Electrocardiogram beat classification using deep convolutional neural network techniques

Electrocardiography (ECG) is utilized as a standard means for monitoring and diagnosing cardiovascular diseases and it has an important role in routine clinical practice. Using this non-invasive tool, the electrical cardiac activities covering the depolarization and repolarization patterns of the heart can be observed on an ECG trace [1]. An ECG trace reflects the time-varying signal changes that depend on the ionic current flow that causes the cardiac fibers to contract and subsequently relax. The changes in potential between the electrodes placed on the surface of the skin can be observed via ECG [2]. A single and routine period of an ECG presents the sequential atrial and ventricular depolarization as well as repolarization. These continuous cycles consist of significant peaks that reflect the morphological structures underlying specific peaks, denoted by P, Q, R, S and T [3]. This means that a heartbeat creates a specific pattern on the ECG trace. This specific pattern occurs when the signal moves away from the baseline heart rate and then returns to the baseline heart rate level. A single sinus rhythm of the ECG traditionally includes the above-mentioned peaks, as shown in figure 12.1.

Zoom In Zoom Out Reset image size

The depolarization of the atria, which causes depolarization waves to spread from the sinoatrial (SA) node to the atria, increases the voltage measured from the heart surface, causing it to diverge from the baseline heart rate, and this wave is referred to as a P-wave [4]. The time between the events of atrial depolarization and ventricular depolarization is the PQ-interval. The QRS-complex is one of the most vital patterns in computerized heart rate analysis and it is frequently used for basic heart rate analysis. The QRS-complex produced the largest amplitude in an ECG cycle [5]. The ventricles become depolarized before their contractions and this depolarization forms the QRS-complex. Atrial repolarization appears just before the QRS-complex. However, the following wave pattern, which is referred to as the QRS-complex, has a wider amplitude. Therefore, atrial repolarization cannot be observed on an ECG trace [6]. The time between the beginning of the ventricular depolarization and end of the ventricular repolarization overlaps the QT-interval on the ECG trace. It has been shown that the QT-interval increases due to the linear increase of the RR-interval in experimental studies. Also, a prolonged QT-interval has been related to late ventricular repolarization and this cardiac ailment may lead to sudden death [7]. The time interval between the S-wave and T-wave is referred to as the ST-interval. High or depressive patterns far from the baseline are consistently considered as symptoms that indicate heart diseases. The relief time elapsed until the next ECG cycle, in other words, the ventricular repolarization, is shown as a T-wave on the ECG signal [8]. ECG signals can have undesired artefacts for different reasons, such as patient or electrode movements, displacements of the electrodes and power-line noise [9]. To achieve useful clinical information, noisy ECG signals must be converted into reliable signals using preprocessing procedures [10, 11]. To this aim, many algorithms have been proposed for R-peak detection [12–14], QRS detection [15–18], QT-interval detection [3, 19], RR-interval detection [20–23] and ST-interval detection [8, 24, 25] to provide useful clinical information.

Cardiac arrhythmia is a medical term and corresponds to a condition which reflects irregularities in heart rate rhythm, where it may be quite slow or fast [1]. Disruptions of the electrical impulses associated with coordination of the heartbeat cause arrhythmia. In other words, arrhythmia is defined as discomfort that includes rhythm irregularity or conduction problems in electrical impulses through the heart. It can occur in the upper or lower chamber of the heart. Arrhythmias of the ventricles may lead to life-threatening events. Various types of arrhythmias, which demonstrate quite different characteristics, have been defined [26]. Each of these types of arrhythmias is associated with a pattern whose morphological characteristics have been identified. Therefore, it is possible to define and classify arrhythmia types using a computer-aided approach. The arrhythmias are categorized into two basic classes. The irregularities associated with the electrical conductivity of the heart is the first class. This category of arrhythmia corresponds to arrhythmias caused by a single irregular vessel shot. The other category includes arrhythmias consisting of a series of irregular heartbeats. These irregularities cause changes in morphology or wave shape and frequency. A large number of these changes can be observed though ECG examination. A normal heart rhythm where each specific peak can be observed properly is defined as the normal sinus rhythm (NSR). During NSR, the triggering pulses from the sinoatria spread in a coordinated manner across the four chambers of the heart. In addition to the NSR, there are also different common arrhythmia types. The source of the NSR beats is the SA node. The regular heart rate can vary due to autonomous inputs occurring in the SA node. Premature beats are a widely observed cardiac disease and are harmless. Patients with this type of arrhythmia feel a beating or missing heartbeat in the chest. Premature beats also include premature atrial contractions (PACs) and premature-ventricular contractions (PVCs) [27]. Supraventricular arrhythmias occur in the atria, with fast heart rates, and some of them are atrial fibrillation (AF), atrial flutter, paroxysmal supraventricular tachycardia (PSVT) and Wolff–Parkinson–White syndrome [28]. Ventricular arrhythmias (VAs) include ventricular flutter, ventricular tachycardia (VT) and ventricular fibrillation (VF). These types of arrhythmias are life-threatening and therefore require quick intervention [29]. Sinus node dysfunction is related to problems associated with the SA node. A slow heart rhythm is observed in this arrhythmia and patients use a pacemaker to tackle this type of arrhythmia [30]. Heart block problems may appear in the AV node or HIS Purkinje system because of incomplete electrical activity [31]. Slow and irregular heartbeats are then observed. To tackle this condition, using a pacemaker is recommended. Arrhythmia diagnosis based on ECG beats has great importance for cardiologists [32]. A broad range of cardiac conditions covering arrhythmia and HRV can be observed on ECG traces [33]. However, even for specialists in the field, it is very difficult to examine, define and categorize a large number of ECG beats. In addition, the possibility of missing vital information is high due to the large scale of the data. As a practical solution, computer-aided systems have been adopted and have become quite important in daily clinical practice [1]. In this view, automatic ECG beat classification has been considered as an important issue in addressing cardiac ailments.

