Transfer learning-based electrocardiogram classification using wavelet scattered features

RS Sabeenian; KK Sree Janani

doi:10.4103/bbrj.bbrj_341_22

ORIGINAL ARTICLE

Year : 2023 | Volume : 7 | Issue : 1 | Page : 52-59

Transfer learning-based electrocardiogram classification using wavelet scattered features

RS Sabeenian, KK Sree Janani
Department of Electronics and Communication Engineering, Sona College of Technology, Salem, Tamil Nadu, India

Date of Submission	28-Nov-2022
Date of Decision	29-Dec-2022
Date of Acceptance	28-Jan-2023
Date of Web Publication	14-Mar-2023

Correspondence Address:
K K Sree Janani
Department of Electronics and Communication Engineering, Sona College of Technology, Salem - 636 005, Tamil Nadu
India

Source of Support: None, Conflict of Interest: None

DOI: 10.4103/bbrj.bbrj_341_22

Abstract

Background: The abnormalities in the heart rhythm result in various cardiac issues affecting the normal functioning of the heart. Early diagnosis helps prevent serious outcomes and to treat them effectively. This work focuses on classifying the various abnormalities with the changes in the heart rhythm and demographic data. The pretrained convolution neural network models classify the wavelet scattered data of different arrhythmic electrocardiograms (ECGs). Methods: The ECG signals of different anomalies from the PhysioNet database are re-sampled and segmented. The sampling is done using the linear interpolation method, which estimates values between the sample points based on nearby data points. The inter-dependence variances among the data points were extracted using wavelet scattering. The one-dimensional (1D) signal data are converted into 2D scalogram images using continuous wavelet transform. Pretrained deep learning models are used to extract features from the scalogram images and classify using a support vector machine classifier. The classification results are analyzed using various performance metrics such as precision, specificity, recall, F-measure, and accuracy. The relationship between the model performance and network depth and learnables is analyzed. Results: The classification results show that the ResNet18 achieves higher accuracy of 98.81% for raw data and 97.05% for wavelet scattered data. No dependency exists between the model depth, network parameters, and performance. The ResNet18 model achieves higher precision, recall, specificity, and F-measure values of 96.49%, 96.42%, 98.24%, and 96.45%, respectively, for wavelet scattered data. Conclusions: The ResNet18 achieves generalized results in classifying dimensionality-reduced data with reduced computational cost and high accuracy. The DenseNet model achieves higher performance metrics for raw data, whereas the ResNet18 model achieves higher performance metrics for wavelet scattered data.

Keywords: Arrhythmia, electrocardiogram, scalogram, transfer learning, wavelet scattering

How to cite this article:
Sabeenian R S, Sree Janani K K. Transfer learning-based electrocardiogram classification using wavelet scattered features. Biomed Biotechnol Res J 2023;7:52-9

How to cite this URL:
Sabeenian R S, Sree Janani K K. Transfer learning-based electrocardiogram classification using wavelet scattered features. Biomed Biotechnol Res J [serial online] 2023 [cited 2023 Jun 10];7:52-9. Available from: https://www.bmbtrj.org/text.asp?2023/7/1/52/371695

Introduction

Nowadays, about 1.5%–5% of the population is affected by cardiac arrhythmias, in which atrial fibrillation is the most common arrhythmia.^[1] Various abnormalities in the morphology and rhythm of electrocardiogram (ECG) are commonly known as arrhythmias. The arrhythmias result in severe cardiac issues if left untreated. The morbidity and mortality rates are higher in cases with arrhythmias. The usage of machine learning and deep learning (DL) models increases at a higher rate facilitating the semi-automated application in a wide range of fields. Data augmentation, denoising, detection, classification, and analysis can be done in the biomedical area using artificial models. The diagnosis and classification had been made for medical issues such as breast cancer,^[2] brain tumors,^[3] and even corona infection.^[4],[5] The considerable number of bio-signals and image collection is significant to proceed further. Since the DL algorithm performs better with massive data, where the collection of huge data is complex in the biomedical field, transfer learning can overcome these limitations, as explained in section 2.2. Transfer learning helps in using DL models with less amount of data. The ECG data used predominantly in the existing studies are the Massachusetts Institute of Technology-Beth Israel (MIT-BIH) Arrhythmia datasets^[6] available in the PhysioNet database. This condition makes a great need for the early diagnosis of arrhythmias. Artificial Intelligence (A.I.)-based models do the work quickly and accurately.^[7] This work focuses on the DL-based classification of ECG signals using the concept of transfer learning.

