Accepted Manuscript

Title: ECG-based Heartbeat Classification for Arrhythmia

Detection: A Survey

Author: Eduardo Joséda S. Luz William Robson Schwartz

Guillermo Cámara Chávez David Menotti

PII: S0169-2607(15)00331-4
DOI: http://dx.doi.org/doi:10.1016/j.cmpb.2015.12.008
Reference: COMM 4033

To appear in: Computer Methods and Programs in Biomedicine

Received date: 27-3-2015

Revised date: 8-11-2015
Accepted date: 17-12-2015

Please cite this article as: Eduardo Joséda S. Luz, William Robson Schwartz, Guillermo
Cámara Chávez, David Menotti, ECG-based Heartbeat Classification for Arrhythmia
Detection: A Survey, <![CDATA[Computer Methods and Programs in Biomedicine]]>
(2015), http://dx.doi.org/10.1016/j.cmpb.2015.12.008

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.
*Highlights (for review)

Paper Title: ECG-based Heartbeat Classification for Arrhythmia Detection: A

Survey

t
Research Highlights:

ip
1. Surveys the feature description methods, and the learning algorithms

cr
employed
2. Also surveys the ECG signal preprocessing and the heartbeat segmentation

us
techniques
3. Description of databases used for methods evaluation indicated by the AAMI
standard

an
4. Discussion of limitations and drawbacks of the methods in the literature
5. Concluding remarks and future challenges are also pointed out.
M
d
p te
ce
Ac

Page 1 of 22
*Manuscript
Click here to view linked References

1
2 ECG-based Heartbeat Classification for Arrhythmia Detection: A Survey
3
4 Eduardo José da S. Luza , William Robson Schwartzb , Guillermo Cámara Cháveza , David Menottia,c,∗
5
a Universidade
6 Federal de Ouro Preto, Computing Department, Ouro Preto, MG, Brazil.
b UniversidadeFederal de Minas Gerais, Computer Science Department, Belo Horizonte, MG, Brazil.
7 c Universidade Federal do Paraná, Department of Informatics, Curitiba, PR, Brazil.
8
9
10
11

t
12

ip
13 Abstract
14 An electrocardiogram (ECG) measures the electric activity of the heart and has been widely used for detecting heart
15 diseases due to its simplicity and non-invasive nature. [REVIEW: By analyzing the electrical signal of each heartbeat,

cr
16 i.e., the combination of action impulse waveforms produced by different specialized cardiac tissues found in the
17 heart, it is possible to detect some of its abnormalities.] In the last decades, several works were developed to produce
18 automatic ECG-based heartbeat classification methods. In this work, we survey the current state-of-the-art methods of

us
19
ECG-based automated abnormalities heartbeat classification by presenting the ECG signal preprocessing, the heartbeat
20
segmentation techniques, the feature description methods and the learning algorithms used. In addition, we describe
21
22 some of the databases used for evaluation of methods indicated by a well-know standard developed by the Association for
23 the Advancement of Medical Instrumentation (AAMI) and described in ANSI/AAMI EC57:1998/(R)2008 (ANSI/AAMI,
24
25
26
an
2008). Finally, we discuss limitations and drawbacks of the methods in the literature presenting concluding remarks and
future challenges, and also we propose an evaluation process workflow to guide authors in future works.
Keywords: ECG-based signal processing, heartbeat classification, preprocessing, heartbeat segmentation, feature
27
extraction, learning algorithms, databases.
M
28
29
30
1. Introduction fatigue. An alternative is to use computational techniques
31
for automatic classification.
32 There are various types of arrhythmias and each type
d

33 A full automatic system for arrhythmia classification

is associated with a pattern, and as such, it is possible from signals acquired by a ECG device can be divided in
34 to identify and classify its type. The arrhythmias can be
te

35 four steps (see Fig. 1), as follows: 1) ECG signal prepro-

classified into two major categories. The first category cessing; 2) heartbeat segmentation; 3) feature extraction;
36
consists of arrhythmias formed by a single irregular heart- and 4) learning/classification. In each of the four steps,
37
beat, herein called morphological arrhythmia. The other an action is taken and the final objective is the discrimi-
p

38
39 category consists of arrhythmias formed by a set of irreg- nation/identification of the type of heartbeat.
ular heartbeats, herein called rhythmic arrhythmias. The The first two steps of a such classification system (ECG
ce

40
41 classification of normal heartbeats and the ones compos- signal preprocessing and heartbeat segmentation) have
42 ing the former group are on the focus of this survey. These been widely explored in the literature [2, 3, 4, 5, 6]. The
43 heartbeats produce alterations in the morphology or wave techniques employed during the preprocessing step directly
44 frequency, and all of these alterations can be identified by influence the final results, and therefore, should be care-
Ac

45 the ECG exam. fully chosen. The results related to the heartbeat segmen-
46 The process of identifying and classifying arrhythmias tation step, in the case of QRS detection, are very close to
47 can be very troublesome for a human being because some- optimal. However, there is still room for exploration and
48 times it is necessary to analyze each heartbeat of the ECG improvements in the steps related to classification (fea-
49 records, acquired by a holter monitor for instance, during ture extraction and learning algorithms). Even though
50 hours, or even days. In addition, there is the possibility the problem of ECG delineation is still open, it is not so
51 of human error during the ECG records analysis, due to
52 useful for the methods in the literature surveyed here.
53 This paper presents a survey of existing studies found
54 ∗ Corresponding author. Address: Universidade Federal do in literature regarding the ECG-based arrhythmia classi-
55 Paraná, Department of Informatics, 81.531-980, Curitiba, PR, fication methods and discusses the main techniques used
56 Brazil. Tel.: +55 41 3361 3655; fax: +55 41 3361 3031. for the construction of these automatic systems as well
Email addresses: eduluz@gmail.com, eduluz@iceb.ufop.br (E.
57 as two main paradigms used for evaluation: inter-patient
J. S. Luz), williamrobschwartz@gmail.com, william@dcc.ufmg.br
58 (W. R. Schwartz), gcamarac@gmail.com, guillermo@iceb.ufop.br and intra-patient [7, 8]. In addition, the most popular
59 (G. Cámara-Chávez), menottid@gmail.com, menotti@iceb.ufop.br, databases and the problems related to the evaluation of
60 menotti@inf.ufpr.br (D. Menotti).
61 Preprint submitted to Computer Methods and Programs in Biomedicine November 8, 2015
62
63
64 Page 2 of 22
65
 1
2
3
4
5
6
7
8
9
10
11

t
12

ip
13
14
15

cr
16
17
18

us
19
20 Figure 1: A diagram of the arrhythmia classification system.
21
22
Analog to
23
24
25
26
source

Isolation High-Pass Band Filter

an
Amplification Low-Pass Band Filter
Digital
Converter
(ADC)
ECG
signal

27
M
28
29 Figure 2: Simplified display of the hardware for the capture of ECG signals. Adapted from [1].
30
31
32 current methods found in literature are also discussed. tion 4 presents the concept of segmenting heartbeats from
d

33 From this discussion, a workflow is proposed to guide the the ECG signals and its commonly employed techniques.
34 evaluation process of future works. Note that this work- Section 5 deals with the key point for the success of ar-
te

35 flow for evaluation process constitutes an important con- rhythmia classification, i.e., the representation of a heart-
36 tribution of this survey work. In the literature, we find beat or the feature extraction process. Section 6 discusses
37 a survey of knowledge-based ECG interpretation [9] re- the most popular learning algorithms found in literature
p

38 viewing methods proposed in the 20th century. Clifford et for arrhythmia classification. Section 7 presents the rec-
39 al. [1] performed an extensive survey on the methods used ommended evaluation standard proposed by AAMI and
ce

40
for ECG signal analysis. Their study focused on the phys- describes the characteristics of the most utilized databases,
41
iology of the signal, as well as its processing techniques, indicated by the standard, to evaluate the classification ar-
42
mainly on the feature extraction and classification. In par- rhythmia methods. Section 8 presents some comments re-
43
44 ticular, Clifford et al. [1] did not focus on the problem of lated to the issue of selecting data for learning/evaluating
Ac

45 evaluating methods, which is the differential of our study, models for arrhythmia classification and its impact on the
46 in addition to a more up-to-date literature review on the final result. Finally, Section 9 discusses the limitations and
47 issue. Moreover, our survey on feature extraction brings a problems of the field and point out future challenges for
48 special review on feature selection. the research community.
49 The remaining of this paper is organized as follows.
50 Section 2 introduces the fundamental aspects of ECG sig-
2. ECG signal
51 nals; the state-of-art is described in Sections 3, 4, 5, and 6;
52 and the evaluation standards developed by the Associ- The heart is a muscle that contracts in a rhythmical
53 ation for the Advancement of Medical Instrumentation manner, pumping blood throughout the body. This con-
54 (AAMI) [10] and the databases recommended for these traction has its beginning at the atrial sine node that acts
55 standards, together with the criticisms related to the sys- as a natural pacemaker, and propagates through the rest
56 tems developed up to date and future challenges, are dis- of the muscle. This electrical signal propagation follows
57 cussed in Sections 7, 8, and 9.
58 a pattern [11]. As a result of this activity, electrical cur-
More specifically, Section 3 deals with the preprocess- rents are generated on the surface of the body, provoking
59 ing techniques most utilized in ECG signals, while Sec-
60 variations in the electrical potential of the skin surface.
61 2
62
63
64 Page 3 of 22
65
 1 These signals can be captured or measured with the aid of chest (V2 to V6) allowing a formation of 12 leads. The 10
2 electrodes and appropriate equipment. electrodes (12 leads) configuration can be seen in Figure 3.
3 The difference of electrical potential between the points From these configurations, several different leads can
4 marked by the electrodes on the skin, usually is enhanced be constructed to visualize the ECG signal. For exam-
5 with the aid of an instrumentation (operational) amplifier ple, Fig. 4 illustrates 3 particular leads: (I) formed by
6 with optic isolation. Then, the signal is submitted to a the electrical potential difference between the LA and RA
7 high-pass filter; and as a second stage, submitted to an electrodes; (II) formed by the electrical potential differ-
8 antialiasing low-pass filter. Finally, it appears in an ana- ence between the LL and RA electrodes; and (III) formed
9 logical to digital converter. The graphical registration of by the electrical potential difference between the LL and
10 this acquisition process is called electrocardiogram (ECG) LA electrodes.
11 (see Fig. 2). Since Augustus Desiré Waller demonstrated

t
12 the first human ECG in 1887, the electrical activity of the

ip
13 heart has been recorded [12]. Even so, the ability to rec-
14 ognize the normal cardiac rhythm and/or arrhythmias did
15
not become routine in medical check-ups until 1960.

cr
16
Nowadays, there are many approaches to measure-
17
18 ment/record ECG. da Silva et al. [13] provided a taxon-
omy of state-of-the-art ECG measurement methods: in-

us
19
20 the-person, on-the-person and off-the-person.
21 Within the in-the-person category, there are equip-
22 ments designed to be used inside human body, such as
23 surgically implanted ones, subdermal applications or even
24
25
26
ingested in the form of pills. These devices are used when
less invasive approach are not applicable.
Contrasting with the in-the-person category, there is
an Figure 3: Typical 10 electrodes configuration.

27 off-the-person category. Devices on this category are de-

M
28 signed to measure ECG without skin contact or with min-
29 imal skin contact. According to [13], this category is
30 aligned with future trends of medical application where
31 pervasive computer systems are a reality. Examples of
32
d

such equipments are the ones based on capacitive devices

33 which measure the electric field changes induced by the
34
body allowing ECG measurement at distance of 1cm or
te

35
more even with clothing between the body and the sen-
36
37 sor [13, 14, 15]. Figure 4: Morphology of the curve for leads I, II and III.
The majority of devices used for ECG measurements
p

38
39 are in the on-the-person category. Devices on this cate- The previously described lead II is one of the most uti-
gory normally require the use of some electrodes attached lized for diagnosing heart diseases. It highlights various
ce

40
41 to the skin surface. Examples of such equipments are bed- segments within the heartbeat, besides displaying three of
42 side monitors and holters. Nowadays, the standard devices the most important waves: P, QRS and T (see Fig. 5).
43 used for heartbeat analysis come from this category. These waves correspond to the field induced by the elec-
44 On equipments belonging into the on-the-person cate- trical phenomena occurring on the heart surface, denom-
Ac

45 gory, three or more electrodes are used to obtain the signal, inated atrial depolarization (P wave), ventral depolariza-
46 in which one of them serves as a reference for the others. tion (QRS complex wave) and repolarization (T wave).
47 Usually, the reference electrode is placed near the right leg. The patterns provoked by arrhythmias can deeply change
48 As such, there can be different visions of the ECG signal, these waves. Meanwhile, lead V and its correlate leads
49 depending on the pair of electrodes chosen to construct (V1, V2) favor the classification of ventricular related ar-
50 the signal. These differentiated visions are given the name rhythmias, since there are electrodes positioned on the
51 of leads. chest, improving the registry of action potentials on ven-
52 A widely used configuration of electrodes is one com- tricular muscle.
53
posed of 5 electrodes [16]: one of the electrodes is posi- Therefore, the leads most utilized for the automatic
54
tioned on the left arm (LA), one on the right arm (RA), heartbeat and arrhythmia classification are leads II and
55
56 one on the left leg (LL), one on the right leg (RL) and one V and the methods that use a combination of these two
57 on the chest, to the right of the external (V or V1). An- leads (and other combinations) are the ones that present
58 other widely employed setup uses 10 electrodes [16], where the best results to date [17]. In this sense, the recent
59 5 extra electrodes (besides V or V1 on the chest and LA, work by Tomasic & Trobec [18] reviews methods work-
60 RD, LL and RA on legs and arms) are positioned on the ing with reduced numbers of leads and approaches for the
61 3
62
63
64 Page 4 of 22
65
 1 synthesis of leads, concluding that the traditional 12 lead makes it unusable for diagnosing cardiac diseases. Archi-
2 system can be synthesized from a smaller number of mea- tectures with adaptive filters [23, 24] were also employed
3 surements [19]. In contrast, another study published by for noise removal from the ECG signals. However, accord-
4 de Chazal [20] demonstrated that similar effectiveness for ing to Thakor & Zhu [25], this technique has constraints
5 ECG arrhythmia classification can be obtained at a lesser and does not offer great advantages over the FIR digital
6 computational cost when using only one lead, compared
Current Controlled Trials in Cardiovascular Medicine 2005, 6:1 http://cvm.controlled-trials.com/content/6/1/1
filters. Xue et al. [26] surmount some of these difficulties
7 with methods using multiple leads [7]. by using adaptive filters based on neural networks such
8 that the noise reduction was significantly improved. This
9 QRS Complex strategy proportioned better detection of the QRS com-
10 R plex, when compared with the same method using linearly
11 adaptive filters.

t
12 In the last decade, many methods based on wavelet

ip
13 TPeak TEnd
transforms have been employed to remove noise, since they
14 Tpe Interval
preserve ECG signal properties avoiding loss of its impor-
15
tant physiological details and are simple from a computa-

cr
16 ST
Segment tional point of view [27, 28, 29]. Sayadi & Shamsollahi [2]
17 PR Interval

18 proposed a modification of the wavelet transform called the

multi-adaptive bionic wavelet transform and it was applied

us
P
19
20
U
to reduce noise and baseline variation of the ECG signal.
21 Tasc. Tdesc. This method presented superior results when compared to
Q
22 the ones based on the traditional wavelet transform.
Other methods have also presented interesting results
S
23
24
25
26
PR
Segment
QT Interval

