Internationa l )ou.

rna~ofJ\dvnncc Research _h, Sdencc and Engineering ~

Volume No.07, Special Issue No.01, Mat'ch 2018 JfARSE
W'\VW.ijarse.con1 ISSN· 23 lfJ-l13,H

SCREENING OF ABNORMAL ECG USING MACHINE

LEARNING TECHNIQUE

Diksha Santosh Gaonkar 1, Purti Savardekar 2

1 Dept. ofElectronics and Telecommunication Engineering, Goa College ojEngineering
2 Dept. of Electronics
and Telecommunication Engineering, Goa College o/Engineering

ABSTRACT

Meteoric advancements in technology and te/emedicine sector have eased prior diagnosis of various diseases.
One of the most recognized medical conditions that need prior diagnoses is cardiac a"hythmia. Since ECG
provides the essential graphical trace of bio potential activity of heart, it has become a primary diagnosis tool
for the assessment of various heart diseases. In this paper we propose a feed-forward back propagation neural
network classification system for arrhythmia detection using bandpass filtering strategy. Proposed method uses
BPNN to classify the ECG signals as normal and abnormal The features are divided into two classes that are
morphological features and DWT based features which are provided as input to the classifier. The performance
of system was evaluated based on the percentage ofaccuracy, Our experimental result on MIT-BIH allhythmia
database showed overall system accuracy of91.3% with the use ofBPNN classifier.

Keywords: ECG, Allhythmia, MIT-BIH, BPNN, DWT

1. Introduction
An electrocardiogram (ECG) is a bioelectrical signal which records the heart's electrical activity versus time
and thus is an important diagnostic tool for evaluation of heart functions. The signals that make the heart's
muscle to contract come from the sino atrial node, the normal pacemaker of heart which is located at the top of
right atrium1 5l_

The ECG waveform generally consists of P, Q, R S, T and U wave. The muscle contraction of the atria
produces the P wave which represents atria depolarization. Ventricular depolarization is characterized by Q, R
and S wave generally termed together as QRS complex which is the most vital feature of the ECG waveform.
The ending of the atrial contraction and the beginning of the ventricular contraction is marked by the R wave
peaks whose magnitude ranges from O. lmV and 1.5mV. At last the ending of ventricular contraction is marked
by the T-peak. The beginning of normal electrical sequence is al ways marked by atrial depolarization followed
by ventricular depolarization. But in case of cardiac arrhythmias this sequence gets disrupted and rhythms tend
to be irregular 111. Over the years several techniques have been developed for ECG signal analysis aiming to
achieve good percentage of accuracy, specificity and sensitivity. These techniques include classification
methods such as support vector machine, RBF neural networks, fuzzy logic, wavelet coefficients and self-
organizing maps. The figure 1 shows the typical ECGwaveform.

837 IP a g e
 Internationa l lournaJ of Advance Resenrch In Science and Engineering '1
1
Volum e No.07, Special Issue No.01 I March 2018 IIARSE
www.ijarse .cnn1 IS\~ ll 19 (H54

----
• • hltn,.

R i',, R 'I
II
ra I, l

lJl r i r
..,,_ I \ u .,._,
I ,·-
p ....'I'
~-
,, ~.· '" ✓-
Q s m~.,T Q y

~ ,.,.,.,,
QTt.1,r•al T,Qt ■i,1.,1

2. DatabaseCollection
Fig 1: Typical ECG wavefom1
) arrhythmia database
For this paper, Massachusetts Institute of Technology Beth Israel Hospital (MIT-BIH
we have selected 46
from physionet is used 1 1_ This database contains total 48 files of30 minutes recordings but
4

based on maximum
files of IO sec duration where 24 are considered as normal class and 22 as abnormal class
classes where first
number of beats present in each record. Database includes division of recordings into two
at random from
part is of 23 records (numbered from 100 to 124 inclusive with some numbers missing) chosen
this set, and second part is of 25 records (numbered from 200 to 234 inclusive, again
with some numbers
file and a binary
missing). The ECG waveform from MIT-Blll Database contains - a text header file, a binary
annotationfile.
3. Methodology
is sectioned into three
The block diagram of proposed technique is depicted in figure 2. The entire methodology
main divisions namely:
ng.
• Preprocessing: It includes baseline wander removal from raw ECG signal followed bydenoisi

called
• Featureextraction:Itincludesextractingandconvertingtheinputdatainforrnationintosetfeatures
feature vectors, by reducing data representationpattem.
i.e.normal
• Classification: It includes the use ofBPNN as a classifier to classify signal into two classes
and abnormal.
and provided for
From the figure 2 it can be seen that the raw ECG signal is loaded from the database
which best resembles
preprocessing followed by feature extraction which involves preparing the precise input
into abnormal and
the original signal and finally the classification stage where the processed signal is classified

normal class.

838 IP age
 lnten1attona JJournnl of Advnncc Hescarch in Science and EngJneerfng ~
1
Volume No. 07, Specia l Issue No.01, Mar ch 2018
IJARSE
w,vw.ijarse.com m~ l w, n1s-i

IJlt OIH Ln1d 1ht tlllnal

1J1,t:1u

l'rtl•l'Ocrulnr,
l.11d1•t ltUt na1t al

fHl111·t F\"lra,1ion

u,,.,a..1eginl furve.1

DWT c.edl'idMb NIN fulw,...