Related work

The studies either use preprocessed data or raw data. The raw data consists of noises such as power line interference, electromagnetic interference, and baseline wander due to the movement and respiration activity of the patients. The noises can be removed using the high pass, low pass, and bandpass filters.^[8] Adaptive filters,^[9] wavelet-based filters,^[10] infinite impulse response filters, and finite impulse response filters are also used for filtering. The selection of filters is essential as it may result in information loss. The classification is done using the whole convolution neural network (CNN) network. The need for vast data is a barrier to using the DL network. These data shortage problem is overcome using transfer learning.^[11] In some work, the feature extraction is done using the pretrained networks,^{[12],[13],[14],[15]} which are later classified using machine learning classifiers. The same network is pretrained on the ImageNet data set, which consists of millions of images. The networks were trained on the openly available PhysioNet Computing in Cardiology (CinC) database and finetuned using the China ECG intelligent competition database to classify the normal and abnormal signals. The study states that the pretrained weight using one lead ECG will work well for finetuning with standard 12 lead ECG data.^[16] The other work uses the ResNet-18V2 network pretrained using the Icentia 11K data set and found using the PhysioNet CinC challenge 2017 data set. The Icentia 11K database of ECG signals is available online. The performance of the DL model improves by 6.5% through pretraining.^[17]

Most of the DL work performs automatic feature extraction. The handcrafted features are sometimes used to train the machine learning models. The ECG signal is analyzed using time and frequency domain features, where the frequency domain features are found to be more efficient. The frequency domain CNN is more efficient when classifying atrial fibrillation than time-domain CNN.^[18] The deep transfer learning-based atrial fibrillation identification using AlexNet performs better with medium-level features achieving an accuracy of 87.9%.^[19] The classification of arrhythmia signals using pretrained DenseNet achieves an accuracy of 98.9% optimized using Adam optimizer.^[20] Another work on classifying arrhythmia using AlexNet achieves an accuracy of 95.67% using the scalogram images. continuous wavelet transform (CWT) is performed to get the two-dimensional (2D) scalogram representations from 1D signals.^[21] The wavelet transform plays a vital role at various stages such as denoising, feature extraction, data compression, 1D to 2D frequency domain transformation, and detecting anomalies.^[10]

Section 3 explains the methodology adopted to classify ECG signals. The various pretrained CNN architectures are explained in detail in subsection 3.2. The achieved results are analyzed in Section 4. The findings are explained in discussion Section 5, and the work is concluded in the last section.

Methods

The classification of the ECG signal involves selecting the dataset, preprocessing the signal segments, converting a 1D signal into 2D images, extracting the extracted features using the pretrained networks, and classifying the extracted features.

Database

The data used in the study include ECG signals from three different databases. The MIT-BIH arrhythmia dataset consists of a 48-h ECG signal. It includes about 17 abnormalities grouped as class Arrhythmia (ARR) (class 1). The second database includes the MIT-Normal Sinus Rhythm (NSR) consisting of ECG signals without abnormalities considered as class NSR (class 3).^[22] The signals are digitalized with 128 samples per second. The third database includes the Beth Israel Deaconess Medical Center coronary heart failure (CHF) datasets, comprising 15 records of individuals with severe CHF belonging to class CHF (class 2).^[23] The CHF signals are digitalized at 360 samples per second. The selected signals from the three databases include 96, 30, and 36 segments of ARR, CHF, and NSR, respectively. The ARR, CHF, and NSR signals are sampled at different sampling frequencies of 360, 128, and 250, respectively. Before proceeding with classification, the signals are re-sampled to a standard frequency of 360 HZ. The linear interpolation-based upsampling is performed for classes 2 and 3 to attain a uniform sampling rate. The interpolation converts the lower sample rate signal into a higher sample rate with the help of various filters. The filters have control over the anti-aliasing phenomenon. It also helps to predict unknown values. The re-sampled signals are segmented into 29,160 segments, each with 1024 samples. The ARR, CHF, and NSR classes of ECG signals include 17,279, 5401, and 6480 segments, respectively. The segment length is selected by considering the existence of two to three beats in each segment. The rhythmic and morphological changes obtained with 2–3 beats of each segment are used for classification.

