An intelligent medical decision support system for diagnosis of heart abnormalities in ECG signals

Abstract

This work describes a novel method to detect a Bundle branch block and myocardial infarction from the multi-lead ECG signal. The clinical characteristics of BBB and MI extracted by using a derivative filter and continuous wavelet transform (CWT). The signal with the frequency below 50 Hz obtained and derivative-based filter applied to extract features. The continuous wavelet transforms also applied to the signals of BBB and MI. The CWT coefficients extracted, and the signals reconstructed from the wavelet to obtain the features. The feature vectors generated from each lead of both the methods computed using parameters such as spectral entropy, mean of peaks, total energy from power spectrum density, form factor, and root mean squared value. The results of both the derivative-based filter and CWT analyzed by applying these features to the classifiers. The accuracy of classification of diseases computed using SVM, KNN, Levenberg-Marquardt Neural Network (LMNN), and scaled conjugate gradient backpropagation network (SCG NN). The best accuracy obtained from the derivative filter and wavelet transform method is 96.4% using LMNN and SCGNN classifier and 96.4% using KNN and LMNN classifier respectively.

Keywords

Bundle branch block myocardial infarction derivative-based filter continuous wavelet transform neural network

1. Introduction

Early diagnosis of heart diseases is essential to give quick treatment to the patient to save their life [1]. The ECG is a rapid tool for the early diagnosis of heart diseases. The major concern of heart diseases is that Bundle Branch Block (BBB) and Myocardial Infarction (MI). The Left bundle branch block along with myocardial infarction diagnosis is very difficult. The diagnosis of delayed MI disease may lead to major complications [2]. It is a challenging issue to diagnose the Bundle Branch Block and Myocardial Infarction based ECG signal. The Bundle Branch Block disturbs the flow of electrical pulses through the bundle of His which leads to the delay in the activation of ventricles that broadens the QRS complex of ECG signal, raises the amplitude of the Q wave and changes in the T waves [1]. Sometimes there is a disappearance of the normal Q waves in Lead I, avL and $V_{6}$ . The ECG criteria for the bundle branch block-based ECG signal is that the duration of the QRS complex is greater than 120 ms, R Wave is broader in lead V5 and V6. The peak time of the R wave is greater than 0.006 s [3].

Figure 1.

Bundle branch block based ECG signal.

Figure 2.

Myocardial infarction based ECG signal [4].

Myocardial Infraction named as a heart attack occurs due to the disruption of the flow of blood to the heart. It caused by the deposition of fats in the blood vessels of the heart. It leads to changes in the morphology of the ST segment and the Q wave of the ECG signal [4]. An elevated ST segment is the major criterion for the diagnosis of myocardial infarction. But other conditions that lead to the elevation of the ST segment are bundle Branch Block and Left Ventricular Hypertrophy [7]. The LBBB pattern confuses with the ECG pattern of MI [8]. Thus, to overcome this issue an algorithm is proposed to diagnose Myocardial infarction and Bundle branch block disease.

A few of the papers which describe the detection of BBB and MI mentioned below. Kora et al. presented an ECG based BBB detection using the Adaptive Bacterial Foraging Optimization technique and Neural Network [8]. Sharma et al. proposed the detection of Myocardial Infarction using multiscale energy and the Eigenspace approach and obtained an accuracy of 96% [9]. Kora et al. applied an improved bat optimization technique and neural network classifier to diagnose Myocardial Infarction [10]. Li et al. described the detection of MI by using Multiple Instance Learning [11]. Hasan et al. reported a technique based on a convolution neural network to classify various cardiovascular diseases [12]. Shi et al. proposed multilayer input based on a deep neural network to classify heartbeats which includes the right bundle branch block and left bundle branch block [13]. Kayikcioglu et al. extracted time-frequency distribution-based features from a multi-lead ECG signal to predict the changes in the ST segment [14]. Han et al. explained a technique that combines the energy entropy and morphological features to diagnose Myocardial Infarction [15].

Liu et al. presented a technique to diagnose MI from the Lead II ECG signal. This technique involves the ECG segments and CNN based algorithm [16]. Mienyea et al. reported improved sparse encoder neural network models to diagnose heart disease [17]. Li et al. proposed a 31-layer residual Neural network to diagnose heartbeats which includes the Left Bundle Branch Block and Right Bundle Branch block [18].