In computer-aided approaches, the two basic procedures of feature extraction and classification are common. The performance of such systems is sensitive and depends on specifically designed features as well as the fine-tuning of the classifiers. Time-domain features have been used for ECG analysis and have been adopted as a useful approach. In addition, the frequency domain holds great clues regarding HRV. Furthermore, the time–frequency domain have been used to describe large-scale ECG data [34]. The morphological and higher-order statistic [35] diagnostic indices have been measured using advanced digital signal processing techniques in the feature extraction stage, such as principal component analysis (PCA) [3, 36, 37], wavelet transforms (WT) [38, 39], Hilbert transforms [40], cross-correlation approaches [41] and the Kalman filter [42]. Support vector machines (SVMs) [43] have ensured high general performance. The extreme learning machine (ELM) has contributed to shortening the training time [44]. Similarly, the k-nearest neighbors (k-NN) method [38], linear discriminant analysis (LDA) [45], artificial neural network (ANNs) [46] and probabilistic neural network (PNN) [47] have also been used for classification purposes. Optimization techniques such as particle swarm optimization (PSO) [48], artificial bee colony [49] and genetic algorithms [50] have been embedded into the proposed models for automatic ECG beat classification tasks [33]. The permutation entropy (PE) and the conditional entropy of ordinal patterns (CEOP) have been applied to ECG signals to realize a binary classification [51]. Deterministic learning schema have been applied to the ECG beat classification task. To this aim, the proposed model exploited the dynamics structure of the ECG beats as a unique feature. In this approach, the classification was carried out without using any feature extraction procedure for the test beats. Also, the model presented good generalization capacity [33]. The enormous differences between individual patients were addressed in another study focused on ECG beat classification. An optimization mechanism was proposed to enhance the achievements of the model. Recurrent neural networks (RNNs) have been employed to learn features automatically and promising results have been reported [52]. To prevent sudden cardiac death, a computer-aided diagnosis model based on spectral entropy and a deep convolutional neural network (DCNN) was introduced. The time–frequency transform and entropy measurement were dealt with in this study. Furthermore, two-directional two-dimensional PCA was utilized to reduce the number of features used. The study points out that the spectral entropy can provide good separation between different ECG beats [34]. A novel model was built on the SVM with rejection schema. Traditional procedures, such as preprocessing, feature extraction and classification, applied in machine learning approaches were followed in the proposed model. As a result, a cost-sensitive classifier was provided [53]. Vector quantization (VQ) has been used in ECG beat classification, since this method increases the classification accuracy while reducing the dimension of the feature set. To this aim, a novel dictionary learning algorithm was presented. This study resulted in an accelerated algorithm and a model with increased accuracy [54]. The ECG beat classification task has also been handled as a binary classification task between normal and abnormal. A method relying on the radial basis function (RBF) system for canceling ectopic beats was proposed. The model yielded good results in terms of the low distortion of the signal recordings [55]. A neuro-fuzzy network model fed with autoregressive model coefficients, higher-order cumulants and the WT has been used for ECG beat classification. The proposed model provided a significant performance enhancement in ECG beat classification [56]. In addition to the normal beat type, four different beat types were addressed in a novel model exploring RNNs trained with the Levenberg–Marquart algorithm. The combined eigenvector method and RNN was reported as a useful tool for ECG beat classification [57]. A combination of PSO and RBF neural networks was used to separate six different ECG beat types. The morphological features were taken into account in the model. As a result, an efficient diagnostic model was provided [58]. Six discriminative ECG beat classification problems were addressed using a PNN and WT. The model was fed with only eleven features and an effective and efficient model was achieved [59]. A novel model, called a modified mixture of experts (MME) network, was used in ECG beat classification with diverse features and a supervised learning approach with the expectation–maximization (EM) algorithm was used in the model. The model used handcrafted features extracted using wavelet coefficients and Lyapunov exponents. The proposed MME approach was found to be a useful tool in categorizing ECG signals for early heart abnormality detection [60]. An automated medical diagnostic tool based on a cross-correlation and least square support vector machine (LS-SVM) was developed for ECG beat classification which covers normal beats, PVC beats and other beats. The performance of the proposed model in this three-class classification scheme was reported to have 95.51% accuracy [61]. In another study, ECG beat classification was performed using a novel independent components (ICs) arrangement strategy and independent component analysis (ICA). The PNN and SVM classifiers were employed to evaluate the proposed method. Eight different ECG beat types were considered in the experiment and a 98.7% classification accuracy was achieved [62].