Figure 4ECG highlighting the common parameters measured when assessing the QT/QTc interval
Normal
an
on noise attenuation. Sameni et al. [30] have proposed
the use of nonlinear Bayesian filters for ECG signal noise
reduction, presenting promising results. A new algorithm
Normal ECG highlighting the common parameters measured when assessing the QT/QTc interval.
27 Figure 5: Fiducial points and various usual intervals (waves) of a based on the Extended Kalman Filter [3], which incorpo-
M
heartbeat. Source [21].
28 rates the parameters of the ECG dynamic model for ECG
29 changes can be provided by use of a visual analogue scale scaled as "1". A classical "borderline" change would be noise reduction and signal compression, yielded a signifi-
30 Although on-the-person isgiventhe
a "5". mainstream on devices
(see Fig. 5). The degree of normality/abnormality in a par-
ticular case is estimated on a scale from 1 to 10, on which: cant contribution because the method showed the greatest
31 aiming heart diseases diagnoses, [13] have shown that data
QT dispersion (QTD)
"1" – is definite abnormal and "10" – is unquestionably
Increased dispersion on the QT interval of the electrocar-
normal. As an example, the flat-to-small negative T-waves effectiveness to date. Note that the works in [2, 30, 3]
32 captured with off-the-person diogram
ofbased
arrhythmias indevices can cardiomyop-
has been proposed as a marker for increased risk
in V5/V6 in the early phase of hypertension could be
be highly
d

patients with hypertrophic

scaled as "7", whereas the large negative T-waves in the
report their results in terms of signal to noise ratio.
correlated to data captured with traditional on-the-person
athy [48], long QT intervals [44], and sustained
same leads, in the case of severe aortic stenosis, would be
33 Techniques for preprocessing the ECG signal are widely
34 based equipments. The authors claim that off-the-person Page 8 of 13
explored, but the choice of which method to use is intrin-
te

(page number not for citation purposes)

35 based equipments can extend preventive medicine prac- sically connected with the final objective of the research.
36 tices by allowing ECG monitoring without interference on Methods focusing on the heartbeat segmentation from the
37 daily routine. In that sense, we encourage researchers to ECG signal (i.e., detection of the QRS complex, other
p

38 build ECG databases based on off-the-person devices to

39 waves or fiducial points aiming at heartbeat delimitation)
evaluate and validate heartbeat classification methods for tend to require a preprocessing that is different from the
ce

40
that category. methods focusing on the automatic classification of ar-
41
42 rhythmias.
43 3. Preprocessing Table 7 sumarizes the main reviewed references of
44 methods aiming at heartbeat classification and this table is
Ac

45 Among all proposals for reducing noise in ECG sig- explored further (Section 8). Those methods follow AAMI
46 nals, the simplest and most widely used is the implemen- instructions and the same protocol to report the results,
47 tation of recursive digital filters of the finite impulse re- but different preprocessing techniques are used. de Chazal
48 sponse (FIR) [22], which was made computationally possi- et al. [7] used two median filters to remove baseline wan-
49 ble with the advance in microcontrollers and microproces- der. One median filter of 200-ms width to remove QRS
50 sors. These methods work well for the attenuation of the complexes and P-waves and other of 600 ms width to re-
51 known frequency bands, such as the noise coming from the move T-waves. The resulting signal is then filtered again
52 electrical network (50 Hz or 60 Hz), since they allow quick with a 12-tap, low-pass FIR filter with 3-dB point at 35
53 and easy application of the reject-band-filter. The problem Hz. Same preprocessing is used in [31, 32, 33, 34, 35, 8, 36].
54 with this approach is that the frequency of the noise is not In [37] signal is preprocessed with 10th order low pass FIR
55
always known, which can be solved by applying filters for filter. Ye et al. [38] used a wavelet-based approach to re-
56
various frequency bands to the signal. However, the indis- move baseline wander [39] and then a band-pass filter at
57
58 criminate use of filters, i.e., high-pass and low-pass ones, 0.5 - 12Hz is applied to maximize QRS complex energy.
59 distorts the morphology of the signal, and many times, Bazi et al. [40] proposed the use of high pass filter for
60 noise artifacts and a notch filter for power network noise.
61 4
62
63
64 Page 5 of 22
65
 1 Lin & Yang [41] uses a second order low pass filter and two fluctuating baseline, nonlinear translations that enhance
2 median filter. In [42], the signal is subtracted by its mean the R peak and adaptive detection threshold were pro-
3 and then normalized. Escalona-Moran et al. [43] used the posed by Pan & Tompkins [49]. More sophisticated meth-
4 raw wave i.e., no preprocessing is applied. ods have also been used, such as methods based on neu-
5 Note that the methods cited in Table 7 use different ral networks [53], genetic algorithms [50], wavelet trans-
6 preprocessing approaches. However, the impact of these form [60, 61, 4], filter banks [46], Quad Level Vector [62],
7 approaches on automatic arrhythmia classification meth- among others. Table 1 displays the performance of some
8 ods is not clear. The considered state-of-the-art methods methods for heartbeat segmentation that use the MIT-BIH
9 do not even apply preprocessing on the signal. Although database for evaluation. Note that the SensitivitySEG
10 some studies exist relating preprocessing techniques with (Se) and P ositive predictivitySEG (+P ) values do not
11 the final performance of the automatic classification of ar- show great differences in the methods studied. It is im-

t
12 rhythmias, such as the work presented in [44], they are portant to highlight that the methods presented in this

ip
13 insufficient in number and more research in this area is table, contemplated a large spectrum of complexity, i.e.,
14 encouraged. It is worth noting that the state-of-the-art from very simple methods to more elaborated ones.
15
methods for automatic arrhythmia classification do not Some algorithms also propose to identify other waves

cr
16
use state-of-the-art preprocessing methods signal to noise associated with heartbeats, such as the P wave and the T
17
18 ratio improvement. wave [4, 63, 64, 65], which can be useful for arrhythmia
classification methods, since more information about the

us
19
20 heartbeats can be obtained.
4. Segmentation
21 Although heartbeat segmentation is not the main fo-
22 Heartbeat segmentation methods (i.e., detection of the cus of this survey, note that this stage is of paramount
23 R peak or the QRS complex) have been studied for more importance in the arrhythmia heartbeat classification pro-
24
25
26
than three decades [49, 53, 46, 54, 55] and the genera-
tions of these algorithms and newly developing methods
reflect the evolution of the processing power of comput-
an
cess, since some errors here are propagated to the following
stages and have a strong impact in the final classification
of the arrhythmia system. However, a large majority of
27 ers. With the facility of using faster processing comput- the reviewed researches herein utilized databases in which
M
28 ers, authors stopped worrying about computational cost the events related to heartbeat segmentation, i.e., the de-
29
and started concentrating on the heartbeat segmentation tection of the R peak or the QRS complex, are identified
30
accuracy. Two measures are usually considered for evalu- and previously labeled, reducing the segmentation stage
31 to a simple search of a labeled event in the database. In
32 ating the accuracy of heartbeat segmentation: sensitivity
d

and positive predictivity, which are defined as: this way, the results reported by these works disregard the
33 impact of segmentation step even though the database la-
34
(1) beling is prone to human errors. Therefore, evaluating the
te

35 SensitivitySEG = T P/(T P + F N ),
impact of different segmentation algorithms on automatic
36
and arrhythmia classification methods can be a promising re-
37
search direction.
p

38
P ositive predictivitySEG = T P/(T P + F P ), (2) Ye et al. [66] proposed a test to investigate the robust-
39
ness of their feature extraction method against one seg-
ce

40 where T P (True Positive), F P (False Positive) and F N

41 mentation issue, the R-peak mislocate error. A Gaussian-
(False Negative) indicate the number of heartbeats cor-
42 distributed artificial jitter was used to add error on R-peak
rectly segmented, number of segmentations that do not
43 annotations. We suggest to other authors to incorporate
correspond to the heartbeats, and number of segmenta-
44 such test in future works aiming automatic heartbeat clas-
Ac

45 tions that were not performed, respectively. sification.

46 For a fair comparison of the methods focusing on
47 the heartbeat segmentation, a standard database needs
48 to be used. The most utilized, and recommended by 5. Feature Extraction
ANSI/AAMI for the validation of medical equipment [10],
49 The feature extraction stage is the key to the success
50 is the MIT-BIH database for arrhythmia analysis [56] - in in the heartbeat classification of the arrhythmia using the
51 this case, used for heartbeat segmentation, although other ECG signal. Any information extracted from the heart-
52 databases are also used, such as that of AHA [57] and
beat used to discriminate its type maybe considered as
53 that of CSE [58]. However, according to Kohler et al. [59],
a feature. The features can be extracted in various forms
54 many of the methods presented in the literature do not
directly from the ECG signal’s morphology in the time do-
55 use a standardized database, or use only part of it, which
main and/or in the frequency domain or from the cardiac
56 makes it difficult to fairly compare methods. rhythm. Most popular methods proposed in literature are
57 An approach widely used for segmentation, due to its discussed in Section 5.1.
58 simplicity and promising results, is based on digital fil-
59 Even though some works regard feature extraction and
ters for the attenuation of the noise and removal of the
60 feature selection as two interchangeable terms, these two
61 5
62
63
64 Page 6 of 22
65
 Table 1: Effectiveness of heartbeat segmentation methods. # and % stand for absolute and percentage numbers. The MIT-BIH Arr. database
1 is used in all methods.
2
3
4 heartbeats TP FP FN error Se +P
Method
(#) (#) (#) (#) (%) (%) (%)
5 Martinez et al. [4] 109428 109208 153 220 0.34 99.80 99.86
6 Moody & Mark [45] 109428 107567 94 1861 1.79 98.30 99.91
7 Li et al. [5] 104182 104070 65 112 0.17 99.89 99.94
Afonso et al. [46] 90909 90535 406 374 0.86 99.59 99.56
8 Bahoura et al. [6] 109809 109635 135 184 0.29 99.83 99.88
9 Lee et al. [47] 109481 109146 137 335 0.43 99.69 99.88
Hamilton & Tompkins [48] 109267 108927 248 340 0.54 99.69 99.77
10 Pan & Tompikins [49] 109809 109532 507 227 0.71 99.75 99.54
11

t
Poli et al. [50] 109963 109522 545 441 0.90 99.60 99.50
12 Moraes et al. [51] N/R N/R N/R N/R N/R 99.22 99.73

ip
Hamilton [52] N/R N/R N/R N/R N/R 99.80 99.80
13
14
15

cr
16 process are in fact different. While feature extraction is de- Doquire et al. [71] confirmed the efficiency of normalized
17 fined as the stage that involves the description of a heart- RR-intervals by means of feature selection techniques.
18 beat, feature selection consists in choosing a subset with Other features extracted from the heartbeat intervals

us
19 the most representative features with the objective to im- are also found in literature, such as other distances be-
20 prove the classification stage. Section 5.2 is dedicated to tween the fiducial points of a heartbeat (here called ECG-
21
describe feature selection approaches. intervals or ECG segments), as can be seen in Fig. 5.
22
Among these intervals, the QRS interval, or the duration
23
24
25
26
5.1. Feature Extraction

Table 2: Typical feature of a normal ECG signal, with a cardiac

frequency of 60 beats per minute (bpm) of a healthy adult. Source [1].
an
of the QRS complex, is the most utilized. Some types of
arrhythmias provoke variations in the QRS interval, mak-
ing it a good discriminating feature [7, 72]. It is worth
27 mentioning that there exist other algorithms available to
M
28 determine these fiducial points, such as the one proposed
29 Feature Normal value Normal variation by Laguna et al. [63]. Table 2 displays the standard val-
30 P wave 110 ms ± 20 ms ues for these intervals, considering a healthy human being
PQ/PR interval 160 ms ± 40 ms
31 QRS Width 100 ms ± 20 ms with no cardiac abnormalities.
32 ± Features extracted from the domain of time/frequency
d

QT interval 400 ms 40 ms
amplitude of P 0.115 mV ± 0.05 mV
33 amplitude of QRS 1.5 mV ± 0.5 mV together with the features of the RR interval appear as
34 ST level 0 mV ± 0.1 mV part of the methods that produced the highest accuracies
te

amplitude of T 0.3 mV ± 0.2 mV

35 in literature to date (see Table 7). The simplest way to
36 extract features in the time domain is to utilize the points
37 of the segmented ECG curve, i.e., the heartbeat, as fea-
p

38 The most common feature found in the literature is tures [73, 74]. However, the use of samples of the curve
39 calculated from the cardiac rhythm (or heartbeat inter- as features is a technique that is not very efficient, since
val), also known as the RR interval. The RR interval is
ce

40
besides producing a vector of the features with high di-
41 the time between the R peak of a heartbeat with respect mensions (depending on the amount of samples used to
42 to another heartbeat, which could be its predecessor or represent the heartbeat), it suffers from several problems
43 successor. With exception of patients that utilize a pace- related to the scale or displacement of the signal with re-
44 maker, the variations perceived in the width of the RR
Ac

45 spect to the central point (peak R).

interval are correlated with the variations in the morphol- Aiming at reducing the dimension of the feature vec-
46 ogy of the curve, frequently provoked by arrhythmias [1].
47 tor, various techniques have been applied directly on the
Thus, the features in the RR interval have a great capacity samples that represent the heartbeat (in the neighbor-
48
to discriminate the types of heartbeats and some authors hood of the R peak) as principal component analysis
49
50 have based their methods only on using the RR interval (PCA) [75, 76, 77], or independent component analysis
51 features [67, 68, 69]. Variations of this feature are used to (ICA) [78, 79, 80], in which new coefficients are extracted
52 reduce noise interference and are very common, e.g., the to represent the heartbeat. Chawla [81] presents a com-
53 average of the RR interval in a patient for a certain time parative study between the use of PCA and ICA to reduce
54 interval [70]. the noise and artifacts of the ECG signal and showed that
55 Lin & Yang [41] have shown that the use of a nor- PCA is a better technique to reduce noise, while ICA is
56 malized RR-interval significantly improves the classifica- better one to extract features. The ICA technique enables
57 tion results. Only normalized RR-intervals are used in statistically separate individual sources from a mixing sig-
58 that work and the results are comparable to the state-of- nal. The ECG is a mix of several action potentials and each
59 the-art methods even under the inter-patient paradigm. action potential could be strongly related to an arrhyth-
60
61 6
62
63
64 Page 7 of 22
65
 1 sampled at 360Hz), was sub-divided and presented in 18
2 samples (see Fig. 6). In the literature, the sub-sampled
3 ECG wave is also called morphology or morphological fea-
4 tures.
5 Recently, random projections have also been employed
6 for such aim, as in [86, 42]. Huang et al. [42] show that fea-
7 tures extracted with random projections produced results
8 comparable to the state-of-the-art methods, even when
9 considering the inter-patient paradigm.
10 Other techniques have also been employed, such as lin-
11 ear predictive coding [87], high order accumulates [88, 89],

t
12 clustering [84, 90, 91], correlation dimension and largest

ip
13 Lyapunov exponent [92, 93], Hermite transform [94], local
14 fractal dimension [95].
15
Although various techniques have been considered,

cr
16 Figure 6: Feature number reduction by means of interpolation.
Source [7]. most of the studies presented in literature use wavelet
17
18 transforms and researchers claim that this is the best
method for extracting features from the ECG signal [44,

us
19
mia class. The rationale behind ICA for ECG heartbeat 96, 97]. The wavelet transform allows information extrac-
20
21 classification is to separate the action potentials sources tion from both frequency and time domains, different from
22 as well as the noise sources. The PCA technique sepa- what is usually achieved by the traditional Fourier trans-
23 rates the sources according to the energy contribution to form [98] which permits the analysis of only the frequency
24
25
26
the signal. The study presented in [81] suggest that noise
sources on this base have low energy and are difficult to
isolate and that the individual sources isolated by ICA are
an
domain. Within the types of wavelet transform, the dis-
crete wavelet transform (DWT) is the most popular for
ECG signal classification due to its easy implementation.
27 promising features for ECG classification. Moreover, it Besides DWT, continuous wavelet transform (CWT)
M
28 has been shown that the combination of these two tech- has also been used to extract features from the ECG sig-
29 niques, i.e., PCA for noise reduction and ICA for feature nals [99], since it overcomes some of the DWT drawbacks,
30 extraction, can offer greater advantages when compared such as the coarseness of the representation and instabil-
31 to using only one of them. Another technique based on ity. However, CWT is not largely used due to the fact
32 PCA, the Kernel Principal Component Analisys (KPCA),
d

that its implementation and its inverse are not available

33 was used by Kanaan et al. [82]. In that work, a compar- in standard toolboxes (such as MATLAB wavelet Tool-
34 ison between PCA and KPCA was performed and it was box) and CWT should be carefully discretized for the use
te

