DIAGNOSIS OF BRADYCARDIA ARRHYTHMIA USING MEMD AND CONVOLUTIONAL NEURAL NETWORKS

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 10 | Oct 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 968
DIAGNOSIS OF BRADYCARDIA ARRHYTHMIA USING MEMD AND
CONVOLUTIONAL NEURAL NETWORKS
CHARUGULLA PAVAN KUMAR 1, ALUGONDA RAJANI2
1M Tech Student, Department of Electronics and Communication Engineering UCEK (A),JNTU Kakinada, Andhra
Pradesh,India,533003.
2Assistant Professor,Department of Electronics and Communication Engineering UCEK(A),JNTU Kakinada, Andhra
Pradesh,India,533003.
------------------------------------------------------------------------***-------------------------------------------------------------------------
Abstract - Heartbeats are crucial to the medical sciences'
study of heart ailments because they reveal significant
details about heart problems and irregular heart rhythms.
Electrocardiogram (ECG) represents the electrical activity
of the heart showing the regular contraction and
relaxation of heart muscle. The heart condition is used to
diagnose by an important tool called Electrocardiography.
The ECG spectrogram is used for diagnosing the heart
diseases. The different types of noises present in ECG signal
are Base-Line Wander, Power-Line Interface, Muscle
Artefacts, Electrode contact noise. One of these is
arrhythmia, in which the heart's regular rhythm is altered
by damage to its muscles and an electrolyte imbalance. A
hybrid technique is utilised to identify and categorise
arrhythmia by combining Multivariate Empirical Mode
Decomposition (MEMD) and Artificial Neural Network
(ANN). Multilayer feed forward neural networks are
utilised for classification, and these networks are trained
utilising back propagation algorithms. Two key properties,
the RR interval and Heart Rate, are retrieved from the ECG
signal for the identification of Arrhythmia when MEMD is
employed to denoise multichannel signals. Tachycardia
and bradycardia are two subtypes of arrhythmia based on
these characteristics. The Extraction of features and
classification is to be done using Convolution neural
network (CNN) classifier and the results obtained using
CNN.
Key Words: Baseline Wander, Powerline Interface,
Muscle Artifacts, Arrhythmia, Tachycardia,
Electrocardiogram, Multivariate Empirical Mode
Decomposition (MEMD), Artificial Neural Network
(ANN),Convolution Neural Network(CNN).
1.INTRODUCTION
1.1 Electrocardiography (ECG) 1
An electrocardiogram (ECG or EKG), a recording of the
electrical activity of the heart, is made using the
electrocardiography technique. When the cardiac muscle
depolarizes and repolarizes throughout each cardiac
cycle, these electrodes detect the minute electrical
changes that result from these processes (heartbeat).
The typical ECG pattern is altered by a number of cardiac
conditions, including irregular heartbeat (like atrial
fibrillation and ventricular tachycardia), inadequate
coronary artery blood flow (like myocardial ischemia
and myocardial infarction), and electrolyte issues (such
as hypokalemia and hyperkalemia).Traditionally, The
term "ECG" has been used to refer to a 12-lead lying-
down ECG, as detailed below. Other tools, like a Holter
monitor, can record the electrical activity of the heart,
despite the fact that some smart watches models may
also record an ECG. ECG signals may be captured using
various equipment and in different settings. Ten
electrodes are placed on the patient's chest and
extremities as part of a standard 12-lead ECG. The
magnitude of the heart's total electrical potential is then
calculated and recorded over time utilising twelve
different angles (or "leads") (usually ten seconds). The
overall amount and direction of the heart's electrical
depolarization at each point in the cardiac cycle may
therefore be quantified. The three main components of
an ECG are the P wave, which denotes depolarization of
the atria, the QRS complex, which denotes depolarization
of the ventricles, and the T wave, which denotes
repolarization of the ventricles. The heart is the most
vital and crucial organ in the human body. The heart
controls a number of biological processes. The heart's
main job is to pump blood to various body parts, which is
the most important thing our body needs to do. Since the
heart emits electrical signals at very low voltages (on the
order of 60mV), which are required to evaluate and
confirm the operations of a healthy heart,
electrocardiogram (ECG) signals are used to record the
electrical activity of the human heart it shown in figure 1