The above papers explain either detection of BBB or MI. The identification of these two ECG signals is a challenging issue. This paper intends to show an algorithm to diagnose BBB and MI-based ECG signal using a derivative-based method and continuous wavelet transform method. The analysis of signal implemented by extracting features which include mean of detected peaks, mean of the continuous wavelet transform coefficients, spectral entropy, power spectral energy, form factor, and root mean square value. Then these features fed to the various classifiers. The reliability of the obtained results verified using 5-fold cross-validation.

The following section explains feature extraction techniques and classifiers to diagnose both Bundle Branch Block and Myocardial Infarction.

Figure 3.

ECG signal with BBB and MI.

Figure 4.

Block diagram of a derivative filter method to detect BBB and MI.

Figure 5.

Block diagram of a continuous wavelet transform method to detect BBB and MI.

Figure 6.

Preprocessing of ECG signal.

2. Materials and methods

2.1 Dataset

The ECG signals involved in this study are from the PTB Diagnostic ECG Database. This database holds 448 ECG records which include BBB, MI, Myocarditis, cardiomyopathy and healthy subjects [10]. These records obtained from 209 males and 81 females. Each ECG record holds 12 conventional leads and 3 frank leads. The sampling frequency of the signal is 1000 Hz. The sample signal from the PTB database is shown in Fig. 3.

2.2 Methodology

The methodology of the proposed work comprises preprocessing using low pass filter, features extraction using the derivative-based method and continuous wavelet transform method and classification using SVM, KNN, LM NN, and SCG NN classifiers. The block diagram of the derivative-based filter and continuous wavelet transform method is shown in Figs 4 and 5 respectively. The feature extraction of derivative-based filter includes differentiation, squaring and normalization of ECG signal. This is explained in detail in Section 2.4. The feature extraction of continuous wavelet transform includes the decomposition of a signal into wavelet coefficients between 10 Hz and 50 Hz and is followed by the reconstruction of the signal. The performance is analyzed using various classifiers.

2.3 Preprocessing

In preprocessing, the ECG signal distorted by noise which includes power line interference, motion artifacts, baseline wandering, and high-frequency noises [10]. In this work, the ECG signal is added with random noise. The fourth-order Butterworth filter designed for the lowpass filter with the cutoff frequency 50 Hz applied to minimize the noise present in the signal. The preprocessing result is shown in Fig. 6.

2.4 Derivative-based filter

The structure of the derivative-based filter is shown in Fig. 4. The derivative filter method improves detection accuracy [19]. The features extracted by using a derivative-based filter technique include differentiation, squaring and normalization operation. The results are shown in Fig. 7. The filtered ECG signal $x(n)$ differentiated to extract the features. It suppresses the peaks of $P$ waves and $T$ waves and also provides details of the QRS complex.

The differentiation of ECG signal implemented by Eq. (1) [19]

$\displaystyle U(n)=1/8(-x[n-2]-2x[n-1]+2x[n+1]+x[n+2])$ (1)

The ECG signal then squared using Eq. (2). The squaring operation attenuates other peaks and allows the QRS peaks as positive peaks irrespective of the polarity of the original ECG signal. It is used to measure the energy of the QRS complex. It amplifies the differentiated ECG signal nonlinearly [20] and then it’s normalized by using the Eqs (3) and (4)

$\displaystyle y(n)=[u(n)]^{2}$ (2) $\displaystyle y(n)=y(n)-\frac{\sum y(n)}{N}$ (3) $\displaystyle y(n)=\frac{y(n)}{\max(|y(n)|)}$ (4)

where $N$ stands for the number of samples in the ECG signal.

Figure 7.

ECG signal obtained from the derivative filter (1) ECG signal (2) Differentiated ECG signal (3) Squared ECG signal (4) Normalized ECG signal.

Figure 8.

ECG signal obtained from continuous wavelet transform (1) ECG signal (2) Reconstructed ECG signal.

This technique applied to all the 12 leads of ECG signal and 60 features were obtained from a single 12 lead ECG signal. Then this fed to the classifier to show BBB and MI-based ECG signals.

2.5 Continuous wavelet transform

Figure 5 shows the structure of the continuous wavelet transform method. The noise exists in the ECG signal removed by a lowpass filter shown in Fig. 6. The continuous wavelet transform method applied to obtain the wavelet coefficients. The signals with the frequency between 10 Hz and 50 Hz were reconstructed using an inverse continuous wavelet transform. This is followed by the estimation of feature extraction parameters. The total features extracted from each 12-lead ECG signal is 72. Then the BBB and MI signals detected using various classifiers.