35 concluded that KPCA is superior to the PCA technique as a CWT analyzer. In addition, even though Addison [99]
36 for classifying heartbeats from the ECG signal. Accord-
37 emphasizes the high computational cost as a disadvantage
ing to Kallas et al. [83], KPCA performs better, due to its for using CWT, it has been employed successfully even on
p

38
nonlinear structure. simple medical equipments for at least a decade. Finally,
39
Özbay et al. [84] used clustering techniques directly in Addison [99] defends the use of DWT, together with CWT,
ce

40
41 points sampled from the curve to reduce from 106 samples because they offer gain over the methodologies used nowa-
42 to 67 clusters/points. The authors also used a clustering days, in which the authors use only one of the transforms.
43 technique to increase the number of features to 212, but According to Güler & Übeyli [44], the choice of the
44 there were no significant differences in the results. mother wavelet function used for feature extraction is cru-
Ac

45 Asl et al. [85] used Generalized Discriminant Analysis cial for the final performance of the classification model.
46 (GDA) to reduce the dimensions of the features of the This choice should be carefully analyzed in order not to
47 heartbeat interval type to classify rhythmic arrhythmias. lose important ECG signal details. Besides the choice of
48 They reported an accuracy close to 100% for this type of the mother wavelet function, the order of filter and level
49 arrhythmia using the MIT-BIH database. However, the of decomposition are parameters that influence the final
50 authors did not take care to separate the heartbeats of results of the arrhythmia classification. Daamouche et
51 the same patient used during training and testing (intra- al. [100] proposed the use of the Particle Swarm Optimiza-
52 patient paradigm), which is a serious concern discussed tion (PSO) technique for optimizing these parameters, and
53 further. The inter-patient paradigm should be considered concluded that this process improve the final results.
54 for a more realistic scenario. In the literature, various statistical features extracted
55 Simpler techniques, such as interpolation, have also
56 from the coefficients of wavelet transform are proposed,
been used to reduce the number of points representing the such as mean, standard deviation, energy [44] and coeffi-
57
heartbeat. An example of this technique is presented by de cient variance [101]. These features have a great advan-
58
59 Chazal et al. [7], in which the heartbeat, originally repre- tage since they are immune to the variations of fiducial
60 sented by 250 samples (approximately 600ms of the curve, point marking. Some authors used techniques to reduce
61 7
62
63
64 Page 8 of 22
65
 1 the space of the features after applying the wavelet trans- and VEB.
2 form, such as in the work of Song et al. [102] who compared
3 the PCA and linear discriminant analysis (LDA) tech- 5.2. Feature Selection
4 niques for dimensional reduction after the use of wavelet According to Llamedo & Martinez [37], many authors
5 transform. Wang et al. [103] and Polat & Güneş [104] have used techniques that reduce the feature space, but
6 also employed PCA to reduce features formed by wavelet few have investigated techniques for feature selection in
7 coefficients and also reported a significant improvement the context of arrhythmia classification. Llamedo & Mar-
8 their results. According to Güler & Übeyli [44], the tinez [37] employed, for the first time in literature, an
9 Daubechies wavelets are the most appropriated mother algorithm for feature selection by using floating sequen-
10 wavelets for ECG heartbeat classification. Among them, tial search for arrhythmia classification. This method in-
11 the Daubechies of order 2 offers the best accuracy. terchanges algorithms executing forward and backward

t
12 Although many techniques have been proposed to ex- searches to obtain a set with the most robust features

ip
13 tract and reduce features from ECG signals aiming heart- and avoid local optima in the feature space. The pro-
14 beat classification, only a few of them have considered the posed method achieved better results than the state-of-art
15
inter-patient paradigm as one can see in Table 7. There- method using only eight selected features.

cr
16
fore, it is difficult to evaluate whether features extracted Recently, Mar et al. [34] also performed feature selec-
17
18 with PCA, ICA, GDA and others are useful to discrimi- tion by using the floating sequential search [108]. In that
nate patients or heartbeats. study, the authors analyzed a set of possibilities of the fea-

us
19
20 The variance of the autocorrelation function is consid- ture selection, searching for a trade-off between the num-
21 ered to be a measure of similarity or coherence between a ber of features and accuracy. The aim of that research was
22 signal and its shifted version [101]. This technique is used to make a specially developed method adequate for ambu-
23 for feature extraction from wavelet coefficients [101, 37], latory monitoring; that is, to be specially useful in real
24
25
26
and have demonstrated to be effective in the discrimina-
tion of arrhythmic heartbeats.
an
world applications. For such aim, an objective function
optimized by a feature selection method, was especially
developed to be an indicator of the quality of the arrhyth-
27 mia classifications from an ECG signal. In addition to the
M
28 linear discriminant (LD) classifier used in previous works,
29 Mar et al. [34] employed a multi-layer perceptron. How-
30 ever, neither of these results were better than those pro-
31 posed by de Chazal et al. [7] and the work of Llamedo &
32
d

Martinez [37] in terms of accuracy. Nonetheless, the focus

33 of Mar et al. [34] work was on the maintenance of accuracy
34
with the use of a reduced number of features.
te

35
Feature selection techniques can bring various benefits
36
37 for the classification methods, such as the increase of the
generalization power of the classification algorithms and
p

38
39 the reduction of the computational cost, due to the fact
that they use a smaller number of features to construct
ce

40
41 the final model [34]. However, in the works analyzed in
42 this survey, these techniques were little explored.
43 Doquire et al. [71] compare wrapper feature selection
44 technique against a filter feature selection technique and
Ac

45 more than 200 types features (dimensions) are considered

46 Figure 7: VCG setup using two heartbeats of MIT-BIH’s record 202. for the task. The wrapper feature selection is used with the
Adapted from [37]
47 weighted LD model using a forward-backward search strat-
48 egy. The filter technique employed is the mutual informa-
49 The vectorcardiogram (VCG) is a representation of the tion in conjunction with ranking approach and weighted
50 ECG signal in two dimensions that integrates information SVM (Support Vector Machines). According to the au-
51 from two leads (see Fig. 7). Features extracted with VCG thors, results have shown that higher figures are obtained
52 were used in [105, 31, 106, 37]. According to Goldberger et when a very small number of features are selected. They
53 al. [107], heartbeat classification categorized as Supraven- stressed that the most important features appears are R-
54 tricular ectopic beat (SVEB) and Ventricular ectopic beat R intervals, the amplitude and length of the T wave, and
55 (VEB) (arrhythmic heartbeats) can be favored by informa- 2nd-order statistics. Also they claimed that the mutual in-
56 tion from leads of type V1, V2 or V4. Because of this, it formation criterion is a powerful tool for feature selection
57 is believed that the features extracted by VCG (combined
58 in this scenario.
with leads II and V1) can help to better discriminate mi- According to Zhang et al. [35], many features are asso-
59 nority and important arrhythmic classes such as SVEB
60 ciated with mathematical interpretation and do not have a
61 8
62
63
64 Page 9 of 22
65
 1 clear meaning to physicians. Usually, the authors employ Moavenian & Khorrami [119] proposed the use of a
2 several combined features and the understanding of which new kernel function for capturing data from SVM. In that
3 feature contributes to detection of which class of heartbeat work, it was used the same methodology for comparing the
4 is also not clear in the literature. Aiming that, Zhang et results obtained from a SVM and a Multilayer Perceptron
5 al. [35] proposed a heartbeat class-specific feature selection Artificial Neural Network (MLP-ANN). While SVM was
6 scheme to allow the investigation of feature contribution more efficient in execution time, both in the training and
7 for each arrhythmia/heartbeat class. Thus, we suggest the in the testing, MLP performed better in terms of accu-
8 incorporation of this approach on works aiming heartbeat racy, Sensitivity (Se), positive prediction (+P ) and false
9 classification. It could bring important contribution to the positive rate (F P R).
10 literature by allowing better understanding of correlation Since SVM presents a negative behavior for imbal-
11 among heart diseases and features extracted from ECG. anced classes, database balancing techniques for the train-

t
12 State-of-art techniques for attribute selection, such as ing phase, which are little explored for this problem, can

ip
13 Genetic Algorithms (GA) [109, 110] and particle swarm be studied in future research, as for example, more so-
14 optimization (PSO) [111, 112] can also provide promising phisticated sampling techniques, i.e., Synthetic Minority
15
results and should be better investigated in future works. Over-sampling Technique (SMOTE) [120].

cr
16
17
18 6.2. Artificial Neural Networks (ANN)
6. Learning Algorithms
The ANN architectures mostly used for arrhythmia

us
19
20 Once the set of features has been defined from the classification are Multilayer Perceptrons (MLP) and Prob-
21 heartbeats, models can be built from these data using ar- abilistic Neural Networks (PNN). According to Yu &
22 tificial intelligence algorithms from machine learning and Chen [101], models constructed with PNN are compu-
23 data mining domains [113, 114, 115] for arrhythmia heart- tationally more robust and efficient than the traditional
24
25
26
beat classification.
The four most popular algorithms employed for this
task and found in the literature are: support vector
an
MLP. However, in [121, 84, 122, 88], it was proposed a
hybrid neuro-fuzzy network methods in order to minimize
the problems of MLP, increasing its generalization and re-
27 machines (SVM) [40, 38, 66], artificial neural networks ducing its training time.
M
28 (ANN) [34, 116, 69] and linear discriminant (LD) [7, 37, Many other approaches based on ANN have been pro-
29
17], and Reservoir Computing With Logistic Regression posed.
30
(RC) [43]. Note that the state-of-the-art method aiming Güler & Übeyli [44] used combined neural networks
31
32 heartbeat classification uses RC algorithm. in order to obtain a more generic method from a more
d

33 Due to their importance for cardiac arrhythmic classi- sophisticated form of cross-validation. However, of all the
34 fication, these four classifiers (SVM, ANN, LD, and RC) articles mentioned in this study, only that of Mar et al. [34]
te

35 are discussed in the next subsections (Sections 6.1, 6.2, 6.3 used MLP with a more fair evaluation protocol by apply-
36 and 6.4). Then, Section 6.5 reviews other techniques that ing the patient division scheme proposed by de Chazal et
37 also have been employed to arrhythmia classification. al. [7]. Thus by using the reported results in the works of
the methods that utilizes ANN as classifier is impossible to
p

38
39 6.1. Support Vector Machines (SVM) makes a fair comparison. Finally, Mar et al. [34] compared
MLP with Linear Discriminants and found that MLP was
ce

40 SVM is one of the most popular classifiers found in lit-

41 erature for ECG-based arrhythmia classification methods. significantly superior.
42 Park et al. [33] used SVM and validated the method ac- Combining classifiers had been little explored for the
43
cording to AAMI standards and the data set split scheme task in question. According to Osowski et al. [91], a com-
44 bination of classifiers not only reduces the overall error in
Ac

proposed by de Chazal et al. [7]. These same authors used

45 the neural networks, but also reduces the incidence of false
46 SVM in a mock-hierarchy configuration to resolve the im-
balance of the MIT-BIH database, and reported promising negatives.
47
48 values. de Lannoy et al. [32] managed to overcome the im-
balance of the MIT-BIH database with SVM, alternating 6.3. Linear Discriminants (LD)
49
50 the objective function for each class (Weighted SVM). Ex- The Linear Discriminant is a statistic method based on
51 pressive gains were reported for the SVEB and F classes. the discriminant functions [114]. Such functions are esti-
52 Various approaches with SVM variations have been mated from a training set of data and try to linearly sepa-
53 proposed, such as a combination of the fuzzy theory to rate the feature vector, being adjusted by the weight vector
54 refine SVM classification [117], combined with an ensem- and a bias. The criteria for calculating the weight vector
55 ble of classifiers [42], genetic algorithms combined with varies according to the model adopted. In [7], the pa-
56 restricted fuzzy SVM [118] and least squares SVM [104]. rameters were determined using the maximum-likelihood
57 Huang et al. [42] used the SVM in a hierarchical manner calculated from training data.
58 with a maximum voting strategy and report significantly Linear discriminants are the classifiers more used in
59 improvements. methods that follow the scheme proposed by de Chazal et
60
61 9
62
63
64 Page 10 of 22
65
 1 al. [7] and recommended by AAMI. The authors of that al. [7], and no one also followed the AAMI recommenda-
2 research claim that the classifier was chosen for its simplic- tions. In addition, the computational cost of these meth-
3 ity and for the fact that they did not want to emphasis the ods was not investigated.
4 classifier, but instead, the proposed features. Amongst its Clustering techniques are widely used along with Ar-
5 advantages, LD can easily overcome problems generated tificial Neural Networks. According to Özbay et al. [84],
6 by the imbalance of the training set (a difficulty presented they can improve the generalization capacity of the neu-
7 by approaches based on SVM). When using the scheme ral networks and diminish the learning time. Some works
8 proposed in [7], it is a great challenge to tune SVM and used unsupervised clustering techniques to agglomerate all
9 MLP classifiers to obtain promising classification effective- of the heartbeats in the record of a given patient into clus-
10 ness for the minority SVEB and VEB classes (see Table 9). ters [131] and the final classification of each cluster, i.e.,
11 Moreover, the LD classifier requires less training time, if the heartbeats of that group, is then defined by a hu-

t
12 compared to SVM and MLP, as it is not iterative. That is, man specialist [138, 130, 73]. Other works in this same

ip
13 it simply calculates statistics from the training data and way [139, 140, 17], using linear discriminant as classi-
14 then, the classification model is defined. fiers and fair evaluation schemes present promising results
15
which are reliable for real-world applications. It is impor-

cr
16 6.4. Reservoir Computing With Logistic Regression (RC) tant to note that this semi-automatic (or patient-specific
17
18 According to Rodan & Tiňo [123], reservoir comput- paradigm) and promising approach still depends on a hu-
ing models are dynamical models aiming to process a time man specialist.