The ECG beat classification task has mainly been addressed in two different ways: the global-classifier and patient-adapting approaches. In the first approach, the experience of specialists is not considered and the classification is performed automatically [52]. The desired performance has not been achieved in this approach since a large-scale inter-individual variability has been observed in ECG waveforms. The second approach focuses on the ECG beat classification task by considering the specific characteristics of each patient's ECG recording to enhance the accuracy of automated systems [51]. However, this enhancement has remained limited since there are different temporal and physical conditions for patients as well as variations in the morphological characteristics of ECG waveforms. ECG signals generally display their characteristic features periodically. Sometimes differences arise in periodic ECG beats and sometimes in the morphological structure of the ECG. These enormous time-varying patterns are marked as temporal or dynamic [63]. For this reason, the recognition of this temporal pattern is very difficult. Also, the presence of notable similarities between different temporal ECG beats is a challenge that remains to be solved [64].

As mentioned above, many different combinations of advanced signal processing and machine learning techniques have already been proposed for the ECG beat classification task [65–67]. Recently, a novel approach called the deep convolution neural network (CNN), a sub-branch of machine learning, has become attractive due to its generalizability on large-scale real-world problems [68]. The great advantage of CNN models is the end-to-end learning scheme [69]. In this way, the tedious and complex feature extraction step is eliminated and the features representing the characteristic and discriminative features of the input data can be obtained using activation maps [70]. Although CNN models require large-scale data and high calculation costs for the training, they have been adopted as an efficient method due to the highly generalizable performance provided.

In the proposed approach, pre-trained CNN models are used and the fine-tuning procedure is performed to transfer domain knowledge. Then, the deep features extracted from the activation maps in the deep CNN architecture are input to one of the traditional machine learning algorithms called SVMs. In the training of the CNN models, raw ECG beat images are considered and five ECG beat types, namely normal (N), left-bundle-branch-block (LBBB), right-bundle-branch-block (RBBB), premature-ventricular-contraction (PVC) and paced beats (PBs) are taken into account in the experiments [71]. As a result, a novel ECG beat classification diagnostic tool with high sensitivity is produced via the proposed model.

The rest of this chapter is organized as follows: the dataset description and methods are provided in section 12.2; the experimental work and results are reported in section 12.3; a comparison is presented in section 12.4; and, lastly, concluding remarks are presented in section 12.5.

12.2.1. The MIT-BIH database

12.2.2. Producing ECG beat images

In this study, as one-dimensional (1D) ECG record signals are converted into two-dimensional (2D) images, we need to crop each ECG beat region and create ECG beat images as a separate input into the machine learning approaches. To do this, we fix the ECG beat image size in a rectangular shape. The first step is to obtain the signal values of the ECG beats by reading the record files. The format of the record files is comma-separated values (CSVs). Each record file has nearly 650 000 records for the 30 min duration and 360 records per second. To label the ECG beat type, we need to read annotate files that have the 'txt' file format. The second step is to determine what the ECG beat image size should be. While producing the ECG beat image, we need a reference point in the signal. The peak value of each signal was selected as a reference point and the width of the ECG beat image is determined by the sum of two parts of the signals. The first part is the points which are before the peak value and the second part is the points which are after the peak value. While producing the ECG beat image, 90 signal values are used for the part before the peak value and 150 signal values for the part after the peak value to define the signal as lossless.

