Biomedical Signal Processing and Control 52 (2019) 128–140

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

Deep Learning Approach to Cardiovascular Disease Classification

Employing Modified ECG Signal from Empirical Mode Decomposition
Nahian Ibn Hasan ∗ , Arnab Bhattacharjee
Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology (BUET), Dhaka 1000, Bangladesh

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

Article history: Multiple cardiovascular disease classification from Electrocardiogram (ECG) signal is necessary for effi-
Received 3 November 2018 cient and fast remedial treatment of the patient. This paper presents a method to classify multiple heart
Received in revised form 23 March 2019 diseases using one dimensional deep convolutional neural network (CNN) where a modified ECG signal
Accepted 6 April 2019
is given as an input signal to the network. Each ECG signal is first decomposed through Empirical Mode
Decomposition (EMD) and higher order Intrinsic Mode Functions (IMFs) are combined to form a mod-
Keywords:
ified ECG signal. It is believed that the use of EMD would provide a broader range of information and
EMD
can provide denoising performance. This processed signal is fed into the CNN architecture that classifies
1-D CNN
Modified ECG
the record according to cardiovascular diseases using softmax regressor at the end of the network. It
Heart disease classification is observed that the CNN architecture learns the inherent features of the modified ECG signal better in
Deep learning comparison with the raw ECG signal. The method is applied on three publicly available ECG databases
Denoising ECG and it is found to be superior to other approaches in terms of classification accuracy. In MIT-BIH, St.-
IMF Petersberg, PTB databases the proposed method achieves maximum accuracy of 97.70%, 99.71%, and
Cardiovascular 98.24%, respectively.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction neural networks or general machine learning algorithms have been

successful in recent years. ECG signal features such as entropy cal-
Cardiovascular Diseases (CVDs) have become the cause of death culation, standard deviation of beat to beat interval, minimum and
of millions of people around the globe in recent years. The common maximum of entropy, wavelet analysis [1], R-peak amplitudes, R–R
CVDs include heart attack, ischemic stroke, hemorrhagic stroke, interval time, QRS duration time, ventricular activation time (VAT)
heart failure, different types of arrhythmia and heart valve prob- [3], dual-tree complex wavelet transformation (DTCWT) [4], etc.
lems etc. Almost all of the abnormal conditions of heart due to these have provided a boost to the ECG based heart disease detection
diseases, many sinus arrhythmias and also the irregularities in the and classification. When an ECG signal is recorded using conven-
nervous system can be detected from the electrocardiogram sig- tional 12 lead ECG recording or any other methods, some noise
nal of a patient. The electrocardiogram (ECG) signal is a record of is added due to signal recording equipment or different lead sig-
the electrical activity of the heart. The standard 12 lead method is nal combination algorithms. As a result some preprocessing task
used frequently for acquiring ECG signal. Time domain, frequency such as outlier rejection, amplitude normalization, FIR filter, slope
domain or other signal processing techniques are used to reveal the sum function (SSF), signal quality index (SQI) [1], bandpass filter-
inherent features of an ECG signal for the purpose of diagnosis of ing [2], median filters [5], segmentation [6], denoising, re-sampling
cardiovascular disease of the patient. and normalizing, R-wave detection [7] have seen a great success in
Deep learning based approach towards heart disease classifica- improving the performance of ECG signal analysis. ECG denoising
tion has been a popular choice recently. Atrial Fibrillation detection, is one of the major tasks before working with this signal. Several
ECG arrhythmia detection [1], myocardial infarction detection [2], techniques like Bayesian filtering [8], Kalman filter method [9],
premature ventricular contraction beat detection [3] and ECG dis- denoising in wavelet domains [10–12] have seen successful efforts
ease classification comparing with the healthy patients using deep in suppressing unexpected signal artifacts in the ECG signal. Learn-
ing features based on machine learning algorithms and CNNs have
added an extra boost to the literature and successful ECG signal
∗ Corresponding author. analysis. Feature extraction using CNN architectures, Boltzmann
E-mail addresses: [email protected] (N.I. Hasan), machine [5], deep belief networks with RBM based algorithms
[email protected] (A. Bhattacharjee).

https://doi.org/10.1016/j.bspc.2019.04.005
1746-8094/© 2019 Elsevier Ltd. All rights reserved.
 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140 129

Fig. 1. An illustration of multiple heart disease classification based on the ECG signal. After EMD analysis, the first three higher order IMF signals are recombined to form a
modified ECG signal which is fed into a 1-D CNN model. The CNN network classifies the patient by learning the inherent features of the modified ECG signal.