The ECG signals are prone to noises. Removing these artifacts is done using a different filter, resulting in a loss of information. Some works preferably use raw signals without any preprocessing.^{[14],[24],[25]} The segments of 1D ECG signals are represented in frequency domain representations in the form of 2D images, as the frequency domain analysis using a scalogram is very efficient than spectrogram representations.^[26] The conversion of 1D to 2D is also done using Hilbert transform and Wigner-Ville distribution.^[27] The scalogram representation of the ECG segments is shown in [Figure 1].

Figure 1: Scalogram representations of ECG signal. ECG: Electrocardiogram

Click here to view

Transfer learning

The concept of transfer learning lies in transferring the gained knowledge to the new application. Concerning the convolutional neural network architectures, the learning happens in terms of achieving optimized learnable such as the weights and costs. The value of weights and costs is continuously updated iteration by iteration until it reaches the optimized value. The training of the model is a time-consuming process. In addition, the data required for training a deep neural network is enormous. Collecting massive amounts of data in biomedicine, such as physiological signals and medical images, is tedious. These two limitations of using a new neural network from scratch are overcome by transfer learning.^[28] In transfer learning, the network with learnable (weights and costs) already optimized using a sizeable available dataset is used to classify ECG signals of interest. Transfer learning gives improved results with limited data and time.^[29] The scalogram features are extracted using pretrained deep neural networks such as AlexNet, ResNet 18, SqueezeNet, ShuffleNet, GoogLeNet, DenseNet201, and Inceptionv3. The various CNN networks are pretrained, extracted features are used for classifying three classes of ECG signals, and the results are comparatively studied as shown in [Table 1].

Table 1: Comparison of pretrained model classification accuracy versus network learnables versus number of layers

Click here to view

Wavelet scattering

The wavelet scattering principle reduces the data dimension, helping speed up the computations.^[30] The wavelet scattering stores the low variance features with minimum inter-class differences and maximum intra-class differences. The wavelet scattering performs data compression by wavelet transform followed by a scattering network, as shown in [Figure 2]. The wavelet scattering differs from other compression models by dealing with the instability of the signal that occurs due to the deformations at higher frequencies. The low variance values of raw ECG signals are extracted using wavelet scattering for the classification. The 1024 sample long segments are reduced to 816 features with essential variance information. The reduction in data reduces the computational cost with less training time.

Figure 2: Wavelet scattering principle

Click here to view

Convolution neural network architectures

AlexNet

AlexNet has eight layers, with the first five convolutions, one max pooling, and three fully connected layers. The AlexNet uses the ReLU activation function instead of other activation functions to achieve improved results. It is also the first architecture to use the dropout function to treat the overfitting problem. The AlexNet comprises 60.9 million learnable pretrained with the ImageNet database and is finetuned to feature extraction and classification of the ECG scalogram images.^[31] ImageNet database consists of 12 lakh images of 100 different categories. In other words, trained learnable is used for extracting features for classification using the support vector machine (SVM) classifier, achieving an accuracy of 97.89%.

ResNet

The architecture of ResNet 18 consists of 71 layers and 11.6 million learnable. The layer count increases as the network go deeper. The central concept of ResNet lies in the audition of residual blocks and identity blocks, as shown in [Figure 3]a which lets the added layer estimate the residual function instead of waiting for the stacked layers to learn.^[32] The identity blogs help in treating the vanishing gradient problem. In addition, the ResNet has batch normalization (BN) layers, improving the network performance. As the model becomes deeper, the complexity increases in computations and training time. The ResNet performs effectively in different applications, which makes the network a preferable network in treating vanishing gradient problems, and reduced training error percentage even with the extensive network layers are the main advantages of using this network. The ResNet50 is used to classify arrhythmias with an accuracy of 98.3%. The ResNet-18 network classifies the ECG signals with an accuracy of 98.81% using the scalogram of raw data.

Figure 3: (a) residual block used in ResNet. (b) Channel shuffle block used in ShuffleNet. (c) Dense block used in DenseNet

Click here to view

GoogLeNet

The architecture reduces the error rate to a great extent, for which the architecture wins the ImageNet large-scale visual recognition challenge 2014 image classification challenge. The architecture varies from the previous by adding a 1 × 1 convolution and using the global average pooling technique. The 1 × 1 convolution is used in the architecture to reduce the number of parameters. As the number of parameters reduces, the number of layers can be increased, making the model deeper. The global average pooling reduces the trainable parameters count and improves accuracy by 0.6%. The inception model has the parallel combination of 1 × 1, 3 × 3, and 5 × 5 convolution layers and a 3 × 3 Max pooling layer between the input and output to handle multi-scale data.^[33] The GoogLeNet model trained on the subset of the ImageNet database is used for classifying ECG time frequency scalogram images. The features from the GoogLeNet are classified using the SVM classifiers, providing an accuracy of 86.25%.