The wavelet transform method can implement in many signal processing applications. This technique based on multiple scales [21]. The continuous wavelet transform disintegrates the signal $x(n)$ into two variables such as scale and time [22]. This is known as coefficients of the wavelet transform. The morlet wavelet used as a mother wavelet in this study. The mother wavelets shifted and dilated to produce daughter wavelets [22]. The dilated parameter is the estimation of the degree of compression. The shifting is the position of wavelet shift from one to another [23]. The wavelet coefficients obtained from the variation of these two parameters.

To examine the signal function, the continuous wavelet transform method transforms time-domain function to frequency domain function and by using inverse continuous wavelet transform (ICWT) the signal again transformed to time-domain function. It recovers the original signal [23].

In this paper, continuous wavelet transform applied to all the 12 leads and disintegrated into wavelet coefficients. The ECG signal plotted in Fig. 8. The inverse continuous wavelet transforms applied to this wavelet to reconstruct the signal between 10 Hz and 50 Hz. The peaks detected and the other features such as spectral entropy, form factor, root mean square value, the total energy in power spectral density are obtained. Then it is applied to the classifier to show BBB and MI-based ECG signals.

2.6 Feature extraction

2.6.1 Spectral entropy

The spectral power distribution of a signal measured using spectral entropy. It is based on the concept of Shannon entropy [24]. The spectral Entropy serves the power distribution of the normalized signal as a probability distribution. This widely used in signal analysis.

For a signal $y(n)$ , the power spectrum $s(m)$ calculated using the formula

$\displaystyle s(m)=|Y(m)|^{2}$ (5)

where $Y(m)$ is the discrete Fourier transform of $y(n)$ . The probability distribution is

$\displaystyle p(m)=\frac{S(m)}{\sum_{i}{S(i)}}$ (6)

The spectral entropy implemented using the formula [24]

$\displaystyle se=-\sum^{N}_{m-1}{p(m)\log_{2}}p(m)$ (7)

The normalized spectral entropy calculated by the following equation

$\displaystyle se_{n}=\frac{\sum^{N}_{m=1}{p(m)\log_{2}p(m)}}{\log_{2N}}$ (8)

where $N$ represents total frequency points.

2.6.2 Form factor

Form Factor is a parameter to analyze the signals. Three variables used to analyze the signal. The first variable is known as an activity. It refers to a variance of the signal. It is denoting by $\sigma^{2}$ . The second variable is known as mobility $M_{x}$ . It is a square root of the ratio between the activity of the first derivative of the signal and the activity of the signal [25].

$\displaystyle M_{x}=[{\sigma}^{2}_{x^{\prime}}/{\sigma}^{2}_{x}]^{\frac{1}{2}}$ (9)

where $x^{\prime}$ represents the first derivative of the signal. The third parameter is known as complexity. It is called a form factor. It defines as the ratio between the mobility of the first derivative of the ECG signal and the mobility of the ECG signal itself [25].

$\displaystyle\text{Form factor}=\frac{M_{x^{\prime}}}{M_{x}}$ (10)

2.6.3 Root mean squared value

The root mean squared value of ECG signal $x(n)$ over its duration of $N$ samples represented by the Eq. (11) [25]

$\displaystyle\text{RMS}=\left[\frac{1}{N}\sum^{N-1}_{n=0}{x^{2}(n)}\right]^{% \frac{1}{2}}$ (11)

But this not recommended for the analysis of nonstationary signals. An RMS value of the signal calculated for a causal window of $M$ samples calculated by the following Eq. (12) [25]

$\displaystyle\text{RMS}(n)=\left[\frac{1}{M}\sum^{M-1}_{K=0}{x^{2}(n-k)}\right% ]^{\frac{1}{2}}$ (12)

where $M$ value is to choose based on the bandwidth of the signal. Where $M\ll N$ .

2.6.4 Power spectrum density

The estimation of the power spectrum says the distribution of power present in the signal over the frequency [26]. It is associated with the correlation function by Fourier transform. The signal which is more correlated consists of the more concentrated power spectrum. The signal which is less correlated leads to more spread in its power spectrum [26]. There are several methods to estimate the power spectrum which includes the non-parametric power estimation method, high-resolution estimation method, and model-based power estimation method. The non-parametric method used in this study.

