 =−1 denotes positive and negative classes,
SVM [36] are maximum margin classifiers that map input vec- respectively, and fl (x) measures the decision value given by the
tors to a higher dimensional space where a maximal separating l SVM. In our problem N = 4, and, consequently, L = 6 due to the
hyperplane is constructed to differentiate two different classes. OAO strategy. The probabilities of the different models P(ym |fl (x))
The main drawback of SVMs is their limitation to binary classifica- are accumulated over their associated class n in ın and then, the
tion problems. In the literature, there are two common alternatives majority voting rule is applied to assign the final decision:
to solve that, namely one-against-all (OAA) and one-against-one
(OAO). Being N the number of classes, in the first alternative N arg maxn ın . (2)
SVM models are constructed, i.e., one per class; whereas in the sec-
ond alternative a SVM is constructed between each pair of classes, 2.4.2. Combining multiple SVMs
resulting in N(N − 1)/2 models. Finally, a voting system is required In this work, an independent OAO SVM model is trained for each
for both alternatives in order to get a final decision. In this work, type of employed feature (see Table 2 ). Let ıtn be the accumulated
the OAO approach was used, since it is more suitable to work with probabilities of the t OAO, considering all the features T = 5. To com-
imbalanced data and requires less time for training than OAA when bine the accumulated probabilities of the different OAO models,
the number of samples is significantly large [37]. some of the most commonly used combination rules were tested:
 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48 45

Table 3 class (N) for all the new incoming data. This classifier would achieve
Description of confusion matrix from AAMI classes: N: normal, S: Supraventricular,
a value of the overall accuracy higher than 89%. In the other hand,
V: Ventricular, and F: fusion.
the mean accuracy would give the same importance to the majority
Algorithm and minority classes. Therefore, these performance measurement
n s v f Sum does not seem appropriate to represent the quality of the classifiers
on this database. To overcome this problem, Mar et al. proposed a
Reference  new index, which they named j index [9], as a combination of two
N Nn Ns Nv Nf
N values: the j index [38] and the Cohen’s Kappa () index [39]:
S Sn Ss Sv Sf
S
V Vn Vs Vv Vf
V
F Fn Fs Fv Ff F j index = w1  + w2 j index, (6)
    
Sum n s v f
where w1 = 1/2, and w2 = 1/8 since  takes values in the [0,1]
range and j index in the [0,4] range. The j index evaluates the
• Product rule: It is a severe rule, since if just one model assigns discrimination of the most important arrhythmias (SVEB, VEB,
a close to zero probability for one class, the final output for this according to the AAMI standard [9]):
class will also be close to zero:
j index = SeS + SeV + PS+ + PV+ , (7)

T
ıtn . (3)
being Se and P+ the sensitivity and positive predictive value of each
t=1
class. Finally, the Cohen’s Kappa () is a measure of agreement that
• Sum rule: Opposite to the product rule, the sum rule presents a globally evaluates the confusion matrix. It was reported as a per-
more relaxed behaviour: formance measurement more robust than the overall accuracy or

T the mean accuracy on imbalance datasets [40]:
ıtn . (4)
Po − Pe
t=1 = ,
1 − Pe
• Majority rule: It adds a vote to each class depending on their rank
Nn + Ss + Vv + Ff
position. This rule does not consider the differences at probability Po =  , (8)
level between the outputs, it only consider the rank order:        
N n+ S s+ V v+ F f