Figure - 1: An ECG signal showing the most important
peaks
The table 1 it represents the Normal ECG signal wave
amplitude and durations.
Table - 1 Normal ECG signal wave amplitude and
durations
Features Amplitude(mV) Duration(sec)
P wave 0.25 0.06-0.08
Q wave 25% of R wave 0.09-0.1
R wave 1.60 0.08-0.12
T wave 0.1-0.5 0.12-0.16
U wave 0.05 0.1
1.2 ECG Spectrogram 2
A spectrogram is a graphic representation of a signal's
frequency spectrum as it evolves over time.
Spectrograms are sometimes referred to as sonographs,
voiceprints, or voicegrams when they are applied to an
audio input. Waterfall displays are what you might refer
to when the data is displayed in a 3D plot. Sonar, radar,
voice processing, seismology, linguistics, music, and
other disciplines frequently use spectrograms.
An ECG signal can also be transferred into a spectrogram.
The ECG spectrogram can be then applied with all of the
required processes. An ECG spectrogram can be
converted into any other domain for a better operations
over the process. The ECG spectrogram contains
information of the signal that the original signal does.
The information in a spectrogram can never deviates
from the original ECG signal captured from the patient.
The converting of a signal into its spectrogram will result
greatly in the vibration analysis. The way the
spectrograms work is, they makes easier for the
implementation of any processes.
1.3 Tachycardia 3
A heart rate that is higher than the typical resting rate is
referred to as tachycardia. Adults are generally
considered to have tachycardia if their resting heart rate
exceeds 100 beats per minute. Heart rates that are
higher than the resting rate might be healthy (like during
activity) or unhealthy (such as with electrical problems
within the heart). Age determines the highest limit of a
typical human resting heart rate. Different age groups
have reasonably well-standardized cutoff levels for
tachycardia; normally, deadlines are 1-2 days:
tachycardia >166 bpm after 3–6 days of tachycardia
>159 bpm.
1.4 Bradycardia 4
A slow rate of heart is known as Bradycardia. For adults
is from 60 to 100 times per minute while they are at rest.
Your heart beats less frequently than 60 times each
minute if you have bradycardia. If the heart doesn't
pump enough oxygen-rich blood to the body and the
pulse rate is exceedingly sluggish, bradycardia can be a
major issue. You might experience this and feel weak,
exhausted, and out of breath. Bradycardia can
occasionally occur without any symptoms or problems.
It's not necessarily dangerous to have a slow heartbeat.
For instance, a resting heart rate of 40 to 60 beats per
minute is typical for some people, especially healthy
young adults and trained athletes. If bradycardia is
severe, a pacemaker implant may be required to assist
the heart it shown table – 2.
Table - 2 Parameters of Arrhythmia
parameters Heart rate(in bpm)
Normal heart rate 60-90 bpm
Abnormal heart rate Less than 60 bpm or
greater than 90 bpm
Tachycardia Greater than 90 bpm
Bradycardia Less than 60 bpm
1.5 MultiVariate Empirical Mode Decomposition
(MEMD) 5
For the adaptive processing of multichannel data, the
multivariate empirical mode decomposition (MEMD) has
lately made significant advances. Despite MEMD's