us
19
20 series signal in two parts: represent the signal through HMM is widely used to audio and speech signal anaysis
21 a non-adaptable dynamic reservoir and a dynamic read- and recognition [141, 142]. Coast et al. [132] used HMM for
22 out from the reservoir. More details regarding RC can be the arrhythmia classification problem, other studies have
23 found in [124]. used this technique to analyze ECG signals. For instance,
24
25
26
The state-of-the-art method for heartbeat classification
uses RC [43]. According to Escalona-Moran et al. [43],
their approach uses a simple nonlinear dynamical element
an
Andreao et al. [143] validated the use of HMM for ECG
analysis in medical clinics (real world).
The Optimum-path Forest (OPF) classifier was used
27 subject to a delayed feedback where each point of the ECG for arrhythmia classification for the first time by Luz et
M
28 signal is sampled and held during one delay time and then al. [134]. In that work, the OPF performance, in terms of
29 multiplied by a binary random mask. The learning process computational cost and overall accuracy, was compared to
30 is accomplished with logistic regression. The technique other three classifiers: Bayesian, SVM and MLP. Experi-
31 appears to be robust to the class imbalance of the dataset. ments showed that OPF obtained, in average, comparable
32
d

Besides, it achieves the best results in the literature to results, revealing it as a promissory approach.
33 date (see Table7). In addition, the authors also claim that Methods that use a decision tree allow an interpreta-
34
the technique is suitable to implement in hardware due to tion of the decisions made by the model [68]. However,
te

35
its low computational cost, which allows the development this type of method is not efficient for continuous features
36
37 of real time applications for heartbeat classification. (belonging to a set of real numbers) [144, 145] and feature
vectors of large dimensions [146]. Thus, methods that use
p

38
39 6.5. Other Techniques decision trees consider only a few features. For example,
Many other methods for arrhythmia classification have in [68], only the features in the RR interval were used
ce

40
41 been developed using other machine learning and data in the decision tree. Meanwhile the hyperbox classifiers,
42 mining algorithms, such as decision trees [125, 126, 68], besides providing high level of interpretation of the classi-
43 nearest neighbors [127, 128, 129], clustering [73, 130, fication rules, are also more efficient for higher dimension
44 131], hidden Markov models [132, 133], hyperbox classi- feature vectors [105]. Mert et al. [147] used a combination
Ac

45 fiers [105], optimum-path forest [134], conditional random technique of bagging and decision tree. According to the
46 fields [8] and rules-based models [135, 67, 136]. authors, the Bagged Decision Tree demonstrated greater
47 Algorithms with a lazy approach, such as the k Nearest accuracy and a better capacity to discriminate the classes.
48 Neighbors (kNN), are not much used for the problem of ar- The methods with the greatest interpretation level are
49 rhythmia classification, since their efficiency is intimately the ones that use a set of rules. The set of rules pre-
50 connected to previous knowledge to perform the classifi- sented by Tsipouras et al. [135, 67, 136] was obtained to-
51 cation of each sample that is represented by the complete gether with cardiologists and are related to a morphologi-
52 training set, which leads to a high computational cost dur- cal tachogram for arrhythmic events. Methods constructed
53
ing the testing phase. This cost can invalidate its use for in conjunction with rules usually present a worser perfor-
54
diagnosis in real time. Mishra & Raghav [95] used a classi- mance, in terms of effectiveness, when compared to other
55
56 fier based on kNN and reported promising results, however methods proposed in literature. However, no test using a
57 the computational cost was not mentioned. In other works, fairer comparison scheme, such as the one proposed by de
58 also based on kNN, in the literature [127, 92, 128, 137, 129], Chazal et al. [7], and the recommendations of AAMI, was
59 no one presented a more fair evaluation protocol for com- done with methods that use a set of rules. This subject is
60 parison of methods as the one proposed by de Chazal et discussed in depth in Section 8.
61 10
62
63
64 Page 11 of 22
65
 1 Using a few discriminative features from previous
2 works [71, 37], de Lannoy et al. [8] proposed the use of
3 weighted Conditional Random Fields for the classification
4 of arrhythmias and compared the approach with SVMs
5 and LDs. The experiments demonstrated that the pro-
6 posed method obtains promising results for the minority
7 arrhythmical classes (SVEB e VEB). However, the rela-
8 tively low efficiency for the normal class (80%) represents
9 a problem when used in real life scenario (inter-patient
10 paradigm), since many healthy heartbeats will be classi-
11 fied as arrhythmic.

t
12

ip
13
14 7. Databases and the AAMI Standard Figure 8: Example of annotations in a MIT-BIH database.
15 Source [107].

cr
16 Various databases are composed of cardiac heartbeat
17 grouped in patients records freely available that permits
18 the creation of a standardization for the evaluation of and some annotations in the center. Noteworthy is the fact
automatic arrhythmia classification methods. This stan- that it is recommended that records of patients using pace-

us
19
20 dardization was developed by AAMI and is specified in makers should not be considered. In this database, 4 pa-
21 ANSI/AAMI EC57:1998/(R)2008 [10] and defined the pro- tients/records have this property and its respective heart-
22 tocol to perform the evaluations to make sure the experi- beats should be removed. In addition, segments of data
23 ments are reproducible and comparable. containing ventricular flutter or fibrillation (VF) should
24
25
26
The use of five databases is recommended by the stan-
dardization:
an
also be excluded from the analysis.
Although various types of cardiac arrhythmias ex-
ist, AAMI recommends that only some types should be
27 • MIT-BIH: The Massachusetts Institute of Technol- detected by equipment/methods. There are 15 recom-
M
28 ogy - Beth Israel Hospital Arrhythmia Database (48 mended classes for arrhythmia that are classified into 5
29 records of 30 minutes each); superclasses: Normal (N ), Supraventricular ectopic beat
30 (SVEB ), Ventricular ectopic beat (VEB ), Fusion beat (F )
31 • EDB: The European Society of Cardiology ST-T
Database (90 records of 2 hours each); and Unknown beat (Q). Table 3 illustrates the 15 classes
32
d

and their symbols, as well as the hierarchy of the 5 groups

33
• AHA: The American Heart Association Database for (superclasses).
34
te

35 Evaluation of Ventricular Arrhythmia Detectors (80

36 records of 35 minutes each); Table 3: Principal types of heartbeats present in the MIT-BIH
database.
37 • CU: The Creighton University Sustained Ventricular
p

38 Arrhythmia Database (35 records of 8 minutes each);

39 Group Symbol Class
N N ou . Normal beat
• NST: The Noise Stress Test Database (12 records of
ce

40 Any heartbeat L Left bundle branch block beat

41 ECG of 30 minutes each, plus 3 records with noise not categorized R Right bundle branch block beat
as SVEB, VEB, e Atrial escape beat
42 excess); F or Q j Nodal (junctional) escape beat
43 A Atrial premature beat
44 The most representative database for arrhythmia is the SVEB a Aberrated atrial premature beat
Ac

45 MIT-BIH, and because of this, it has been used for most of Supraventricular
ectopic beat
J
S
Nodal (junctional) premature beat
Supraventricular premature beat
46 the published research. It was also the first database avail- VEB V Premature ventricular contraction
47 able for this goal and has been constantly refined along the Ventricular
ectopic beat
E Ventricular escape beat
48 years [148]. F
F
Fusion of ventricular
49 The majority of the heartbeats recorded in these Fusion beat
P ou /
and normal beat
Paced beat
50 databases have annotations associated with the type of Q
f Fusion of paced and normal beat
51 Unknown beat
heartbeat or the events. These heartbeat annotations, as U Unclassifiable beat
52 much for the class and for the fiducial points (e.g., point R,
53
maximum amplitude of the heartbeat) are fundamental for
54
the development and evaluation of automatic arrhythmia The measures recommended by AAMI for evaluating
55
classification methods. methods are: Sensitivity (Se), Positive predictivity (+P ),
56
The ANSI/AAMI EC57:1998/(R) 2008 standard also False positive rate (F P R) and Overall accuracy (Acc).
57 Sensitivity and Positive Predictivity are also known in lit-
58 specifies how annotations should be done in the databases.
An example can be seen in Fig. 8, in which there is the erature as recall and precision, respectively; the overall
59 accuracy can be strongly distorted by the results of the
60 lead II at the upper part of the figure, lead V1 at the lower,
61 11
62
63
64 Page 12 of 22
65
 1 majority class. In this way, the first three measures are 7.3. AHA
2 the most relevant for comparing the methods, since the The AHA database3 consists of 155 records, each one
3 classes for the heartbeat types are extremely imbalanced composed of two leads, sampled at 250 Hz with 12-bit
4 in available databases. resolution. Each recording is three hours long and only
5 Calculation of the measures is based on the definitions the final 30 minutes have been annotated. The database
6 presented in Table 4. Note that in sections (a), (b), and was created to evaluate ventricular arrhythmia detectors.
7 (c) of Table 4, formulas and schemas to compute Se, +P , However, the database does not differentiate normal sinus
8 F P R and Acc are given for V , S and N classes, respec- rhythm from supraventricular ectopic beats (SVEB ).
9 tively. Observe that according to this table, it is not neces-
10 sary to penalize VEB +P with the false positives F v and 7.4. CU
11 Q v (as highlighted in the schema of Table 4(a)), mean-

t
12 The CU database4 is composed of 35 eight-minute ECG
while, for SVEB +P , Q v (as highlighted in the schema of

ip
13 recordings, sampled at 250 Hz with 12-bit resolution. The
Table 4(b)) also does not need to enter the calculation.
14 database was intended to evaluate algorithms aiming at
The standard also suggests that the results should be
15 detecting episodes of sustained ventricular tachycardia,
presented in a global manner, considering that each heart-

cr
16 ventricular flutter, and ventricular fibrillation. It pro-
beat has the same weight (gross statistics) and in per sam-
17 vides reference annotation files to aid users to locate these
ple basis. A set of results is exemplified in Table 5.
18 events on the recordings. More information regarding this
Next, we briefly discuss the five databases recom-

us
19 database can be found in [150].
20 mended by the standard, presenting the number of records,
21 sample frequency, resolution and finality of each.
7.5. NSD
22
23 7.1. MIT-BIH The NSD database5 includes 12 half-hour ECG record-
24
25
26
This database1 is presented in majority of the publica-
tions found in literature. It is unique since it contemplates
the five arrhythmia groups proposed by AAMI as described
an
ings and 3 half-hour noise recordings. The noise inserted
in the recordings are typical interferences found in am-
bulatory care services, such as baseline wander, muscle
27 in Table 3. artefact (EMG) and electrode motion artefact. According
M
28 This database contains 48 records of heartbeats at to Goldberger et al. [107], the electrode motion artefact
29 360Hz for approximately 30 minutes of 47 different pa- is considered to be the most troublesome, since it can be
30 tients. Each record contains two ECG leads and in the easily misinterpreted as ectopic beats. Also, it cannot be
31 majority of them the principal lead (lead A) is a modifi- easily removed by filters.
32 The ECG recordings available in the NSD database
d

cation of lead II (electrodes on the chest). The other lead

33
(lead B) is usually lead V1, modified, but in some records, were created based on two clean recordings from MIT-BIH
34
this lead is known to be V2, V5 or V4 [107]. Generally, (118 and 119). The noise was artificially inserted in the
te

35 signals. This database is more detailed described in [151].

36 lead A is used to detect heartbeats, since the QRS complex
37 is more prominent in this lead. Lead B favors the arrhyth-
mic classification of the types SVEB and VEB [107]. More
p

38 8. Heartbeats selection problem for evaluation of

39 information regarding this database can be found in [148]. methods
ce

40
41 7.2. EDB The AAMI standard specifies a protocol for tests and
42 The EDB database2 is a collection of 90 records ac- evaluation of arrhythmia classification methods. It also
43 quired from 79 subjects, sampled at 250 Hz with 12-bit stipulates which databases should be used. However, it
44 resolution. These records were extracted from 70 men (be- does not specify which patients/heartbeats should be used
Ac

45 tween 30 and 84 years old) and 8 women (between 55 and to construct the model to be classified (training phase) and
46 71 years old). As all of these subjects were suffering from which patients/heartbeats should be used for evaluation
47 a specific cardiac disease (i.e., myocardial ischaemia), the methods, i.e., the testing phase, which may render biased
48 database was originally built to allow ST-segment and T- results. For instance, de Chazal et al. [7] demonstrated
49
wave analysis. that the use of heartbeats from the same patient for both
50 the training and the testing makes the evaluation process
The heartbeats were recorded for a two hour duration
51 biased. This is because the models tend to learn the par-
52 and each of them contains two signals (i.e., two leads).
Two cardiologists made the annotations for the record and ticularities of the patient’s heartbeat during the training,
53
the AAMI standard was used. More information regarding obtaining expressive numbers during the test (very close to
54
55 this database can be found in [149].
56 3 The AHA database can be obtained in https://www.ecri.org/.
57 1 The complete information regarding the database as well 4 The CU database as well as the annotation files can be obtained
58 as its usage and data annotation/labelling can be found in in [107].
5 The NSD database and its annotation files can be obtained
59 http://www.physionet.org/.
2 The EDB database can be obtained in [107]. in [107].
60
61 12
62
63
64 Page 13 of 22
65
 Table 4: Calculations for method evaluations. Source [7]. (a), (b), and (c) highlight the calculation of measures for V , S, and N , respectively.
1
2
3 (a) (b) (c)

4 Algorithm Algorithm
n s
Algorithm
v f q sum
n s v f q n s v f q
5 N Nn Ns Nv Nf N q ΣN

Reference
N Nn Ns Nv Nf Nq N Nn Ns Nv Nf Nq

Desired
Desired

S Sn Ss Sv Sf Sq ΣS
6 S
V
Sn
Vn
Ss Sv Sf
Vs Vv Vf
Sq
Vq
S
V
Sn
Vn
Ss Sv Sf
Vs Vv Vf
Sq
Vq
V Vn Vs Vv Vf V q ΣV
F Fn Fs Fv Ff F q ΣF
7 F Fn Fs Fv Ff Fq F Fn Fs Fv Ff Fq
Q Qn Qs Qv Qf Qq ΣQ
Q Qn Qs Qv Qf Qq Q Qn Qs Qv Qf Qq
8 Σ

9 T NV = Nn + Ns + Nf + Nq
+Sn + Ss + Sf + Sq
T NS = Nn + Nv + Nf + Nq
+V n + V v + V f + V q
TN
T PV
=
=
Nn
Vv
10 +F n + F s + F f + F q +F n + F v + F f + F q T PS
T PF
=
=
Ss
Ff
+Qn + Qs + Qf + Qq +Qn + Qv + Qf + Qq
11

t
F NV = Vn+Vs+Vf +Vq F NS = Sn + Sv + Sf + Sq T PQ = Qq
T PV = Vv T PS = Ss Sp = T N/ΣN
12

ip
F PV = N v + Sv F PS = Ns + V s + F s V EB Se : see Table 4(a)
V EB Se = T PV /(T PV + F NV ) SV EB Se = T PS /(T PS + F NS ) SV EB Se : see Table 4(b)
13 V EB + P = T PV /(T PV + F PV ) SV EB + P = T PS /(T PS + F PS ) F Se = T PF /ΣF
V EB F P R = F PV /(T NV + F PV ) SV EB F P R = F PS /(T NS + F PS ) Q Se = T PQ /ΣQ
14 V EB Acc =
T PV +T NV
SV EB Acc =
T PS +T NS T N +T PS +T PV +T PF +T PQ
T PV +T NV +F PV +F NV T PS +T NS +F PS +F NS Acc =
15 Σ

cr
16 Abbreviations: Acc: Accuracy; F : Fusion heartbeat group (superclass); F P R: False positive rate; N : Normal heartbeat group (superclass); +P : Positive predictivity; Q:
Unknown heartbeat group (superclass); Se: Sensitivity; Sp: Specificity; S & SVEB: Supraventricular ectopic heartbeat group (superclass); V & VEB: Ventricular ectopic
17 heartbeat group (superclass); TN : True negative; and TP: True positive.
18

us
19 Table 5: Exhibition example of results according to the AAMI standard
20
21 N SVEB VEB F Q
Record Acc
22 Se/+P /F P R Se/+P /F P R Se/+P /F P R Se/+P /F P R Se/+P /F P R
101 99.5 99.7 99.8 60.0 33.3 12.5 0.4 – – 0.0 – – 0.0 0.0 – 0.0
23
24
25
26
106
108
109
...
...
215
220
72.5
97.2
95.9
...
...
3.0
98.5
97.5
98.3
97.4