T Pe = 2 ,
tn , (5)
t=1
where Po is the observed probability, being equal to the overall
where tn contains a vote inversely proportional to the rank accuracy, and Pe corresponds with the chance agreement. Note that
position of class n in ıt . the term Pe takes into consideration the number of samples of each
class. Assuming equally distributed data over the four classes, Pe
Finally, once the accumulated probabilities are combined, the will be a constant, and hence  and the overall accuracy will be
final decision is selected using the majority voting rule (Eq. (2)). linearly dependent.
We used the j index for the evaluation, since this index takes
3. Experimentation into account, in a single score, the misclassification and the imbal-
ance that is present between all the considered classes, thanks to
The data were standardized (z-score), i.e., subtracting the mean the included  index, and at the same time emphasises the discrim-
and dividing the standard deviation of the training data. Since the ination of the most important arrhythmias (SVEB and VEB), thanks
MIT-BIH database is highly imbalanced, several weights equal to to the j index.
the ratio between the two classes that compose each l model were
employed to compensate these differences. The Radial Basis Func-
tion (RBF) kernel was fixed for all the experimentation process. The 3.2. Experiment 1: Features evaluation
same values for C and  were selected for all the L models. The
gamma value was fixed to 1/size(features). To adjust the penalty An OAO SVM model was independently trained for each fea-
parameter C, a 10-fold cross-validation strategy was performed ture in order to compare their single performance. Table 4 shows
over the training dataset (DS1), varying C over the grid {0.001, 0.01, the results that were obtained for the different models over the
0.1, 1, 10, 100}. Once the best parameters were selected, the models evaluation set (DS2) from the MIT-BIH database. The included per-
were trained again over the full training set (DS1) and tested over formance measurements are: the sensitivity (Se) and the positive
the evaluation set (DS2) following the inter-patient division [2]. predictive value (P+ ) for each class, the overall accuracy (Acc), the
mean Se and P+ of the four classes, the j index, the Cohen’s kappa (
3.1. Performance measurements index) and the j index. The best results regarding the most impor-
tant arrhythmias (SVEB and VEB) are obtained by the HOS feature,
Following the AAMI specifications, the performance measure- which achieved the best j index, followed very closely by our mor-
ments were computed from the confusion matrix (Table 3 ). They phological descriptor. Conversely, the wavelet presents a low score
include some particularities in the measurements computation [2], for j index. The model of this feature classifies most of the beats
e.g., they do not reward or penalize a classifier for the classification as class N or VEB, achieving the highest overall accuracy (Acc) due
of ventricular fusion (F) as (VEBs), Fv. The confusion matrix pro- to the imbalance, but at the same time it obtain a low score for
vides a complete description of any classification results. However, mean Se due to the minority classes (SVEB and F). In regards to the
due to imbalance of the MIT-BIH database, the overall accuracy or j index, i.e., considering the four classes with an emphasis on the
the mean accuracy do not represent well how good a classifier is. discrimination of the most important arrhythmias, R-R is the best
Suppose we have a classifier that only assigns the output of normal descriptor. Finally, the LBP obtain the worst j index score by far.
46 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48

Table 4
Results of the OAO SVM classifiers trained with the different features over MIT-BIH (DS2). Best feature per measurement in bold.

N SVEB VEB F Average

Features Se P+ Se P+ Se P+ Se P+ Se P+ Acc j index  index j index

R-R 0.769 0.989 0.505 0.264 0.802 0.472 0.874 0.058 0.738 0.446 0.762 2.044 0.368 0.439
HOS 0.572 0.977 0.719 0.106 0.736 0.690 0.765 0.045 0.698 0.455 0.589 2.251 0.216 0.389
Wavelet 0.857 0.953 0.106 0.079 0.959 0.426 0.013 0.056 0.484 0.378 0.826 1.570 0.384 0.388
Our Morph. 0.468 0.958 0.707 0.112 0.771 0.610 0.030 0.001 0.494 0.420 0.494 2.201 0.154 0.352
LBP 0.744 0.921 0.005 0.008 0.524 0.293 0.003 0.000 0.319 0.307 0.686 0.846 0.132 0.172

Table 5
All the possible combinations of the five features tested with the single SVM model and the ensembles of SVMs over MIT-BIH (DS2). Ensembles of SVMs are combined with
product, sum, and majority rules. Best results per configuration and measurement in bold.

Features Single-SVM Product rule Sum rule Majority rule

+ + +
R-R Wavelets HOS LBP Our Morph. Acc Se P j index Acc Se P j index Acc Se P j index Acc Se P+ j index

