2019 International Conference on Automation, Computational and Technology Management (ICACTM)

Amity University

A Survey of Algorithms for Feature Extraction and

Feature Classification Methods
Prachi Lamba Kirti Rawal
Electronics and Communication Electronics and Communication
Lovely Professional University Lovely Professional University
Phagwara, India Phagwara, India
[email protected] [email protected]

Abstract— Recording of Electrocardiogram (ECG) signal machine learning techniques for classification the ECG beats
is a non-invasive method for examining the electrical actions of are not only used to reduce the doctor’s load but also improve
the heart. The characterization of ECG beat is an important the accuracy of diagnosis. The proper treatment of patient can
step in finding the location of various heart ailments. The be done by recognizing the abnormalities of heart.
abnormal beats are the symptoms of several diseases such as
myocardial infarction and ischemic heart diseases. The
classification of beats in healthy and diseased subjects aids the ECG beat classification is well known for classifying
researchers to detect various abnormalities in case of any variety of diseases particularly in the diagnosis of heart
arrhythmia. The arrhythmias of heart like structural, diseases. The different classes of ECG beats are Normal (N),
circulatory or electrical can be diagnosed by classifying the Paced (P), Ventricular ectopic beat (V), Supra Ventricular
abnormal beats. Thus, the classification of ECG beats is ectopic beat (S), Fusion (F), Unknown (Q), Left bundle
necessary in detecting and diagnosing the diseases. In this branch (LBBB or L) and Right bundle branch (RBBB or R)
paper, the computational methodologies adopted for ECG beat [2] as shown in TABLE 1.
classification and the issues related to it are presented. Further,
the feature extraction and feature classification methods of
ECG signal have been reviewed. The primary step of ECG beat classification is the
collection of Raw ECG data from the standard database
Keywords—Beat classification, ECG, Feature Extraction, followed by Pre- processing of the ECG signal. In pre-
Feature classification. processing, the noise in the signal which occurs due to power
line interference, baseline wandering, muscle noise etc is
I. INTRODUCTION removed. After pre-processing, the next step is to extract the
features from the signal known as Feature extraction. It
In order to measure the electrical activity of heart, ECG determines the amplitude as well as interval of the signal
signal is recorded by placing the electrodes on the skin. Using using various techniques [3] .Some of the techniques which
these electrodes, the heart produces tiny electrical changes are used for extracting ECG features are DWT (Discrete
that appears from the heart muscles during each heartbeat. wavelet transform), PCA (Principal component analysis),
Heart muscles follow the pattern of polarization and ICA (Independent component analysis), LDA (Linear
depolarization. Therefore, it is a medical test which identifies discriminate analysis) etc. After extraction, the last step is
cardiac abnormalities that occurred due to the contraction of Feature classification. It is used to classify the normal and the
the heart [1]. The machine which is used to record the abnormal beats by using various methods. Some of the
patient’s ECG is known as electrocardiograph. The ECG methods used for classifying the ECG beats are SVM
signal is used to analyze the variations in heart rate, the size (Support vector machine), KNN (K-nearest neighborhood),
and position of the heart chambers, heart's muscle cells etc. ANN (Artificial neural network), RBF (Radial basis
Further, ECG is also used determine the effects of function) etc.
cardiovascular medications, and the capacity of embedded
pacemakers. Therefore, ECG analysis is preferred for most of
the cases in detecting and diagnosing the heart related
diseases.

The features and parameters of ECG signal changes due

to any illness or irregularities in the heart rhythm. A specialist
or radiologist usually records the ECG of heart patients or the
subjects who might be at the danger of the coronary illness.
Thus, ECG signal analysis is not limited to diagnose the heart
diseases, but it also includes the subjects having
hypertension, elevated cholesterol, diabetes, overweight,
smoking propensities and coronary illness [1].