The flowchart for producing the ECG beat images is shown in figure 12.2. The preprocessing block of the flowchart in figure 12.2 has parameters such as sig1, wden, sym8, peak values, beat types and size of the images. sig1 is the raw signal of a patient record, wden is a function that denoises the raw signal using WT, sym8 is the name of the wavelet function, lev is the level of wavelet function and its value is six in this process, peak value and beat type are obtained through annotate files. The peak value is the highest point in each beat and the beat type is the label.

Zoom In Zoom Out Reset image size

There are 23 symbols including beat and non-beat types in the MIT-BIH ECG database. The symbols of the beat types and non-beat types are !, ], ∣, [, ∼, +, A, a, E, e, F, f, J, j, L, . or N, / or p, Q, q, R, S, V and x. The beat types used in this study are L, N, p, R and V. They are called the left-bundle-branch-block (LBBB) beats, normal (N) beats, paced beats (PBs), right-bundle-branch-blocks (RBBB) beats and premature-ventricular-contraction (PVC) beats. These five beat types are selected due to the huge number of their occurrence. Sample beat type images that are produced using the preprocessing block are shown in figure 12.3.

Zoom In Zoom Out Reset image size

In total, 8001 LBBB (symbol L) beat samples, 8067 N (symbol N) beat samples, 7018 PB (symbol P) samples, 7247 RBBB (symbol R) beat samples and 7125 PVC (symbol V) beat samples were used in the experiments to evaluate the proposed method.

12.2.3. Convolutional neural networks (CNNs)

The deep learning approach consists of multiple processing layers combined with multiple abstraction structures to learn the representations of the data. Various preprocessing procedures, dimension reduction and feature selection are performed to reveal these features [73]. To reduce the cost at this stage, it is necessary to avoid the dependence on features. In this way, when designing classifiers and other prediction systems, it will be easier and less costly to extract useful information (discriminative features) from data using artificial intelligence [74]. In this context, the subject of deep learning is closely related to representative learning. Deep learning algorithms have a wide range of applications and have been applied successfully in many different fields. Deep learning is considered a sub-branch of machine learning and performs feature extraction transformation procedures through non-linear layers. The layers feed each other, in other words, the output of one layer can be the input of another layer. Deep learning algorithms can either be supervised or unsupervised and these algorithms generally have a non-linear learning structure. Thus, the representation of the data is performed using distinctive features obtained at different levels [75]. There is a hierarchical learning representation in deep learning and this representation is based on the abstraction of layers at different levels. The data applied as an input in deep learning are passed through a series of processes and learning is realized. Convolution is performed to reveal the distinctive features of the data. This process results in activation maps representing local distinctive features. The power of activation maps to represent data may vary. In other words, some distinctive features may be retained in different activation maps.With layers such as pooling and dropout, calculation costs are reduced and overfitting is prevented [76].

The convolution layer, pooling layer, normalization and fully connected layer form the basic CNN model [77]. Layers are used consecutively to create numerous models for many different tasks. The convolution layers are operated to reveal the discriminative features in the input data. In the equation below, ${X}_{i}^{l-1}$ indicates the features from the previous layers, ${k}_{{ij}}^{l}$ displays the kernels that are learnable and the bias is represented by ${b}_{j}^{l}$ [78]. The output values produced by the feature map are calculated as follows:

$$\begin{eqnarray}&&{X}_{j}^{l}=f\left({\sum }_{i\in {M}_{j}}{X}_{i}^{l-1}\,* \,{k}_{{ij}}^{l}+{b}_{j}^{l}\right),\end{eqnarray} \tag{ 12.1 }$$

where the input map selection represented by ${M}_{j}$ and $f(.)$ indicates an activation function. With the pooling layer, the calculation cost is reduced and different approaches, such as average pooling or maximum pooling, can be used for this purpose. In CNN models, another advantage of pooling layers is to prevent overfitting. This process is also supported by layers such as dropout layers. In these layers, the downsampling is performed as defined in

$$\begin{eqnarray}&&{{\rm{X}}}_{j}^{l}=\mathrm{down}\left({X}_{j}^{l-1}\right).\end{eqnarray} \tag{ 12.2 }$$