• • 0.848 0.518 0.465 0.486 0.866 0.568 0.459 0.525 0.857 0.560 0.450 0.501 0.893 0.531 0.472 0.531
• • 0.874 0.552 0.531 0.555 0.837 0.865 0.581 0.637 0.836 0.863 0.579 0.635 0.853 0.781 0.577 0.588
• • 0.902 0.484 0.470 0.499 0.836 0.516 0.438 0.450 0.832 0.497 0.427 0.425 0.834 0.474 0.432 0.389
• • 0.792 0.617 0.470 0.526 0.879 0.673 0.579 0.669 0.873 0.667 0.573 0.656 0.899 0.616 0.578 0.650
• • 0.841 0.440 0.473 0.400 0.835 0.665 0.519 0.519 0.821 0.672 0.504 0.500 0.923 0.502 0.524 0.571
• • 0.909 0.456 0.438 0.458 0.820 0.438 0.345 0.333 0.821 0.442 0.349 0.340 0.823 0.395 0.324 0.278
• • 0.844 0.428 0.466 0.385 0.779 0.500 0.458 0.422 0.765 0.520 0.462 0.434 0.873 0.447 0.461 0.429
• • 0.909 0.448 0.442 0.457 0.768 0.542 0.472 0.385 0.761 0.536 0.459 0.370 0.833 0.443 0.463 0.360
• • 0.779 0.647 0.505 0.511 0.765 0.646 0.510 0.512 0.745 0.661 0.510 0.504 0.845 0.594 0.532 0.549
• • 0.912 0.431 0.411 0.429 0.746 0.422 0.424 0.328 0.728 0.429 0.423 0.330 0.820 0.378 0.401 0.282

• • • 0.901 0.522 0.530 0.565 0.901 0.815 0.606 0.690 0.896 0.818 0.603 0.679 0.916 0.687 0.595 0.706
• • • 0.926 0.491 0.503 0.558 0.869 0.503 0.411 0.462 0.869 0.500 0.408 0.457 0.867 0.471 0.399 0.425
• • • 0.883 0.507 0.500 0.518 0.926 0.650 0.572 0.711 0.921 0.651 0.566 0.703 0.917 0.640 0.562 0.688
• • • 0.930 0.525 0.565 0.592 0.874 0.759 0.581 0.630 0.870 0.759 0.578 0.616 0.876 0.706 0.576 0.617
• • • 0.884 0.696 0.558 0.640 0.921 0.782 0.635 0.742 0.907 0.801 0.622 0.719 0.890 0.706 0.586 0.668
• • • 0.930 0.491 0.509 0.569 0.911 0.626 0.562 0.670 0.903 0.618 0.560 0.656 0.892 0.608 0.550 0.611
• • • 0.923 0.470 0.476 0.514 0.861 0.469 0.403 0.417 0.853 0.470 0.394 0.408 0.864 0.449 0.400 0.398
• • • 0.876 0.442 0.491 0.425 0.845 0.607 0.517 0.563 0.831 0.650 0.516 0.555 0.845 0.601 0.518 0.564
• • • 0.912 0.458 0.443 0.469 0.837 0.445 0.374 0.366 0.825 0.444 0.371 0.358 0.811 0.407 0.344 0.294
• • • 0.917 0.454 0.434 0.476 0.830 0.592 0.507 0.524 0.823 0.600 0.508 0.521 0.824 0.568 0.500 0.501

• • • • 0.933 0.509 0.547 0.604 0.902 0.602 0.528 0.581 0.893 0.590 0.506 0.548 0.913 0.553 0.534 0.587
• • • • 0.900 0.523 0.532 0.567 0.945 0.703 0.664 0.773 0.943 0.736 0.674 0.771 0.943 0.640 0.620 0.745
• • • • 0.926 0.494 0.510 0.562 0.908 0.525 0.480 0.552 0.902 0.515 0.465 0.530 0.897 0.509 0.468 0.517
• • • • 0.940 0.505 0.584 0.627 0.930 0.707 0.625 0.732 0.920 0.727 0.614 0.712 0.906 0.655 0.589 0.653
• • • • 0.922 0.470 0.471 0.508 0.866 0.502 0.469 0.480 0.856 0.506 0.458 0.469 0.890 0.468 0.458 0.467

• • • • • 0.933 0.509 0.551 0.606 0.938 0.625 0.617 0.707 0.933 0.621 0.596 0.692 0.934 0.621 0.587 0.704