Cmuifiation

I
SHrllfrhi \buormal Clo<1

4. ECG SignalPreprocessing
Fig 2 : Block diagram of arrhythmia detection system

Preprocessing is the initial stage in ECG signal analysis l J_ ECG being a non-stationary signal is often
3

contaminated with various sought of noise artifacts which corrupt the raw ECG. It is therefore necessary to
eliminate these noise artifacts for further analysis. This preprocess of ECG involves removal of various trends
affecting the ECG and denoising using band pass filtering strategy.

4. 1 Removal of baseline wander

The most common noise artifact responsible for contaminating the ECG signal is the Baseline Drift. It generally
results from respiration, motion and changes in electrode impedance and lies between O. l 5Hz and 0.3Hz. It can
actually mask the necessary information from the ECG and cause irregularities in beat morphology. In this
paper, the baseline drift of the ECG signal is eliminated by using the moving average filter to achieve the
smoothed signal. Then the smoothed signal is subtracted from the original signal to get rid of the baseline drift
from the ECG signal. Hence the signal thus obtained is free from baseline

4.2 Denoising

With the eradication of baseline drift the ECG signal is free from DC offset but still contains some noise. In this
paper we try to implement band-pass filtering approach for noise rejection as it improves SNR and reduces
influence of these noise sources Since the desired pass band 5-15Hz is unable to achieved using bandpass filters
directly for our chosen sample rate we use the cascade of low pass and high pass filter to achieve 3dB pass band

from 5-l2Hz.

It is the cascade of low pass filter and high pass filter. The low pass filter eliminates 50Hz power line noise and
electromyogram noise, having cutoff of l lHz and with the gain 36 followed by processing delay of 6 samples.

The difference equation for low pass filter is:

y(nT) =2y(nT-T)-y(nT-2T)+x(nT)-2x(nT -6T)+x(nT-12T) ......... .. ......... . ......... .... ........ ... .... .. (1) The

839 I p a g C
 In Sdcnre and C1111t11ccring IL
International Journal of Advance llcscarch '1
, MOtch 2018
Volume No ,07, S1,ecial Issue No ;OJ IIARSE
w,vw.ijarse.co1n 1•,<;"J • 1, w, ne ,
, _ _ . T wav e
high pass filter eliminat es lower froqucn cy components wlu ch nrc molion ortlfo cts, fl wave and

filter is:
The difference equation for high puss

y(oT) = 32x(nT-l6T)-[y(nT-T)+
x(nT)-x(nT-32T)J. " ..... . ..... .... .. . ..... . .. ... ... ..... ..... . ..... .... (2)

.·1.1/t~,~ ,. i,J..,, .L.l,,l "ar•Ut.

1
;t I
.. I i r , r·. I
I

~ ; ~ ~ ; ;
:~

•• /f', • •
~ ••l " t •IO,t

____ .... --~ ' __ ,

I., " - .J,_, - '-~~~-<J.-·- ~◄ --· ...tJ ... _._.,..
<f
~
i.' ~
~
iU
iJ J ;, J.
iSf.J"1 1;)'loJO,

removal
Fig 4: ECG signal with baseline drift

Band Pass FIitered

0,5

-0,5
650 700 750 800 850
500 550 600

Fig 5: Dcnoised ECG signul

5. FeatureExtraction
re extraction of ECG
oval of base line wan ders , noise reduction and peak detection, the featu
Followed by rem set extracts the
as use it in the precedin g stage of ECG signal analysis. The feature
waveform is essential so to
ation task. It is thus the prevalent
ropriate info rmation from the inpu t data in order to perform the classific
app
ed of ECGwaveform.
hod, two types of features are consider
step in pattern recognition. In this met

1. Morphological Features
atures
2. Wavelet Coefficients based StatisticalFe
840 IP n g e
 International Journal of Advonce Research In Science and Englnecring ~
Volume No.07,. Special Issue No.01, March 201 8 IJARSE
www.ijarse.com i ~~ l 119 RY ..4

Selecting npproprintc fcntmes pl oys n very crucinl role in pattern recognition task.

ln this work . the considered morphological features include QRS co mplex, maximum R interval , mi nim um R
interval . QT intervaJ , ST inte1val , heart rat e (HR), maxi mum amplitude of R wave, max.i mum amplitude of T
wave. maxi mum amplitltde of S wave and maximum amplitude of Q wave.

Apart from morphological features wavelet· coeffic ients based stati sti cal features are al so obtai ned. Si nce DWT
coefficients represent the energy di stribution of the signal th ey can thus be provided as feature vectors
representing the signal to the classifi er. Considering the dimensional reduction issue the statistics of wavel et
coeffi cients are used

The wavelet coefficient based features utilized are as follows:

a Mean values of the detail and approximate coefficients at eachlevel.

b. Variance values of the detail and approximate coefficients at eachlevel.

c. Standard deviation values of the detail and approximate coefficients at eachlevel.

So at the end of feature extraction stage total 48 wavelet coefficients based features and IO morphological
features for each ECG signal was obtained which comprised of S8 features in total.

6. Neural Network
Artificial neural network (ANN) is a massively parallel-distributed processor resembling the human brain that
has a natural propensity for storing experimental knowledge and making it available for use.

In this paper neural network is used in pattern recognition providing input units as feature vectors and output
units as the class to be classified. Each corresponding feature vector is served to the input layer whose output is
considered to be corresponding element in the vectors. Hidden layer is responsible for calculation of weighted
sum of its input thus providing net activation of its scalar.

1.1 Feed-forward back propagation neural network

The term back propagation refers to the manner in which the gradient is computed for nonlinear multilayer
networks. Standard back propagation is a gradient decent algorithm which is used to find the
weights.Backpropagation algorithm uses feedfordward architecture containing hidden layers.