In order to prevent the life of numerous heart patients,

early diagnosis is necessary to cure the diseases. For early
and timely diagnosis, the ECG signal is classified by using
several machine learning techniques. These computers-based Fig. 1. Flowchart of ECG beat classification system

338
978-1-5386-8010-0/19/$31.00 ©2019 IEEE

Authorized licensed use limited to: VIT University. Downloaded on October 20,2020 at 03:42:53 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Automation, Computational and Technology Management (ICACTM)
Amity University

Once the classification is done, then on the basis of ST. ECG beats includes P wave, QRS complex and T wave
classification of normal/abnormal beats, the diagnosis can be in one ECG signal.
done [4]. The steps used for classifying the ECG beats are
shown in Fig. 1. In spite of the huge developments in this P wave and QRS complex lasts for 0.08 seconds and
field there are still lots of improvement needs to be done in their duration is not more than 0.10 seconds. P wave occurs
order to make the extraction and classification system more due to depolarization from SA node throughout atria. But in
robust, reliable and accurate. So, in this paper we tried to case of QRS complex, Q and S both are immediately
present the summary of the work done so far in order to deflecting downwards. Q arises before ventricular
facilitate the researchers to have an idea of existing work. The contraction while S arises after ventricular contraction and R
comprehensive analysis and the basics related to ECG beat occur due to polarization of ventricles. R beat deflected in
classification can help the beginners to understand the positive direction, but Q and S beat deflected in negative
research area. direction. For such beat, ventricular systole & atrial diastole
takes place. T wave lasts for 0.16 seconds and it occur due to
This paper incorporates the fundamental concepts of repolarization of ventricles or when ventricular diastole takes
ECG and its characterization (segment 1), foundation place. Lastly, U wave is a kind of short wave which occurs
information (segment 2) and a detailed study of ECG include after the occurrence of T wave, but it is not mostly observed.
extraction and classification of ECG beats. In this area,
diverse strategies are utilized for highlighting the extraction III. LITERATURE SURVEY OF ECG BEAT CLASSIFICATION
of ECG features, classification of ECG beats (segment 3) and METHODS
the various databases used for ECG analysis (segment 4). At
last, the paper presents conclusion (segment 5). Numerous methodologies are adopted by the
investigators for the ECG beat classification. It includes 3
The different types of beats which are used to major stages namely, pre-processing stage, feature extraction
classify the normal and abnormal subjects are shown in and feature classification. MIT-BIH Arrhythmia database
TABLE I. (The Massachusetts institute of Technology Beth Israel
hospital arrhythmia database) [5] is most commonly used in
TABLE I CLASSIFICATION OF ECG BEATS most of the papers. Based on various classifiers, some of the
parameters are calculated which gives the information about
Name of the performance of classification of beats.
Explanation
the Beat
A typical resting pulse ranges from 60 to 100 BPM Ref. [6, 7 and 8] extracted some ECG features with the
N (Beats per Minute). help of XWT (Cross wavelet transform), DWT and ICA
techniques respectively and classified the beats by SVM
A postponement or blockage of electrical driving classifier having accuracy of 94.8%, 98.94% respectively and
L forces to left side of the heart.
classification rate for [8] is 99.75%. Ref. [9 and 11] used
A postponement or blockage of electrical driving CWT (Continuous wavelet transform) for feature extraction
R forces to right side of the heart. and classification of arrhythmia beats is done by RBF and
SVM respectively having accuracy of 98.92% and 95.03%
When heart rhythm is interrupted, irregular or too
P slow.
respectively. Ref. [10] determined the characteristics of
cardiac beats by EMD (Empirical mode decomposition),
A beat that is pre-mature, narrow in width and is DWPD (Discrete Wavelet Packet Decomposition) and KNN
different from normal patient’s beat which may which yielded an average accuracy, sensitivity, selectivity of
S even lead to heart attack or heart stroke. Also,
called as PAC (Premature atrial contraction).
99.2% and average specificity of 99.8%. Ref. [12] extracted
ECG parameters by using GA (Genetic algorithm) which was
A beat which shows irregular heart rhythm due to broadly classified into temporal, spectral, complexity and
premature heartbeat. It occurs due to skipped beat wavelet features. They further used ANN and SVM, among
or stopped beat, when heart is in a fluttering
V them ANN gave better results when provided with 100%
condition. Also, called as PVC (Premature
ventricular contraction). training. Ref. [13] diagnosed different arrhythmias by DT
(Decision tree) and KNN resulted in accuracy of 96.3%,
When ventricular and supra ventricular electrical 93.3% and sensitivity of 99.3%, 97.5% respectively. Analysis
impulses coincide at the same time to produce a of different arrhythmia beats was done with the help of DWT
F
hybrid complex beat.
and SVM methods [14], and achieved specificity and NPV
(Negative predictive value) of 92%. Ref. [15] examined the
performance of normal and abnormal heartbeats with DSNT
II. BACKGROUND KNOWLEDGE (Discrete sinc transform), DCT and SVM got the accuracies
of 95.45%, 90.91% for recognition, with specificity 100%,
An ECG tracing consists of several components that 91.67%, and sensitivity 90% respectively.
indicate the electrical events during one heartbeat. ECG
signal comprises of various ECG beats which in turn contains Ref. [16, 17, and 18] applied DWT and SVM for the
some ECG features such as peaks, intervals and segments. detection of ECG beats and obtained accuracy 99.59%, 96%
Peaks namely P, Q, R, S, T and U while that of intervals and 88.24% respectively. For extraction of ECG beats,
namely PR, RR, QRS, ST, QT and segments namely PR and various authors [19, 20, 21, 22 and 23] utilized DWT and