excellent efficiency in time-frequency analysis of
nonlinear and non-stationary signals, its wider
applicability has been constrained by high computing
load and over-decomposition.
For breaking down non-linear and non-stationary signals
into a sequence of Intrinsic Mode Functions, Huang et al.
devised the multivariate empirical mode decomposition
(MEMD) in 1998. (IMFs). IMF records the signal's
repetitive activity at a specific time frame. The empirical
mode decomposition breaks down a time signal into a
collection of basis signals similarly to the Fourier or
wavelet transforms, however unlike those
transformations, the basis functions are obtained
directly from the data. As a result, the results maintain
the signal under consideration's complete non-
stationarity. The instantaneous frequency and amplitude
of the signal can be calculated when the Hilbert
transform is used on the IMFs. This method is known as
the Hilbert-Huang transform (HHT).
Multiscale non-linear, non-stationary signals are broken
down into a number of adaptive, entirely data-driven
AMFM zero mean signals, known as Intrinsic Mode
Functions (IMF). This process is known as mutlivariate
empirical mode decomposition (MEMD). The
fundamental premise of EMD is that any signal is made
up of several IMFs, each of which represents an
embedded distinctive oscillation on a distinct time scale.
2 Convolutional Neural Network (CNN) 1
In deep learning, a Convolutional Neural Network (CNN)
is a class of artificial neural network, most commonly
applied to analyze 1D signals like all of the Bio-medical
signals such as, ECG, EMG, EEG and voice and speech
signals and 2D as well as 3D signals such as, visual
imagery. Convolutional neural networks are
distinguished from other neural networks by their
superior performance with bio-medical signals, image,
speech, or audio signal inputs. The three basic types of
layers in a CNN are convolutional, pooling, and fully
connected (FC). The convolutional layer, along with
convolutional layers or pooling layers, is the first layer of
a convolutional network, while the fully-connected layer
is the last layer and it shown in figure 2. The CNN
becomes more complicated with each layer, detecting
more chunks of the signals as well as the picture. The
CNN begins to detect greater features or forms of the
item as it advances through the layers, eventually
identifying the desired object.
2.1 Convolutional layer 2
The convolutional layer, which makes up the majority of
the computation in a CNN, is its core component. Input
data, a filter, and a feature map are among the things it
requires. Assume that a signal with a 1D matrix of values
will be used as the input. In order to detect if the feature
is present, we also have a kernel or filter that traverses
the signal's receptive fields. This procedure is known as
convolution. The quantity of filters affects the output's
depth.
As an illustration, three separate filters might produce
three different feature maps, providing a depth of three.
The kernel's traversal of the input matrix is measured by
its stride. A longer stride yields a lower output
notwithstanding the rarity of stride values of two or
higher. Usually, zero-padding is used when the filters
don't fit the input signal. By setting any elements that
aren't a part of the input matrix to zero, this produces an
output that is bigger or more evenly proportioned. Three
types of padding are available: These terms include valid
padding and no padding. The last convolution is
discarded in this case if the dimensions do not match.
The input layer and the output layer are made to be the
same size using the same padding. Full padding
increases the output size by adding zeros to the
boundary input. Each convolution operation in a CNN is
followed by a Rectified Linear Unit (ReLU) modification
on the feature map, which gives the model more
nonlinearity.
2.2 Pooling Layer 3
Down sampling, sometimes referred to as pooling layers,
reduces the dimensionality of the input by lowering the
number of parameters. The pooling operation sweeps a
filter across the whole input, much like the convolutional
layer, with the exception that this filter has no weights.
As a substitute, the kernel fills the output array by
applying an aggregation function to the values in the
receptive field. Pooling may be broken down into two
categories: When the filter passes over the input, max
pooling chooses the input pixel with the greatest value to
transmit to the output array. As a side note, this method
is applied more frequently than traditional pooling.
Average pooling is used to get the average value within
the receptive field. This value is provided to the output
array when the filter traverses the input. Many pieces of
information are lost due to the pooling layer, but the
CNN gains a number of advantages as a result. They
improve effectiveness, decrease complexity, and lower
the risk of over-fitting.
2.3 Fully-Connected Layer 4
The full-connected layer is what its name suggests it to
be. As was previously stated, partly linked layers do not
directly link the input image's pixel values to the output
layer. In contrast, every node in the output layer of the
fully-connected layer is directly linked to every node in
the layer above it. The characteristics that were collected