7. ExperimentalResults
In this work, the entire system was evaluated on MIT-BIH arrhythmia database sectioned into two classes
namely abnormal and normal . Among the total 48 files of 30 min long only 46 files of 10 sec duration were
utilized. The total S8 number of features were divided into two separate sections. These include 48 DWT
coefficients based statistical features and 10 morphological features of ECG signal. Simltlation and training of
the network involved 32 samples from the database with 7 samples provided for testing and 7 samples for
validation set. The classifier was designed with 3 hidden layers containing 20 neurons in first hidden layer, 2S
neurons in second hidden layer and 31 in third hidden layer respectively . The output layer is responsible for

841 IPage
 International Journa l of Advanc~ Hesearch in Science and Engineering -~
Volum.o No.07, Special Issue No.0 1, March 2018 JJA~ E
www.1jaf'se.com ,ssu.J.r11tJ.,n.:M
. rornng
docidiug th o fo rgets mnt rix ns ( I , 0) nnd (0 , I) 101 · to normal and abnormal class.
· torm1J or con1us1
Figuro 6 depicts the si mul11ti o111·ost1l1s o/' tli c 110urnI noIwork 111 r ·on matri·x.

Alt Conru ■ lon Matrix

..
::i
t
211
00.9%
4
8.7%
11
I
1#.l I

ra )
u
a. 2 0 H fl)!M
:
1;
,(),0% 30,1~ "·O,,,
a . I! ;,;f

·r
0
·"
1
%
,,
I•
..
'T1 ,,~ ,
~ I

2
,,
,1.,v,
817"

Target Class

Fig 6: Confusion matrix of BPNN

With the reference to the confusion matrix there is no misclassification carried out for normal ECG samples.
That means out of 28 samples provided to the network all 28 samples were classified correctly by BPNN
are
.Whereas 4 abnormal samples were classified wrongly by BPNN as normal samples, that means 14 samples
classified correctly out of total 18 samples. Therefore the overall classification accuracy of9l.3% was achieved
using back propagation neural network.

8. Conclusion
n
This paper addresses an effective feedfordwad back propagation neural network based system for classificatio
of cardiac arrhythmia into two separate classes as normal and abnormal. This work was carried out on 46 files
of MIT-BIH arrhythmia database consisting of 48 files 30 mins long. With total 58 features comprising of
IO

morphological features and 48 DWT coefficients based features the overall system accuracy of 91.3% was
achieved using 3 hidden layers with 20,25 and 31 number of neurons in first , second and third hidden layer
respectively

Acknowledgment
I am highly indebted to my guide Miss Purti Savardekar for her guidance and constant supervision regarding the
project.

I would also like to express my sincere gratitude towards the HOD and staff of electronics and
telecommunication department Goa College of Engineering Farmagudi, where the entire project was carried
out.

References

[I] K Amtul Salam, G. Srilakshmi, "An algorithm for ECGanalysisof arrhythmia

detection",Electrical,
Computers and Communication Technologies (ICECCT), .IEEE International Conference, March 2015

842 IP a g e
 International Journal of Advance Research inScience and Engineering 4
Volume No.07, Special Issue No.Oli March 2018 IJA ~ E
www.ijarse.com ,ssN; z319.03s1

[2] Ali Bahararni Rad,Trygve Etestol,Kjersti Engan, Unai lrusta,Jan Johansen,Lars Wik,Aggelos K.
Katsaggelos, " ECG based classification of Resuscitation Cardiac Rhythms for Retrospective Data Analysis",
IEEE Transaction on Biomedical Engineering ,volume:64,Issue: 10,October2017.

[3] K K Makesheva, Y. A. Altay, "Noise-resistant processing of electrocardiosignal in MATLAB",

European Journal ofScientific Researc, vol. 13, no. I, pp. 898-903,2016.

[4] "TheMIT -BIHArrhythrniaDatabase,http://physionet.ph. bi~-ac.i1/physio bank/database/mitdb/

[SJ Hari Mohan Rai,Anurag Trivedi, "ECG signal Classification using Wavelet Trasform and
Backpropagation neural network",Computers and Devices for communication (CODEC), 5th IEEE International
Conference, December 2012.

[6] Hilmy Assodiky, Jwan Syarif, Tessy badriyah, "Deep learning algorithm for arrhythmia
detection",Knowiedge Creation and Intelligent Computing (IES-KCIC) ,IEEE International Electronics
Symposium,September 2017.
[7] Miad Faezipour,Adnan Saeed,Hlaing Minn,lakshman Tamil, "A Patient Adaptive profiling Scheme for
ECG Beat Classification",IEEE Trans. On Information Technology in Biomedicine , volume:14,
Issue:S,September 2010.