3.1
99.4
100.0
99.7
99.5
...
...
99.0
99.0
0.0
21.7
27.5

0.6
21.3
–

–
0.0 27.5
75.0 6.4 2.5
0.0 3.7
...
...
100.0 0.1 97.1
78.7 87.1 0.6
an
0.0
0.0
0.0

0.0 –
– –
–
–
–
...
...
0.0
0.0
0.0

10.0
0.0
–
0.0
0.0
–
–
–
...
...
0.0
0.0
0.0

90.0 – 60.0
– – 0.0
–
–
–

–
–
–
–
–
–
–
...
...
0.0
0.0
0.0

0.0
0.0
27 223 79.8 98.6 94.7 20.0 86.3 13.2 16.3 0.0 – 30.0 0.0 – 0.0 – – 0.0
M
Gross 75.5 83.0 98.1 14.4 45.8 4.3 19.2 0.0 0.0 4.5 50.0 0.0 10.0 0.0 0.0 0.0
28
29
30
31 100%). As previously mentioned, this heartbeat division primarily divided in two parts: odd and even numbered
32 protocol is called in the literature intra-patient scheme or records. The final record selection was achieved by ex-
d

33 paradigm [8, 42]. However, in a clinical environment, a changing some of the records between the parts so as to
34 fully automatic algorithm/method will find heartbeats of balance the classes. The heartbeat distribution of the sets
te

35 patients different from those they used to learning in the can be seen in Table 6. Observe that the two sets have
36 training phase. approximately the same number of heartbeats per class,
37
Intending to specify a protocol, the work of de Chazal with approximately 100 thousand heartbeats. It is worth
p

38
et al. [7] proposed a division of the heartbeats from the mentioning that records 201 and 202 are from the same pa-
39
MIT-BIH database into two sets so that the database be- tient, but are in different sets. The other records pertain
ce

40
41 comes more coherent with reality. The first set is composed to only one patient.
42 of all heartbeats of records: 101, 106, 108, 109, 112, 114,
43 115, 116, 118, 119, 122, 124, 201, 203, 205, 207, 208, 209, Table 6: Heartbeat distribution by classes of sets/parts as proposed
by de Chazal et al. [7].
44 215, 220, 223 and 230, called Dataset 1 (DS1). While the
Ac

45 second is composed of all heartbeats of records: 100, 103,

46 105, 11, 113, 117, 121, 123, 200, 202, 210, 212, 213, 214, Set N SVEB VEB F Q Total
47 219, 221, 222, 228, 231, 232, 233 and 234, called Dataset 2 DS1 45866 944 3788 415 8 51021
48 (DS2). The authors only used DS1 to construct the clas- DS2 44259 1837 3221 388 7 49712
DS1+DS2 90125 2781 7009 803 15 100733
49 sification model, while DS2 was reserved for evaluation.
50 In this way, they guaranteed that the created model had
51 no contact with the heartbeats pertaining to DS2, i.e.,
52 heartbeats from DS1 and DS2 come from different indi- de Chazal et al. [7] concluded that for a more realis-
53 viduals. Such division protocol is called in the literature tic evaluation, the DS1 set must be used for training and
54 inter-patient scheme or paradigm [7, 8, 42]. Note that the DS2 set for testing, making heartbeat classification a
55 significantly more difficult task, and consequently, reduc-
only the MIT-BIH database was used for the creation of
56 ing the performance of the presented classifying method.
the sets, since it is the only one indicated by the AAMI
57 They also concluded that the minority classes (SVEB and
standard that contemplates all 5 of the superclasses for
58
arrhythmias. VEB ), where the most problematic arrhythmias are found,
59
According to de Chazal et al. [7], these records were suffered more with this type of protocol.
60
61 13
62
63
64 Page 14 of 22
65
 1 Tables 7 and 8 lists the main works, considered by us, 5. Classification:
2 published in literature, grouped according to the scheme of • During training, use a k-patient cross validation
3 heartbeat selection: intra-patient, where heartbeats of the to define model parameters as proposed in [7].
4 same patients probably appear in the training as well as in
• Investigate the database imbalance impact on
5 the testing dataset; and the second scheme, where authors
chosen classifier, by reporting results with and
6 took the precaution to construct and evaluate the clas-
without use of techniques to compensate the
7 sification using heartbeats from different patients (inter-
imbalance;
8 patient), following the protocol proposed by [7].
9 Works that do not fit into inter-patient category, do 6. Evaluation:
10 not permit fair comparisons with the results in the litera- • Present the results according to AAMI recom-
11 ture, once a great majority of the authors did not follow mendations to allow literature comparison;

t
12 the same protocol for the evaluations. As one can see in

ip
13 Another paradigm found in the literature is the
Table 8, it is also difficult to assess which technique con- patient-specific paradigm [139]. This class of work relies
14 tribute to heartbeat classification, since methods with dif-
15 on semi-automatic heartbeat classification and is out of
ferent approaches achieves very high (>98%) accuracies.

cr
16 scope of this work.
Thus, the reported results grouped in Table 8 cannot be Luz & Menotti [161]7 reimplemented some models that
17
18 taken into consideration from a clinical point of view, since presented an overall accuracy of nearly 100% and were
the reported values by these works are probably different not concerned about the heartbeat selection scheme. Af-

us
19
20 in a real life scenario in terms of accuracy. terwards, they re-evaluated the results produced by the
21 Unfortunately, the great majority of the works in the methods with the objective of reporting experiments in
22 literature does not concern on following the division de- accordance with the protocol recommended by AAMI and
23 fined in [7] or any other inter-patient protocol that im- using the division scheme proposed in [7]. The reported
24
25
26
posed the non-usage of heartbeats from the same patient
in the training and testing [160], as shown in Table 8.
Aiming to standardize the evaluation process consider-
an
results following and not following the AAMI standard and
the protocol proposed in [7] can be seen in Tables 9 and 10,
respectively. Observe that the chosen methods for this ex-
27 ing a clinical point of view and AAMI recommendations, periment are reasonably recent and contemplate the use of
M
28 we suggest future works to follow a workflow6 : various classifiers and various forms of feature extraction.
29
1. Database selection: Analyzing the values of Tables 9 and 10, it can be observed
30
that the results obtained by the same classification method
31 • Use MIT-BIH ARRDB with inter-patient
using a scheme of random selection (in which heartbeats
32 scheme proposed in [7] to allow unbiased lit-
d

were randomly chosen to compose the training and testing

33 erature comparison;
34 sets) are significantly superior to the values obtained with
• Use INCART database to assess generalization
te

35 experiments using the division proposed in [7].

power of the method, as proposed in [37]; The results showed in Tables 9 and 10 suggest that to
36
37 2. Preprocessing. Run all process with at least 2 filter- perform a fair evaluation of ECG-based heartbeat classifi-
ing scheme besides the proposed filtering method by
p

38 cation methods, heartbeats of the same patient should not

39 the authors: be present in both training and testing sets, since it is not
a realistic scenario. Otherwise, the classifiers will learn nu-
ce

40 • Signal filtering as proposed in [7] to allow liter-

41 ature comparison; ances of patients in the training set and as such, the eval-
42 uation of a method on the testing set using heartbeats of a
• Use the Raw Signal, i.e., no filtering. This
43 patient whose heartbeats are present in the training set as
44 should work as a ground truth;
well, is biased, even if the heartbeats of the same patient
Ac

45 3. Segmentation: are different. Although some works in literature strongly

46 • Add jitter to R location annotation as proposed draw attention to this bias problem [7, 160, 161, 37], few
47
in [66] to test the robustness of the method authors have taken the precaution of following a protocol,
48
against segmentation errors; as proposed by AAMI, to report the results and evalu-
49
50 4. Feature extraction: ate the methods, which makes it difficult to make a fair
comparison of the works published in literature.
51 • Use feature selection to report which proposed
52 features improve the results;
53 9. Concluding Remarks and Future Challenges
54 • Use a class-oriented feature selection to assess
55 which feature is more suitable to which disease Researchers have raised several problems related to the
56 as proposed in [35]. This could result in an automatic classification of cardiac arrhythmias [161, 7, 34,
57 important contribution to the literature; 37], which are discussed in the next paragraphs.
58
6 The proposed workflow for the evaluation process is an important 7 The content of this part of this section is an overview of the work
59
60 contribution of this survey work published in [161].
61 14
62
63
64 Page 15 of 22
65
 1
2 Table 7: Methods which used inter-patient paradigm. Artificial Neural Network (ANN); Principal Component Analysis (PCA); Floating
Feature Selection (FFS); Independent Component Analisys (ICA); Back Propagation Neural Network (BPNN); Hermite Basis Function (HBF);
3
high order statistics cummulants (HOSC); Linear Discriminants (LD); Sequential forward floating search (SFFS); Importance Weighted Kernel
4 Logistic Regression (IWKLR); Conditional Random Fields (CRF); Reservoir Computing (RC); $ Authors optimize their result for 3 classes
5 (N,SVEB,VEB); # Where confusion matrix was not given, some values could not be computed.
6
7
8 Work Feature set Classifier Effectiveness
9 de Chazal et al., 2004 [7] ECG-Intervals, Weighted LD Acc = 83%;
10 Morphological SeN = 87%; +PN = 99%;
SeS = 76%; +PS = 38%;
11

t
SeV = 77%, +PV = 82%;
12

ip
Soria & Martinez, 2009 [31] RR-Intervals, Weighted LD Acc = 90%;
13 VCG, morphological SeN = 92%, +PN = 85%;
14 + FFS SeS = 88%, +P = 93%;
15 SeV = 90%, +P = 92%

cr
Llamedo & Martinez, 2011 [37] Wavelet, Weighted LD Acc = 93%;
16 VCG SeN = 95%; +PN = 98%;
17 + SFFS SeS = 77%; +PS = 39%;
18 SeV = 81%, +PV = 87%;

us
19 Mar et al., 2011 [34] Temporal Features, Weighted LD Acc = 89%;
20 Morphological, MLP SeN = 89%; +PN = 99%;
statiscial features SeS = 83%; +PS = 33%;
21 + SFFS SeV = 86%, +PV = 75%;
22 # Bazi et al., 2013 [40] Morphological, SVM Acc = 97% (DS1)
23
24
25
26
Luz et al., 2012 [134]
Wavelet

features
proposed in
[102, 101, 79]
an IWKLR, DTSVM

SVM,
ANN,
Bayesian,
Acc = 92% (DS2)

SeN = 84% SpSV EB = 18%

SpV EB = 72%

27 [70, 44] OPF

M
28 Ye et al., 2012 [38] Morphological, Wavelet, SVM Acc = 86.4%
29 RR interval, ICA, SeN = 88%; +PN = 97%;
30 PCA SeS = 60%; +PS = 53%;
SeV = 81%, +PV = 63%;
31 de Lannoy et al., 2010 [32] ECG-Intervals, weighted SVM Acc = 83%;
32
d

morphological, SeN = 80%;

33 HOS, SeS = 88%;
34 HBF coeficients SeV = 78%;
te

35 Park et al., 2008 [33] HOS, HBF Hierarchical SVM Acc = 85%;
SeN = 86%;
36 SeS = 82%;
37 SeV = 80%;
p

38 Zhang et al., 2014 [35] RR-intervals, Combined SVM Acc = 86%;

39 morphological features, SeN = 89%; +PN = 99%;
ECG-intervals and segments SeS = 79%; +PS = 35%;
ce

40
SeV = 85%, +PV = 92%;
41 Escalona-Moran et al., 2015 [43] Raw wave RC Acc = 98%;
42 SeN = 96%; +PN = 91%;
43 SeS = 79%; +PS = 96%;
44 SeV = 96%; +PV = 99%;
Ac

45 # Huang et al., 2014 [42] Random projection Ensemble of SVM

RR-intervals SeN = 99%; +PN = 95%;
46 SeS = 91%; +PS = 42%;
47 SeV = 94%, +PV = 91%;
48 $ Lin & Yang, 2014 [41] normalized RR-interval weighted LD Acc = 93%;
49 SeN = 91%; +PN = 99%;
50 SeS = 81%; +PS = 31%;
SeV = 86%, +PV = 73%;
51
de Lannoy et al., 2012 [8] RR-intervals, ECG-segments weighted CRF Acc = 85%;
52 morphological, HBF, HOS AccN = 79%;
53 AccS = 92%;
54 AccV = 85%;
55 Zhang & Luo, 2014 RR-intervals, Combined SVM Acc = 87%;
morphological features, SeN = 88%; +PN = 98%;
56
ECG-intervals and segments, SeS = 74%; +PS = 59%;
57 wavelets coeff. SeV = 88%, +PV = 82%;
58
59
60
61 15
62
63
64 Page 16 of 22
65
 1 Table 8: Methods which used Intra-patient paradigm. Neural Network (NN); Principal Component Analysis (PCA); Generalized Discriminant
2 Analyses(GDA); Error correcting output codes (ECOC); Neural Network Adaptative Activation Funcion (NNAAF); Ant Colony Optimization
(ACO) based clustering; Independent Component Analisys (ICA); Fuzzy C-Means (FCM); adaptive wavelet network (AWN); Probabilistic
3 neural Network (PNN); Back Propagation Neural Network (BPNN); Particle swarm optimization (PSO); Continues Wavelet Transform
4 (CWT); Discrete Wavelet Transform (DWT); Discrete Cosine Transform (DCT); Fuzzy C-Means type 2 (FCMT2); Mixture of Gaussian
5 (MOG) ; Qualitative feature selection (QFS); Hidden Markov modeling (HMM); linear predictive coding (LPC); Burgs maximum entropy
6 (BME); self- organizing maps (SOM); Hermite Basis Function (HBF); High order statistics cummulants (HOSC); Higher order statistics
7 (HOS); Linear Discriminants (LD); Self-organizing cerebellar model articulation controller (SOCMAC) network; Extreme Learning Machine
(ELM); Local fractal dimension (LFD); linear discriminant analysis (LDA).
8
9
10 Work # cl. Feature set Classifier Effectiveness
11

t
Chen et al., 1996 [152] 2 RR-interval Set of rules Acc = 95%
12

ip
Lagerholm et al., 2000 [138] 16 HBF, SOM clustering Acc = 98%
13 Dokur & Olmez, 2000 [98] 10 Fourier, Wavelet + FSDP MLP, RCE, Acc = 96%
14 Novel hybrid NN
15 Osowski & Linh, 2001 [88] 6 HOSC fuzzy NN Acc = 96%

cr
Tsipouras et al., 2002 [135] 9 RR-interval Deterministic automata Acc = 96%
16 Mehmet, 2004 [122] 4 HOSC, Wavelet Min. Dist, kNN, Bayes Acc = 98%
17 Cristov & Bortonal, 2004 [106] 2 Heartbeat-Intervals, VCG NN Acc = 99%
18 Guler & Ubeyli, 2005 [44] 4 Wavelets (statistics) Combined NN Acc = 96%

us
19 Song et al., 2005 [102] 6 Wavelet coef., LDA SVM Acc = 99%
20 RR-Intervals
Karimifard et al., 2006 [153] 7 HBF kNN Acc = 99%
21 Özbay et al., 2006 [84] 10 Raw-wave MLP, Fuzzy Cluster, Acc = 99%
22 FCNN
23
24
25
26
Tsipouras et al., 2007 [136]
Bortolan et al., 2007 [105]