This method estimates PSD directly from the signal itself. It is also known as a periodogram. It defined using Eq. (13) [26]

$\displaystyle P_{xx}=\frac{1}{N}\left|\sum^{N-1}_{n=0}{x(n)e^{-j2\pi fn}}% \right|^{2}=\frac{1}{N}{|X(f)|}^{2}$ (13)

where $x(n)$ stands for the input signal $X(f)$ is the Fourier transform of the input signal. The total energy calculated by summing all the values of the PSD.

$\displaystyle\text{Total energy∼{}}E=\sum^{i-1}_{i=0}{p_{xx}(f)}$ (14)

3. Classification

In this work, four classifiers such as Support Vector Machine (SVM), K-nearest neighbor (KNN), SCG NN and LM NN employed for diagnosis and classification of BBB and MI.

3.1 SVM classifier

This method discriminates the features by computing the perfect hyperplane. This hyperplane separates the feature vectors one from the other class [27]. It also employs a nonlinear mapping which transforms the input features to high dimension [27]. In this paper, the SVM classifier is employed using the Gaussian kernel function. The Gaussian kernel equation is represented by $k(x,y)={e^{\frac{-\|x-y\|}{{2\sigma}^{2}}}}^{2}$ [28]. The benefit of the Gaussian kernel is that it is more flexible.

3.2 $K$ nearest neighbour classifier

The $K$ nearest neighbor’s classifier computes the least distance between training and testing data. The evaluation of the classifier is depending on the value of $k$ and distance. In this work, this parameter computed based on cross-validation [29]. The database disintegrated into samples of $k$ . For testing the data, the single sample is employed as a data for verification and the balance $k-1$ samples are employed as a training [29]. This process replicates $K$ times to evaluate the performance [29].

3.3 Neural network classifier

The LMNN and SCGNN neural network techniques were employed to classify the BBB and MI-based ECG signals. Artificial Neural Network (ANN) is an essential segment of artificial intelligence [30]. It is one of the excellent machine learning techniques. It uses backpropagation NN to find solutions for the approximate feature dataset [30]. The structure of back propagation model is shown in Fig. 9.

Figure 9.

Backpropagation neural network [33].

Table 1

Sample features extracted from derivative based filters

Peaks		Total energy in PSD		Spectral entropy		Form factor		Root mean square
BBB	MI	BBB	MI	BBB	MI	BBB	MI	BBB	MI
0.0400	0.0166	1457	2811	0.675	0.681	2.061	1.869	0.120	0.016
0.0266	0.0197	6279	31353	0.702	0.689	2.399	1.948	0.025	0.017
0.0021	0.0390	2668	25655	0.697	0.692	1.830	1.800	0.016	0.016
0.0142	0.0158	6128	25278	0.678	0.695	3.122	1.796	0.078	0.015

Table 2

Sample features extracted from CWT

Peaks		Total energy in PSD		Spectral entropy		Form factor		Root mean square		Wavelet coefficients
BBB	MI	BBB	MI	BBB	MI	BBB	MI	BBB	MI	BBB	MI
0.0723	0.035	1943	7613	0.586	0.598	1.3050	1.3702	0.1394	0.087	0.0382	0.015
0.0255	0.041	6613	7271	0.618	0.592	1.5142	1.2097	0.0813	0.085	0.0156	0.044
0.0121	0.038	1107	8466	0.620	0.594	1.2055	1.2149	0.0332	0.086	0.0215	0.050
0.0341	0.020	6232	1908	0.596	0.595	1.2447	1.3082	0.0789	0.092	0.0209	0.020

Figure 10.

Boxplot of a derivative filter of Lead III ECG signal.

The Levenberg-Marquardt (LM) back-propagation technique is a combination of gauss-newton and gradient descent approach [31]. It is the quickest learning feedforward network. At every iteration, the gradient descent technique reduces the solution by selecting the parameters that bring the value of the function minimum [31]. The summation of squared errors minimized by the interpolation between the direction of steepest descent and the direction estimated by the newton method. The convergence speed raised in the LM NN technique [31]. The training algorithm of the LM NN as follows as [30]

Step 1: Step 1:

Calculate the $u(w)=\sum^{n}_{i=1}(t_{i}-o_{i})^{2}$ where $t_{i}$ and $o_{i}$ are the targets and output value.

Step 2:

Start with a $\mu=$ 0.01.

Step 3:

Find $u(w_{m}+\Delta w_{m})$ .

Step 4:

If $u(w_{m}+\Delta w_{m})\geqslant u(w_{m})$ , raise the $\mu$ by 10 times and go to 3 ${}^{\text{rd}}$ step.