843 IPage
Nt1tio11al Co11fere11ce 011 Emer,:l11,: Treml~ i11 Scie11ce and Tec/mology
Removal of ECG Noise Artifacts Using Signal
Processing Techniques
Diksha Santosh Gaonkar 1, Purti Sav11rdekar2
12
• Dept of Electronics and Telecommunication Department
12
• Goa College of Engineering
Abstract- ECG is considered to be the most vital physiological
parameter for examining the anatomic and physiologic facets
of the heart muscles. It is a graphical trace providing the
elccrrical activity of the hea,1 and is th11s keenly 11sed for
clinical diagnosis. But over time, the ECG signals tend to
change their statistical prope11ies and hence are referred to
have non-stationary behavior. Since the tools 11sed in ECG
recordings are based on analog circuit,y, in certain
exhausting and peripatetic conditions, the recorded ECG is
contaminated with various types of noise arlifacts. The main
anificial artifacts include power-line interference, baseline
drift and electrode contact noise. The main biological artifacts
are motion artifacts and muscles artifacts (EMG noise). This
paper presents an approach to filter the noise artifacts in the
ECG signal and thus facilitate the detection of the QRS
complex using the adaptive threshold The experimental
results also present the evaluation parameters before and
after filtering of the ECG signal.
Keywords- ECG; MIT-BIH; Baseline Drift; SNR
L INTRODUCTION
The Electrocardiogram (ECG) is the electrical
manifestation of the contractile activity of the heart that can be
recorded rapidly and automatically. The ECG signals can be
measured painlessly, using 12 lead or 3 lead electrode system.
This is a noo-invasive procedure that is used to record the
electrical activity of heart.
The contraction and relaxation of heart muscles
causes the blood to flow in and out of the heart. During each
cardiac cycle, electrical impulses are generated by a group of
tissues in the heart called as the sino atrial node also knooo as
the pacemaker of the heart 121. This electrical activity impulses
spread all through the heart and causes rhythmic contraction
and relaxation of the heart muscles. By placing electrodes at
specific points on the body these electrical impulses can be
detected. The electrocardiogram signal is characterized by six
peaks and valleys labeled with successive letters of the
alphabet: P, Q, R, S, T and U rsJ _
Page \ 13
ISSN IONUNEJ: 2395-1052
A
p T
Q s
PR ST
segment . ORS
interval
segment
PR
Interval
QT interval
Figure l : ECG Waveform
The tools of ECG recording procedure are non-
invasive and hence are based on analog circuitry and manual
handling. Therefore a typical ECG is invariably corrupted by
electrical interference from surrounding equipment (effects of
the electrical mains supply), movements artifacts, baseline
drift and instrumentation noise which are the major noise
sources contaminating the raw ECG. It is necessary to analyze
the ECG signal correctly and remove the processing
techniques which include digital filter and adaptive filter
approaches have been proposed by the authors for the same l7l_
This paper presents the filtering strategy for noise removal
from the raw ECG and thus facilitates the detection of QRS
complex using thresholding rules. It also slio,vs a table
depicting the evaluation parameters before and after filtering
highlighting the fact that noise removal and baseline
cancellation is necessary for processing the ECG signal.
II. DATABASE COLLECTION
For this paper, MIT-BIH Arrhythmia Database directory of
ECG signal from physionet is used 161 . Total 48 files of IO sec
recording each in raw unit are utilized out of which 22 are
females and 26 are males. This database is divided into two
parts, first 23 files (numbered from 100m to 124m with some
numbers missing) and second part of 25 files (numbered from
200m to 234m again with some numbers missing). The ECG
waveform from MIT-BIH Database contains - a te.'<t header
file, a binary file and a binary annotation file. The ECG
signals are stored in 212 format, \\bich means each one
www.ijsart.com
 N11 ti111111I Cmifcrc11cc 011 l:.'mcrgl11g 1'rrml.~l11 Sclc11cc and 'f'ec/1110/ogy
sampl e imposes number of leads timl)s 12 bcaLr; to be stored
and binary ftl e contains bMt ru1n otutions.
111.METIIODOLOGV
l11e raw ECG signal mny be contaminated with
nirious types of noise artifacts like baseline drift, power-line
interference, muscles noise, electrode contact und polarization
noise. In most of the ECG recordings the respiration and body
movements can give rise to baseline drift in the signal.
Tims, the common problem in ECG signal is the
elimination of baseline drift and suppression of other noise
artifacts.
A. Removal of baseline drift
The most common noise artifact responsible for
contaminating the ECG signal is the Baseline Drift. It
generally results from respiration, motion and changes in
electrode impedance and lies between 0.15Hz and 0.3Hz. It
can actually mask the necessary information from the ECG
and cause irregularities in beat morphology. In this paper, the
baseline drift of the ECG signal is eliminated by using the
moving average filter to achieve the smoothed signal. Then
the smoothed signal is subtracted from the original signal to
get rid of the baseline drift from the ECG signal. Hence the
signal thus obtained is free from baseline drift.
6 6 7 fJ U 10
Samplus
Figure 2: Raw ECG signal
Figure 3: ECG signal with baseline drift removal
B. Removal ofnoise artifacts
With the eradication of baseline drift the ECG signal
is free from DC offset but still contains some noise. To get rid
of this noise various schemes have been proposed earlier like
Page I 14
•
t
ISSN [ONLINE!: 2395-1052
digi tal filt c-r ing technique and usc of adaptive filters. In this
paper we try to implement band-pass filterin g approach for
noise rejecti on as it improves SNR and reduces influence of
these noise sources. Jt also permjts the usc of lower thresholds
that v.ouJd be possibl e on filtered ECG signal and thus reduces
the noise caused due to fal se positives that mimic the QRS
complex.
The approach involves linear as ~11 as nonlinear
filtering for noise rejection. Since the desired pass band 5-15
Hz cannot be achieved using bandpass filters directly for our
chosen sample rate, we use the cascade of low pass and high
pass filters to achieve 3dB pass band from 5-12 Hz.
• Bandpass Filtering:
It is the cascade of low pass filter and high pass filter.
The low pass filter eliminates 50 Hz power line noise and
electromyogram noise, having cutoff of 11 Hz, wi-th the gain
36 and a delay of 6 samples.
The difference equation for low pass filter is:
y(nT)=2y(nT-T)-y(nT-2T)+x(nT)-2x(nT-6T)+x(nT-12T). .. (l)
The high pass filter eliminates lower frequency components
which are motion artifacts, P wave and T wave.
The difference equation for high pass filter is:
y(nT)=32x(nT-l 6T)-[y(nT-T)+x(nT)-x(nT-3Zf)] ........ (2)
Band P- Filtered
500 550 600 650 700 150 800 SSC
Figure 4: bandpass filter output
• Derivative Operator:
The next step is differentiation. After filtering, the
signal is differentiated to provide QRS slope informatioo. Five
point derivative is been implemented as it approximates ideal
derivative over de and 30 Hz with delay of two samples.
The difference equation for derivative operation is:
y(nT)=(l /8T)[-x(nT-2T)-
2x(nT+T)+2x(nT+T)+x(nT+2T)] ..... .(3)
www.ijsart.com
 "'"';"""' O ,r,frrt't1cr n11 Bmcrglfl![ Trrmlt 111 Sde11cc anti 1eclmnlogy l.<JSN [ONLINE!: 2395--1052