Ubeyli, 2007 [154]

Yu & Chen, 2007 [101]
4
2

4
5
RR-interval

hyperbox+GA
DWT
ICA, RR-interval
an
VCG and Morphological
Fuzzy Expert System
Fuzzy Clustering

SVM, ECOC
PNN
Acc = 96%
Acc = 99%

Acc = 99%
Acc = 99%
27 Ceylan & Osbay, 2007 [76] 10 DWT FCM, NN Acc = 99%
M
28 Yu & Chen, 2007 [101] 6 Wavelet (statistics) PNN Acc = 99%
29 RR-interval
30 Minhas & Arif, 2008 [137] 6 Wavelet, RR-interval, PCA kNN Acc = 99%
Lin et al., 2008 [96] 7 Morlet Wavelet AWN Acc = 90%
31 Korurek & Nizam, 2008 [127] 6 RR-interval, ECG-segments ACO-based Cluster, Acc = 94%
32
d

kNN
33 Yu & Chou, 2008 [79] 8 RR-interval, ICA PNN, BPNN Acc = 98%
34 Asl et al., 2008 [85] 6 HVR, GDA SVM Acc = 100%
te

35 Ceylan et al., 2009 [90] 10 PCA, DWT FCMT2, ANN Acc = 99%
Wen et al., 2009 [73] 16 RR-interval, raw-wave SOCMAC-based Cluster Ac = 98%
36 Yu & Chou, 2009 [80] 8 ICA SVM Acc = 98%
37 Kim et al., 2009 [77] 6 RR-interval, PCA ELM Acc = 98%
p

38 Ye et al., 2010 [70] 15 Wavelet, ICA, SVM Acc = 99%

39 PCA, RR-interval
Ozbay & Tezel, 2010 [74] 10 ECG-wave NNAAF Acc = 98%
ce

40
Mishra & Raghav, 2010 [95] 6 LFD Nearest Neighbor Acc = 89%
41 Korurek & Nizan, 2010 [127] 6 RR-interval, QRS-width, ACO, kNN Acc = 90%
42 Wavelet, PCA
43 Lanata et al., 2011 [128] 6 HOS MOG, kNN Acc = 85%
44 Yeh et al., 2012 [131] 5 Morphological, RR-interval clustering Acc = 94%
Ac

45 QFS
Kallas et al., 2012 [83] 3 KPCA SVM Acc = 97%
46
Khazaee, 2013 [155] 3 Heartbeat intervals PSO + SVM Acc = 97%
47 morphology amplitudes
48 Wang et al., 2013 [103] 8 PCA, LDA PNN Acc = 99%
49 Kumar & Kumaraswamy, 2013 [69] 3 RR-intervals CART, RBF, Acc = 92%
50 MLP, IOAW-FFNN
Chen et al., 2014 [156] 6 RR-intervals SVN, NN Acc = 100%
51
Mert et al., 2014 [147] 6 RR-intervals, HOS, Bagged Decision Tree Acc = 99%
52 2nd order LPC coeff.
53 Ahmed & Arafat, 2014 [157] 11 Heartbeat intervals MLP, SVM, TreeBoost Acc = 98%
54 morphology amplitude, HOS
55 Sarfraz et al., 2014 [78] 11 RR-intervals, QRS power BPNN Acc = 99%
ICA coeff.
56
Tran et al., 2014 [158] 7 RR-intervals, HBF Ensemble of classifiers Acc = 98%
57 Alickovic & Subasi, 2015 [159] 5 autoregressive (AR) modeling SVM, MLP, Acc = 99%
58 RBF, kNN
59
60
61 16
62
63
64 Page 17 of 22
65
 1
Table 9: Results obtained by methods considering a set of randomly neously considered as state-of-the-art, when in truth, they
chosen data for the training and testing, based on the MIT-BIH
2 database for arrhythmia. could just be specialized with the heartbeat of only a single
3 patient.
4 Researchers from the machine learning community
5 Method Acc
N
Se/+P
SVEB
Se/+P
VEB
Se/+P
F
Se/+P
Q
Se/+P have shown that the size/diversity of the databases used
6 Ye et al. [70]
(%)
96.5
(%)
98.7
(%) (%) (%)
96.3 72.4 94.5 82.6 97.8 65.6 88.6 95.8
(%)
99.3
for the construction of methods impacts more than the
7 Yu & Chou [79]
Yu & Chen [101]
95.4
81.1
96.9
85.2
97.3 73.8 88.4 92.3 94.3 51.0 73.4 94.1
81.2 0.0 0.0 70.0 79.2 0.0 0.0 0.0
80.8
0.0
choice of the learning algorithm and/or employed tech-
8 Güler & Übeyli [44]
Song et al. [102]
89.1
98.7
93.2
99.5
90.3 0.0 0.0 81.6 74.6 0.0 0.0 0.0
98.9 86.4 94.3 95.8 97.4 73.6 90.2 0.0
0.0
0.0
niques [164]. Efforts to create new databases or even to in-
9 crease the size of existing ones, as well as creating standard
10 evaluation protocols, have been made in several research
11 areas involving pattern recognition, specially to avoid un-

t
Table 10: Results obtained by methods according the division record
12 scheme proposed in [7] based on the MIT-BIH database for arrhyth- fair comparisons between methods [165].

ip
13 mias.
We believe that one major obstacle to achieving ad-
14 vances in the research focusing on fully-automatic classifi-
15
cation of heartbeats (arrhythmias) in ECG is the reduced

cr
N SVEB VEB F Q
Method Acc
16 Se/+P Se/+P Se/+P Se/+P Se/+P
(%) (%) (%) (%) (%) (%) number of available databases. Therefore, we suggest to
17
the research community dedicated to study the heartbeat
Ye et al. [70] 75.2 80.2 78.2 3.2 10.3 50.2 48.5 0.0 0.0 0.0 0.0
Yu & Chou [79] 75.2 78.3 79.2 1.8 5.9 83.9 66.4 0.3 0.1 0.0 0.0
18
classification problem that they encourage/stimulate the
Yu & Chen [101] 73.9 81.5 74.2 0.0 0.0 21.0 59.4 0.0 0.0 0.0 0.0

us
Güler & Übeyli [44] 66.7 69.2 72.1 0.0 0.0 78.8 43.8 1.8 0.5 0.0 0.0
19
extension of databases dedicated for this end.
Song et al. [102] 76.3 78.0 83.9 27.0 48.3 80.8 38.7 0.0 0.0 0.0 0.0
20
21 We also suggest the use of new trends to capture the
22 Results presented in literature usually use the MIT- ECG signal, such as off-the-person approaches, for the
23 BIH database (also known as MIT-BIH ARR DB) that elaboration of new databases. Nonetheless, we believe that
24
25
26
is extremely unbalanced. However, this aspect has been
ignored by authors that use the intra-patient scheme. Au-
thors that followed a more realistic approach and opted
an
the creation of such databases would be a great challenge
because, besides the financial costs involved, they would
have to be incorporated into standards such as AAMI stan-
27 not to mix heartbeats for the training and testing (inter- dards to reach the desired audience.
M
28 patient scheme), reported great difficulty in obtaining As few authors use the same evaluation scheme for
29 promising results for the heartbeat arrhythmia classes tests, it is difficult to make a fair comparison between the
30 SVEB and VEB. As such, there exist innumerous pro- methods. It is also difficult to asses the real contribution
31 posed methods in the literature that do not follow a more of the methods since the intra-patient scheme favors the
32
d

fair evaluation protocol. reported figures. Another challenge would be study and
33 re-implement intra-patient methods published in the liter-
[REVIEW: Several authors employ semi-automatic
34
ature, following a heartbeat selection scheme without bias
te

35 approaches [139, 162, 163] to improve the reported re-

sults. According to [162], semi-automatic approaches (inter-patient approach), an initial work pointing towards
36
37 can improve the results in over 40% even when a small this direction appears in [161]. Furthermore, analyze their
number of heartbeats are selected for adaptation. The impact under a non-biased view is of the paramount im-
p

38
39 drawback of such approaches is that they demand in- portance for the literature of this subject.
tervention of experts. However, expert intervention is
ce

40
41 common in clinical environment and therefore this ap-
Acknowledgments
42 proach is a promising research direction.]
43 Regarding fully-automatic approaches, we stress that The authors would like to thank UFOP, UFMG,
44 even the protocol proposed by de Chazal et al. [7], con- UFPR, FAPEMIG, CAPES and CNPq for the financial
Ac

45 sidered to be the most fair presented in the literature, has support. The authors also would like to thank the review-
46 some problems that were previously related by Llamedo & ers and the editors for their valuable comments and contri-
47 Martinez [37] and Mar et al. [34]. The imbalance between butions that helped to increase significantly the readability
48 the classes led the authors to add two records of the same and organization of the present survey.
49 patient in the two already-mentioned sets. Since these
50 records, 201 and 202, are from the same patient and belong
51 References
to sets DS1 and DS2, respectively, results slightly better
52
than expected might be achieved. In addition, records 201 [1] G. D. Clifford, F. Azuaje, P. McSharry, Advanced Methods
53
54 and 202 significantly concentrate on a large part of heart- And Tools for ECG Data Analysis, 1st Edition, Artech House

55 beats class SVEB. Another important weakness of the pro- Publishers, 2006.
[2] O. Sayadi, M. B. Shamsollahi, Multiadaptive bionic wavelet
56 tocol proposed in [7] is the use of the imbalanced record 232 transform: Application to ECG denoising and baseline wan-
57 in DS2. That record contains more than 75% of heartbeats dering reduction, EURASIP Journal on Advances in Signal
58 of the class SVEB. As such, methods that achieve correct Processing 2007 (14) (2007) 1–11.
59 classification for the heartbeats of this record can be erro-
60
61 17
62
63
64 Page 18 of 22
65
 1 [3] O. Sayadi, M. B. Shamsollahi, ECG denoising and compres- ECG enhancement by adaptive cancellation of electrosurgi-
sion using a modified extended Kalman filter structure, IEEE cal interference, IEEE Transactions on Biomedical Engineering
2 Transactions on Biomedical Engineering 55 (9) (2008) 2240– 30 (7) (1983) 392–398.
3 2248. [25] N. V. Thakor, Y.-S. Zhu, Applications of adaptive filtering to
4 [4] J. P. Martinez, R. Almeida, S. Olmos, A. P. Rocha, P. La- ECG analysis: noise cancellation and arrhythmia detection,
5 guna, A wavelet-based ECG delineator: evaluation on stan- IEEE Transactions on Biomedical Engineering 38 (8) (1991)
6 dard databases, IEEE Transactions on Biomedical Engineering 785–794.
51 (4) (2004) 570–581. [26] Q. Xue, Y. H. Hu, W. J. Tompkins, Neural-network-based
7 [5] C. Li, C. Zheng, C. Tai, Detection of ECG characteristic points adaptive matched filtering for QRS detection, IEEE Trans-
8 using wavelet transforms, IEEE Transactions on Biomedical actions on Biomedical Engineering 39 (4) (1992) 317–329.
9 Engineering 42 (1) (1995) 21–28. [27] B. N. Singh, A. K. Tiwari, Optimal selection of wavelet ba-
10 [6] M. Bahoura, M. Hassani, M. Hubin, DSP implementation of sis function applied to ECG signal denoising, Digital Signal
11 wavelet transform for real time ECG wave forms detection Processing 16 (3) (2006) 275–287.

t
and heart rate analysis, Computer Methods and Programs in [28] S.-W. Chen, H.-C. Chen, H.-L. Chan, A real-time QRS
12

ip
Biomedicine 52 (1) (2007) 35–44. detection method based on moving-averaging incorporating
13 [7] P. de Chazal, M. O’Dwyer, R. B. Reilly, Automatic classifica- with wavelet denoising, Computer Methods and Programs in
14 tion of heartbeats using ECG morphology and heartbeat in- Biomedicine 82 (3) (2006) 187–195.
15 terval features, IEEE Transactions on Biomedical Engineering [29] A. E. Zadeh, A. Khazaee, V. Ranaee, Classification of the elec-

cr
16 51 (7) (2004) 1196–1206. trocardiogram signals using supervised classifiers and efficient
[8] G. de Lannoy, D. François, J. Delbeke, M. Verleysen, Weighted features, Computer Methods and Programs in Biomedicine
17 conditional random fields for supervised interpatient heartbeat 99 (2) (2010) 179–194.
18 classification, IEEE Transactions on Biomedical Engineering [30] R. Sameni, M. B. Shamsollahi, C. Jutten, G. D. Clifford,

us
19 59 (1) (2012) 241–247. A nonlinear Bayesian filtering framework for ECG denoising,
20 [9] M. Kundu, M. Nasipuri, D. K. Basu, Knowledge-based ECG IEEE Transactions on Biomedical Engineering 54 (12) (2007)
21 interpretation: A critical review, Pattern Recognition 33 (3) 2172–2185.
(2000) 351–373. [31] M. L. Soria, J. P. Martinez, Analysis of multidomain features
22 [10] ANSI/AAMI, Testing and reporting performance results of for ECG classification, in: Computers in Cardiology, 2009, pp.
23
24
25
26
cardiac rhythm and ST segment measurement algorithms,
American National Standards Institute, Inc. (ANSI), Associa-
tion for the Advancement of Medical Instrumentation (AAMI),
ANSI/AAMI/ISO EC57, 1998-(R)2008 (2008).
an
[32]
561–564.
G. de Lannoy, D. François, J. Delbeke, M. Verleysen, Weighted
SVMs and feature relevance assessment in supervised heart
beat classification, in: Biomedical Engineering Systems and
[11] M. S. Spach, J. M. Kootsey, The nature of electrical propa- Technologies (BIOSTEC), 2010, pp. 212–223.
27
gation in cardiac muscle, American Journal of Physiology – [33] K. S. Park, B. H. Cho, D. H. Lee, S. H. Song, J. S. Lee, Y. J.
M
28 Heart and Circulatory Physiology 244 (H) (1983) 3–22. Chee, I. Y. Kim, S. I. Kim, Hierarchical support vector ma-
29 [12] E. Besterman, R. Creese, Waller–pioneer of electrocardiogra- chine based heartbeat classification using higher order statis-
30 phy, British Heart Journal 42 (1) (1979) 61–64. tics and hermite basis function, in: Computers in Cardiology,
31 [13] H. P. da Silva, C. Carreiras, A. Lourenço, A. Fred, R. C. das 2008, pp. 229–232.
Neves, R. Ferreira, Off-the-person electrocardiography: perfor- [34] T. Mar, S. Zaunseder, J. P. Martínez, M. Llamedo, R. Poll, Op-
32
d

mance assessment and clinical correlation, Health and Technol- timization of ECG classification by means of feature selection,
33 ogy 4 (4) (2015) 309–318. IEEE Transactions on Biomedical Engineering 58 (8) (2011)
34 [14] Y. M. Chi, T.-P. Jung, G. Cauwenberghs, Dry-contact and 2168–2177.
te

35 noncontact biopotential electrodes: methodological review, [35] Z. Zhang, J. Dong, X. Luo, K.-S. Choi, X. Wu, Heartbeat clas-
36 IEEE Reviews in Biomedical Engineering 3 (2010) 106–119. sification using disease-specific feature selection, Computers in
[15] R. C. Martins, D. Primor, T. Paiva, High-performance ground- Biology and Medicine 46 (2014) 79–89.
37
less EEG/ECG capacitive electrodes, in: Medical Measure- [36] Z. Zhang, X. Luo, Heartbeat classification using decision level
p

38 ments and Applications Proceedings (MeMeA), 2011 IEEE In- fusion, Biomedical Engineering Letters 4 (4) (2014) 388–395.
39 ternational Workshop on, 2011, pp. 503–506. [37] M. Llamedo, J. P. Martínez, Heartbeat classification using fea-
ce

40 [16] T. Barill, The Six Second ECG: A Practical Guidebook to ture selection driven by database generalization criteria, IEEE
41 Basic ECG Interpretation, nursecom, 2003. Transactions on Biomedical Engineering 58 (3) (2011) 616–625.
42 [17] M. Llamedo, J. P. Martinez, An automatic patient-adapted [38] C. Ye, B. V. K. Kumar, M. T. Coimbra, Combining general
ECG heartbeat classifier allowing expert assistance, IEEE multi-class and specific two-class classifiers for improved cus-
43 Transactions on Biomedical Engineering 59 (8) (2012) 2312– tomized ECG heartbeat classification, in: International Con-
44
Ac

2320. ference on Pattern Recognition (ICPR), 2012, pp. 2428–2431.