SqueezeNet

SqueezeNet offers accurate image classification with minor network architecture. The smaller network has many advantages over the deep networks. It includes reduced server communication, decreased bandwidth requirement, and easy field-programmable gate array deployment. The minor architecture uses 1 × 1 filters instead of 3 × 3 filters, which reduces the parameter by nine times since the 1 × 1 filter parameters are nine times fewer than the 3 × 3 filter. Another concept is used in achieving higher accuracy with the delayed down-sampling process.^[34] Other versions of SqueezeNet are available, namely SqueezeNet with simple bypass and SqueezeNet with complex bypass. These two models differ from the basic one by including skip connections inspired by residual networks. The SqueezeNet network trained on ImageNet, a database that contains millions of images, is used to classify the ECG scalogram images.

ShuffleNet

ShuffleNet directly concentrates on speed or memory access cost to level with the computational complexity of the model. The central concept lying in ShuffleNet is equal channel width between input and output, group convolution, network fragmentation, and element-wise operations.^[35] The channel shuffle block used in ShuffleNet is shown in [Figure 3]b. That simple structure without any point-wise group convolutions resembles XceptionNet. Group convolution improves the whole network performance, resulting in smaller models. In the case of a multiple group convolution layer, the information flow between the group is made possible with channel shuffle. The ShuffleNet network performs faster when compared with AlexNet.^[35] The speed and fewer million Floating Point Operations per second (mFLOP) requirements make the network suitable for mobile applications. The ShuffleNet, pretrained on the ImageNet data set, is used to classify scalogram images. The classification with an accuracy of 96.85% is achieved.

Inceptionv3

Inceptionv3 is the third version of Google Inception CNN which helps to achieve a deep neural network without increasing the number of parameters. The model differs from its previous version by certain modifications such as using smaller convolutions, spatial factorization into asymmetric convolutions, auxiliary classifiers, and reducing grid size. The Inception V3 performs better than the visual geometry group, GoogLeNet, and BN-Inception networks in terms of error rate. The error rate is notably reduced in the InceptionV3 model.^[36] Pretrained Inception V3 model used consists of 315 layers and 28.3 learnable. The classification accuracy achieved using the Inception V3 model is 97.4%

DenseNet201

DenseNet is the network with additional dense blocks facilitating a dense connection between the layers. These dense networks made it possible to receive information from all preceding layers. It makes it possible to have a thinner network with fewer channels.^[37] It uses features of different complexity levels. The dense block used in the network is shown in [Figure 3]c. The advantage of DenseNet has reduced parameter that helps to avoid learning redundant feature maps.^[20] Also, the DenseNet has very thin layers with a small count of new feature maps.^[37] The DenseNet201 used here has 708 layers with 20 million learnables which is very small for 708 layers. The DenseNet is pretrained using the ImageNet database to classify ECG signals' scalogram images, achieving an accuracy of 98.02%. [Figure 4]c shows the raw data classification confusion matrix using the DenseNet pretrained model. [Figure 4]a and [Figure 4]b shows the confusion matrix of ResNet classification using raw and wavelet scattered data. The classification with the ResNet model using wavelet scattered scalogram images achieves accuracies that match the classification's accuracies using raw data with a deviation of only 0.18%.

Figure 4: (a) Confusion matrix of ResNet18 for raw data classification. (b) Confusion matrix of ResNet18 for wavelet scattered data classification. (c) Confusion matrix of DenseNet for raw data classification

Click here to view

Results

The classification of 3 classes of ECG signals, namely, ARR, CHF, and NSR using an SVM classifier with the features extracted using DL models employing transfer learning, are analyzed in this section.

The various DL CNN networks work on 2D images. Hence the ECG signals are converted into 2D scalogram images using CWT. The classification accuracy achieved using features from pretrained networks like AlexNet, ResNet-18, GoogLeNet, SqueezeNet, ShuffleNet, Inceptionv3, and DenseNet are 97.89%, 98.81%, 95.61%, 96.07%, 96.85%, 97.46%, and 98.02%, respectively. [Table 1] shows the pretrained networks, number of layers, total learnables, and their classifying overall accuracies. [Figure 5] shows the graphical plot of those parameters.

