Cardiovascular disease is a common disease that seriously threatens human health, especially the health of middle-aged and older people. It is characterized by high prevalence, high disability, and high mortality. Nowadays, the world is facing with the aging population. The increasing aggravation of cardiovascular disease has become a major public health problem (Mc Namara et al., 2019). ECG analysis is an effective way of evaluating heart health. Therefore, the identification and classification of ECG signals are essential to cardiovascular diseases. Not only for early prevention but also necessary for timely detection and proper treatment. It is of considerable significance to study the classification of related ECG signals (Guo et al., 2016; Yin et al., 2016).

The electrocardiogram is a visual time series that records the electrical activity generated by each cardiac cycle of the heart in real-time and is now widely used in heart rate detection (Homaeinezhad et al., 2012). This non-invasive detection method is easy to operate and has become an essential tool for assisting doctors in analyzing pathology. At this stage, the judgment of cardiovascular disease mainly depends on the doctor's experience. However, there are many types of heart diseases, and long-term manual detection makes it easy to cause false detection. How to quickly and accurately analyze specific diseases has become a new problem (Song et al., 2014). In addition, the traits of ECG signals include random, low-frequency, and susceptible, resulting in the diagnosis results are unstable. Intelligent automatic recognition and classification of ECG signals have become an inevitable choice to improve the efficiency and accuracy of ECG recognition (Yu and Chen, 2007).

With the maturity of Artificial Intelligence (AI) technology, many machine learning methods are used in the ECG signal feature detection, aiming at solving the problems related to large amounts of ECG signal feature data and a heavy load of manual detection. The typical methods are neural networks (NN) (Jiang and Kong, 2007), support vector machine (SVM) (Osowski et al., 2004), path forest (Luz et al., 2013), Independent Component Correlation Algorithm (ICA) (Sarfraz et al., 2014). Regarding neural networks, Jiang and Kong (2007) propose an evolutionary block-based neural network (BbNNs) for the classification of ECG signals. The BbNN is composed of a group of two-dimensional modular networks with flexible structure and internal configuration. The Hermite transform coefficient and the time interval between adjacent two R peaks are used as the input of BbNN. Osowski et al. (2004) introduce Hermite function as a feature extraction method in the SVM classifier, and also use higher-order statistics (HOS) to better extract features. Luz et al. (2013) introduced an optimal path forest classifier (OPF) to compare the performance of 6 distance metrics, six feature extraction algorithms, and three classifiers in two variants of the same data set. Although the accuracy of OPF is not as good as that of SVM, OPF is more efficient than SVM in terms of calculation time during training and testing phases. Sarfraz et al. (2014) use the basic function of a typical ECG signal obtained by ICA for pattern recognition. The features obtained by ICA are used as the training set. Although these methods have good performance and have achieved certain results, it is difficult to put into practice due to the severe deficiency of requiring manual design features.

In recent years, the machine learning and deep learning network has not only made remarkable achievements in the fields of image processing, audio recognition and many other fields (Wong et al., 2015a,b, 2016; Kandala et al., 2019; Pławiak et al., 2019, 2020), it has also been commonly used in the assisted diagnosis of heart disease based on ECG signals(Zubair et al., 2016; Acharya et al., 2017a,b; Yildirim et al., 2018; Gao et al., 2019; Atal and Singh, 2020; Pławiak and Acharya, 2020). Pławiak and Acharya (2020) use a deep genetic ensemble of classifiers to classify long-duration ECG signal (10 s). Gao et al. (2019) implemented an effective long short-term memory (LSTM) recurrence network model to classify 8 types of heartbeats. Atal and Singh (2020) proposed an optimization-based deep convolutional neural network to classify five different heartbeats. Compared with traditional neural networks, deep learning network can automatically extract features, recognize intricate data patterns, and eliminate complex signal preprocessing. Deep learning network also has a stronger nonlinear fitting ability, which has a better effect in identifying single-lead, multi-class, and unbalanced ECG datasets (Acharya et al., 2017a). Convolutional neural network (CNN) is a feedforward neural network that has been widely researched and used in deep learning, which has been applied successfully for the classification of arrhythmia (Zubair et al., 2016; Acharya et al., 2017a,b; Yildirim et al., 2018; Gao et al., 2019; Atal and Singh, 2020; Pławiak and Acharya, 2020) ECG signals. In the previous literature (Zubair et al., 2016; Acharya et al., 2017a,b; Yildirim et al., 2018; Atal and Singh, 2020), most of the works focus on the recognition of five main macro classes, namely Non-ectopic (N); Supraventricular ectopic (S); Ventricular ectopic (V); Fusion (F); Unknown (Q).