[6], stacked sparse auto-encoder (SAE), multi-scale deep feature The mean of the two signals (M = (xu (t) + xl (t))/2) is subtracted from
learning (MDFL, Softmax-based multiple class classifier) [2], deep the original signal (x(t)) to obtain a first proto IMF signal G1 (t),
dictionary learning (DDL) [13], K-nearest neighbor (KNN), SVM and
G1 (t) = x(t) − M(t) (1)
Multi-Layer Perceptron (MLP) networks [14], recurrent neural net-
works, density-based clustering technique [4], ensemble methods This sifting process is applied on the proto IMF signal G1 (t) until
like incremental bagging, incremental boosting [14], these meth- a threshold point is reached where the conditions for an IMF are
ods have aided in the effective analysis and classification of ECG fulfilled [16]. In this paper the maximum limit for the number of
signals. Besides, in recent times, the usage of one dimensional CNN sifting iterative process is set to 3000. The threshold values for stop-
has provided some interesting results for time series data [15]. Mul- ping criterion of this sifting iterative method, corresponding to  1
tiple arrhythmias classification using ECG signal is still a challenge [18] is 0.05,  2 [18] is 0.5 and the tolerance for stopping criterion
as the ECG signals corresponding to many arrhythmias show simi- corresponding to ˛ [18] is 0.05. After this operation first IMF signal
lar features which require complex techniques for classifying using (I1 ) is obtained. The residue signal (R1 (t)) is,
traditional signal processing methods. In this paper, we introduce
deep learning based multiple heart disease classification method R1 (t) = x(t) − I1 (t) (2)
where the features are selected from Empirical Mode Decomposi- The sifting process is again applied on the residue signal (R1 (t))
tion of ECG signal. The features from EMD analysis are trained in a for second, third and etc IMFs. Generally,
one dimensional CNN model to learn the inherent features corre-
sponding to each disease. The method is applied to three publicly Ri−1 (t) − Ii (t) = Ri (t); i = 1, 2, 3, 4, . . ., N (3)
available databases with various training methodology. In this paper, the total number of IMFs is N = 6. Finally a residue
signal (RN (t)) and 6 IMFs (I1 (t), I2 (t), . . ., I6 (t)) are obtained such that,

2. Proposed Method

N
x(t) = Ii (t) + RN (t) (4)
i=1
In this section, first, we present a brief overview of our proposed
method of classifying multiple heart diseases from ECG signal and here, N = 6, and Ii (t) is the ith order IMF. In the above method of EMD
in the consecutive sections we present the details of every step. A analysis, lower order IMFs are fast oscillatory modes and higher
flow chart of the proposed method is presented in Fig. 1. The ECG order IMFs are slower oscillatory modes [16]. Fig. 2 shows the IMF
signal is processed with Empirical Mode Decomposition (EMD). In signals and the residual signal of an ECG signal from two databases.
this paper, instead of working with the raw ECG signal, a modi-
fied ECG signal is formed by summing the first three Intrinsic Mode 2.2. Formation of Modified ECG
Functions (IMFs). Although CNN networks have been very success-
ful in recent times for two dimensional or multidimensional signals, The raw ECG signal is empirically mode decomposed and first
we used one dimensional CNN model for heart disease classification three IMF signals are summed to construct a modified ECG signal
based on the modified ECG signal. shown in Fig. 3. Combination of first three IMFs after EMD analysis
removes most of the noise and artifacts and constructs a denoised
ECG signal [16,19,20].
2.1. Empirical Mode Decomposition (EMD) The modified and denoised ECG signal (x) is defined as,

xmod = I1 (t) + I2 (t) + I3 (t) (5)

EMD decomposes a signal in time domain into separate Intrinsic
Mode Functions (IMFs) [16,17]. The decomposed signals are of the here, I1 (t), I2 (t) and I3 (t) are the first three IMFs. In the raw ECG
same length and each IMF has an equal number of extrema and zero signal, the P, Q, S, T peaks of the ECG are not clearly separable due
crossings with its envelopes serving as a simple oscillatory mode. to noise. But, the modified ECG signal can be explored easily to
Even though the IMFs are not necessarily orthogonal, they are suf- separate Q, R, S peaks as shown in Fig. 3. The lower order IMFs
ficient to describe the signal. EMD is most useful for natural signals thus contain noise information. Hence, summing only the first three
(i.e. ECG) which are non-stationary and non-linear. Unlike Fourier IMFs removes most of the noise artifacts from the raw ECG signal.
Transforms and wavelet decomposition, EMD provides intrinsic The noise between two ECG cycles has also been minimized as illus-
features of a signal using the IMFs. A signal can be denoised in the trated in Fig. 3 for two different ECG signals from two databases.
EMD time domain by discarding lower-order IMF signals on the Moreover, it has been reported in [10] that the combination of
basis of the assumption that noise and signal are unrelated in fre- the first three IMFs are better suited for denoising performance.
quency bands [16]. Say, from a signal x(t), the upper envelope curve Furthermore, the denoising performance of the modified ECG is
(xu (t)) and the lower envelope curve (xl (t)) are obtained by joining investigated in this paper following the algorithm reported in [21].
the maxima and minima points of the original signal, respectively. Section 3.3 illustrates the detailed analysis.
130 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140

Fig. 2. Empirical Mode Decomposition (EMD) of an ECG signal from database [24] in (a) and from [22] in (b). The topmost plot of each sub-figure is the original raw signal
followed by six IMF signals and the residual portion after EMD.