Here in the related procedure is symbolized by the function 'down'. It can be stated that the number of local distinctive features is reduced [79]. The activation maps processed in previous layers are combined in fully connected (FC) layers. The deep features obtained as a result of all the operations applied in the deep architecture are combined in the FC layers [80]. Optimization algorithms play an important role in the training of deep networks, and their use is inevitable for proper learning to be achieved. Equation (12.3) below represents an optimization algorithm. For optimization purposes, stochastic gradient descent with momentum (SGDM) and adaptive moment estimation (ADAM) are preferred frequently due to their performance on real-world problems [81]:

$$\begin{eqnarray}&&{V}_{t}=\beta {V}_{t-1}+a{\nabla }_{w}L(W,X,y),\end{eqnarray} \tag{ 12.3 }$$

where $W$ denotes the weights, and $L$ and α show the loss function and learning rate, respectively. During the training of the CNN, the new weights are computed as follows:

$$\begin{eqnarray}&&W=W-a{V}_{t}.\end{eqnarray} \tag{ 12.4 }$$

The ADAM optimization algorithm performs weights within each iteration, taking into account the average of the second momentum. It also uses the average first moment in the RMSProp method [82].

12.2.4. Deep transfer learning (DTL)

12.2.5. Support vector machines

The SVM is the most popular traditional supervised machine learning approach and is used by researchers all over the world [87]. Due to its easy implementation and robustness, it is possible to come across an SVM in almost all machine learning-based classification studies. The basic principle of the SVM is to linearly separate two classes from each other by drawing a margin (a line, a plane or a hyperplane) between them. The goal of the SVM method is to maximize the margin between two classes [88]. When the datasets become non-linear, researchers solved this issue by adding some new kernel structures to the SVM to overcome these non-linear datasets. Some of these SVM kernels are linear, quadratic, cubic, fine Gaussian, medium Gaussian and coarse Gaussian.

12.2.5.1. The linear SVM kernel

$$\begin{eqnarray}&&K(x,y)={x}^{T}y+c,\end{eqnarray} \tag{ 12.5 }$$

where K denotes the kernel, x is the input vector (the feature vector of one class), y is the input vector (the feature vector of another class), c is a constant term and T denotes the transpose. The terms defined in equation (12.5), i.e. $K,{x},{y},{c},{T}$, are also the same for equations (12.6)–(12.9).

12.2.5.2. The quadratic SVM kernel

$$\begin{eqnarray}&&K(x,y)={(\alpha {x}^{T}y+c)}^{2},\end{eqnarray} \tag{ 12.6 }$$

where α denotes the slope term. The quadratic SVM kernel is a type of polynomial SVM that has a polynomial degree of 2.

12.2.5.3. The cubic SVM kernel

$$\begin{eqnarray}&&K(x,y)={(\alpha {x}^{T}y+c)}^{3},\end{eqnarray} \tag{ 12.7 }$$

cubic SVM is a type of polynomial SVM that has a polynomial degree of 3.

12.2.5.4. The fine Gaussian SVM kernel

$$\begin{eqnarray}&&K(x,y)={\rm{\exp }}\left(-\frac{{\parallel x-y\parallel }^{2}}{{2\sigma }^{2}}\right)\end{eqnarray} \tag{ 12.8 }$$

$$\begin{eqnarray}&&K(x,y)={\rm{\exp }}{(-\gamma \parallel x-y\parallel }^{2}),\end{eqnarray} \tag{ 12.9 }$$

where γ depends on σ which is the Gaussian term. The fine Gaussian SVM kernel scale set is the result of the square root of P divided by 4, where P is the number of predictors. Equation (12.9) includes all types of Gaussian SVM kernels, such as fine, medium and coarse.

12.2.5.5. The medium Gaussian SVM kernel

The medium Gaussian SVM kernel scale set is the square root of P, where P is the number of predictors.

12.2.5.6. The coarse Gaussian SVM kernel

The coarse Gaussian SVM kernel scale set is the result of the square root of P multiplied by 4, where P is the number of predictors.

12.2.6. Performance metrics

The most basic tool used to assess the performance of a model is the confusion matrix, also called the error matrix. This matrix includes the true-positive, false-positive, false-negative and true-negative parameters. Using these indices, various performance metrics such as accuracy can be calculated. In this study, only the accuracy metric is taken into account and this reflects the system's ability to accurately estimate the ECG beats. It can be calculated as follows:

$$\begin{eqnarray}&&\mathrm{Accuracy}=\frac{\mathrm{NoTCB}}{\mathrm{NoACB}}\end{eqnarray} \tag{ 12.10 }$$

$$\begin{eqnarray}&&\mathrm{NoTCB}=\mathrm{NoACB}-\mathrm{NoFCB},\end{eqnarray} \tag{ 12.11 }$$

where NoTCB is the number of true classified beats, NoACB is the number of all classified beats and NoFCB is the number of false classified beats. All the numbers of classified beats belong to only one class.