There is very little effort devoted to classify the micro-classes of the ECG signal, hence it serves as our main motivation to study the micro-classification heartbeats, of five types, i.e., Normal (NOR), Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Atrial Premature (AP), Premature Ventricular Contraction (PVC). The contribution of this paper is two-fold, in which the proposed algorithm is endowed with an ability to effectively process the non-filtered dataset with its potential anti-noise features, and secondly this paper presents an analysis of micro-classes of the ECG signal that compares some techniques of machine learning such as BP and Random Forest. The results can be served as a good source of benchmark literature to other researchers in the same field for future research work.

Section ECG Data Processing of the paper mainly introduces the ECG dataset used in this study and provides a detailed description of the data segmentation and preprocessing. In section Methodology, the architecture of the proposed algorithm and the specific experiment design is outlined. In section Result and Discussion, the performance and robustness of the proposed network is evaluated on the MIT-BIH Arrhythmia database and compared with BP, Random Forest, and several benchmarked CNN networks. Finally, section Conclusion summarizes the paper.

The proposed CNN algorithm is trained on a PC with Intel i5-7300HQ processors with 16GB of RAM and GTX1050 as GPU. It takes ~4236 s to complete training, and ~11 h to complete ten-fold cross validation. The average classification time of a single sample is 0.242 milliseconds. The implementation of the algorithm is using the MATLAB Deep Learning Toolbox and MATLAB Neural Network Toolbox.

The confusion matrix in Table 4 shows the classification results of the proposed CNN network. The overall classification accuracy rate of the five micro-class classification of heartbeats reaches 97.41%, and the positive rate and specificity of each category are over 90%, which has fully illustrated the effectiveness of the model. However, despite the data enhancement, the total number of C4 samples in the data set is still relatively small compared to the other four categories. Hence, the model's sensitivity to the C4 type is quite low.

Tables 5–7 illustrate the confusion matrix of BP neural network, Random Forest network, and the compared CNN network, respectively. The result shows the accuracy rate of the proposed CNN algorithm is 97.41%, which is 10.16% higher than the BP neural network, 1.69% higher than the Random Forest, and 3.34% higher than the compared CNN network. Furthermore, the sensitivity and specificity of the proposed CNN network are higher than the other three networks. Compared with the traditional machine learning methods like BP network and the random forests, the CNN network has the weight-sharing feature, which significantly accelerates the optimization process. The proposed CNN network also shows better performance on extracting local features, which is essential to classify different heartbeat types. The proposed CNN network uses the average-pooling layer instead of the max-pooling layer of the compared CNN network. The average-pooling layer can preserve the overall feature of the input data, which will be beneficial to classify the ECG signal. Also, the proposed CNN network has one more alternating convolution and pooling layer than the compared CNN network. The filter size and number of each convolution layer are also larger than the compared CNN network.

Table 8 depicts the classification performance of the different networks applied to the original and denoising data sets. For the proposed CNN network, the accuracy rates of the original and de-noised data are 96.9 and 97.2%, respectively, and the accuracy rate of the classification of the original data is only decreased by 0.3%, which shows that the network proposed in this paper has a degree of noise resistance. The accuracy of BP neural network classification of raw data is dropped 3.3% compared to the de-noised data. The random forest is 0.5% lower, and the compared CNN network is also 0.5% lower, all of which are suffered more loss than the proposed CNN network.

Table 9 lists the standard deviations of the sensitivities, specificities, and positive rates of four different networks. According to Table 9, the standard deviations of the four metrics of the proposed CNN network are less than the BP neural network, random forest, and the compared CNN network. The result indicates that the model's performance against multiple classifications is relatively stable, and the recognition effect of each classification is consistent, which shows the robustness of the model.

Table 10 shows the existing literature of ECG classification. The dataset used in these literatures is not exactly the same, but the comparison is useful because classification is all on the same MIT-BIH database. Li and Zhou (2016) applied random forest classifier to recognize five main classes (N, Q, S, V, F), which achieved a 94.61% accuracy rate. Osowski et al. (2004) used an SVM classifier that achieved 98.18% accuracy on 13 classes of heartbeats. Martis et al. (2013) obtained 94.52% performance on five main classes (N, Q, S, V, F) in their studies. Pławiak and Acharya (2020) used a deep genetic ensemble of classifiers to classify long-duration ECG signal, which achieved 94.6% of accuracy on 17 arrhythmia classes in the MIT-BIH database. Gao et al. (2019) implemented an effective long short-term memory (LSTM) recurrence network model to classify 8 types of heartbeats (N, LBBB, RBBB, APC,NESC, ABERR, NPC, AESC). Atal and Singh (2020) proposed an optimization-based deep convolutional neural network, achieving 93.19% accuracy on five main classes (N, Q, S, V, F). Acharya et al. (2017a) achieved 95.22% of accuracy on the classification of two types of heartbeats only (Normal and MI) with an 11-layer CNN network. The architecture of this CNN network was 4 alternating convolutions and max-pooling layers, followed by 3 fully-connected layers. In another work, Acharya et al. (2017b) used 9-layer CNN network to classify 5 main classes (N, Q, S, V, F) and they achieved 94.03% of accuracy. For fair comparison, CNN model of Acharya et al. (2017b) is reconstructed to benchmark with the proposed model. Zubair et al. (2016) obtained a 92.70% classification performance rate for five main classes (N, Q, S, V, F) using an 8-layer CNN-based network. Zubair et al. (2016) implemented 3 alternating convolution and max-pooling layer in the CNN, followed by one MLP layer. Generally speaking, the proposed CNN network achieved relatively high accuracy on the 5 micro-classes of heartbeats classification.