339

Authorized licensed use limited to: VIT University. Downloaded on October 20,2020 at 03:42:53 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Automation, Computational and Technology Management (ICACTM)
Amity University

PCA technique and further used different classifiers for beat minimization of within-class distance is the key concept of
classification and gave better results in [23]. An Accuracy of LDA which concludes in the dimensionality reduced space.
99.08% was achieved in [24] who classified 5 different heart
beat types with the help of DWT and RF classifier. Ref. [25] C. Feature classification methods
classified 5 types of beats with ST (Stockwell transform),
BFO (Bacteria foraging optimization) algorithm and LMS Feature classification is the process of grouping of features
(Least mean square) based SVM resulted in average accuracy based on some criteria which categorizes data based on
and sensitivity of 98.2%, 74.7% (S-detection) and 98.6%, different classes and it includes various techniques such as:
91.7% (V-detection).
1) SVM:
IV. DATABASE AND METHODS SVM classifier is a sort of administered learning model
which is utilized to break down information for arrangement
This section includes ECG databases, various feature and relapse. It is characterized utilizing an isolating hyper
extraction methods, feature classification methods and plane which groups all the preparation vectors into two classes.
performance measures which are discussed below: The best decision will be the hyper plane that departs the most
extreme edge from the two classes. The edge is the separation
A. ECG Database between the hyper plane and the nearest component from hyper
MIT-BIH arrhythmia database have been used by the plane. Hence, the margin is maximized between the two classes
researchers for carrying out their study based on ECG or which mean seperability is also maximized.
other cardiac disease analysis or diagnosis. According to the
AAMI (Association for the advancement of Medical 2) RBF:
Instrumentation) recommendation, this database was taken RBF is a classification approach which is used for
from PhysioBank ATM that is available at physioNet website functional approximation and classification that uses nonlinear
[5] contains various data sets of biomedical signals. Most of activation functions like Sigmoidal and Gaussian Kernel.
the researchers [6-11, 12, 13-18, 20-23, 24 and 25] used MIT- Therefore, the response of Gaussian function is positive for all
BIH database and European database for analyzing the ECG the values of x and it decreases to zero as |x→0|. RBF as the
signals [19]. name suggests, is found to be radically symmetric as it
produces identical output for all the value of the input from the
B. Feature extraction methods centre of the kernel.
Feature extraction is the method of gathering
discriminative information from a set of samples. Features 3) KNN:
are required to be computed for extraction of clinically KNN is a simplest algorithm which is used for
valuable information from the textures. The features can be classification as well as regression of k nearest neighbour data
considered as supplement to the visual skills of the sets given. The KNN model splits the given data into a number
researchers that are relevant to the diagnostic problem but of classes to predict the classification of a new dataset. It acts
may not be visually perceptible. The distinguishing and as a “clustering model”. As a classifier, it predicts the
potent features are extracted by using different feature membership of a class based upon the majority vote of its k
extraction methods. Some of the feature extraction methods nearest neighbours. In case of regression, the mean (average)
are explained below: of its k nearest neighbours depicts its class. It is determined
using Euclidean distance.
1) DWT:
DWT deals with the decomposition of signal into a set V. CONCLUSION
of mutually orthogonal wavelet basis function which have
smaller bandwidths and slower sample rates. It consists In this paper, the computational methodologies adopted
mainly of two frequency ranges i.e low frequency leads to for ECG beat classification and the issues related to it are
Approximation component (cA) and high frequency leads to presented. Additionally, the review of feature extraction and
Detail components (cD) of the signal. feature classification methods of ECG signal has been
presented in this paper. Classification of beats in normal and
2) PCA: abnormal class plays an important role in early diagnosis of
PCA is a statistical technique and a dominant tool which heart related diseases. Abnormal beats can be the cause of the
is used to analyze and compress a large amount of data that cardiac ailments like Ischemic heart disease or related
captures the essence of original data. This method indicates Arrhythmias or can be the vital sign provider for life
various similarities and differences present in various threatening. This diagnosis helps in detecting the cardiac
samples of data. Hence, it focuses on calculating the diseases at an early stage so as to reduce the mortality rate
correlation between the samples. caused by cardiac arrhythmias, cardiac arrests, and overall
cardiac deaths. A considerable amount of work has been
3) LDA: already done by different researchers to extract the features
LDA is a straightforward arrangement procedure which from the ECG signal and then classify it. Investigation of the
depends on the idea of looking for direct mix of factors that ECG signal relies upon precise recognition of different
best isolates among two classes. It mainly focuses on features of ECG signal. DWT, PCA, ICA, LDA and Pan-
conversion from high dimensional data into low dimensional Tompkins algorithm are some of the techniques which have
data. Maximization of between-class distance and been applied to extract the ECG features. Similarly, SVM,