Figure 5: Graphical representation of model performance versus learnable versus number of layers

Click here to view

The performance of various networks is analyzed by calculating the different performance metrics. The metrics include precision, recall, specificity, and F-measure. Recall gives the model's efficiency in predicting the true positives, precision gives the model efficiency in predicting values correctly, and specificity metric gives the model's ability to predict the true negatives correctly. The F1 score explains the model efficiency over the correctness of positive and negative predictions. The performance metrics for the different networks are graphically represented in [Figure 6]. The bar graph demonstrates a minor variation in metric values with high specificity across all networks. The classification accuracies of three different ECG classes are shown in [Table 2].

Figure 6: Bar graph representation of performance metrics obtained using pretrained CNN models. CNN: Convolution neural network

Click here to view

Table 2: Classification accuracies of different classes of electrocardiogram signals using convolution neural network models

Click here to view

The accuracy deflection appears to be affected by the distribution of the data set. The graphical representation of the accuracies of three ECG classes for all pretrained networks is shown in [Figure 7]. [Table 3] displays the classification accuracies of raw data scalogram images and wavelet scattered values. Wavelet scattering aids in dimensionality reduction while retaining as much information as possible. The wavelet scattered values achieve less accuracy, as shown in [Table 3], by comparing all the values that DenseNet performs the classification process with a higher accuracy of 98.02% next to ResNet18, with an accuracy of 98.81% using raw ECG data. The pretrained models with high classification accuracies are employed in classifying the scalograms of the wavelet scattered low variance values. [Table 3] shows the classification accuracies of the five best-performing pretrained models using raw data (accuracy 1) and the scalogram images of scattered wavelet data (accuracy 2).

Figure 7: Graphical representation of the classification accuracies of three ECG classes. ECG: Electrocardiogram

Click here to view

Table 3: A comparative analysis of classification accuracies using raw data and wavelet scattered data

Click here to view

The values are graphically represented using a bar graph, as shown in [Figure 8]. The pretrained AlexNet, ResNet, DenseNet, inceptionetv3, and GoogLeNet achieve a classification accuracy of 88.34%, 97.05%, 91.39%, 90.74%, and 88.37%, respectively. The DenseNet model achieves a higher precision, recall, specificity, and F-measure value of 97.68%, 97.71%, 98.81%, and 97.7%, respectively, for raw data.

Figure 8: Bar graph representation of the classification accuracies using raw data and wavelet scattering data

Click here to view

Discussion

Various observations done around the classification process are discussed in this section. ECG signals' frequency domain 2D representation is used to finetune the pretrained model. The features extracted from the scalogram images with the help of CNN architecture are used for classification using an SVM classifier. On comparative analysis, the classification's performance based on the raw signals' scalogram achieves high classification accuracy, with the scalogram of wavelet scattered features. The comparative analysis of network learnables and the classification performance shows no relationship, whereas the layer depth has less effect on the performance. Despite the layer count, the ResNet18 achieves good performance because of the residual block. In [Figure 5], it can be seen the dissimilarity between the curves of performance and count of learnables and the deviations that exist between the classification performance and the layer depth. The dependency of the model's performance on the network parameters and depth is not linear, as shown in [Table 1]. The result shows the effect of architecture on the model's performance despite the depth of the model. From [Figure 7], class NSR achieves maximum accuracies using all pretrained networks. Class ARR achieves less accuracy among the three classes. The classification and accuracy of class ARR deviate from the other two classes due to the availability of extensive data set underclass compared to the other two classes, showing the need for data balancing. The model achieves higher accuracy compared with other related works and explains the efficiency of frequency domain CNN architectures over time domain CNN architectures. The ResNet18 model achieves the highest metric values for raw data, whereas the DenseNet model achieves higher performance metrics for wavelet scattered data. It can be observed that the recall metric attains the minimum value in comparison with other metrics in all models other than the ResNet model. The classification of wavelet scattered values using transfer learning is a new approach differing from the existing model. ResNet18 classifies ECG signals with accuracies of 98.81% for raw data and 97.05% for wavelet scattered data with better generalization.