Statistical variance test is carried out to study the differences of the classification performance of the proposed CNN algorithm as well as the different approaches for the raw and denoised dataset. Analysis of variance is a collection and representation of statistical model. The associated estimation procedures are used to analyze the differences among group means in a sample. The results are recorded in Tables 11–13. The p-value is the probability of obtaining test results at least as extreme as the actually observed results, under the assumption that the null hypothesis is correct. In the statistical significance test, the smaller the p-value, the stronger the evidence we should reject the null hypothesis. For analysis in Table 11, the null hypothesis is that the denoised data and raw data perform equally well. Significance level is set to 0.05. The p-value obtained from Table 11 shows 0.011 for accuracy, 0.006 for sensitivity, 0.009 for specialty and 0.027 for positive prediction rate. It clearly demonstrates all of the p-values are <0.05, therefore it has the full evidence of rejecting the null hypothesis. Thus, the effect of denoising is very significant in the classification of the micro-heartbeats type.

On the other hand, in Table 12, the null hypothesis is under raw data, all four classifiers perform equally well. Significance level is still set to 0.05. The obtained p-values are too miniscule for accuracy, sensitivity, specialty, and positive prediction rate. All of the p-values are < 0.05, therefore it distinctly shows that the null hypothesis should be rejected. Thus, under raw data, it is obvious to interpret that the proposed CNN network performs much better than the rest of the classifiers.

In Table 13, the null hypothesis is under denoised data, all four classifiers perform equally well. Significance level is still set to 0.05. The obtained p-values are too miniscule for accuracy, sensitivity, specialty and positive prediction rate. All of the p-values are < 0.05, therefore it distinctly shows that the null hypothesis should be rejected. Thus, under denoised data, it is obvious to interpret that the proposed CNN network performs much better than the rest of the classifiers.

Cardiovascular disease is a major health problem in today's world. The early diagnosis of cardiac arrhythmia highly relies on the ECG. Unfortunately, the expert level of medical resources is rare, visually identify the ECG signal is challenging and time-consuming. Different from the existing literatures in which most of them classify the five main classes, such as the Non-ectopic, Supraventricular ectopic, Ventricular ectopic, Fusion, and Unknown, in the MIT-BIH Arrhythmia database, our paper pays more attention to specific micro-classes, namely the Normal, Left Bundle Branch Block, Right Bundle Branch Block, Atrial Premature Beats, Premature Ventricular Beats. Compared with the BP neural network, random forests, and other CNN networks, it is worth to highlight that the proposed CNN network has relatively higher accuracy and robustness. The proposed CNN network shows an outstanding performance in the overall classification accuracy of 97.41%, sensitivity of 97.05%, specificity of 99.35%, and positive prediction rate of 97.21% on the classification of the micro-classes of Arrhythmia dataset.

The advantages of the proposed CNN network have been put to evidence. It is endowed with an ability to effectively process the non-filtered dataset with its potential anti-noise features. Besides that, ten-fold cross-validation is implemented in this work to further demonstrate the robustness of the network. In addition, this paper presents an analysis of the classification of micro-classes of the ECG signal with comparison to some techniques of machine learning such as BP and Random Forest. We add values to the research community by discussing the results of the classification of less popular micro-classes of Arrhythmia that can be served as a good source of benchmark literature to other researchers in this field for further research. One possible setback of the proposed solution is that it is computationally intensive to train the network, due to deep learning series is often attributed to large scale data required for training.

As for future work, it would be interesting to explore the use of optimization techniques to find a feasible design and solution. The limitation of our study is that we have yet to apply any optimization techniques to optimize the model parameters and we believe that with the implementation of the optimization, it will be able to further elevate the performance of the proposed solution to the next level.

Publicly available datasets were analyzed in this study. This data can be found here: https://physionet.org/content/mitdb/1.0.0/.

SYW: conceptualization, methods and materials, and supervision and review. MW and YL: data curation. MW, YL, and WY: formal analysis. WY: resources. MW: validation. MW, YL, WY, and SYW: writing. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.