3.3. Experiment 2: Comparison of single SVM vs. combination of that include the R-R interval, which was the best single feature.
multiple SVMs The higher j index = 0.773 score was achieved by the configuration
R − R, Wavelets, HOS, and Our Morph . with an ensemble of SVMs
The goal of this experiment is: evaluate if an ensemble of OAO using the product rule. On the other hand, the best single SVM con-
SVMs, combining the decision of the previous models, improves the figuration of j index = 0.640 was achieved by R − R, HOS, and Our
results over a single OAO SVM model trained with all the features Morph.
together. Table 5 contains the results obtained for all the possible
configurations of the employed features for the two alternatives.
3.4. Experiment 3: Comparison with the state-of-the-art
For the ensemble case, three combination rules were employed:
the product, the sum and the majority rule. As we previously said,
The goal of this experiment is to compare the result of our
we compare the results of the methods with the j index mea-
best configurations against other classification approaches, which
surement. The values of the overall accuracy (Acc), the mean Se
also employed the MIT-BIH public database with the same inter-
and P+ are also displayed. Results in Table 5 show that, in general,
patient division. As indicated, this is a well-known public database
ensembles of SVMs produce superior scores than a single SVM,
used as reference for validation of computational proposals of the
especially when the product rule is employed. However, there is
issue. Table 6 includes the comparison of our best configuration
an exception when the LBP descriptor is present. Note that single
of ensemble of SVMs and single SVM, next to some of the best
SVM models have only superior j index values than their ensemble
state-of-the-art methods. Results on Table 6 show that our ensem-
approaches when the LBP descriptor is present. This is due to the
ble achieves more than a 10% of improvement regarding to the j
fact that when the ensemble of SVMs is used, all the features add
index in comparison with the Zhang et al. method [22], which is
the same amount of confidence to the final decision. This causes a
the second highest one. On the other hand, our approach with a
deterioration of the performance if a feature is significantly worst
single SVM behaves similar to the state-of-the-art methods. As we
than the rest. On the other hand, when a single SVM is employed,
can see, looking at class level, the highest positive predictive value
the training process itself may discard the feature dimension that
is obtained by our ensemble method for all classes, except for the
behave worst. In general, the more features are added the better j
normal class (N). This means that our method tends to be more
index is achieved. Not surprisingly, the best configurations are those
conservative at assigning abnormal classes (SVEB, VEB, F) as the
 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48 47

Table 6
Results on MIT-BIH (DS2) comparing our best configurations against state-of-the-art methods.

N SVEB VEB F Average

Classifier Se P+ Se P+ Se P+ Se P+ Se P+ Acc j index  index j index

Our Ensemble SVM 0.959 0.982 0.781 0.497 0.947 0.939 0.124 0.236 0.703 0.664 0.945 3.165 0.755 0.773
Zhang et al. [22] 0.889 0.990 0.791 0.359 0.855 0.927 0.938 0.137 0.868 0.604 0.883 2.934 0.592 0.663
Our Single SVM 0.895 0.982 0.670 0.349 0.933 0.849 0.286 0.055 0.696 0.559 0.884 2.800 0.579 0.640
Mar et al. [9] 0.896 0.991 0.832 0.335 0.868 0.759 0.611 0.166 0.802 0.564 0.899 2.798 0.599 0.649
Chazal et al. [2] 0.871 0.992 0.760 0.385 0.803 0.866 0.894 0.086 0.832 0.570 0.862 2.767 0.532 0.612