from the levels above and their corresponding filters are
used in this layer to carry out the classification process.
FC layers frequently produce a probability between 0
and 1 using a softmax activation function to classify
inputs properly. ReLu functions are commonly employed
in convolutional and pooling layers.
Figure - 2: Convolutional Neural Network Architecture.
3 Proposed Method
The raw ECG signal is taken from the dataset that
contains multiple samples of ECGs of different people
having different types of health conditions that maybe,
arrhythmia that maybe, tachycardia or bradycardia and
maybe perfect rhythms of heartbeats. The taken raw ECG
signal maybe affected from noise that maybe any of the
commonly occurring noises, such as, baseline wander,
power line interface and muscle artifacts.
The ECG sample must be noise free or it should be
minimized to a desired level for a better results, for that,
we have applied multivariate empirical mode
decomposition (MEMD). Multi-Variate Empirical Mode
Decomposition (MEMD) will represents the signal in
Intrinsic Mode Functions (IMFs). We can remove the
noise from the signal by decomposing the signal in its
IMFs.
The noise free ECG signal will then be converted into
spectrograms of containing same data with easy
representation. The spectrogram is a de-noised sample
means, the sample will be ready for extracting the P or Q
or R or S or T waves. The denoising of the signal makes
the peaks to be visible directly for detecting. The
spectrogram will then be transferred to the next stage
where, the classification of the different kinds of
arrhythmias are done using the convolutional neural
networks architecture.
The noise removed ECG spectrogram will then, be,
transferred to, the Convolutional Neural Networks
(CNN) for classification of different kinds of health
conditions. The dataset will be divided into training and
testing for process. The CNN will be fed with the training
samples which contains majority of the samples of the
dataset.
The CNN will be trained using the training samples over
many iterations for a better performance. The training
samples contains all kinds of health conditions. The
different kinds of health conditions are fed to the CNN
with labels. The data provided to the CNN training phase
will be labelled such that, it will learn the patterns of the
different kinds of the health conditions.
Figure - 3: proposed method block diagram.
The trained CNN network will then, be, tested using the
test data that was separated during the previous phase
of the process. The CNN is classifying the data at better
rate than any other existing methods such as, ANN.
Comparison of various parameters of CNN with the
existing ANN. It is evident that, our proposed CNN has
performed better than existing ANN in various
parameters.
4 Physionet Database
The physionet database contains datasets of different
types of heart conditions. That maybe, arrhythmias of
any kind like, tachycardia or bradycardia etc. The
arrhythmia condition dataset has been taken from the
physionet database. The different conditions are normal
or arrhythmia and arrhythmia are also of two kinds like,
tachycardia as well as, bradycardia etc. Physionet often
provides the samples of patients of different conditions
that might be, healthy or unhealthy. Not only heart
related condition signals but all types of signals that
include Electro-Myography (EMG), Electro-
Encephalography (EEG) and many more such as, X-Ray
samples, CT scan samples and even MRI scan samples to
everyone for free.

5 ResNet 101 Layers
Let's now describe this block with a recurring name. A
ResNet is made up of multiple blocks, one for each layer.
This is due to the fact that ResNets typically increase the
number of operations within a block to go deeper, while
the total number of layers—four—remains constant.
With the exception of the final operation in a block, it
shown in figure 4 which lacks the ReLU, an operation in
this context refers to a convolution, batch normalisation,
and ReLU activation to an input.
Figure- 4: ResNet 101 Network
As a result, the PyTorch implementation makes a
distinction between blocks with 2 operations, known as
Basic Blocks, and blocks with 3 operations, known as
Bottleneck Blocks. Although we are already using the
term layer for a group of blocks, each of these processes
is typically referred to as a layer.
We can confirm that the kernel size is [3x3, 64] and the
output size is [56x56] by looking at the table from the
paper again. We can observe that, as we previously
indicated, the volume's size remains constant within a
block. This is due to the use of a padding of 1 and a stride
of 1. Let's see how this applies to the 2 [3x3, 64] that is
shown in the table as a whole block.
6 Simulation Results
The raw ECG signal which are corrupted by different
kinds of noises are taken from the dataset of ECG signals
from the Physionet Database and represented in the
below waveform. The raw ECG signal will always be
corrupted because of the many reasons, so, it has be de-
noised for that reasons. To make sure it doesn’t affect
our process.
Figure- 5: Raw ECG Signal
The application of EMD to a raw ECG signal will
decompose the signal into its IMFs. The IMFs are used to
do any operation to the ECG signal.
Figure- 6: Decomposition into IMFs after application of
EMD
The application of MEMD to a raw ECG signal will
decompose the signal into its IMFs. Unlike EMD, an
MEMD will be applied to many parameters of the raw
ECG signal. The IMFs are used to remove noise from the
raw ECG signal