2.3. Convolutional Neural Network Architecture fully connected layers. The final output of the network has a soft-
max regressor with a specific number of classes which vary among
The complete network architecture for modified ECG signal clas- different databases.
sification is illustrated in Fig. 4. The Convolutional Neural Network The first convolutional layer is fed by a 1000 × 1 sequence and
(CNN) architecture is designed to handle one-dimensional data the layer modifies the input sequence with 96 kernels of size 11 × 1
and all the convolution operations in the convolutional layers are with a stride of 4. Consequently, a feature space of size 248 × 96
performed on the 1-D sequence. The kernel size in each layer is is produced. Then a Maxpooling layer converts the feature vector
modified to be applied on the 1-D sequence. The first five layers space to 124 × 96 space. The second convolutional layer converts
of the network are convolutional layers and are followed by three the feature space to a shape of 124 × 256 with 256 kernels of shape
 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140 131

Fig. 3. Modified ECG signal is formed by summing the first three IMF signals after EMD. The peak amplitudes of a typical ECG signal are better detectable in the modified
ECG signal from [24] in (a) and from [22] in (b).

5 × 1 and a stride of 1. The third, fourth and fifth convolutional lay- 30 × 256 and then flattened to feed to the first fully connected layer
ers convert the respective input feature vector space to a shape of containing 4096 neurons. The second fully connected layer does
61 × 256, 61 × 384, 61 × 256 respectively. The corresponding ker- also contain 4096 neurons. Finally a ‘softmax’ activation function
nel numbers, kernel shapes, and strides are shown elaborately in classifies the signal to desired classes.
Fig. 4. The third, fourth and fifth convolutional layers do not have The convolutional layers use bias vectors and initialized with
any intermediate pooling layers or normalization layers in between zero values. The kernels are initialized uniformly. No kernel regu-
them. After the first and second convolutional layers, batch normal- larizer, bias regularizer or activity regularizer are used in separate
ization layers are introduced. Throughout the architecture, ReLU is convolutional layers. Additionally, the kernels and biases have no
acting as the non-linear activation function. The final pooling layer constraint function applied to them. The batch normalization lay-
converts the feature space of last convolutional layer into a shape of ers have a momentum of moving mean and variance equal to 0.99.
132 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140

Fig. 4. Illustration of the proposed neural network architecture with a total of 5 convolutional layers, 3 fully connected layers, total trainable parameters of 49,434,376. The
input of the network is 1000 × 1 dimensional. The number of neurons in the subsequent layers are 23808-31744-23424-15616-4096-4096-Number of classes.

Table 1
Train and test data-sets specifying the number of sequences per class.

Database PTB diagnostic ECG database (7 classes) [22] MIT-BIH arrhythmia database (4 classes) [23] St.-Petersburg arrhythmia database (9 classes) [24]

Training Bundle branch block = 26,980 Sinus bradycardia = 22,702 Healthy = 20,225
Valvular heart disease = 26,980 Normal sinus rhythm = 19,901 Acute MI = 18,648
Myocarditis = 26,980 Paced rhythm = 22,676 AV nodal block = 18,642
Healthy control = 26,980 Atrial fibrillation = 15,287 Coronary artery disease = 18,564
Dysrhythmia = 26,980 Transient ischemic attack = 18,543
Myocardial infarction = 26,980 Earlier MI = 18,531
Cardiomyopathy = 26,980 WPW = 18,468
Sinus node dysfunction = 18,466
Atrial fibrillation = 18,466

Testing Bundle branch block = 11,485 Sinus bradycardia = 9730 Healthy = 8650
Valvular heart disease = 12,532 Normal sinus rhythm = 6082 Acute MI = 7986
Myocarditis = 12,805 Paced rhythm = 4471 AV nodal block = 7989
Healthy control = 12,454 Atrial fibrillation = 2564 Coronary artery disease = 7956
Dysrhythmia = 12,343 Transient ischemic attack = 7945
Myocardial infarction = 11,993 Earlier MI = 7938
Cardiomyopathy = 12,799 WPW = 7912
Sinus node dysfunction = 7912
Atrial fibrillation = 7912

Besides, the beta weights and moving means are initialized to zeros training approaches, comparison among the training methods and
while gamma weights and moving variances are initialized to ones. the learning procedures are described in detail.
No beta regularizer or gamma regularizer are used separately in
each batch normalization layer separately. Also, beta and gamma
weights are free of any kind of constraints. Moreover, the weights 3.1. Database
are scaled by the gamma initializer.
The slope of the negative part of activation function ReLU is set to The proposed method is evaluated on ‘The PTB Diagnos-
zero and the threshold value for threshold activation is set to zero tic ECG Database’ [22], ‘MIT-BIH Arrhythmia Database’ [23],
as well. The activation function returns element-wise maximum ‘St.-Petersburg Institute of Cardiological Technique’s 12-Lead
value compared to zero value, i.e. (maximum(x, 0); here x = output Arrhythmia Database’ [24] in the PhysioNet [25]. Each database has
data from the previous layer). The pooling layers have different a different number of classes and data recording criteria. The PTB
stride number and pooling size in different layers specified in Fig. 4. database contains 549 records digitized at a frequency of 1000 Hz.
The dense layers use bias vectors and initialized to zeros. Also, the The MIT-BIH database contains 48 records, each of which is 30 min-
kernels are initialized uniformly. No regularizer (i.e., kernel regular- utes long and digitized at a frequency of 36 Hz. The St.-Petersberg
izer, bias regularizer or activation regularizer) is used in any densely database contains 75 annotated recordings sampled at 257 Hz. A
connected layer. Besides, the biases and kernels are free from any comprehensive database description is shown in Table 1 along with
kind of constraint function. The final two densely connected layers the considered disease labels and the corresponding number of seg-
are followed by dropout layers with a dropout percentage of equal mented data. The CNN architecture requires a sequence of shape
to 50% of the incoming nodes from the previous layer. 1000 × 1 as input to the first layer. Therefore, a sequence of 1000
data points is sampled from the modified ECG signal for feeding
into the neural network model. There is overlapping between two
3. Results and Discussion consecutive sequences for preserving inter-sequence and intra-
sequence features. The overlapping is done in a random manner
In this section the database used in the experiment, denoising in different records. This fulfills two purposes. First, the data aug-
performance of the modified ECG, classification results of different mentation task is completed. Second, the training procedure does
 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140 133