Table 7 the segmentation step and one single lead (lead II) for the fea-
Confusion matrix over (DS2) MIT-BIH of our best configuration: ensemble of SVM
ture extraction. Instead of that, other state-of-the-art methods may
(R − R, W, HOS, Our. Morph .) using the product rule.
require many leads [9,22] and a more complex segmentation step
Algorithm [2,9,22] that includes the computation of the position and duration
n s v f Total of P, QRS, and T waves. The highest complexity in the segmentation
implies a higher error probability during this step.
Reference
N 42,244 1540 99 150 44,033 Possible future works, include the use of multiple leads, and
S 427 1601 21 1 2050 also the addition of more sophisticated data fusion methodolo-
V 90 75 3051 4 3220 gies, employing techniques such as Dempster–Shafer theory of the
F 256 2 82 48 388 evidence [41]. Ideally, each classifier model from a certain feature
Total 43,017 3218 3253 203 49,691 descriptor behaves better than the others at specific cases, hence,
a system that assigns more confidence to the right model at those
cases, would increase the performance of the system.
normal class (N) than the other methods. Regarding the sensitivity, All the code developed in this work is publicly available on the
our methods achieved higher values comparing to the state-of-the- repository.1
art methods for the majority classes, N and VEB. But at the same
time, the lowest sensitivity for class F was achieved by our meth- Acknowledgements
ods, causing their low value of mean sensitivity. This must be due
to the inclusion of the Wavelets and Our Morphology features, which This work was partially supported by the Research Project
obtained considerably lower sensitivity than the RR or HOS features RTC-2016-5143-1, financed by the Spanish Ministry of Economy,
for this class (see Table 4). Looking at the confusion matrix (Table Industry and Competitiveness and the European Regional Develop-
7 ), it is noticeable that most of the F beats were misclassified as ment Fund (ERDF). Also, this work has received financial support
class N and VEB. However, note also that due to the highly imbal- from the ERDF and the Xunta de Galicia, Centro singular de inves-
anced data samples from F class correspond only with a 1% of the tigación de Galicia accreditation 2016–2019, Ref. ED431G/01; and
total samples from MIT-BIH. Considering more appropriate mea- Grupos de Referencia Competitiva, Ref. ED431C 2016-047.
surements for this database that take into account the imbalance
of the data, like  index and j index, our methods present good
References
results, especially our ensemble of SVMs approach.
[1] E.J. da S. Luz, T.M. Nunes, V.H.C. de Albuquerque, J.P. Papa, D. Menotti, ECG
arrhythmia classification based on optimum-path forest, Expert Syst. Appl. 40
4. Conclusions (9) (2013) 3561–3573, http://dx.doi.org/10.1016/j.eswa.2012.12.063.
[2] P. de Chazal, M. O’Dwyer, R.B. Reilly, Automatic classification of heartbeats
A new approach for ECG classification based on an ensemble using ECG morphology and heartbeat interval features, IEEE Trans. Biomed.
Eng. 51 (7) (2004) 1196–1206, http://dx.doi.org/10.1109/TBME.2004.827359.
of SVMs was proposed. All the experiments were performed on
[3] S. Meek, F. Morris, Introduction. I – Leads, rate, rhythm, and cardiac axis, BMJ
the MIT-BIH public database, following an inter-patient scheme 324 (7334) (2002) 415–418, http://dx.doi.org/10.1136/bmj.324.7334.415.
division. In order to evaluate the results the j index, which [4] J. Pan, W.J. Tompkins, A real-time QRS detection algorithm, IEEE Trans.
Biomed. Eng. BME-32 (3) (1985) 230–236, http://dx.doi.org/10.1109/TBME.
were proposed as an adequate performance measurement for this
1985.325532.
database, has been employed. We tested several feature descrip- [5] Y.C. Yeh, W.J. Wang, QRS complexes detection for ECG signal: the difference
tors, including: R-R intervals, wavelets, HOS, LBP, and our own operation method, Comput. Methods Prog. Biomed. 91 (3) (2008) 245–254,
morphological descriptor. In the first experiment, a SVM model http://dx.doi.org/10.1016/j.cmpb.2008.04.006.
[6] H. Li, X. Wang, L. Chen, E. Li, Denoising and R-Peak detection of
was trained for each descriptor, being R-R intervals the one that electrocardiogram signal based on EMD and improved approximate envelope,
obtained the highest j index. For the second experiment, we eval- Circ. Syst. Signal Process. 33 (4) (2014) 1261–1276, http://dx.doi.org/10.1007/
uated the improvement of the ensemble of SVMs against a single s00034-013-9691-3.
[7] H. Li, X. Wang, Detection of electrocardiogram characteristic points using
SVM. All the possible combinations of the five feature descrip- lifting wavelet transform and Hilbert transform, Trans. Inst. Meas. Control 35
tors and the three combination rules, the product, the sum, and (5) (2013) 574–582, http://dx.doi.org/10.1177/0142331212460720.
the majority were tested. The obtained results show that, in the [8] G.B. Moody, R.G. Mark, The impact of the MIT-BIH arrhythmia database, IEEE
Eng. Med. Biol. Mag. 20 (3) (2001) 45–50, http://dx.doi.org/10.1109/51.
majority of the cases, our approach combining multiple SVM mod- 932724.
els is superior than concatenating all the features and training a [9] T. Mar, S. Zaunseder, J.P. Martnez, M. Llamedo, R. Poll, Optimization of ECG
single SVM model. Only when the LBP descriptor is employed, classification by means of feature selection, IEEE Trans. Biomed. Eng. 58 (8)
(2011) 2168–2177, http://dx.doi.org/10.1109/TBME.2011.2113395.
which was the worst single descriptor, our ensemble approach does
[10] A.S. Al-Fahoum, I. Howitt, Combined wavelet transformation and radial basis
not improve. For the best configuration, employing an ensemble neural networks for classifying life-threatening cardiac arrhythmias, Med.
of SVMs using R-R interval, wavelets, HOS and our morphologi- Biol. Eng. Comput. 37 (5) (1999) 566–573, http://dx.doi.org/10.1007/
BF02513350.
cal descriptor, combined with the product rule, the score obtained
for j index is over a 10% better than the previous machine learn-
ing approaches of the state-of-the-art. Additionally, it must be
emphasised that our method only requires the QRS detection for 1
https://github.com/mondejar/ecg-classification.
48 V. Mondéjar-Guerra et al. / Biomedical Signal Processing and Control 47 (2019) 41–48