Step 5:

If $u(w_{m}+\Delta w_{m})\leqslant u(w_{m})$ , reduces $\mu$ by 10. times and update $w_{m}=w_{m}+\Delta w_{m}$ and jump to step 3.

The SCG NN technique is also a feed-forward network. It implements conjugates-based search direction instead of a linear search [8]. It determines the optimal distance for each iteration [32].

In this work, the feature datasets obtained from the derivative-based filter and continuous wavelet transform method applied to the ANN. The neurons presented in the hidden layer of ANN are 10. The 70% of the dataset is for training and 15% of the dataset is for validation and 15% employed for testing the dataset. The performance of the LM NN and SCG NN model evaluated on these datasets. The LMNN and scaled conjugate gradient (SCG NN) shows a better result than the other classifiers.

4. Results and discussion

The derivative based filter and continuous wavelet transform techniques were employed to discriminate BBB and MI-based ECG signals. The features extracted are the means of peaks, spectral entropy, form factor, root mean square value, wavelet coefficients and the total energy in power spectral density. The features extracted from derivative filters are 60 ( $5\times 12=$ 60) from each 12 lead ECG signal. The features extracted from CWT are 72 ( $6\times 12=$ 72). The features extracted are employed for training the classifier. These features used as input for SVM, KNN, LMNN, and SCNN classifiers. The sample features extracted from the derivative-based filter and continuous wavelet transform method are shown in Tables 1 and 2.

The Figs 10 and 11 shows the box plot of the BBB and MI Lead III ECG signal features obtained from the derivative-based filter and continuous wavelet transform method. Box plot used to analyze the range and distribution of the given data set and also it used to compare the various data sets [33]. In this plot, the dataset divided into four or more equal parts called quartile. Box plot comprises minimum value, maximum value, first quartile, second quartile, and third quartile. The first quartile is the middle number lies between the minimum value and the median of the given dataset. The second quartile represented as the median of the given dataset. The third quartile is the middle number lies between the median and the maximum value of the dataset. The values of BBB based ECG signal various from MI-based ECG signal. It shows for both techniques in Tables 3 and 4.

Table 3
Experimental results of I ${}^{\text{st}}$ , II ${}^{\text{nd}}$ , III ${}^{\text{rd}}$ quartile and maximum values of feature vectors of BBB and MI ECG signal from lead III using the derivative-based method

Values from feature vectors	Peaks		Total energy		Spectral entropy		Form factor		RMS value
	BBB	MI	BBB	MI	BBB	MI	BBB	MI	BBB	MI
I ${}^{\text{st}}$ quartile	0.0037	0.0073	50175.57	25530.93	0.6795	0.6573	1.9390	1.7989	0.0223	0.0160
II ${}^{\text{nd}}$ quartile	0.0128	0.0132	75701.11	26050.99	0.6878	0.6808	2.1465	1.8123	0.0275	0.0161
III ${}^{\text{rd}}$ quartile	0.0227	0.0188	599653.2	31996.99	0.6977	0.6895	2.6028	1.9524	0.0774	0.0179
Maximum value	0.0400	0.0390	1457794	1111374	0.7037	0.6954	3.1221	3.3801	0.1207	0.1054

Table 4

Experimental results of I ${}^{\text{st}}$ , II ${}^{\text{nd}}$ , III ${}^{\text{rd}}$ quartile and maximum values of feature vectors of BBB and MI ECG signal of lead III using the wavelet method

Values from feature vectors	Peaks		Total energy		Spectral entropy
	BBB	MI	BBB	MI	BBB	MI
I ${}^{\text{st}}$ quartile	0.0340	0.0339	651865.1	674412.4	0.5951	0.5765
II ${}^{\text{nd}}$ quartile	0.0377	0.0405	909327.8	753401.8	0.6007	0.5922
III ${}^{\text{rd}}$ quartile	0.0463	0.0551	1134189	896405.2	0.6101	0.5962
Maximum value	0.0723	0.0810	2239069	2749701	0.6207	0.6046
Values from feature vectors	Form factor		RMS value		Wavelet coefficients
	BBB	MI	BBB	MI	BBB	MI
I ${}^{\text{st}}$ quartile	1.2666	1.2137	0.0807	0.0821	0.0201	0.0217
II ${}^{\text{nd}}$ quartile	1.3075	1.3034	0.0953	0.0868	0.0243	0.0322
III ${}^{\text{rd}}$ quartile	1.3442	1.3815	0.1061	0.0946	0.0324	0.0451
Maximum value	1.5142	1.5899	0.1496	0.1658	0.0529	0.0587

Figure 11.