Figure- 7: Decomposition into IMFs after application of
MEMD
The R peaks are detected after the application of MEMD
to the raw ECG signal which was decomposed into its
IMFs.
Figure- 8: Detected R peaks
The filtered ECG signal was plotted in the below
waveform. After the application of MEMD to the raw ECG
signal the noise will be filtered out and the signal will be
filtered in many parameters.
Figure- 9: Filtered ECG signal
Figure- 10: Confusion Matrix.
Confusion Matrix: Confusion Matrix is used to calculate
the Performance of the Classifier, In the Confusion
Matrix it is represented in the matrix form. It Compares
the True Label i.e.;Correctly Predicted to the Actually
Predicted Values it shown if figure 10. Confusion Matrix

is of a NXN Matrix here N represents the number of
Classes considered for Classification.
Figure- 11: Detected the Diseases.
The normal rhythm spectrogram representing the
normality of the particular person who has been
classified as normal it shown in figure 12.
range of the particular person who has been classified as
bradycardia it shown in figure 13.
Figure- 13: Bradycardia rhythm spectrogram
a. Accuracy : It is the number of correctly
classified cases divided by total number of
instances
Accuracy =
b. Sensitivity :It is the probability of True
Positives in the Class.
Sensitivity =
c. Specificity:It is the probability of True
Negatives in the Class.
Specificity =
d. Precision:It is defined as how accurately
correctly predicted to the total positive
predictions.
Precision = .
Figure- 12: Normal rhythm spectrogram
The bradycardia rhythm spectrogram representing the
7. Parametric Evaluation:

Table -3 Comparison of CNN with existing ANN
It is evident that, our proposed CNN has performed
better than existing ANN in various parameters.
8 Conclusion
Finally, we can conclude that, the application of MEMD
removed the noise from many parameters which made
the PQRST waves to be easily detected by the
Convolutional Neural Network (CNN). The CNN needs
true peaks as clear as possible, for a better training over
the different kinds of health conditions associated with
the Human Heart. The CNN has a better accuracy and
sensitivity than the existing methods such as, Artificial
Neural Networks (ANN).
9 Discussion
The Convolutional Neural Network (CNN) worked better
than any other existing methods or techniques. The CNN
achieved a greater accuracy than the artificial neural
networks. The conversion of an ECG signal into
spectrograms made our signal easy for processing. The
Multi-Variate Empirical Mode Decomposition (MEMD) is
very great at reducing the noise in the spectrograms of
ECG signals. MEMD worked better than any other
existing methods or techniques.
REFERENCES
[1] Srinivasan, Neil T, and Richard J Schilling. ‘‘Sudden
Cardiac Death and Arrhythmias.” Arrhythmia &
electrophysiology review, vol. 7, no.2, 2018.
[2] Sanamdikar, S.T., Hamde, S.T. and Asutkar, V.G.,
(2015, Jun). ‘‘A literature review on arrhythmia analysis
of ECG signal,” International Research Journal of
Engineering and Technology (IRJET), vol.2, no.3, June
2015.
[3] “A Human ECG Identification System Based on
Ensemble Empirical Mode Decomposition”, National
Natural Science Foundation of China, Journal List
Sensors (Basel) v.13(5); 2013 May, PMC3690084.
[4] A. Agrawal and D. H. Gawali, ‘‘Comparative study of
ECG feature extraction methods,” 2nd IEEE International
Conference on Recent Trends in Electronics, Information
& Communication Technology (RTEICT), pp. 2021-2025,
Bangalore, India, 19-20 May, 2017.
[5] Reddy, K.G., Vijaya, D.P. and Suhasini, S., ‘‘ECG Signal
Characterization and Correlation to Heart
Abnormalities,” International Research Journal of
Engineering and Technology (IRJET), vol.4, no.5, May
2017.
[6] T. T. Khan, N. Sultana, R. B. Reza and R. Mostafa, ‘‘ECG
feature extraction in temporal domain and detection of
various heart conditions,” International Conference on
Electrical Engineering and Information Communication
Technology (ICEEICT), pp. 1-6, Savar, Bangladesh, 21-23
May, 2015.
PARAMETERS ANN CNN
ACCURACY 89.583333 93.750000
SENSITIVITY 89.189189 94.927536
SPECIFICITY 97.058824 96.323529

DIAGNOSIS OF BRADYCARDIA ARRHYTHMIA USING MEMD AND CONVOLUTIONAL NEURAL NETWORKS

More Related Content

Similar to DIAGNOSIS OF BRADYCARDIA ARRHYTHMIA USING MEMD AND CONVOLUTIONAL NEURAL NETWORKS (20)

More from IRJET Journal (20)

Recently uploaded (20)

DIAGNOSIS OF BRADYCARDIA ARRHYTHMIA USING MEMD AND CONVOLUTIONAL NEURAL NETWORKS