C. Hl'ul11atlo11 Parameters before and (ljter jilterotfng

SN"R Cron Corttb tion Pu.k Jigrul to noist

Gl tlo
Bdou. Ma Befo,e Mer Befon After
Ehtrin FiJtmr:1 Fdtmn Filterin filumg Rtmn
Figure 5: Derivntivc operator output g g g g g

Squaring: . 0.1988d . 0.0066 2.43306d 7J 583d

0.959d B 0 052 1 86 B 8
After differentiation the signal is squared point by B 8
point This emphasizes the higher frequency components and
attenuates the lower frequency components.
JV. CONCLUSION
The equation for squaring operation is:
This paper presents a filtering approach to cope ap
with the noise artifacts affecting the ECG. Three evaluation
y(nT) = [~"(nT)f ...... .. .. .. .. .... .. .. .... ...... .. ... (4)
parameters were compared for their values before and after
filtering the ECG signal. The procedure of bandpass filtering,
finding the derivative, squaring and integrating was found to
significantly improve the signal in terms of the signal to noise
ratios with the improved SNR of 0.1988dB and peak SNR of
7.3583dB and cross correlation of0.006686.

1-<00 1500 1600 1700 1800 1900 2000 2100 2200

V. ACKNOWLEDGMENT

Figure 6: Squared output I express my sincere gratitude to my guide Miss Purti

• Integration:
The squared signal is then passed through moving
window integrator. It acts as a smoother and performs as a
moving window integrator, with sample rate of 200
samples/sec and window width of 30 samples (150ms).
Savardekar for the necessary guidance and support for the
preliminary study for this research \fflrk.
I am also thankful for the HO.D Dr.HG. Vuani of
Electronics and Telecommunication Department for his
valuable assistance and timely cooperatim. In addition, this
research -oork carried out in Goa College of Engineering
Farmagudi Goa.

The equation for moving average integration is: REFERENCES

y(n) = (x[n-(N-1)]) =x[n-(N-2)]+.. .. +x[N]/N .. ...... ....... (5) [I] Sazia Javed,Noor Atinah Ahmad, ''An adaptive noise
cancellation model for removal of noise from modelled
Where N is length of Moving Average Filter. ecg signal",Region 10 Symposium 2014, IEEE , April
2014.
[2] Hari Mohan Rai,Anurag Trivedi, "ECG signal
Moving Average tiller
Classification using Wavelet Trasform and
Backpropagation neural netoork",Computers and Devices
for communication (CODEC), 5th IEEE International
Conference, December 2012.
[3] M. Z. U. Rahman, R A. Shaik, et al., Adaptive noise
500 1()XJ 1500 2000 2500 3000 3500 removal in the ECG using the Block LMS algorithm,
IEEE proceeding of 2nd International Conference ICAST,
Figure 7: Integration output 2009.
[4] Miad Faezipour,Adnan Saeed,Hlaing Minn,lakshman

Page I 15 www.ijsart.com
Natio11al Cm,fere11ce 011 Emerging 1're11ds in Science am/ 1'eclmology ISSN [ONLINEJ: 2395-1052

Tamil , ''A Patient Adaptive profiling Scheme for ECG

Beat Classification",IBEE Trans. On Information
Technology in Biomedicine volume:14,
Issue:5,September 2010.
15] K. K. Makesheva, Y. A. Altay, "Noise-resistant
processing of electrocardiosignal in
MATLAB" , E11ropea11 Journal of Scientific Reseat'C, vol.
13, no. I, pp. 898-903, 2016.
[6] "The MIT-BIHA.rrhythmiaDatabase,
http://physi onet. ph.bi u.ac.il/physiobank/database/mitdh/
[7] Review of methods for processing ECG signal, 2017
[online] Available:
h11p:llopenbooks.ifmo.n1/mljile/4951/4951 .pdf
[8] K Amtul Salam, G. Srilakshmi , "An algorithm for ECG
analysis of arrhythmia detection", Electrical, Computers
and Communication Technologies (ICECCT) , IEEE
International Conference, March 2015
[9] Sani Saminu, Nalan Ozkurt, Lbrah.im Abdullahi Karaye,
"Wavelet feature extraction for ECG beat dassification",
Adaptive Science and Technology (ICAST), rEEE
International Conference, October 2014

Page l 16 www.ijsart.com
►

Volume XX I fssuc XX I Month XXXX (Y car)

IJSTE - lnlornolionnl Jounwl of' Scluncc Techno logy & IJ11ginccri11gl
ISSN : XXXX -XXXX