A workstation equipped with the NVIDIA Quadro M4000 GPU was used in the experiments. The MATLAB software on the MIT-BIH arrhythmia database was considered in all coding. Five ECG beat types namely normal (N) beats, left-bundle-branch-block (LBBB) beats, right-bundle-branch-block (RBBB) beats, premature-ventricular-contraction (PVC) beats and paced beats (PBs) were used in the classification stage of the work. The MIT-BIH arrhythmia database contains 40 ECG records and all records were used in the experiments. In the experiments, 8067 N beat samples, 8001 LBBB beat samples, 7247 RBBB beat samples, 7125 PVC beat samples and 7018 PB samples were used.

The ECG signals were converted into ECG images by taking into account the annotations prepared by the experts. The signal-to-image conversion was handled using a code that was written by the chapter authors. First, the resolutions of the constructed images were adjusted to 224 × 224 px and then were converted to gray-scale images. Five well-known pre-trained deep CNNs, called VGG16, VGG19, ResNet18, ResNet50 and ResNet101, were considered in the experiments. In addition, five SVM methods, namely the linear SVM, quadratic SVM, cubic SVM, medium Gaussian SVM and coarse Gaussian SVM were used in the classification. In the linear SVM, the linear kernel function was used and the epsilon value of the SVM method was set to 0.01. In the quadratic SVM, the quadratic function was used as the kernel function and the epsilon value was adjusted to 0.2. The cubic kernel function was used in the cubic SVM and the epsilon value was assigned as 0.02. For both the medium and coarse Gaussian SVMs, the Gaussian kernel function was used and the epsilon parameters were adjusted to 0.1 and 0.2. The performance of the proposed method was evaluated by using the accuracy score. A ten-fold cross-validation procedure was also performed in the experiment to yield objective results. At the end of the ten-fold cross-validation procedure, the mean accuracy scores were calculated. Table 12.1 shows the obtained accuracy scores. The rows of table 12.1 show the SVM classifier types, while the columns show the pre-trained CNN models. The last row and column show the average accuracy scores.

From table 12.1, it is seen that only the VGG16 and VGG19 models produced average accuracy scores above 99.0%. VGG16 achieved 99.32% and VGG19 produced a 99.27% average accuracy score. In addition, the ResNet18 model obtained a 98.98% average accuracy score, and the ResNet50 and ResNet101 models both produced 98.65% average accuracy scores. The highest accuracy score of 99.6% was also produced by both the VGG16 and VGG19 models. When the obtained results are evaluated from the viewpoint of the classifiers, it is seen that the best mean accuracy score of 99.42% was yielded by the cubic SVM classifier. The second best mean accuracy score of 99.40% was yielded by the quadratic SVM, and the fine and medium Gaussian SVMs produced 99.18% and 99.16% average accuracy scores, respectively. Linear SVM produced a 99.00% average accuracy score and, finally, the worst average accuracy score of 97.68% was produced by the coarse Gaussian SVM classifier.

Figure 12.4 shows the confusion matrix for the ResNet50 features and the linear SVM classifier. The labels 1, 2, 3, 4 and 5 show N, LBBB, RBBB, PVC and PB, respectively. From figure 12.4, it is observed that 171 N beat samples were classified incorrectly. The correct classification rate of the N beat samples was 97.88%. 142 LBBB beat samples were misclassified and the accuracy of the LBBB beats was 98.23%.

Zoom In Zoom Out Reset image size

The obtained correct classification rate for RBBB was 99.37% and only 46 misclassified RBBB samples can be seen in figure 12.4. The highest misclassification was obtained for PVC beats. In total, 214 PVC beat samples were classified incorrectly and the accuracy of the PVC beat was 97.00%. Finally, the highest accuracy score of 99.52% was obtained for the PB class. Only 34 samples were misclassified. Figure 12.5 shows the distribution of ResNet50 features. The x-axis shows the first column of the ResNet50 features and the y-axis shows the second column of the ResNet50 features.

Zoom In Zoom Out Reset image size

Figure 12.6 illustrates the training period of the fine-tuning of the ResNet50 model. The top panel of figure 12.6 displays the training and testing accuracies (blue for training and black for testing) and the bottom panel of figure 12.6 shows the loss values for both the training and testing samples (orange for training and black for testing). Table 12.2 provides the results of the fine-tuning of the pre-trained deep CNNs on ECG beat categorization.

Zoom In Zoom Out Reset image size

As seen in table 12.2, all the fine-tuned deep CNN models achieved classification accuracy scores above 99.00%. The highest accuracy score of 99.66% was yielded by the VGG19 model and a 99.49% accuracy score, which was the second best result, was obtained by the ResNet101 model. Accuracy scores of 99.39%, 99.36% and 99.04% were obtained by the ResNet18, ResNet50 and VGG16 models, respectively.