Fig. 5. Average Pearson’s Correlation Coefficient of raw ECG signal and proposed modified ECG signal for each disease class in (a) MIT-BIH database [23], (b) St.-Petersburg
database [24], (c) PTB database [22]. The coefficient value is higher for modified ECG signal than that of raw ECG signal signifying better signal quality and regularity throughout
the database.

not become dependent on a specific overlapped sampling method tional layers. The neuron biases are initialized with constant zero.
throughout the entire database. A single segmented sequence con- The learning rate is same for all the layers. The model is run on the
tains four to five periodic rhythms (depending on the sampling rate TESLA K40 Graphics Processing Unit (GPU).
of the database) so that any sequence fed into the network can learn Fig. 7 shows the data validation accuracy improvement over
both periodic and inter periodic features. the training process using different signals as input to the CNN.
The usage of the learning rate scheduler makes the training time
3.2. Learning Methodology much shorter than that without a scheduler. Too low learning rate
makes the training process slower. Hence, initially, the learning
The model is trained using stochastic gradient descent (SGD) rate is 0.01 and gradually decreased according to the scheduler
[26] algorithm with a batch size of 32 sequences, momentum of 0.9. as the model approaches near the optimized position following
The SGD is followed to optimize the neural network weights during the gradient descent algorithm. The model trained with the pro-
training. The weights are updated after each sample passes through posed modified ECG learns the features of interest faster compared
the network during the training rather than changing the gradient to other signals such as the raw ECG or individual IMF. The time
based on all of the available training samples. The algorithm has needed to reach the highest accuracy is the shortest for proposed
been reported in [26]. modified ECG.
Learning rate scheduler is applied for faster learning the features
of interest. In the scheduler, the initial learning rate is set to 0.01 3.3. Denoising Performance of Modified ECG
and dropped by 50% from its previous value with a step of 5 epochs.
Fig. 6 shows the details of this scheduling scheme. The final learn- The template matching algorithm using Pearson’s correlation
ing rate, at which the training accuracy is almost at its peak point, is coefficient, reported in [21], is utilized to investigate the denoising
approximately 0.0007 which is achieved at near 20th epoch. At this capability of the modified ECG signal. According to [21], a denoised
epoch, the validation accuracy is almost saturated as illustrated in ECG signal has higher correlation coefficient than that of a noisy
Fig. 7. The kernels are initialized uniformly in all of the convolu- ECG signal. The algorithm is applied to both raw and modified ECG.
134 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140

Table 2
Comparison of classification accuracy of the proposed method in Fig. 1, raw ECG signal training scheme and individual IMF training scheme.

Training methodology PTB diagnostic ECG MIT-BIH arrhythmia St.-Petersburg arrhythmia

database(7 classes) (%) database (4 classes) (%) database (9 classes) (%)

Raw ECG signal 95.03 97.08 99.29

IMF 1 83.08 96.35 98.29
IMF 2 90.01 91.97 94.70
IMF 3 81.46 88.88 91.00
IMF 4 64.92 84.57 89.38
IMF 5 51.85 82.62 86.58
IMF 6 44.11 76.22 76.31

Proposed method 98.24 97.7 99.71

Table 3
The total number of records corresponding to true and false predictions of the proposed method for St.-Petersburg database [24].

True label Predicted label

AV nodal block Acute MI Atrial Coronary Earlier MI Healthy Sinus Node Transient Ischemic
(AV NB) fibrilation Artery Disease (EMI) Dysfunction Attack (TIA)WPWTrue
(AF) (CAD) (SND) Label

AV NB 7985 (99.96%) 0 0 3 0 0 0 0 0
Acute MI 0 7967 0 4 2 7 1 0 5
(99.76%)
AF 0 0 7827 1 0 0 0 9 0
(99.87%)
CAD 0 2 6 7899 (99.37%) 4 36 2 0 0
EMI 1 0 0 5 7930 0 0 2 0
(99.90%)
Healthy 1 0 0 9 8 8604 1 24 0
(99.50%)
SND 0 0 0 0 0 1 7911 (99.99%) 0 0
TIA 0 0 0 21 1 18 0 7903 (99.50%) 0
WPW 0 0 0 1 0 29 0 3 7879
(99.58%)