Boxplot of CWT of Lead III ECG signal.

Thus, the box plot differentiates the BBB and MI-based ECG signals by using a derivative-based filter and continuous wavelet transform method. The confusion matrix of both the methods shown in Figs 12 and 13. The performance of the experiments evaluated based on the term’s such as accuracy, sensitivity, and specificity. These terms calculated by using True Positive values, True Negative values, False Positive values, and False Negative values. True Positive (TN) values denote the number of identical events. False Negative values denote the number of non-identical events. True Negative value denotes the event that considered as not defectives. False Positive Values denotes that the non-identical event detected as an identical event. The accuracy is determined by the term $(P+Q)/(P+Q+R+S)$ . The sensitivity $=P/(P+S)$ and specificity $=Q/(Q+R)$ where $\textit{TP}=P$ , $\textit{TN}=Q$ , $\textit{FP}=R$ , and $\textit{FN}=S$ . The evaluated values indicated in Table 5.

The receiver operating characteristic (ROC) curve for SVM and KNN technique compared to evaluate the classifier performance. The area under the ROC curve for SVM, KNN of both the techniques evaluated in this work. It shows that the area under the curve for SVM, KNN, SG NN and LM NN using derivative filters were 0.92, 0.95,0.96 and 0.96 respectively. For the CWT method, the area under the curve for SVM, KNN, SCG NN, LM NN were 0.93, 0.97, 0.93, 0.96 respectively. The 5-fold cross-validation technique performed for choosing the training and testing of classifiers.

Table 5

Evaluation of the performance of classifiers to detect BBB and MI ECG signal

Classifiers	Derivative based filter			Continuous wavelet transform
	Sensitivity %	Specificity %	Accuracy %	Sensitivity %	Specificity %	Accuracy %
SVM classifier	91.7	91.7	91.7	91.7	91.7	91.7
Fine KNN classifier	93.8	97.2	95.2	93.8	100	96.4
SCG NN	93.8	100	96.4	93.8	91.7	92.9
LM NN	93.8	100	96.4	93.8	100	96.4

Figure 12.

Confusion matrix obtained based on derivative based method.

Figure 13.

Confusion matrix obtained based on the continuous wavelet transform method.

Table 5 indicates the comparative solutions of sensitivity, specificity, the accuracy of the derivative filter and the CWT method to characterize BBB and MI-based ECG signal using SVM, KNN, LMNN, and SCNN classifier. As referred to the table, both the methods achieved the highest classification accuracy of 96.4% using the LM NN classifier. The comparative results to Diagnose Myocardial Infarction disease with the ECG signal from the PTB database is shown in the following Table 6.

Table 6

Comparative analysis of diagnosis of MI from the PTB database

Techniques	Sensitivity %	Specificity %	Accuracy %
Sharma et al. [7]	93	96	99
Cruces et al. [34]	98	97	–
Kora et al.	88.2	87.2	87.9

This study performs an experimental work using derivative filter and CWT for extracting BBB and MI-based ECG signal features and SVM, KNN and LMNN and SCG NN classifiers for detection of BBB and MI-based signals. From the study, it reveals that in both the methods the LMNN classifier outperformed than other technique.

Bundle Branch Block and Myocardial Infarction are cardiovascular diseases. This disease causes heart failure which leads to a high mortality rate. In India, the death rate due to cardiovascular disease increased to 272 per 1 lakh population [35]. The diagnosis of these two diseases using various imaging modalities provides high cost and time-consuming [36, 37, 38]. This research work implies the automatic diagnosis of these diseases using a machine learning algorithm. This proposed work implemented in MATLAB 2018 software installed in a computer having an Intel Core i3 processor with 4 GB RAM. The proposed work provides an accuracy of 96.4%. This research work carried out with the real-time ECG signal acquired from a local hospital. This research also highlights the doctor’s view of the automatic diagnosis of these diseases. They suggested that this work may help general physicians to diagnose heart disease at an early stage at a low cost.