In the final experiments, a novel CNN model was constructed and trained in an end-to-end fashion by the chapter authors and used in ECG beat classification. The developed CNN model is depicted in figure 12.7 and, as can be seen, the developed CNN model was composed of 15 layers. The network started with the input layer and there were three convolution layers namely conv_1, conv_2 and conv_3 in the model, and batch normalization, ReLU and pooling layers followed each convolution layer. A fully connected layer, softmax layer and classification layer were also used for the classification of the ECG beats. The conv_1, conv_2 and conv_3 layers contained 64, 32 and 18 filters of size 3 × 3. The max operator was used in the pooling layers.

Zoom In Zoom Out Reset image size

The layers of the proposed architecture and the parameters and values used with the layers are given in table 12.3. The SGDM was chosen as the optimization algorithm. The initial learning coefficient was set to 0.01 and 20 was used for the batch size value. The network trained more than 1200 iterations.

The end-to-end learning scheme of the proposed CNN model is given in figure 12.8. The top panel of figure 12.8 shows the training and testing accuracies (blue for training and black for testing) and the bottom panel of figure 12.8 demonstrates the loss values for both the training and test samples (orange for training and black for testing). The obtained accuracy score was 98.68% and the training procedure was completed in 876 iterations.

Zoom In Zoom Out Reset image size

A comparison is presented in table 12.4, taking into account the methods, databases and results obtained in studies related to ECG beat classification.

Thomas et al [89] used the dual-tree-complex WT and MLNs for the classification of cardiac arrhythmias. Five types of ECG beats were considered. The model followed three key stages that wer: preprocessing, feature transformation and classification. Amplitude normalization and band-pass filtering were applied to the raw ECG signals in the preprocessing stage. Dual-tree complex wavelet-based features and four other extracted features were used in the feature set. Finally, the MLNs were employed to determine the five different ECG beat types. The model achieved 97.68% classification accuracy. Kaur et al [90] used wavelet coefficients and feature reduction methods for efficient cardiac dysfunction classification. The WT was computed with db4. All WT coefficients at levels 4, 5, 6 and 7 were taken into account and a total of 43 coefficients constituted the feature vector. Then, factor analysis was utilized to reduce the size of the feature set without rotation and with orthogonal rotation. In this way, the number of features was decreased and 23 features were obtained. The model achieved 99.06% accuracy. Rai et al [91] used Daubechies WT and RBFNN for five types of ECG beat classification. A preprocessing step covering noise elimination, baseline wander removal, R-peak detection and QRS-complex extraction was applied to the raw ECG signals. Each ECG sample was represented by 21 points, which were input to the model. In this study, the performances of the BPNN, MLP and SVM classifiers were also examined and the model yielded a 99.60% classification score. Khalaf et al [92] used spectral correlation and the SVM classifier for cardiac arrhythmia classification. For this purpose, the spectral correlation was utilized. The proposed system used Otsu's thresholding, and statistic features which included PCA and Fisher scores. The SVM classifier was used to determine the ECG beat types and the model produced a 98.60% classification accuracy. Dong et al [33] introduced a new approach based on the deterministic learning algorithm. This system ensured different features and the model was performed on the MIT-BIH database. The main advantage of the system was that no feature extraction procedure was required for the test beats. The model yielded a 97.78% classification score. Osowski and Ling [35] carried out ECG beat classification based on higher-order statistical features and a fuzzy neural network. The model produced an accuracy of 98.00% using a multi-layer perceptron and Fourier analysis. Zidelmal et al [53] used an SVM with rejection. The frequency information, RR intervals, QRS morphology and AC power of QRS detail coefficients were also considered in the analysis. The beats were collected into two classes and the model achieved 98.80% classification. Liu et al [54] used the vector quantization method for classifying eight different ECG beat types. To de-noise the raw ECG signal, a median filter was used in the preprocessing step. Wave boundary detection was also performed. The dictionary learning algorithm was employed to describe the signals. As a result, the model achieved 94.50% classification success using VQ k-medoids. Mateo et al [55] introduced a system using RBFNN to cancel out ectopics.

As seen in table 12.4 and mentioned in the studies [35, 53–56, 89–92], ECG beat classification has been realized using advanced signal processing techniques that include WT, high-ordered statistical features, PCA, ICA, VQ, factor analysis, PSO and specific morphological feature extraction processes. These approaches mainly followed three key stages: preprocessing, feature transformation (feature extraction and feature selection) and finally classification. The quality of both the preprocessing and feature extraction steps is critical to ensuring a consistent and robust diagnostic tool, since the performance of the next steps of the model depends on these inital steps. In other words, the selected features have a great impact on the performance of the classifiers.