are reported in Table 2, where it is evident that the higher order

IMFs (IMF 1, IMF 2, IMF 3) provide better results than the lower
order IMFs. Based on the observation, we realized that the higher
order IMFs are comprised of signal features that are crucial for the
classification of ECG signals and it is expected that the combination
of these higher order IMFs might boost the classification accuracy
rather than the individual performances. This leads us to consider
a modified ECG by combining the first three IMF signals expect-
ing better results. Hence, the modified ECG signal is fed into the
CNN model for classification. The overall procedure is illustrated
in Fig. 1. Table 2 shows the detailed classification result on three
databases. The modified ECG signal, the raw ECG signal, and 6 IMF
signals are trained separately using the methodology in Fig. 1. The
proposed modified ECG outperforms individual IMF accuracies as
Fig. 6. Learning rate scheduler implementation in the training procedure. The train- anticipated. Besides, the proposed method performs better than
ing accuracy is at its maximum at a learning rate approximately equal to 0.0007. the raw ECG signal as shown in Table 2. The detailed classification
results using the modified ECG signal are reported in Table 3–5 for
The analysis has been performed by taking windows of 5000 sam- databases [24,22,23], respectively indicating true and false predic-
ples from the database records. The values of average correlation tions. As demonstrated in Table 3, the proposed method predicts
coefficients per disease are shown in Fig. 5 on three databases. It can with almost 100% accuracy for most of the test classes except the
be observed from the figure that the modified ECG shows a higher ‘Coronary artery disease’, ‘Transient ischemic attack’, and ‘Healthy’
correlation coefficient than the raw ECG signal for all the diseases patients. This results in a total accuracy of 99.71%. In Table 4, the
in each database. Hence, it can be concluded that the modified ECG maximum false prediction is only 3.75% for the class ‘Dysrhythmia’.
has better denoising performance compared to the raw ECG signal. This class is being falsely predicted with the other three major dis-
eases. Consequently, this class had a significant effect on the overall
accuracy of the whole database reaching a value of 98.24%. How-
3.4. Results of Proposed Modified ECG Training Method ever, in Table 5, there is a maximum of 5.34% false predictions for
the class ‘AFIB’ and the overall accuracy is 97.70%.
First, the classification results on the test set of all databases are One interesting study could be to investigate the performance of
shown using the proposed method. In the next section, classifica- different combinations of IMFs. For a total 6 IMFs at hand, there are
tion accuracy is presented employing other training approaches. 63 different possible combinations. Performance for all of the pos-
Individual IMF obtained after EMD are trained separately to find sible combinations has been investigated for the PTB database. The
out the performance of each IMF in the classification of ECG sig- IMF combinations with accuracies greater than 80% are presented
nals. The resultant testing accuracies corresponding to each IMF
 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140 135

Table 4
The total number of records corresponding to true and false predictions of the proposed method for PTB diagnostic database [22].

True label Predicted label

Bundle branch Cardiomyopathy Dysrhythmia Healthy control Myocardial Myocarditis Valvular heart
block (BBB) infarction (MI) disease (VHD)

BBB 12,177 (98.1%) 18 174 27 14 2 1

Cardiomyopathy 83 12,633 219 11 97 1 0
(96.85%)
Dysrhythmia 23 40 12,182 75 327 4 5
(96.25%)
Healthy 3 1 17 12,681 51 3 2
Control (99.40%)
MI 56 37 101 118 11,820 20 4
(97.24%)
Myocarditis 3 0 2 9 3 12,967 0
(99.87%)
VHD 1 0 0 1 3 0 12,539
(99.96%)

Table 5
The total number of records corresponding to true and false predictions of the proposed method for MIT-BIH database [23].

True label Predicted label

AFIB Normal P SBR

AFIB 5071 (94.66%) 104 180 2

Normal 317 7423 (95.37%) 36 7
P 150 9 7987 (98.05%) 0
SBR 0 3 0 9732 (99.97%)

Table 6
Experimental result of separate training methodology according to Fig. 8. (Five best combinations (rows 1–5) along with the group of all IMFs (row 6) and the proposed
method (row 7) accuracy.)

PTB diagnostic ECG database (7 classes)(%) MIT-BIH arrhythmia database (4 classes)(%) St.-Petersburg database (9 classes)(%)

IMF(1, 2, 3) = 90.98 IMF(1, 2, 3) = 94.96 IMF(1, 2, 4) = 97.47

IMF(1, 2, 4) = 87.79 IMF(1, 2, 4) = 94.55 IMF(1, 2, 5) = 97.30
IMF(1, 2, 5) = 85.60 IMF(1, 2, 5) = 94.58 IMF(1, 2, 3) = 97.08
IMF(2, 3, 4) = 85.36 IMF(1, 2, 6) = 93.58 IMF(1, 2, 3, 4, 5) = 96.24
IMF(1, 2, 3, 4, 6) = 84.42 IMF(1, 3, 5) = 92.87 IMF(1, 3, 5) = 95.72
IMF(1,2,3,4,5,6) = 70.36 IMF(1,2,3,4,5,6) = 87.67 IMF(1,2,3,4,5,6) = 92.75

Proposed method = 98.24 Proposed method = 97.70 Proposed method = 99.71

Table 7
Comparison between the accuracy of the proposed method and accuracies after parallel training according to the scheme in Fig. 9.