ECG Analysis And Classification Techniques - A Survey

First Author Second Author

Diksh a Santosh Gaonk ar Purli Savardekar

/\ssislant Prof Dept. of rncctronics and Teleco mmunication
Dept. of Electronics and Telecommunication Engineering engi neering
Goa College of Engineering Goa College of Engineering
Ponda-Goa, India Ponda -Goa, India
Email : dikshn gaon ker l 3(algmuil. com

Abstract
the electrical activity of heart muscles. Therefore ECG is
Electrocardiogram (ECG) provides a graphical representation of
s. But due to its non-sinusoidal nature, its analysis and
widely used by the physicians to diagnose various heart disease
ed signals processing techniques for ECG signal analysis are
interpretation is a challenging task. Many of the recently develop
accurate classification strategy. This paper provides a literature
mostly based on feature extraction algorithm in combination with
survey of such ECG analysis schemes.
or
rk (ANN); Support Vector Machine (SVM); k-nearest neighb
Keywords: Electrocardiogram (ECG); Aritificial neural netwro
II. INTRODUCTION
n of bio-potential versus time. It is considered to be the
Electrocardiogram (ECG) is a graphical record indicating the variatio
es of heart in clinical practice.
prominently available diagnostic tool to monitor electrical activiti wherein the number of
The normal ECG wavefonn consist s of P-QRS-T waves. It is obtained by electrodes placement method
141 cycle in an ECG signal. Variations in the time duration of these
electrodes might vary from 3-12 . Figure I show one cardiac mia 181 . In recent times, a number of
condition called arrhyth
waves and segments from their normal ranges indicate a pathological 121151161 .These techniques .,.
ed for analyzing ECG signal and thus leading to arrhythmia detection
techniques have been propos ""
s, SVM, Bayesian and time domain based methods with each
involve various approaches based on ANN, digital signal analysi signal.
ng and classifying the ECG
method having their own advantages and disadvantages for analyzi
re review of various ECG analysi s schemes proposed earlier for ECG classification, feature
This paper provides a literatu
ed as follows . Section 2 provides the overview of the work proposed
extraction and arrhythmia detection. Its remainder is section
3 gives the comparative table assessing the perfonnance of
earlier in literature regarding ECG analysis and classification. Section
4 concludes the paper with few remarks.
various techniques put forth by authors regarding the same, and Section
R -R ln!Arrval

Q s
P-Jl
Interval
Q ,T lat..-val T -Qh1tu >11l

Figure l: A sample ECG wave.

 111. LITERATURE SURVEY
us signnl processing odvanccd schemes have been proposed in recent ycnrs for accurate aml quick ECG 1 · Th 1.5
r.iurneroofthe poper focuses on various npproochcs put forth by the nulhors e.irlier in literature rdr ru,aJyzing the ~~cis~
s. aJ d
~ sign an
secuonno the curdinc di scnso. ,
dctrCll•oe
· dA T . d'. 121
Han Mohan Rru ~ nurog nve I m proposed the use of llnck propagati on neural network for classification of ECG
,,nrcfonll using dts~rcte wavelet lrn~i~form. llic whole m~lhodology was divided into lhrce basic parts Preprocessing, fea~ure
•\trnction and classificotton. De-noi.srng wus done by us111g Doubcchi es wavelet of order four to eliminate different noise
;truciures followed by usage of •~iovmg average filter fo r baseline wu11der removal. The total 64 number of features (48 DWf
16
bnsed features nnd morphological) features were provided us input to the BPNN classifier with twenty number of neurons in
the hidden layer and network of two neurons al the output layer which decided tlie target matrix as (0 I) and (I O) referring co
00 rnuil and abnormal . class. Tiie res~~ of this method reached tl1e accuracy
of 96% for normal ciass data ~d I 00% fo r
abnormal class data usmg BPNN classther.

A norel method to detect arrhythmia from ECG signal using diITerent concepts as discrete wavelet transform (DWf), adapti ve
1east mean square (ALMS) and Support vector machine (SVM) was put forth by K.Amtul Salam and G Srilakshrni in !3J. Their
proposed ~aper focused on auton1'.1ted ~T seg~en~ nnalysi~ '.111~ QRS detection of arrhythmia. De-noising of ECG signal was
done by uttltzmg the approach of 1t~rat1011 modtfymg to mmuruze the error using LMS algorithm, Then using the feast square
0 ptinnzatton method a pattern adaptwe ,.vavelet was generated which was used for discrete wavelet
transform analysis of ECG
signal. At fast w1th the use of SVM classifier boili morphological features and DWT based features best results were obtained.
111
RidhiSaini, Namita ~indal and ~uneet Bansal in described an approach for ECG signal classification by extracting features
from ongmal ECG signals and sixth level wavelet transformed ECG signals. This paper involves two methods to classify the
ECG signal. In first method, the statistical features were computed directly from raw ECG signals and then principal
component analysis (PCA) was applied before classification. In second method, the ECG signal was decomposed using wavelet
tranSform with Daubechies order four wavelet up to level six. Then wavelet coefficients at detailed level four were extracted
and statistical features of these coefficients were computed. K- Nearest neighbor (kNN) classifier provided the lower accuracy
in comparison with the first method and which is an improvement of31.25 %.
191
Sani Saminu, Nalan Ozkurt and Ibrahim Abdullahi Karaye in proposed a robust ECG feature extraction technique suitable for
mobile devices by e>.1racting only 200 samples between RR intervals as equivalent R-T interval. Preprocessing stage for removal
of noise structures was carried out along with the QRS complex detection using well known and acceptable Pan Tomkins
algorithm based on analysis of slope, amplitude and width of QRS complexes. Feature extraction part involved R-R interval
features without applying any operation or transform were used as benchmark. Further R-R interval decomposition was done to
achieve a time -frequency representation of the signal using DWT. At last statistical parameters of DWT coefficients were
calculated and used as hybrid feature for training and testing using neural network classifier.