Asgharzadeh-Bonab et al [34] proposed a model based on the spectral entropy and deep CNN. The noise reduction and QRS segmentation were applied as the preprocessing procedures in the proposed model. Then the model followed time–frequency analysis, feature reduction, and classification steps. The images obtained using the two-directional two-dimensional principal component analysis were applied as the input to deep CNN. The model ensured an average accuracy of 98.33%. Wang et al [52] proposed a model fed with morphological and temporal vectors and uses a global recurrent neural network. The model applied a single RNN for automatic feature learning and classification. An end-to-end learning schema was ensured in this study. As a result, a diagnostic tool with high accuracy was ensured. Li et al [93] offered a deep residual CNN model from 2 lead ECG. The efficient features were learned directly ECG samples. The 1D convolutional layer was used in the architecture. As a result, a successful model was ensured. Acharya et al [94] introduced two architectures of CNN performed on the three different databases. They did not consider the QRS-complex in the experiment. Also, feature extraction and feature selection were not involved. The model achieved an accuracy of 94.90%. Yao et al [95] suggested an attention-based time-incremental CNN model. The model performed on a database consisting of a normal and eight different beat types and ensured 81.20% classification accuracy.

Considering the related studies [34, 52, 93–95], it is seen that in addition to the models following traditional preprocessing, feature transformation and classification processes, deep CNN models have also been used to classify different ECG beat types, thanks to their efficient end-to-end learning schemes. In this context, it has been observed that both deep CNN models, LSTM-based models, ResNet models and RNN models have been used efficiently. Deep CNN networks push the raw data, which is received as inputs, forward to many non-linear layers throughout the network. In this process, the deep CNN models perform the learning and classification processes directly on the input data using functions such as convolution, pooling and dropout. Thanks to this end-to-end learning scheme, the labor-intensive extraction process can be eliminated from the diagnostic models. Another great feature of the deep CNN models is that they offer a hierarchical learning scheme. Activation maps obtained from different layers can reveal many distinctive features of the input data. Sharing weights in the network also provides stronger generalization performance. In the proposed method, the performance of the pretrained CNN models was first examined to classify the ECG beats. Later, instead of extracting features manually, the deep features obtained from the activation maps were inputted into the SVM classifier and quite efficient results were obtained. In addition to all their advantages, deep CNN models need a lot of data for proper training. As this experimental study focused on ECG beats, sufficient data were available. However, for many clinical applications, accessing such a large-scale dataset may not be possible or may take a lot of time. In addition, sufficient hardware resources with high capacity must be available for the training of deep CNN models. Moreover, compared to traditional machine learning methods, deep CNN models take a longer time to train.

The vast majority of studies were carried out on the MIT-BIH database. In addition, different databases are also available through PhysioNet for ECG beat classification. In the studies, it can be observed that five classes are generally used for ECG beat classification problems, and binary classification has also been realized by separating the beats as normal and abnormal. Moreover, some researchers evaluated the problem as having four, seven or eight classes. These differences also show that it is not possible to make a one-to-one comparison. In related studies, it is seen that the classification performance is generally over 95%. It is also worth noting that there is a significant increase in classification success with the RNN and CNN based methods that use end-to-end learning schemes. The classification success has been over 99% using these models. Asgharzadeh-Bonab et al [34] achieved a 99.33% classification score while Wang et al [52] reported an accuracy of 99.80%. Similarly, for the proposed model we achieved 99.66% classification. As a result, from table 12.4, it is observed that the proposed method outperformed most other state-of-the-art methods.

In this chapter, three deep CNN approaches are considered in ECG beat classification. These approaches achieve an end-to-end learning scheme of a new CNN model, fine-tuning of pre-trained CNN models, and extraction of deep features and their classification using a traditional or deep machine learning approach such as SVM. The obtained results present the following conclusions:

• 1.  
  Saving ECG beats as images and applying the CNN approach to them for classification purposes produced quite efficient results. All deep CNN based approaches produced reasonable accuracy scores.
• 2.  
  Fine-tuning of a pre-defined model and deep feature extraction produced better results than end-to-end training of the new CNN model. This might have happened because the pre-trained models were trained on 25 million images.
• 3.  
  The cubic kernel function produced better accuracy scores than the other kernel functions.

In future works, dense CNN models will be developed and trained on ECG beat classification. In addition, time–frequency images are considered for comparison with the presented results.