IMF signal parallel training PTB diagnostic ECG MIT-BIH arrhythmia St.-Petersburg
methodology database (7 classes)(%) database (4 classes)(%) database (9 classes)(%)

IMF(1, 2, 3) parallel train 91.77 97.58 97.98

IMF(1, 2, 4) parallel train 89.18 97.55 97.75
IMF(1, 2, 3, 4, 5, 6)parallel train 85.14 95.39 97.70

Proposed method 98.24 97.7 99.71

in Fig. 10 as a bar chart. It is evident from the figure that the combi- 3.5. Results of Other Training Approaches
nation of the first three IMFs, which is considered in the proposed
method, performs better than all other combinations considered. Beside the proposed method analysis, all IMFs have been trained
The average time complexity added to the system due to decom- in a group using different training schemes. In this section, these
posing the raw ECG through EMD and constructing the modified approaches are analyzed and compared with the proposed method.
ECG is approximately 0.1593 s which is effective enough for real- It is investigated if training any combination of IMF signals
time diagnosis of a broad range of ECG signals of cardiovascular through the method shown in Fig. 8 can provide better results than
patients. This processing time is calculated on a system of Intel the proposed method in Fig. 1. The best 6 combinations with greater
(R) Xeon (R) 64 bit CPU E5-2680 v4 @ 2.40GHz with a capacity of accuracies than the others are shown in Table 6 for this methodol-
3300 MHz. ogy along with the accuracy of the proposed method. The proposed
methodology shows a better result in this case too. However, as
136 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140

Fig. 7. Database validation accuracy improvement throughout the training procedure for (a) PTB diagnostic ECG database [22], (b) MIT-BIH Arrhythmia database [23], (c)
St.-Petersburg Arrhythmia database [24]. In every case, the CNN learns the features of the proposed modified ECG signal faster than the individual IMF signal during the
training phase.
 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140 137

Fig. 8. Separate training of IMF signals. The prediction of individual CNN model is combined using the maximum voting criterion to predict the final class. Each of the CNN
models is depicted in Fig. 4.

the lower order IMFs are included in any group, the classification what is reported in [2]. But the proposed method utilizes multiple
accuracy drops. After comparing the results of each IMF signals in databases and multiple classes in each database. The use of differ-
Table 2 and the accuracies of combined IMFs in Table 6, it is indi- ent sources of databases makes the proposed methodology more
cating that combining all IMFs according to Fig. 8 provides a worse robust and generalized than other methods in the literature. Hence,
result than individual IMF. The lower order IMFs degrade the result the proposed modified ECG based method has a broader range of
if combined with the higher order IMFs. cardiovascular disease classification capability.
Next, we considered parallel training of IMF signals to experi-
ment if learning interdependence of IMF signals can provide better 4. Conclusion
results. The IMF signals are trained in parallel mode simultaneously
and concatenated at the last softmax classifying layer as shown in In conventional multiple heart disease classification schemes,
Fig. 9. This layer is followed by three fully connected layers with a the raw ECG signal is employed in feature learning and classifi-
dropout rate of 20%. These fully connected layers have kernels ini- cation. But in this paper, a modified ECG signal is formed using
tialized uniformly with a maximum weight constraint of 3 in each EMD analysis and then forwarded to the classification task. First,
hidden layer. The activation function used in these three fully con- the modified ECG is formed by summing the first three IMF signals
nected layers is ReLU. Three different combinations of IMFs such obtained through EMD analysis and it is shown that the modified
as (IMF1, IMF2, IMF3), (IMF1, IMF2, IMF4) and (IMF1, IMF2, IMF3, ECG has better denoising property. Next, a one dimensional CNN
IMF4, IMF5, IMF6) are trained in parallel mode. The results are network is developed to learn the features of the modified signal
shown in Table 7. By comparing Table 6, and Table 7 it is evident for classification purpose. Apart from this, the proposed method is
that the parallel mode of training according to Fig. 9 provides better compared with other approaches to training with different combi-
results than the separate mode of training as in Fig. 8. But still, it nations of IMF signals. It has been shown that the proposed method
cannot surpass the result of the proposed method. achieves a better result than all other approaches. The proposed
Table 8 compares the result of the proposed method with the approach is evaluated on three publicly available databases and so
reported results in the literature using similar databases. Even this makes the method robust and capable of classifying a broader
though it is reported by Wu et al. [2] an MI detection accuracy of range of cardiovascular diseases. Through detailed investigation it
99%, the proposed method achieves an accuracy of 97.24% in PTB has been shown that the modified ECG exhibits superior perfor-
database depicted in Table 4. On the other hand, both ‘Acute MI’ mance in terms of classification accuracy (98.24% in PTB database,
and ‘Earlier MI’ have been recognized almost perfectly in database 97.70% in MIT-BIH and 99.71% in St.-Petersberg database) and train-
[24] as shown in Table 3. Consequently, the combined accuracy ing time [30].
of detecting ‘Myocardial Infarction(MI)’ is 99% which is similar to
138 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140

Fig. 9. Parallel mode of training of IMF signals. The IMF signals are trained simultaneously in parallel streams and the final softmax regressor of each stream are concatenated.
Following the concatenation layer, there are two fully connected layers along with dropout layers. The final softmax regressor performs the ultimate categorical classification.

Table 8
ECG classification results of different methods in the literature reported on similar databases. The accuracies of the proposed method are reported for the overall database.