Sean Shensheng Xu, Man-Wai Mak and Chi-Chung Cheung in !7J described a method to enhance the performance of heart
arrhythmia classification by selecting relevant features from ECG signal followed by application of dimension reduction on the
feature vectors. Feature extraction part involved the usage of Fisher discriminant ratio (FDR) to select relevant features and
Principal Component Analysis to reduce the dimension of feature vectors. For classifying heart arrhythmia two popular
classifiers Support vector machine (SVM) and Deep neural network (DNNs) were used. Results showed that for classifying
normal against abnormal heart arrhythmia , the best combination of feature pre-processing and classification was Fisher
discriminant ratio (FDR) and Deep neural network (DNN) while pre training DNN was an essential step for training DNN
classifiers, especially when number of training samples were very limited.
141
Maid Faezipour, Adnan Saeed, Suma Chandrika Bulusu, Mehrdad Nourani, Hlaing Minn and Lakshman Tamil together in
provided a mechanism to extract precise fiducial ECG points and implemented a novel local ECG beat classifier to profile each
patient's normal cardiac behavior as well. Denoising part included the use of wavelet detrend for Baseline wander removal
followed by usage of UWT sym5 wavelet for wide band noise suppression. For feature extraction QRS complexes were
accurately detected using wavelet peak and valley detector along with search-back algorithm and slope detection technique.
Classification was done using feature repetition profiling which included feature hash generation and hash memory counter. This
classification technique provided high abnormality classification accuracy.
161 presented a robust arrhythmia classifier based on the
Nyoke Goon Chia, Yuan Wen Hau, Mohd Najeb Jamaludi in
combination of wavelet transform and timing features, as well as support vector machine classification technique. The pre-
processing stage involved the usage of band-pass filters followed by ECG segmentation, normalization and removal of baseline
wander. Feature extraction was carried out using different combination of wavelet transform, statistical and tinting interval
feature. Finally support machine (SVM) classifier was used as a classifier.
151 described a method for arrhythmia detection based on deep learning
Hilmy Assodiky, Iwan Syarif and Tessy Badriyah in
algorithm. The paper was divided in three sections; data ex1raction, pre-processing and performance measurement. The data was
extracted from ECG signal. Pre-processing part involved data cleaning, data reduction and data transformation followed by the
 0
netics algorithm (GA) nnd pllrticl o swur111 op timi1.ntio11 (PSO) for fea ture selection. For classi ficati on part various
ll5 ge1 of ~en approaches bnsed on Nnrvo Bnyos, k-NN. ANN, deep Icorni ng o.nd SVM wllh leave-one-ou t cr-oss val1da1ion method
c1ossesfilucadt1od
ie . The best results were achieved using deep learning with PSO search fcuture selection
11'er

Tr'f 'O
omi Rod together ,,~th Bft ostol,Kjcrsti Engan , Unoi ltustn . Jnn Tcrje Kvaloy , Jo Kramer - Johan1en . Lars Wrk
; Ii Bnh~los K. Kn~uggelos in. ' dov?lopcd ECO bnsed nlgorilhm for O,e rotrospcclive and ,u1oma1ic clao,i ficaJion of
IJJld1 sc1ln 1 cnrdioc rhytJ1m. Their work is n comprehensive tcchni cnl descriplion of BCO based fi ve claas resusci1.1tion rhythm
~ggt·on · d · •
res ~. Featnre exlrnct1on was one using discrete wnvelet lro.nsfonn (DW1) follo\ved by feature selection scheme ba'led on a
c1assifi~ss-rnlidotion technique nnd multiple ex peri ments wi th different classification approaches .At last classifiers based on
nestt-.d_c decision theot}'. k-nenres1 neighbor, k-locnl hyper plmie distance nearest neighbor , ANN and ensemble of decision fte!!s
On!•esi~ied.
i1\'re s111 The best results were obtained for ANN classifier with Buycs ion regulari zation backpropogation training algorithm.

IV. RESULT

Authors Pre-processing technique

Modeling Scheme Accuracy
Ali Baharami Et al . Featttre extraction : DWT combined with 12 ANN classifier with Bayesian 78.5%
(2017) classification features for shock/no-shock regularization backpropoga:tion training
algorithm
Hilmy Assodik."y Et Pre-processing: data cleaning, data DNN 76.51%
al. (2017) reduction, data transformation.
Featttre selection: GA and PSO
Sean Shensheng Feature pre-processing: FDR and PCA DNN 82.96%
Xu Et al. (2017)

t
-- Nyoke Goon Chia
Et al. (2017)
Denoising: Band-pass filtering.
Feature extraction: Wavelet transforms.
SVM 87.93%

I
.....-- Ridhi Saini Et al. Feature extraction: Daubechies 4 wavelet up kNN 87.65%

'
I
(2015) to level 6 decomposition.
Data reduction: PCA

I K. Am.tu! Salam Et Denoising: LMS algorithm. SVM 98.65%

al. (2015) Feature extraction: QRS pattern adaptive ·

I
wavelet.