5. Conclusion

In this paper, the derivative filter and continuous wavelet transform method adopted to characterize the ECG signal with the BBB and MI. A group of feature vectors includes mean of peaks, coefficients of wavelet transforms, spectral entropy, the total energy of power spectrum density, form factor, Root mean squared value exploited to discriminate the ECGs with the BBB and MI. The generated feature vectors denote the BBB and MI ECG signals. This uses as an input to the classifiers. The accuracy achieved using derivative filters is 91.7%, 95.2%, 96.4% and 96.4% using SVM and KNN, SCGNN and LMNN classifiers. The accuracy achieved using CWT techniques is 91.7%, 96.4%, 92.9% and 96.4% using SVM, KNN, LMNN, and SCG NN classifiers. The derivative-based method and the CWT method shows the best classification accuracy of 96.4% using the LMNN classifier.

In future work, various other techniques can implement with different classifiers and its performance will be tested for comparison to diagnose cardiac diseases such as Bundle Branch Block and Left Ventricular Hypertrophy. The deep learning algorithm can also be implemented to diagnose Myocardial Infarction, Bundle Branch Block and Left Ventricular Hypertrophy.

References

Kora

Krishna

. Hybrid firefly and particle swarm optimization algorithm for the detection of bundle branch block. International Journal of the Cardiovascular Academy. 2016; 2(1): 44-48.

Idris

Hatahet

Edris

. Acute myocardial infarction in the setting of left bundle branch block: Chapman’s sign. The American Journal of Emergency Medicine. 2019; 37(10): 1991-e5.

Gupta

Shelke

. Acute myocardial infarction with bundle branch block (Rbbb or Lbbb)-clinical characteristics, complications & prognostic significance-a hospital based study. IOSR J Dent Med Sci (IOSRJDMS). 2017; 16: 43-48.

Liu

Wang

Huang

Zheng

Luo

Zhou

. A novel electrocardiogram parameterization algorithm and its application in myocardial infarction detection. Computers in Biology and Medicine. 2015; 61: 178-184.

Chien

Koninklijke Philips

. Electrocardiography to differentiate acute myocardial infarction from bundle branch block or left ventricular hypertrophy. United States patent US 9, 782, 102. 2017.

Pollak

Brady

. Electrocardiographic patterns mimicking ST segment elevation myocardial infarction. Cardiology Clinics. 2012; 30(4): 601-615.

Tabas

Rodriguez

Seligman

Goldschlager

. Electrocardiographic criteria for detecting acute myocardial infarction in patients with left bundle branch block: A meta-analysis. Annals of Emergency Medicine. 2008; 52(4): 329-336.

Kora

Kalva

. Detection of bundle branch block using adaptive bacterial foraging optimization and neural network. Egyptian Informatics Journal. 2017; 18(1): 67-74.

Sharma

Tripathy

Dandapat

. Multiscale energy and eigenspace approach to detection and localization of myocardial infarction. IEEE Transactions on Biomedical Engineering. 2015; 62(7): 1827-1837.

10.

Kora

Kalva

. Improved bat algorithm for the detection of myocardial infarction. SpringerPlus. 2015; 4(1): 1-8.

11.

Sun

Yang

. ECG analysis using multiple instance learning for myocardial infarction detection. IEEE Transactions on Biomedical Engineering. 2012; 59(12): 3348-3356.

12.

Hasan

Bhattacharjee

. Deep learning approach to cardiovascular disease classification employing modified ECG signal from empirical mode decomposition. Biomedical Signal Processing and Control. 2019; 52: 128-140.

13.

Shi

Qin

Xiao

Zhao

Liu

. Automated heartbeat classification based on deep neural network with multiple input layers. Knowledge-Based Systems. 2020; 188: 105036.

14.

Kayikcioglu

Akdeniz

Kose

Kayikcioglu

. Time-frequency approach to ECG classification of myocardial infarction. Computers & Electrical Engineering. 2020; 84: 106621.

15.

Han

Shi

. Automated interpretable detection of myocardial infarction fusing energy entropy and morphological features. Computer Methods and Programs in Biomedicine. 2019; 175: 9-23.

16.

Liu

Wang

Chang

Xing

Zhou

. A simple and effective method for detecting myocardial infarction based on deep convolutional neural network. Journal of Medical Imaging and Health Informatics. 2018; 8(7): 1508-1512.

17.

Mienye

Sun

Wang

. Improved sparse autoencoder based artificial neural network approach for prediction of heart disease. Informatics in Medicine Unlocked. 2020; 18: 100307.

18.

Zhou

Wan

Mou

. Heartbeat classification using deep residual convolutional neural network from 2-lead electrocardiogram. Journal of Electrocardiology. 2020; 58: 105-112.

19.

Arzeno

Deng

Poon

. Analysis of first-derivative based QRS detection algorithms. IEEE Transactions on Biomedical Engineering. 2008; 55(2): 478-484.