Approach Database Network architecture Accuracy or precision

methodology

Wu [2] PTB diagnostic data-set DNN with SAE 99%

(myocardial infarction
detection (only one class))
Banerjee [27] PTB diagnostic data-set Wavelet transformation 97.6%
Kiranayaz [28] MIT-BIH arrhythmia (patient CNN 95%
specific ECG classification)
Pasolli [29] MIT-BIH arrhythmia database Margin sampling, posterior 96.91% with 1000 samples
probability sampling, query by
committee
Macek [14] MIT-BIH arrhythmia, PTB C5.0 decision tree algorithm Error rate = 51.33 % (max)
diagnostic (7 classes)
Proposed method (Modified MIT-BIH [23] (4 classes), PTB 1-D CNN PTB = 98.24%, MIT-BIH = 97.70%,
ECG) database [22] (7 classes) and St.-Petersburg = 99.71%
St.-Petersberg [24] (9 classes)
 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140 139

Fig. 10. Effect of different combinations of IMFs on the model performance in the PTB database [22]. The proposed combination of the first three IMFs (IMF 1, IMF 2, IMF 3)
performs better than other combinations considered. The x-axis labels denote the IMF combinations.

Conflicts of interest the 2000 IEEE International Conference on Acoustics, Speech, and Signal
Processing, vol. 6, IEEE, 2000, pp. 3578–3581, http://dx.doi.org/10.1109/
ICASSP.2000.860175 (Cat. No. 00CH37100).
None. [12] O. Sayadi, M.B. Shamsollahi, ECG denoising with adaptive bionic wavelet
transform, in: 2006 International Conference of the IEEE Engineering in
Acknowledgment Medicine and Biology Society, IEEE, 2006, pp. 6597–6600, http://dx.doi.org/
10.1109/IEMBS.2006.260897.
[13] A. Majumdar, R. Ward, Robust greedy deep dictionary learning for ECG
This research did not receive any specific grant from funding arrhythmia classification, in: 2017 International Joint Conference on Neural
agencies in the public, commercial, or not-for-profit sectors. Networks (IJCNN), IEEE, 2017, pp. 4400–4407, http://dx.doi.org/10.1109/
IJCNN.2017.7966413.
[14] J. Macek, Incremental learning of ensemble classifiers on ECG data, in: 18th
References IEEE Symposium on Computer-Based Medical Systems (CBMS’05), IEEE, 2005,
pp. 315–320, http://dx.doi.org/10.1109/CBMS.2005.69.
[1] S.P. Shashikumar, A.J. Shah, Q. Li, G.D. Clifford, S. Nemati, A deep learning [15] S. Ha, J.-M. Yun, S. Choi, Multi-modal convolutional neural networks for
approach to monitoring and detecting atrial fibrillation using wearable activity recognition, in: 2015 IEEE International Conference on Systems, Man,
technology, in: 2017 IEEE EMBS International Conference on Biomedical & and Cybernetics, IEEE, 2015, pp. 3017–3022, http://dx.doi.org/10.1109/SMC.
Health Informatics (BHI), IEEE, 2017, pp. 141–144, http://dx.doi.org/10.1109/ 2015.525.
BHI.2017.7897225. [16] B. Weng, M. Blanco-Velasco, K.E. Barner, ECG denoising based on the
[2] L. Zhou, Y. Yan, X. Qin, C. Yuan, D. Que, L. Wang, Deep learning-based empirical mode decomposition, in: 2006 International Conference of the IEEE
classification of massive electrocardiography data, in: 2016 IEEE Advanced Engineering in Medicine and Biology Society, IEEE, 2006, pp. 1–4, http://dx.
Information Management, Communicates, Electronic and Automation Control doi.org/10.1109/iembs.2006.259340.
Conference (IMCEC), IEEE, 2016, pp. 780–785, http://dx.doi.org/10.1109/ [17] N.E. Huang, An adaptive data analysis method for nonlinear and
IMCEC.2016.7867316. nonstationary time series: the empirical mode decomposition and Hilbert
[3] T.J. Jun, H.J. Park, N.H. Minh, D. Kim, Y.-H. Kim, Premature ventricular spectral analysis, in: Wavelet Analysis and Applications, Springer, 2006, pp.
contraction beat detection with deep neural networks, in: 2016 15th IEEE 363–376, http://dx.doi.org/10.1007/978-3-7643-7778-6 25.
International Conference on Machine Learning and Applications (ICMLA), [18] G. Rilling, P. Flandrin, P. Goncalves, et al., On empirical mode decomposition
IEEE, 2016, pp. 859–864, http://dx.doi.org/10.1109/ICMLA.2016.0154. and its algorithms, IEEE-EURASIP Workshop on Nonlinear Signal and Image
[4] C. Zhang, G. Wang, J. Zhao, P. Gao, J. Lin, H. Yang, Patient-specific ECG Processing, vol. 3, NSIP-03, Grado (I) (2003) 8–11.
classification based on recurrent neural networks and clustering technique, [19] A.-O. Boudraa, J.-C. Cexus, et al., Denoising via empirical mode decomposition,
in: 2017 13th IASTED International Conference on Biomedical Engineering Proc. IEEE ISCCSP 4 (2006) 2006.
(BioMed), IEEE, 2017, pp. 63–67, http://dx.doi.org/10.2316/P.2017.852-029. [20] G. Singh, G. Kaur, V. Kumar, Ecg denoising using adaptive selection of IMFS
[5] Z. Wu, X. Ding, G. Zhang, X. Xu, X. Wang, Y. Tao, C. Ju, A novel features learning through EMD and EEMD, in: 2014 International Conference on Data Science &
method for ECG arrhythmias using deep belief networks, in: 2016 6th Engineering (ICDSE), IEEE, 2014, pp. 228–231, http://dx.doi.org/10.1109/
International Conference on Digital Home (ICDH), IEEE, 2016, pp. 192–196, icdse.2014.6974643.
http://dx.doi.org/10.1109/ICDH.2016.048. [21] C. Orphanidou, T. Bonnici, P. Charlton, D. Clifton, D. Vallance, L. Tarassenko,
[6] Y. Yan, X. Qin, Y. Wu, N. Zhang, J. Fan, L. Wang, A restricted Boltzmann Signal-quality indices for the electrocardiogram and photoplethysmogram:
machine based two-lead electrocardiography classification, BSN (2015) 1–9, derivation and applications to wireless monitoring, IEEE J. Biomed. Health
http://dx.doi.org/10.1109/BSN.2015.7299399. Inform. 19 (3) (2015) 832–838, http://dx.doi.org/10.1109/jbhi.2014.2338351.
[7] C. Yuan, Y. Yan, L. Zhou, J. Bai, L. Wang, Automated atrial fibrillation detection [22] R. Bousseljot, D. Kreiseler, A. Schnabel, Nutzung der ekg-signaldatenbank
based on deep learning network, in: 2016 IEEE International Conference on cardiodat der ptb über das internet, Biomed. Techn./Bimed. Eng. 40 (s1)
Information and Automation (ICIA), IEEE, 2016, pp. 1159–1164, http://dx.doi. (1995) 317–318, http://dx.doi.org/10.1515/bmte.1995.40.s1.317 [database].
org/10.1109/ICInfA.2016.7831994. URL https://www.physionet.org/physiobank/database/ptbdb/.
[8] R. Sameni, M.B. Shamsollahi, C. Jutten, G.D. Clifford, A nonlinear Bayesian [23] George B. Moody, R.G. Mark, The impact of the MIT-BIH arrhythmia database,
filtering framework for ECG denoising, IEEE Trans. Biomed. Eng. 54 (12) IEEE Eng. Med. Biol. Mag. 20 (3) (2001) 45–50, http://dx.doi.org/10.1109/51.
(2007) 2172–2185, http://dx.doi.org/10.1109/tbme.2007.897817. 932724 (database), URL https://www.physionet.org/physiobank/database/
[9] O. Sayadi, M.B. Shamsollahi, ECG denoising and compression using a modified mitdb/.
extended Kalman filter structure, IEEE Trans. Biomed. Eng. 55 (9) (2008) [24] T. Viktor, A. Khaustov, St.-Petersburg Institute of Cardiological Technics 12 –
2240–2248, http://dx.doi.org/10.1109/tbme.2008.921150. lead arrhythmia database, Circulation-Electronic 101 (i23) (2000) e215–e220,
[10] M.A. Kabir, C. Shahnaz, Denoising of ECG signals based on noise reduction URL https://www.physionet.org/physiobank/database/incartdb/ (database).
algorithms in EMD and wavelet domains, Biomed. Signal Process. Control 7 [25] A.L. Goldberger, L.A. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark, J.E.
(5) (2012) 481–489, http://dx.doi.org/10.1016/j.bspc.2011.11.003. Mietus, G.B. Moody, C.-K. Peng, H.E. Stanley, Physiobank, physiotoolkit, and
[11] N. Nikolaev, Z. Nikolov, A. Gotchev, K. Egiazarian, Wavelet domain wiener physionet: components of a new research resource for complex physiologic
filtering for ECG denoising using improved signal estimate, in: Proceedings of
140 N.I. Hasan, A. Bhattacharjee / Biomedical Signal Processing and Control 52 (2019) 128–140