[11] S. Osowski, T.H. Linh, ECG beat recognition using fuzzy hybrid neural [26] L. Breiman, Bagging predictors, Mach. Learn. 24 (2) (1996) 123–140, http://dx.
network, IEEE Trans. Biomed. Eng. 48 (11) (2001) 1265–1271, http://dx.doi. doi.org/10.1023/A:1018054314350.
org/10.1109/10.959322. [27] Y. Freund, R.E. Schapire, A decision-theoretic generalization of on-line
[12] G. de Lannoy, D. François, J. Delbeke, M. Verleysen, Weighted SVMs and learning and an application to boosting, J. Comput. Syst. Sci. 55 (1) (1997)
Feature Relevance Assessment in Supervised Heart Beat Classification, 119–139, http://dx.doi.org/10.1006/jcss.1997.1504.
Springer, Berlin, Heidelberg, Berlin, Heidelberg, 2011, pp. 212–223, http://dx. [28] T.G. Dietterich, G. Bakiri, Solving multiclass learning problems via
doi.org/10.1007/978-3-642-18472-7 17, ISBN:978-3-642-18472-7. error-correcting output codes, J. Artif. Int. Res. 2 (1) (1995) 263–286.
[13] C.C. Lin, C.M. Yang, Heartbeat classification using normalized RR intervals and [29] R.P.W. Duin, D.M.J. Tax, Experiments with classifier combining rules, in:
morphological features, Math. Probl. Eng. (2014) 10, http://dx.doi.org/10. Multiple Classifier Systems, Springer, Berlin, Heidelberg, Berlin, Heidelberg,
1155/2014/712474. 2000, pp. 16–29, ISBN:978-3-540-45014-6.
[14] S. Chen, W. Hua, Z. Li, J. Li, X. Gao, Heartbeat classification using projected and [30] B. Waske, J.A. Benediktsson, Fusion of Support Vector Machines for
dynamic features of ECG signal, Biomed. Signal Process. Control 31 (2017) classification of multisensor data, IEEE Trans. Geosci. Remote Sens. 45 (12)
165–173, http://dx.doi.org/10.1016/j.bspc.2016.07.010. (2007) 3858–3866, http://dx.doi.org/10.1109/TGRS.2007.898446.
[15] E.J. da S. Luz, W.R. Schwartz, G. Cmara-Chvez, D. Menotti, ECG-based [31] J. Kittler, M. Hatef, R.P.W. Duin, J. Matas, On combining classifiers, IEEE Trans.
heartbeat classification for arrhythmia detection: a survey, Comput. Methods Pattern Anal. Mach. Intell. 20 (3) (1998) 226–239, http://dx.doi.org/10.1109/
Prog. Biomed. 127 (Suppl. C) (2016) 144–164, http://dx.doi.org/10.1016/j. 34.667881.
cmpb.2015.12.008. [32] A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark,
[16] K.N. Rajesh, R. Dhuli, Classification of imbalanced ECG beats using re-sampling et al., PhysioBank, PhysioToolkit, and PhysioNet: components of a new
techniques and AdaBoost ensemble classifier, Biomed. Signal Process. Control research resource for complex physiologic signals, Circulation 101 (23) (2000)
41 (2018) 242–254, http://dx.doi.org/10.1016/j.bspc.2017.12.004. e215–e220.
[17] H. Khorrami, M. Moavenian, A comparative study of DWT, CWT and DCT [33] Testing and Reporting Performance Results of Cardiac Rhythm and ST
transformations in ecg arrhythmias classification, Expert Syst. Appl. 37 (8) Segment Measurement Algorithms, Association for the Advancement of
(2010) 5751–5757, http://dx.doi.org/10.1016/j.eswa.2010.02.033. Medical Instrumentation, 1998.
[18] R.J. Martis, U.R. Acharya, K. Mandana, A. Ray, C. Chakraborty, Application of [34] P. de Chazal, Detection of supraventricular and ventricular ectopic beats using
principal component analysis to ECG signals for automated diagnosis of a single lead ECG, 2013 35th Annual International Conference of the IEEE