Conclusions

Cardiac arrhythmias are a common medical issue which occurs at a higher rate in recent years. These conditions should be diagnosed early and treated to prevent significant complications. This work is to classify those arrhythmias signals using A.I. The ARR, CHF, and NSR signals from the database are re-sampled and segmented into 29,160. CWT converts the 1D ECG signal segments into 2D scalogram images. The pretrained CNN models extract features from the 2D scalogram images of raw data and are classified using the SVM classifier. The relation between accuracy, model depth, learnables, and data processing are comparatively analyzed. The classification accuracies of the three classes are comparatively analyzed. The transfer learning using pretrained dense net architecture produces second high accuracy of 98.02% in classifying ECG signals. The pretrained model with the top five accuracies is used to classify ECG signals from the wavelet scattered images. The ResNet18 model provides generalized performance with the highest classification accuracy of 98.81% for raw data and 97.05% for wavelet scattered data. The ResNet18 classifies dimensionality-reduced data with reduced computational cost and high accuracy. The DenseNet model achieves higher performance metrics for raw data, whereas the ResNet18 model achieves higher performance metrics for wavelet scattered data.

Limitation of the study

The deviations in classification accuracies of each class from [Table 2] show the importance of data balancing. The classification is done using the raw ECG signals, but the probability of getting affected by noise is high in the case of physiological signals. Data augmentation increases the data quantity and helps to face the data shortage problem. Data augmentation and effective denoising of signals will be done in future work to improve classification accuracy.

Financial support and sponsorship

Nil.

Conflicts of interest

There are no conflicts of interest.