340

Authorized licensed use limited to: VIT University. Downloaded on October 20,2020 at 03:42:53 UTC from IEEE Xplore. Restrictions apply.
2019 International Conference on Automation, Computational and Technology Management (ICACTM)
Amity University

KNN, RBF and ANN are some of the classifiers which have [12] Karthika, J. S., Jan Mary Thomas, and Jubilant J. Kizhakkethottam,
“Detection of life-threatening arrhythmias using temporal, spectral and
been used for ECG classification. However, from the survey
wavelet features,” In Computational Intelligence and Computing
it is examined that DWT with the combination of SVM has Research (ICCIC), pp. 1-4, December 2015.
been verified with several measures which provide promising [13] Acharya, U. Rajendra, Hamido Fujita, Muhammad Adam, Oh Shu Lih,
classification results in case of arrhythmia. Therefore, it is Tan Jen Hong, Vidya K. Sudarshan, and Joel EW Koh, “Automated
characterization of arrhythmias using nonlinear features from
concluded that there is a lot of scope of research in extracting
tachycardia ECG beats,” In 2016 IEEE International Conference on
the features of ECG signal and then classification of the Systems, Man, and Cybernetics (SMC), pp. 000533-000538, October
following ECG signal. The classification of the ECG beats is 2016.
the open issue of research. ECG beat classification is [14] Dewangan, Naveen Kumar, and S. P. Shukla, “ECG arrhythmia
classification using discrete wavelet transform and artificial neural
important to diagnose the possible illness of human being
network,” In Recent Trends in Electronics, Information &
along with ECG signal analysis. Communication Technology (RTEICT), pp. 1892-1896, May 2016.
[15] Gad, Mohamed Moustafa Azmy, “Feature extraction of
REFERENCES electrocardiogram signals using discrete sinc transform,”
In Biomedical Engineering International Conference (BMEiCON), pp.
1-4, December 2016.
[1] Jambukia, Shweta H., Vipul K. Dabhi, and Harshadkumar B. Prajapati, [16] Ranaware, Preeti N., and Rohini A. Deshpande, “Detection of
“Classification of ECG signals using machine learning techniques: A arrhythmia based on discrete wavelet transform using artificial neural
survey,” In Computer Engineering and Applications (ICACEA), pp. network and support vector machine,” In Communication and Signal
714-72, March 2015. Processing (ICCSP), pp. 1767-1770, April 2016.
[2] Ahmed, Waqas, and Shehzad Khalid, “ECG signal processing for [17] Venkatesan, C., P. Karthigaikumar, Anand Paul, S. Satheeskumaran,
recognition of cardiovascular diseases: A survey,” In Innovative and R. Kumar, “ECG Signal Preprocessing and SVM Classifier-Based
Computing Technology (INTECH), pp. 677-682, August 2016. Abnormality Detection in Remote Healthcare Applications,” IEEE
[3] Luz, E.J.D.S., Schwartz, W.R., Cámara-Chávez, G. and Menotti, D., Access 6, pp. 9767-9773, 2018.
“ECG-based heartbeat classification for arrhythmia detection: A [18] Sabeeha, S., and C. Shiny, “ECG-based heartbeat classification for
survey,” Computer methods and programs in biomedicine, pp.144- disease diagnosis,” In Computing Methodologies and Communication
164, April 2016. (ICCMC), pp. 1113-1117, July 2017.
[4] Houssein, E.H., Kilany, M. and Hassanien, A.E., “ECG signals [19] Desai, Usha, “Automated detection of cardiac health condition using