45 [18] I. Tomaŝić, R. Trobec, Electrocardiographic systems with re- [39] D. Zhang, Wavelet approach for ECG baseline wander cor-
46 duced numbers of leads – synthesis of the 12–lead ECG, IEEE rection and noise reduction, in: Annual International Confer-
47 Reviews in Biomedical Engineering 7 (2014) 126–142. ence of the IEEE Engineering in Medicine and Biology Society
[19] J. C. Principe, Editorial, IEEE Reviews in Biomedical Engi- (EMBC), 2005, pp. 1212–1215.
48 neering 7 (2014) 1–2. [40] Y. Bazi, N. Alajlan, H. AlHichri, S. Malek, Domain adaptation
49 [20] P. de Chazal, Detection of supraventricular and ventricular methods for ECG classification, in: International Conference
50 ectopic beats using a single lead ECG, in: Annual International on Computer Medical Applications (ICCMA), 2013, pp. 1–4.
51 Conference of the IEEE Engineering in Medicine and Biology [41] C.-C. Lin, C.-M. Yang, Heartbeat classification using normal-
52 Society (EMBC), 2013, pp. 45–48. ized RR intervals and morphological features, Mathematical
[21] C. Pater, Methodological considerations in the design of trials Problems in Engineering 2014 (2014) 1–11.
53 for safety assessment of new drugs and chemical entities, Trials [42] H. Huang, J. Liu, Q. Zhu, R. Wang, G. Hu, A new hierarchical
54 6 (1) (2005) 1–13. method for inter-patient heartbeat classification using random
55 [22] P. Lynn, Recursive digital filters for biological signals, Medical projections and RR intervals, Biomedical Engineering Online
56 and Biological Engineering and Computing 9 (1) (1979) 37–43. 13 (2014) 1–26.
57 [23] E. R. Ferrara, B. Widraw, Fetal electrocardiogram enhance- [43] M. A. Escalona-Moran, M. C. Soriano, I. Fischer, C. R. Mi-
ment by time-sequenced adaptive filtering, IEEE Transactions rasso, Electrocardiogram classification using reservoir comput-
58 on Biomedical Engineering 29 (6) (1982) 458–460. ing with logistic regression, IEEE Journal of Biomedical and
59 [24] M. Yelderman, B. Widrow, J. M. Cioffi, E. Hesler, J. A. Leddy, Health Informatics 19 (3) (2015) 892–898.
60
61 18
62
63
64 Page 19 of 22
65
 1 [44] I. Güler, E. D. Übeyli, ECG beat classifier designed by com- ical Engineering Conference, 1997, pp. 455–457.
bined neural network model, Pattern Recognition 38 (2) (2005) [66] C. Ye, V. Bhagavatula, M. T. Coimbra, Heartbeat classifica-
2 199–208. tion using morphological and dynamic features of ECG signals,
3 [45] G. B. Moody, R. G. Mark, Development and evaluation of a IEEE Transactions on Biomedical Engineering 59 (10) (2012)
4 2-lead ECG analysis program, in: Computers in Cardiology, 2930–2941.
5 1982, pp. 39–44. [67] T. P. Exarchos, M. G. Tsipouras, D. Nanou, C. Bazios, Y. An-
6 [46] V. X. Afonso, W. J. Tompkins, T. Q. Nguyen, S. Luo, toniou, D. I. Fotiadis, A platform for wide scale integration
ECG beat detection using filter banks, IEEE Transactions on and visual representation of medical intelligence in cardiology:
7 Biomedical Engineering 46 (2) (1999) 192–202. the decision support framework, in: Computers in Cardiology,
8 [47] J. Lee, K. Jeong, J. Yoon, M. Lee, A simple real-time QRS de- 2005, pp. 167–170.
9 tection algorithm, in: Annual International Conference of the [68] T. P. Exarchos, M. G. Tsipouras, C. P. Exarchos, C. Pa-
10 IEEE Engineering in Medicine and Biology Society (EMBC), paloukas, D. I. Fotiadis, L. K. Michalis, A methodology for
11 Vol. 4, 1996, pp. 1396–1398. the automated creation of fuzzy expert systems for ischaemic

t
[48] P. S. Hamilton, W. J. Tompkins, Quantitative investigation of and arrhythmic beat classification based on a set of rules ob-
12

ip
QRS detection rules using the MIT/BIH arrhythmia database, tained by a decision tree, Artificial Intelligence in Medicine
13 IEEE Transactions on Biomedical Engineering 33 (12) (1986) 40 (3) (2007) 187–200.
14 1157–1165. [69] R. G. Kumar, Y. S. Kumaraswamy, Investigation and classi-
15 [49] J. Pan, W. J. Tompkins, A real-time QRS detection algorithm, fication of ECG beat using input output additional weighted

cr
16 IEEE Transactions on Biomedical Engineering 32 (3) (1985) feed forward neural network, in: International Conference on
230–236. Signal Processing, Image Processing & Pattern Recognition
17 [50] R. Poli, S. Cagnoni, G. Valli, Genetic design of optimum linear (ICSIPR), 2013, pp. 200–205.
18 and nonlinear QRS detectors, IEEE Transactions on Biomed- [70] C. Ye, M. T. Coimbra, B. V. K. V. Kumar, Arrhythmia de-

us
19 ical Engineering 42 (11) (1995) 1137–1141. tection and classification using morphological and dynamic
20 [51] J. C. T. B. Moraes, M. M. Freitas, F. N. Vilani, E. V. Costa, features of ECG signals, in: Annual International Confer-
21 A QRS complex detection algorithm using electrocardiogram ence of the IEEE Engineering in Medicine and Biology Society
leads, in: Computers in Cardiology, 2002, pp. 205–208. (EMBC), 2010, pp. 1918–1921.
22 [52] P. Hamilton, Open source ECG analysis, in: Computers in [71] G. Doquire, G. de Lannoy, D. François, M. Verleysen, Feature
23
24
25
26
Cardiology, 2002, pp. 101–104.
[53] Y. H. Hu, W. J. Tompkins, J. L. Urrusti, V. X. Afonso, Ap-
plication of artificial neural networks for ECG signal detection
and classification, Journal of Eletrocardiology 26 (supplement)
an selection for interpatient supervised heart beat classification,
Computational Intelligence and Neuroscience 2011 (2011) 1–9.
[72] M. Korürek, B. Doğan, ECG beat classification using particle
swarm optimization and radial basis function neural network,
(1990) 66–73. Expert Systems with Applications 37 (12) (2010) 7563–7569.
27
[54] Y.-C. Yeh, W.-J. Wang, QRS complexes detection for ECG [73] C. Wen, T.-C. Lin, K.-C. Chang, C.-H. Huang, Classification
M
28 signal: The difference operation method, Computer Methods of ECG complexes using self-organizing CMAC, Measurement
29 and Programs in Biomedicine 91 (3) (2008) 245–254. 42 (3) (2009) 399–407.
30 [55] O. Sayadi, M. B. Shamsollahi, A model-based bayesian frame- [74] Y. Özbay, G. Tezel, A new method for classification of ECG ar-
31 work for ECG beat segmentation, Physiological Measurement rhythmias using neural network with adaptive activation func-
30 (3) (2009) 335–352. tion, Digital Signal Processing 20 (4) (2010) 1040–1049.
32
d

[56] Massachusetts Institute of Technology, MIT-BIH ECG [75] F. Castells, P. Laguna, L. Sörnmo, A. Bollmann, J. M.
33 database, available at: http://ecg.mit.edu/ (2011). Roig, Principal component analysis in ECG signal process-
34 [57] American Heart Association, AHA database, available at: ing, EURASIP Journal on Applied Signal Processing 2007 (1)
te

35 http://www.ahadata.com/ (1998). (2007) 98–98.

36 [58] J. H. van Bemmel, J. L. Williams, Standardisation and vali- [76] R. Ceylan, Y. Özbay, Comparison of FCM, PCA and WT tech-
dation of medical decision support systems: The CSE project, niques for classification ECG arrhythmias using artificial neu-
37
Methods of Information in Medicine 29 (4) (1990) 261–262. ral network, Expert Systems with Applications 33 (2) (2007)
p

38 [59] B.-U. Kohler, C. Hennig, R. Orglmeister, The principles of 286–295.

39 software QRS detection, IEEE Engineering in Medicine and [77] J. Kim, H. S. Shin, J. Shin, M. Lee, Robust algorithm for
ce

40 Biology Magazine 21 (1) (2002) 42–57. arrhythmia classification in ECG using extreme learning ma-
41 [60] S. Kadambe, R. Murray, G. F. Boudreaux-Bartels, Wavelet chine, BioMedical Engineering OnLine 8 (1) (2009) 1–12.
42 transform-based QRS complex detector, IEEE Transactions on [78] M. Sarfraz, A. A. Khan, F. F. Li, Using independent com-
Biomedical Engineering 46 (7) (1999) 838–848. ponent analysis to obtain feature space for reliable ECG ar-
43 [61] Y. Jung, W. J. Tompkins, Detecting and classifying life- rhythmia classification, in: IEEE International Conference on
44
Ac

threatening ECG ventricular arrythmias using wavelet decom- Bioinformatics and Biomedicine (BIBM), 2014, pp. 62–67.
45 position, in: Annual International Conference of the IEEE En- [79] S.-N. Yu, K.-T. Chou, Integration of independent component
46 gineering in Medicine and Biology Society (EMBC), Vol. 3, analysis and neural networks for ECG beat classification, Ex-
47 2003, pp. 2390–2393. pert Systems with Applications 34 (4) (2008) 2841–2846.
[62] H. Kim, R. F. Yazicioglu, P. Merken, C. van Hoof, H.-J. Yoo, [80] S.-N. Yu, K.-T. Chou, Selection of significant independent
48 ECG signal compression and classification algorithm with quad components for ECG beat classification, Expert Systems with
49 level vector for ECG holter system, IEEE Transactions on In- Applications 36 (2) (2009) 2088–2096.
50 formation Technology in Biomedicine 14 (1) (2010) 93–100. [81] M. Chawla, A comparative analysis of principal compo-
51 [63] P. Laguna, R. Jané, P. Caminal, Automatic detection of wave nent and independent component techniques for electrocar-
52 boundaries in multilead ECG signals: Validation with the CSE diograms, Neural Computing and Applications 18 (6) (2009)
database, Computers and Biomedical Research 27 (1) (1994) 539–556.
53 45–60. [82] L. Kanaan, D. Merheb, M. Kallas, C. Francis, H. Amoud,
54 [64] B. Celler, P. de Chazal, Selection of parameters from power P. Honeine, PCA and KPCA of ECG signals with binary SVM
55 spectral density, wavelet transforms and other methods for the classification, in: IEEE Workshop on Signal Processing Sys-
56 automated interpretation of the ECG, in: IEEE International tems (SiPS), 2011, pp. 344–348.
57 Conference on Digital Signal Processing (ICDSP), 1997, pp. [83] M. Kallas, C. Francis, L. Kanaan, D. Merheb, P. Honeine,
71–74. H. Amoud, Multi-class SVM classification combined with ker-
58 [65] J. S. Sahambi, S. N. Tandon, R. K. P. Bhatt, DSP based ST- nel PCA feature extraction of ECG signals, in: International
59 segment analysis: the wavelet approach, in: Southern Biomed- Conference on Telecommunications (ICT), 2012, pp. 1–5.
60
61 19
62
63
64 Page 20 of 22
65
 1 [84] Y. Özbay, R. Ceylan, B. Karlik, A fuzzy clustering neural net- [105] G. Bortolan, I. I. Christov, W. Pedrycz, Hyperbox classifiers
work architecture for classification of ECG arrhythmias, Com- for ECG beat analysis, in: Computers in Cardiology, 2007, pp.
2 puters in Biology and Medicine 36 (4) (2006) 376–388. 145–148.
3 [85] B. M. Asl, S. K. Setarehdan, M. Mohebbi, Support vector [106] I. Christov, G. Bortolan, Ranking of pattern recognition pa-
4 machine-based arrhythmia classification using reduced fea- rameters for premature ventricular contractions classification
5 tures of heart rate variability signal, Artificial Intelligence in by neural networks, Phisyological Measurement 25 (5) (2004)
6 Medicine 44 (1) (2008) 51–64. 1281–1290.
[86] I. Bogdanova, F. Rincón, D. Atienza, A multi-lead ECG clas- [107] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff,
7 sification based on random projection features, in: IEEE In- P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K.
8 ternational Conference on Acoustics, Speech and Signal Pro- Peng, H. E. Stanley, Physiobank, physiotoolkit, and physionet:
9 cessing (ICASSP), 2012, pp. 625–628. Components of a new research resource for complex physio-
10 [87] F. M. Ham, S. Han, Classification of cardiac arrhythmias using logic signals, Circulation 101 (23) (2000) 215–220, database
11 fuzzy ARTMAP, IEEE Transactions on Biomedical Engineer- and tools available at: http://www.physionet.org/.

t
ing 43 (4) (1996) 425–429. [108] P. Pudil, J. Novovicova, J. Kittler, Floating search methods in
12

ip
[88] S. Osowski, T. H. Linh, ECG beat recognition using fuzzy feature selection, Pattern Recognition Letters 15 (11) (1994)
13 hybrid neural network, IEEE Transactions on Biomedical En- 1119–1125.
14 gineering 48 (11) (2001) 1265–1271. [109] I.-S. Oh, J.-S. Lee, B.-R. Moon, Hybrid genetic algorithms for
15 [89] S. Osowski, L. T. Hoai, T. Markiewicz, Support vector feature selection, IEEE Transactions on Pattern Analysis and

cr
16 machine-based expert system for reliable heartbeat recogni- Machine Intelligence 26 (11) (2004) 1424–1437.
tion, IEEE Transactions on Biomedical Engineering 51 (4) [110] J. Yang, V. Honavar, Feature subset selection using a genetic
17 (2004) 582–589. algorithm, IEEE Intelligent Systems and their Applications
18 [90] R. Ceylan, Y. Özbay, B. Karlik, A novel approach for clas- 13 (2) (1998) 44–49.