References

1.	Lakshminarayan K, Anderson DC, Herzog CA, Qureshi AI. Clinical epidemiology of atrial fibrillation and related cerebrovascular events in the United States. Neurologist 2008;14:143-50.
2.	Parekh DH, Dahiya V. Predicting breast cancer using machine learning classifiers and enhancing the output by combining the predictions to generate optimal F1-score. Biomed Biotechnol Res J (BBRJ) 2021;5:331.
3.	Sabeenian RS, Vijitha V. Identification and categorization of brain tumors using ensemble classifiers with hybrid features. Biomed Biotechnol Res J (BBRJ) 2021;5:357.
4.	Anand R, Sowmya V, Menon V, Gopalakrishnan A, Soman KP. Modified VGG deep-learning architecture for COVID-19 classification using chest radiography images. Biomed Biotechnol Res J (BBRJ) 2021;5:43-9.
5.	Rezayi S, Ghazisaeedi M, Kalhori SR, Saeedi S. Artificial intelligence approaches on X-ray-oriented images process for early detection of COVID-19. J Med Signals Sens 2022;12:233-53. [Full text]
6.	Moody GB, Mark RG. The impact of the MIT-BIH arrhythmia database. IEEE Eng Med Biol Mag 2001;20:45-50.
7.	Yilmaz EÇ, Ozdemir S. Baby crying analyzing and solution using matlab graphical user interface; interdisciplinary collaboration between engineering and nursing. Biomed Biotechnol Res J (BBRJ) 2022;6:410.
8.	Pal A, Srivastva R, Singh YN. CardioNet: An efficient ECG arrhythmia classification system using transfer learning. Big Data Res 2021;26:100271.
9.	Merdjanovska E, Aleksandra R. Comprehensive survey of computational ECG analysis: Databases, methods and applications. Expert Syst Appl 2022;203:117206.
10.	Hong S, Zhou Y, Shang J, Xiao C, Sun J. Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review. Comput Biol Med 2020;122:103801.
11.	Jang JH, Kim TY, Yoon D. Effectiveness of transfer learning for deep learning-based electrocardiogram analysis. Healthc Inform Res 2021;27:19-28.
12.	Singh P, Sharma A. “Atrial Fibrillation Classification Using Transfer Learning.” In 2021 IEEE Bombay Section Signature Conference (IBC), IEEE; 2021. p. 1-5.
13.	Cao M, Zhao T, Li Y, Zhang W, Benharash P, Ramezani R. ECG heartbeat classification using deep transfer learning with convolutional neural network and STFT technique 10.48550/arXiv.2022;2206.14200.
14.	Salem M, Taheri S, Yuan JS. “ECG Arrhythmia Classification Using Transfer Learning from 2-Dimensional Deep CNN Features.” In 2018 IEEE Biomedical Circuits and Systems Conference (BioCAS), IEEE; 2018. p. 1-4.
15.	Asif RN, Abbas S, Khan MA, Sultan K, Mahmud M, Mosavi A. Development and Validation of Embedded Device for Electrocardiogram Arrhythmia Empowered with Transfer Learning. Computational Intelligence and Neuroscience, vol. 2022, Article ID 5054641, 15 pages, 2022. https://doi.org/10.1155/2022/5054641.
16.	Chen L, Xu G, Zhang S, Kuang J, Hao L. Transfer learning for electrocardiogram classification under small dataset. In: Machine Learning and Medical Engineering for Cardiovascular Health and Intravascular Imaging and Computer Assisted Stenting. Cham: Springer; 2019. p. 45-54.
17.	Weimann K, Conrad TO. Transfer learning for ECG classification. Sci Rep 2021;11:5251.
18.	Kent M, Vasconcelos L, Ansari S, Ghanbari H, Nenadic I. Transfer learning application of a novel frequency shift convolutional neural network method for atrial fibrillation classification. Europace 2022;24 Suppl 1:pp.euac053-017, https://doi.org/10.1093/europace/euac053.017.
19.	Ghaffari A, Madani N. Atrial fibrillation identification based on a deep transfer learning approach. Biomed Phys Eng Express 2019;5:035015.
20.	Ullah H, Bu Y, Pan T, Gao M, Islam S, Lin Y, et al. “Cardiac Arrhythmia Recognition Using Transfer Learning with a Pre-trained DenseNet.” In 2021 IEEE 2^nd International Conference on Pattern Recognition and Machine Learning (PRML), IEEE; 2021. p. 347-53.
21.	Gaddam PG, Sreehari RV. Automatic classification of cardiac arrhythmias based on ECG signals using transferred DL convolution neural network. J Phys Conf Ser 2021;2089:012058.
22.	Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation 2000;101:E215-20.
23.	Baim DS, Colucci WS, Monrad ES, Smith HS, Wright RF, Lanoue A, et al. Survival of patients with severe congestive heart failure treated with oral milrinone. J Am Coll Cardiol 1986;7:661-70.
24.	Gajendran MK, Khan MZ, Muazzam A, Khattak K. “ECG Classification Using Deep Transfer Learning”. In: 2021 4^th International Conference on Information and Computer Technologies (ICICT), IEEE; 2021. p. 1-5.
25.	Shi H, Wang H, Qin C, Zhao L, Liu C. An incremental learning system for atrial fibrillation detection based on transfer learning and active learning. Comput Methods Programs Biomed 2020;187:105219.
26.	Singh SA, Majumder S. Short and noisy electrocardiogram classification based on deep learning, In Deep learning for data analytics. United states, Academic Press;2020. p. 1-19.
27.	Zhang Y, Li J, Wei S, Zhou F, Li D. Heartbeats classification using hybrid time-frequency analysis and transfer learning based on ResNet. IEEE J Biomed Health Inform 2021;25:4175-84.
28.	Tan C, Sun F, Kong T, Zhang W, Yang C, Liu C. A survey on deep transfer learning. In: International Conference on Artificial Neural Networks, Lecture Notes in Computer Science, Cham: Springer; 2018;11141:270-9.
29.	Bizzego A, Gabrieli G, Esposito G. Deep neural networks and transfer learning on a multivariate physiological signal dataset. Bioengineering (Basel) 2021;8:35.
30.	Liu Z, Yao G, Zhang Q, Zhang J, Zeng X. Wavelet scattering transform for ECG beat classification. Computational and Mathematical Methods in Medicine, vol. 2020, Article ID 3215681, 11 pages, 2020. https://doi.org/10.1155/2020/3215681.
31.	Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM 2017;60:84-90.
32.	Huggins PS, Chen HF, Belhumeur PN, Zucker SW. Deep Residual Learning for Image Recognition, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp. 770-8, doi: 10.1109/CVPR.2016.90.
33.	Szegedy, Liu, Jia, Sermanet, Reed, Anguelov, et al. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition 2015: p. 1-9.
34.	Iandola Forrest N, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William J. Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5 MB model size. arXiv 2016:1602.073600.
35.	Zhang, Xiangyu, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. Proceedings of the IEEE conference on computer vision and pattern recognition ;2018: p. 6848-56.
36.	Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition; 2016: p. 2818-26.
37.	G. Huang, Z. Liu, L. Van Der Maaten and K. Q. Weinberger, “Densely Connected Convolutional Networks,” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017: p. 2261-9, doi: 10.1109/CVPR.2017.243.