signals, Circulation 101 (23) (2000) e215–e220, http://dx.doi.org/10.1161/01. Conference of the IEEE Engineering in Medicine and Biology Society (EMBC),
CIR.101.23.e215 [database website]. IEEE, 2015, pp. 2608–2611, http://dx.doi.org/10.1109/EMBC.2015.7318926.
[26] H. Robbins, S. Monro, A stochastic approximation method, Ann. Math. Stat [29] E. Pasolli, F. Melgani, Active learning methods for electrocardiographic signal
(1951) 400–407, http://dx.doi.org/10.1214/aoms/1177729586. classification, IEEE Trans. Inf. Technol. Biomed. 14 (6) (2010) 1405–1416,
[27] S. Banerjee, M. Mitra, Application of cross wavelet transform for ECG pattern http://dx.doi.org/10.1109/titb.2010.2048922.
analysis and classification, IEEE Trans. Instrum. Measure 63 (2) (2014) [30] N. Hasan, Cardiovascular ḋisease ċlassification ėmploying ėmd, URL https://
326–333, http://dx.doi.org/10.1109/TIM.2013.2279001. github.com/NahianHasan/Cardiovascular Disease Classification Employing
[28] S. Kiranyaz, T. Ince, R. Hamila, M. Gabbouj, Convolutional neural networks for EMD, (2019).
patient-specific ECG classification, in: 2015 37th Annual International