cardiac health, Expert Syst. Appl. 39 (14) (2012) 11792–11800, http://dx.doi. Engineering in Medicine and Biology Society (EMBC) (2013) 45–48, http://dx.
org/10.1016/j.eswa.2012.04.072. doi.org/10.1109/EMBC.2013.6609433.
[19] H. Li, D. Yuan, X. Ma, D. Cui, L. Cao, Genetic algorithm for the optimization of [35] Y. Kaya, M. Uyar, R. Tekin, S. Yldrm, 1d-local binary pattern based feature
features and neural networks in ECG signals classification, Scientific Reports extraction for classification of epileptic EEG signals, Appl. Math. Comput. 243
(2017), http://dx.doi.org/10.1038/srep41011. (2014) 209–219, http://dx.doi.org/10.1016/j.amc.2014.05.128.
[20] W. Lu, H. Hou, J. Chu, Feature fusion for imbalanced ECG data analysis, [36] C. Cortes, V. Vapnik, Support-vector networks, Mach. Learn. 20 (3) (1995)
Biomed. Signal Process. Control 41 (2018) 152–160, http://dx.doi.org/10. 273–297, http://dx.doi.org/10.1023/A:1022627411411.
1016/j.bspc.2017.11.010. [37] J. Milgram, M. Cheriet, R. Sabourin, “One against one” or “one against all”:
[21] F. Melgani, Y. Bazi, Classification of electrocardiogram signals with Support which one is better for handwriting recognition with SVMs? in: G. Lorette
Vector Machines and Particle Swarm Optimization, IEEE Trans. Inf. Technol. (Ed.), Tenth International Workshop on Frontiers in Handwriting Recognition.
Biomed. 12 (5) (2008) 667–677, http://dx.doi.org/10.1109/TITB.2008.923147. Université de Rennes 1, La Baule (France): Suvisoft, 2006 https://hal.inria.fr/
[22] Z. Zhang, J. Dong, X. Luo, K.S. Choi, X. Wu, Heartbeat classification using inria-00103955.
disease-specific feature selection, Comput. Biol. Med. 46 (Suppl. C) (2014) [38] M.L. Soria, J.P. Martinez, An ECG classification model based on multilead
79–89, http://dx.doi.org/10.1016/j.compbiomed.2013.11.019. wavelet transform features 2007 Computers in Cardiology (2007) 105–108,
[23] H. Li, X. Feng, L. Cao, E. Li, H. Liang, X. Chen, A new ECG signal classification http://dx.doi.org/10.1109/CIC.2007.4745432.
based on wpd and apen feature extraction, Circ. Syst. Signal Process. 35 (1) [39] J. Cohen, A coefficient of agreement for nominal scales, Educ. Psychol. Meas.
(2016) 339–352, http://dx.doi.org/10.1007/s00034-015-0068-7. 20 (1) (1960) 37–46, http://dx.doi.org/10.1177/001316446002000104.
[24] H. Li, H. Liang, C. Miao, L. Cao, X. Feng, C. Tang, et al., Novel ECG signal [40] M. Fatourechi, R.K. Ward, S.G. Mason, J. Huggins, A. Schlgl, G.E. Birch,
classification based on KICA nonlinear feature extraction, Circ. Syst. Signal Comparison of evaluation metrics in classification applications with
Process. 35 (4) (2016) 1187–1197, http://dx.doi.org/10.1007/s00034-015- imbalanced datasets, 2008 Seventh International Conference on Machine
0108-3. Learning and Applications (2008) 777–782, http://dx.doi.org/10.1109/ICMLA.
[25] T.G. Dietterich, Ensemble Methods in Machine Learning, Springer, Berlin, 2008.34.
Heidelberg, Berlin, Heidelberg, 2000, pp. 1–15, http://dx.doi.org/10.1007/3- [41] T. Denoeux, A k-nearest neighbor classification rule based on
540-45014-9 1, ISBN:978-3-540-45014-6. Dempster–Shafer theory, IEEE Trans. Syst. Man Cybern. 25 (5) (1995)
804–813, http://dx.doi.org/10.1109/21.376493.