20.

Choudhary

Hamde

. A simple and robust algorithm for the detection of QRS complexes. In 2015 International Conference on Industrial Instrumentation and Control (ICIC). 2015; 870-872.

21.

Sumathi

Beaulah

Vanithamani

. A wavelet transform based feature extraction and classification of cardiac disorder. Journal of Medical Systems. 2014; 38(9): 1-1.

22.

Amin

Davis

Thomas

Holloway

. Analysis of wave slam induced hull vibrations using continuous wavelet transforms. Ocean Engineering. 2013; 58: 154-166.

23.

Yadav

. A brief description of wavelet and wavelet transforms and their applications. International Journal of Statistics and Applied Mathematics. 2018; 3(1): 266-271.

24.

Tian

Zhang

Yang

Zheng

Shi

. Spectral entropy can predict changes of working memory performance reduced by short-time training in the delayed-match-to-sample task. Frontiers in Human Neuroscience. 2017; 11: 437.

25.

Rangayyan

. Biomedical signal analysis. John Wiley & Sons; 2015.

26.

Khazaee

Ebrahimzadeh

. Classification of electrocardiogram signals with support vector machines and genetic algorithms using power spectral features. Biomedical Signal Processing and Control. 2010; 5(4): 252-263.

27.

Zeng

. Support vector machines application research in pattern classification. Telkomnika Indonesian Journal of Electrical Engineering. 2015; 13(1).

28.

Hofmann

. Support vector machines-kernels and the kernel trick. Notes. 2006; 26(3): 1-6.

29.

Pan

Tompkins

. A real-time QRS detection algorithm. IEEE Transactions on Biomedical Engineering. 1985; (3): 230-236.

30.

Ibrahim

Jihad

Kamal

. Determining optimum structure for artificial neural network and comparison between back-propagation and levenberg-marquardt training algorithms. International Journal of Engineering Science. 2017; 14887.

31.

Damodara

Alphones

Chen

Lou

Martin

. Flare performance modeling and set point determination using artificial neural networks. International Journal of Energy and Environmental Engineering. 2020; 11(1): 91-109.

32.

Nayak

Kumar

Choudhury

. Scaled conjugate gradient backpropagation algorithm for selection of industrial robots. International Journal of Computer Application (2250-1797). 2017; 7(6).

33.

Vignesh

Pavithra

Dinakaran

Thirumalai

. Data analysis using box and whisker plot for stationary shop analysis. in: 2017 International Conference on Trends in Electronics and Informatics (ICEI). 2017; 1072-1076.

34.

Cruces

Arini

. A novel method for cardiac vector velocity measurement: Evaluation in myocardial infarction. Biomedical Signal Processing and Control. 2016; 28: 58-62.

35.

Prabhakaran

Jeemon

Roy

. Cardiovascular diseases in India: Current epidemiology and future directions. Circulation. 2016; 133(16): 1605-1620.

36.

Deepak

Arnav

Jain

Tushar

Ashish

Victor

. Optimized cuttlefish algorithm for diagnosis of Parkinson’s disease. Cognitive Systems Research. 2018; 52: 36-48.

37.

Deepak

Shrish

Ashish

Aboul

Victor

. Improved diagnosis of Parkinson’s disease based on optimized crow search algorithm. Computers and Electrical Engineering. 2018; 68: 412-424.

38.

Aditya

Aman

Divya

Deepak

Ashish

Arun

Joseph

, A novel deep learning based multi-model ensemble methods for prediction of neuromuscular disorders. Neural Computing and Applications (Springer). 2018.

An intelligent medical decision support system for diagnosis of heart abnormalities in ECG signals

Abstract

Keywords

1. Introduction

2.1 Dataset

2.2 Methodology

2.3 Preprocessing

2.4 Derivative-based filter

2.6 Feature extraction

2.6.1 Spectral entropy

3.1 SVM classifier

3.2 K nearest neighbour classifier

3.3 Neural network classifier

Table 3 Experimental results of I st , II nd , III rd quartile and maximum values of feature vectors of BBB and MI ECG signal from lead III using the derivative-based method

References

3.2 $K$ nearest neighbour classifier

Table 3
Experimental results of I ${}^{\text{st}}$ , II ${}^{\text{nd}}$ , III ${}^{\text{rd}}$ quartile and maximum values of feature vectors of BBB and MI ECG signal from lead III using the derivative-based method