classification: a review,” International Journal of Intelligent linear techniques,” In Recent Trends in Electronics, Information &
Engineering Informatics, pp.376-396, 2017. Communication Technology (RTEICT), pp. 890-894, May 2017.
[5] https://physionet.org/ (Internet Source). [20] El-Saadawy, Hadeer, Manal Tantawi, Howida A. Shedeed, and
[6] Jacob, Neenu, and Liza Annie Joseph, “Classification of ECG beats Mohamed F. Tolba, “Electrocardiogram (ECG) heart disease diagnosis
using cross wavelet transform and support vector machines,” using PNN, SVM and Softmax regression classifiers,” In Intelligent
In Intelligent Computational Systems (RAICS), pp. 191-194, Computing and Information Systems (ICICIS), pp. 106-110, December
December 2015. 2017.
[7] Desai, Usha, Roshan Joy Martis, C. Gurudas Nayak, K. Sarika, and G. [21] Elhaj, Fatin A., Naomie Salim, Taqwa Ahmed, Arief R. Harris, and
Seshikala, “Machine intelligent diagnosis of ECG for arrhythmia Tan Tian Swee, “Hybrid classification of Bayesian and extreme
classification using DWT, ICA and SVM techniques,” In India learning machine for heartbeat classification of arrhythmia detection,”
Conference (INDICON), pp. 1-4, December 2015. In Student Project Conference (ICT-ISPC), pp. 1-4, May 2017.
[8] Barhatte, Alka S., Rajesh Ghongade, and Abhishek S. Thakare, “QRS [22] Akin, Zahide Elif, and Süleyman Bilgin, “Classification of normal
complex detection and arrhythmia classification using SVM,” beat, atrial premature contraction and ventricular premature
In Communication, Control and Intelligent Systems (CCIS), pp. 239- contraction based on discrete wavelet transform and artificial neural
243, November 2015. networks,” In Medical Technologies National Congress (TIPTEKNO),
[9] Harkat, A., R. Benzid, and L. Saidi, “Features extraction and pp. 1-4, October 2017.
classification of ECG beats using CWT combined to RBF neural [23] Martis, Roshan Joy, U. Rajendra Acharya, and Lim Choo Min, “ECG
network optimized by cuckoo search via levy flight,” In Electrical beat classification using PCA, LDA, ICA and discrete wavelet
Engineering (ICEE), pp. 1-4, December 2015. transform,” Biomedical Signal Processing and Control 8, pp. 437-448,
[10] Shahnaz, C., T. B. Anowar, R. H. M. Rafi, I. Ahmmed, and S. A. September 2013.
Fattah, “Cardiac beat classification based on wavelet analysis of [24] Emanet, Nahit, “ECG beat classification by using discrete wavelet
empirical mode decomposed ECG signals,” In TENCON 2015-2015 transform and Random Forest algorithm,” In Soft Computing,
IEEE Region 10 Conference, pp. 1-6, November 2015. Computing with Words and Perceptions in System Analysis, Decision
[11] Shahbudin, S., S. N. Shamsudin, and H. Mohamad, “Discriminating and Control, ICSCCW, pp. 1-4, September 2009.
ECG signals using Support Vector Machines,” In Computer [25] Das, Manab K., and Samit Ari, “Patient-specific ECG beat
Applications & Industrial Electronics (ISCAIE), pp. 175-180. IEEE, classification technique,” Healthcare technology letters 1, no. 3, pp.
April 2015. 98-103, September 2014.

341

Authorized licensed use limited to: VIT University. Downloaded on October 20,2020 at 03:42:53 UTC from IEEE Xplore. Restrictions apply.