us
19 sification of ECG arrhythmias: Type-2 fuzzy clustering neu- [111] X. Wang, J. Yang, X. Teng, W. Xia, R. Jensen, Feature se-
20 ral network, Expert Systems with Applications 36 (3) (2009) lection based on rough sets and particle swarm optimization,
21 6721–6726. Pattern Recognition Letters 28 (4) (2007) 459–471.
[91] S. Osowski, T. Markiewicz, L. T. Hoai, Recognition and clas- [112] S.-W. Lin, K.-C. Ying, S.-C. Chen, Z.-J. Lee, Particle swarm
22 sification system of arrhythmia using ensemble of neural net- optimization for parameter determination and feature selection
23
24
25
26
works, Measurement 41 (6) (2008) 610–617.
[92] M. I. Owis, A. H. Abou-Zied, A. B. M. Youssef, Y. M. Kadah,
Study of features based on nonlinear dynamical modeling in
ECG arrhythmia detection and classification, IEEE Transac-
an of support vector machines, Expert Systems with Applications
35 (4) (2008) 1817–1824.
[113] R. O. Duda, P. E. Hart, D. G. Stork, Pattern Classification,
2nd Edition, Wiley-Interscience, 2000.
tions on Biomedical Engineering 49 (7) (2002) 733–736. [114] C. M. Bishop, Pattern Recognition and Machine Learning, 1st
27
[93] E. D. Übeyli, Adaptive neuro-fuzzy inference system for classi- Edition, Springer, 2006.
M
28 fication of ECG signals using Lyapunov exponents, Computer [115] S. Theodoridis, K. Koutroumbas, Pattern Recognition, 4th
29 Methods and Programs in Biomedicine 93 (3) (2009) 313–321. Edition, Elsevier, 2009.
30 [94] W. Jiang, G. S. Kong, Block-based neural networks for person- [116] E. D. Übeyli, Combining recurrent neural networks with eigen-
31 alized ECG signal classification, IEEE Transactions on Neural vector methods for classification of ECG beats, Digital Signal
Networks 18 (6) (2007) 750–1761. Processing 19 (2) (2009) 320–329.
32
d

[95] A. K. Mishra, S. Raghav, Local fractal dimension based ECG [117] N. Özcan, F. Gurgen, Fuzzy support vector machines for ECG
33 arrhythmia classification, Biomedical Signal Processing and arrhythmia detection, in: IEEE International Conference on
34 Control 5 (2) (2010) 114–123. Pattern Recognition (ICPR), 2010, pp. 2973–2976.
te

35 [96] C. Lin, Y. Du, T. Chen, Adaptive wavelet network for multiple [118] J. A. Nasiri, M. Naghibzadeh, H. S. Yazdi, B. Naghibzadeh,
36 cardiac arrhythmias recognition, Expert Systems with Appli- ECG arrhythmia classification with support vector machines
cations 34 (4) (2008) 2601–2611. and genetic algorithm, in: IEEE European Symposium on
37
[97] Y. Kutlu, D. Kuntalp, Feature extraction for ECG heartbeats Computer Modeling and Simulation (EMS), 2009, pp. 187–
p

38 using higher order statistics of WPD coefficients, Computer 192.

39 Methods and Programs in Biomedicine 105 (3) (2012) 257– [119] M. Moavenian, H. Khorrami, A qualitative comparison of ar-
ce

40 267. tificial neural networks and support vector machines in ECG

41 [98] Z. Dokur, T. Ölmez, ECG beat classification by a novel hy- arrhythmias classification, Expert Systems with Applications
42 brid neural network, Computer Methods and Programs in 37 (4) (2010) 3088–3093.
Biomedicine 66 (2–3) (2001) 167–181. [120] N. V. Chawla, K. W. Bowyer, W. P. Kegelmeyer, Smote: Syn-
43 [99] P. S. Addison, Wavelet transforms and the ECG: a review, thetic minority over-sampling technique, Journal Of Artificial
44
Ac

Physiological Measurement 26 (5) (2005) 155–199. Intelligence Research 16 (1) (2002) 321–357.
45 [100] A. Daamouche, L. Hamami, N. Alajlan, F. Melgani, A wavelet [121] Y. P. Meau, F. Ibrahim, S. A. L. Narainasamy, R. Omar, In-
46 optimization approach for ECG signal classification, Biomedi- telligent classification of electrocardiogram (ECG) signal using
47 cal Signal Processing and Control 7 (4) (2012) 342–349. extended kalman filter (EKF) based neuro fuzzy system, Com-
[101] S.-N. Yu, Y.-H. Chen, Electrocardiogram beat classification puter Methods and Programs in Biomedicine 82 (2) (2006)
48 based on wavelet transformation and probabilistic neural net- 157–168.
49 work, Pattern Recognition Letters 28 (10) (2007) 1142–1150. [122] E. Mehmet, ECG beat classification using neuro-fuzzy net-
50 [102] M. H. Song, J. Lee, S. P. Cho, K. J. Lee, S. K. Yoo, Support work, Pattern Recognition Letters 25 (15) (2004) 1715–1722.
51 vector machine based arrhythmia classification using reduced [123] A. Rodan, P. Tiňo, Minimum complexity echo state network,
52 features, International Journal of Control, Automation, and IEEE Transactions on Neural Networks 22 (1) (2011) 131–144.
Systems 3 (4) (2005) 509–654. [124] M. Lukoševičius, H. Jaeger, Reservoir computing approaches
53 [103] J.-S. Wang, W.-C. Chiang, Y.-L. Hsu, Y.-T. C. Yang, ECG to recurrent neural network training, Computer Science Re-
54 arrhythmia classification using a probabilistic neural network view 3 (3) (2009) 127–149.
55 with a feature reduction method, Neurocomputing 116 (20) [125] V. Mahesh, A. Kandaswamy, C. Vimal, B. Sathish, ECG ar-
56 (2013) 38–45. rhythmia classification based on logistic model tree, Journal of
57 [104] K. Polat, S. Güneş, Detection of ECG arrhythmia using a dif- Biomedical Science and Engineering 2 (6) (2009) 405–411.
ferential expert system approach based on principal compo- [126] J. Rodriguez, A. Goni, A. Illarramendi, Real-time classifica-
58 nent analysis and least square support vector machine, Applied tion of ECGs on a PDA, IEEE Transactions on Information
59 Mathematics and Computation 186 (1) (2007) 898–906. Technology in Biomedicine 9 (1) (2005) 23–34.
60
61 20
62
63
64 Page 21 of 22
65
 1 [127] M. Korürek, A. Nizam, A new arrhythmia clustering technique arrhythmia beats, Neural Computing and Applications 24 (2)
based on ant colony optimization, Journal of Biomedical Infor- (2014) 317–326.
2 matics 41 (6) (2008) 874–881. [148] G. B. Moody, R. G. Mark, The impact of the MIT-BIH ar-
3 [128] A. Lanatá, G. Valenza, C. Mancuso, E. P. Scilingo, Ro- rhythmia database, IEEE Engineering in Medicine and Biology
4 bust multiple cardiac arrhythmia detection through bispec- Magazine 20 (3) (2001) 45–50.
5 trum analysis, Expert Systems with Applications 38 (6) (2011) [149] A. Taddei, G. Distante, M. Emdin, P., G. B. Moody, C. Zeelen-
6798–6804. berg, C. Marchesi, The european ST-T database: standard for
6 evaluating systems for the analysis of ST-T changes in ambu-
[129] V. Tavakoli, N. Sahba, N. Hajebi, A fast and accurate method
7 for arrhythmia detection, in: Annual International Confer- latory electrocardiography, European Heart Journal 13 (1992)
8 ence of the IEEE Engineering in Medicine and Biology Society 1164–1172.
9 (EMBC), 2009, pp. 1897–1900. [150] F. M. Nolle, F. K. Badura, J. M. Catlett, R. W. Bowser, M. H.
10 [130] F. Sufi, I. Khalil, Diagnosis of cardiovascular abnormalities Sketch, CREI-GARD, a new concept in computerized arrhyth-
from compressed ECG: A data mining-based approach, IEEE mia monitoring systems, in: Computers in Cardiology, Vol. 13,
11

t
Transactions on Information Technology in Biomedicine 15 (1) 1986, pp. 515–518.
12

ip
(2011) 33–39. [151] G. B. Moody, W. E. Muldrow, R. G. Mark, A noise stress
13 [131] Y.-C. Yeh, C. W. Chiou, H.-J. Lin, Analyzing ECG for car- test for arrhythmia detectors, in: Computers in Cardiology,
14 diac arrhythmia using cluster analysis, Expert Systems with Vol. 11, 1984, pp. 381–384.
15 Applications 39 (1) (2012) 1000–1010. [152] S.-W. Chen, P. M. Clarkson, Q. Fan, A robust sequential de-

cr
[132] D. A. Coast, R. M. Stern, G. G. Cano, S. A. Briller, An ap- tection algorithm for cardiac arrhythmia classification, IEEE
16 Transactions on Biomedical Engineering 43 (11) (1996) 1120–
proach to cardiac arrhythmia analysis using hidden markov
17 models, IEEE Transactions on Biomedical Engineering 37 (9) 1124.
18 (1990) 826–836. [153] S. Karimifard, A. Ahmadian, M. Khoshnevisan, M. S. Nam-

us
19 [133] P. R. Gomes, F. O. Soares, J. H. Correia, C. S. Lima, ECG bakhsh, Morphological heart arrhythmia detection using her-
20 data-acquisition and classification system by using wavelet- mitian basis functions and kNN classifier, in: Annual Inter-
domain hidden markov models, in: Annual International Con- national Conference of the IEEE Engineering in Medicine and
21 Biology Society (EMBC), 2006, pp. 1367–1370.
ference of the IEEE Engineering in Medicine and Biology So-
22 ciety (EMBC), 2010, pp. 4670–4673. [154] E. D. Übeyli, ECG beats classification using multiclass support
23
24
25
26
[134] E. J. d. S. Luz, T. M. Nunes, V. H. C. De Albuquerque, J. P.
Papa, D. Menotti, ECG arrhythmia classification based on
optimum-path forest, Expert Systems with Applications 40 (9)
(2012) 3561–3573.
an
[155]
vector machines with error correcting output codes, Digital
Signal Processing 17 (3) (2007) 675–684.
A. Khazaee, Heart beat classification using particle swarm op-
timization, International Journal of Intelligent Systems and
[135] M. G. Tsipouras, D. I. Fotiadis, D. Sideris, Arrhythmia classi- Applications (IJISA) 5 (6) (2013) 25.
27 [156] H. Chen, B.-C. Cheng, G.-T. Liao, T.-C. Kuo, Hybrid classi-
fication using the RR-interval duration signal, in: Computers
M
28 in Cardiology, 2002, pp. 485–488. fication engine for cardiac arrhythmia cloud service in elderly
29 [136] M. G. Tsipouras, C. Voglis, D. I. Fotiadis, A framework for healthcare management, Journal of Visual Languages & Com-
30 fuzzy expert system creation-application to cardiovascular dis- puting 25 (6) (2014) 745–753.
eases, IEEE Transactions on Biomedical Engineering 54 (11) [157] R. Ahmed, S. Arafat, Cardiac arrhythmia classification using
31
(2007) 2089–2105. hierarchical classification model, in: International Conference
32
d

[137] F. A. Minhas, M. Arif, Robust electrocardiogram (ECG) beat on Computer Science and Information Technology (CSIT),
33 classification using discrete wavelet transform, Physiological 2014, pp. 203–207.
34 Measurement 29 (5) (2008) 555–570. [158] H. L. Tran, V. N. Pham, H. N. Vuong, Multiple neural network
te

35 [138] M. Lagerholm, C. Peterson, G. Braccini, L. Edenbrandt, integration using a binary decision tree to improve the ECG
L. Sornmo, Clustering ECG complexes using hermite functions signal recognition accuracy, International Journal of Applied
36
and self-organizing maps, IEEE Transactions on Biomedical Mathematics and Computer Science 24 (3) (2014) 647–655.
37 [159] E. Alickovic, A. Subasi, Effect of multiscale pca de-noising in
Engineering 47 (7) (2000) 838–848.
p

38 [139] P. de Chazal, R. B. Reilly, A patient-adapting heartbeat clas- ECG beat classification for diagnosis of cardiovascular diseases,
39 sifier using ECG morphology and heartbeat interval features, Circuits, Systems, and Signal Processing 34 (2) (2015) 513–
533.
ce

40 IEEE Transactions on Biomedical Engineering 53 (12) (2006)

2535–2543. [160] T. Ince, S. Kiranyaz, M. Gabbouj, A generic and patient-
41
[140] M. Llamedo, J. P. Martinez, Analysis of a semiautomatic al- specific electrocardiogram signal classification system, in:
42 ECG Signal Processing, Classification and Interpretation,
gorithm for ECG heartbeat classification, in: Computers in
43 Cardiology, 2011, pp. 137–140. Springer, 2012, pp. 79–98.
44 [161] E. Luz, D. Menotti, How the choice of samples for building
Ac

[141] M. Gales, S. Young, The application of hidden markov mod-

45 els in speech recognition, Foundations and Trends in Signal arrhythmia classifiers impact their performances, in: Annual
46 Processing 1 (3) (2007) 195–304. International Conference of the IEEE Engineering in Medicine
[142] L. R. Rabiner, A tutorial on hidden markov models and se- and Biology Society (EMBC), 2011, pp. 4988–4991.
47 [162] P. de Chazal, Heartbeat classification system using adaptive
lected applications in speech recognition, Proceedings of IEEE
48 77 (2) (1989) 257–286. learning from selected beats, in: Computing in Cardiology
49 [143] R. V. Andreao, B. Dorizzi, J. Boudy, ECG signal analy- Conference (CinC), 2014, 2014, pp. 729–732.
50 sis through hidden markov models, IEEE Transactions on [163] S. Kiranyaz, T. Ince, M. Gabbouj, Real-time patient-specific
51 Biomedical Engineering 53 (8) (2006) 1541–1549. ECG classification by 1D convolutional neural networks, IEEE
[144] U. M. Fayyad, K. B. Irani, On the handling in decision tree Transactions on Biomedical Engineering (2015) 1–12 accepted
52 for publication.
of continuous-valued attributes generation, Machine Learning
53 8 (1992) 87–102. [164] M. Banko, E. Brill, Scaling to very very large corpora for nat-
54 [145] J. R. Quinlan, Improved use of continuous attributes in c4.5, ural language disambiguation, in: Annual Meeting on Associ-
55 Journal of Artifficial Intelligence Research 4 (1996) 77–90. ation for Computational Linguistics, 2001, pp. 26–33.
56 [146] J. R. Quinlan, Induction of decision trees, Machine Learning [165] A. Torralba, A. A. Efros, Unbiased look at dataset bias, in:
1 (1986) 81–106. IEEE Conference on Computer Vision and Pattern Recogni-
57 tion (CVPR), 2011, pp. 1521–1528.
[147] A. Mert, N. Kılıç, A. Akan, Evaluation of bagging ensemble
58 method with time-domain feature extraction for diagnosing of
59
60
61 21
62
63
64 Page 22 of 22
65