ANN 99.84%
Sani Saminu Et al. Feature ex tractioil'· DWT decomposition and
(2014) statistical features.
BPNN 97.8%
Hari Mohan Rai Et . . . Daubechies 4 wavelet up to level
Deno1smg.
al. (2012) 5 decomposition.
Feature ex traction·· morphological features
and wavelet coefficients based features.
Feature repetition profiling 97.42%
Denoising: Wavelet detrend UWT sym5
Maid Faezipour
wavelet.
(2010)
Feature extraction: wavelet peak valley
. se arch back algorithm, slope
detectwn,
detection. 1 ..J

-
 CONCLUSION
trical ncu·v1•1y of 11eart and th J b · h · I tool · drag ·
·des the reco rding of the elec . ·
ature vario us a J us ios ecom e t c_ v,ta '?
nosis of abnormal heart
01'1
have pr~p osed earl~ er in liter anal ysis and class ification
~cG pr AUtl1ors
15 · ed paper provides a detatled survey of , . I proaches and tronsformatron s for ECO ECO class ifica tion.
,lt1'1hn ues and algo rithm s proposed in literature.
ely pre- proc essin g fieatuvraeneo uts lel ~h11iq h
11is proP_ons,,olves steps nam . X roe 1011 and c/ nss,' (iicat,·on. r erefore the algorithm developed for ECG
,
' . .
· racy and finst execution .
i

· wil' I provide l11gh degr ee of accu • . , necessary for elim

time . Pre -proccsstng inating lhe noise
~3,11 c . • ,ust hi d . ' ,s
~ ,~is n
l)f1 ~· tl1at is gener
all ac eve by usmg various filters , ·1
y
stfll'' ~e authors for feature extra ction inclu de Disc
rithm
rete
etc.
;s
I

sc1:~f
:
811
1
/ffe
r~
ct the feature exl~action
1
requ ired for
~'.orm (~WT): Dt~crete Cosin~ T~ansform
class ification. Meth ods
(DCT), Princ ipal
Tom kins algo Neural Network(A NN)
11-(ed b~nent Anal ysis (PCA),_ Pan- hin t;v or class1ficat1on ,~vo!ves A~tfic,al
0
Sup port Vec tor Mac
coniPeJJJ'est Neighbor Clas sifier (kN N),
t bl ( M). ' Deep L~ammg Algo nthm Backpropagation Neural
prop osed pape r also men tione d the class ification techniques proposed
. ~-N 1
k (BPNN)etc. The a e companng the various ECG
Ne11or
ea.rf1er.
ACKNOWLEDGEMENT
fi th . ursory study for
5 my sincere thanks to my guide Miss Purti Savard ekar or e necessary guidance and support for the prec
I e\-pres rch work. · .
. resea · f El · n Department for his valuable
(b)s b Of gratitude is also owe.d to the H: O. ·D' Dr. · H· G· V'iraru O ectrorucs and Telecommumcat1o gineering Farmagudi Goa.
A de ed out in Goa College ofEn
t
nce and
.
nme ly coop eration. In addition, this research work carri
as~sta
REFERENCES
commmrication and
let transform and kNN classifier,Computing
al,Pu neet B ~ ~Clas sifica tion of heart diseases from ECG using wave
Ridhi Saini,Namita Bind .
Ill auwroarion (ICCCA), IEEE International conference, May 2015 gation netll'al network",Computers and
Devices for
di, "ECG s
_ ignal Class ificat ion using Wavelet Trasform and Backpropa
n/ f1ari Mohan Rai,A nurag ~rive mber 2012 .
5 IEEE International Conference, Dece Cardiac
i- coIDIDurucation (CODEC), ~ ECG base& classification of R=i tatio n
seu,Lars Wik,Aggelos K. Katsaggelos, ober 2017.
l,Kjer s_ti Enga n, Unai Irusta ,Jan Johan 0,Oct
/'I Ali BaharaIID Rad, T ryg~e Etesto l Engineering ,volume:64J ssue: l
., RbythmS for Rd:rospective Data Anal
ysis", IEEE Transaction on Biomedica ification"JEEE Trans. On Infonnation
,laksh man Tami l, "A Patient Adap tive profiling Scheme for ECG Beat Class
Miad FaezipoW',A~ Saeed ,Hlai ng Minn
Issue:5,September 20 I0.
141
Ta:hnology in B1omedicme , volume: 14, tion",Knowledge Creation and Intelligent
Computing (IES-KCIC)
Syai: if, Tessy b_adr iyah, Deep
~ learning algorithm for arrhythmia detec
/5/ HihnYAssodil')', Iwan
,September 2017. ine classificationn,Signal
' ,IEEE International Electrorucs Symposmm wavelet transform and support vector mach
Yuan Wen Hau, Moh d Najeb Jama ludin, "Robust arrhythmia classifier using . . . . .
/~ Nroke Goon Chia., ation al Collo quium,March 2017. a ans
IEEE Intern n" .Multmu:rli
I processing and its Applicatio n (CSP A ),
-Chu ng Cheu ng, "Deep neural networks versus suppo
rt vector machines for ECG arrhythmia classificabo
/7) Sean shensheng Xu,M an-W ai Mak, Chi
al Conference )uly 2017. Technologies (ICECCT), IEEE
Expo Workshops (ICMEW), IEEE Internation Electrical, Computers and Communication
, G. Srilak shm.i , ~An algor ithm for ECG analysis ofarrhythmia detection",
118/ KAmtnl Salam
lnlernational Conference, March 2015 ce and Technology (ICAS T), IEEE
ECG beat classification",Adaptive Scien
u,Nala n Ozkw 1,Ibra him Abdu llahi Karaye,"Wavelet teatun: e>.1ration for
1[9) Sani Samin
International Conference October 2014 .