Deep learning in ECG diagnosis: A review☆
Introduction
Cardiovascular disease (CVD) is a collective term for disorders related to the heart or blood vessels. According to the statistics provided by American Heart Association in 2019, CVDs had become a global dominant cause for death. More than 17.6 million death was caused in 2016 and it is estimated that this figure will reach 23.6 million in 2030 [1]. CVDs may cause blockage of blood vessels and formation of blood clots, which can lead to cerebral or cardiac ischemic necrosis, giving rise to stoke and myocardial infarction. Due to the long-term poor blood pumping of the heart, all organs in the body may be congested and deprived of oxygen, suffering from different degree of damage [2].
Electrocardiogram (ECG) is one of the most commonly used tools for clinical diagnosis in cardiovascular health due to its simplicity, low cost and non-invasive nature. For example, for the patients who suffer from acute heart failure, only in 7.5% of patients have normal ECG [3]. ECG is an object of analysis that detected by an ECG machine that can reflect electric potential change generated by the heart during each cardiac cycle. The heart contracts in a rhythmical manner with regular excitement of myocardium, pumping blood throughout the body. In the process myocardium contraction, slight current is generated by heart and conducted to body surface, causing potential changes in each part of the body. An ECG is be obtained by measuring the potential change through the electrodes from different parts of the body tissue, and recorded with an electrocardiograph or a vector electrocardiograph. In this way, the abnormal rhythm and activity of heartbeat can be shown. Hence, heart diseases or diseases that damage myocardial function can be diagnosed, including arrhythmia, myocardial infarction, coronary heart disease and part of medical co-morbidities such as diabetes and high blood pressure. ECG can also serve as a predictor of coronary heart disease, cardiovascular disease and congestive heart failure. Early detection of such diseases is necessary, because some of them are associated with increased risk of stroke or even sudden-death. There are studies demonstrating that ECG is of importance in predicting both short- and long-term outcomes. For example, for patients suffering from myocardial infarction, the sooner is the abnormal heart rhythm detected, the greater is the chance of avoiding threats of life and recovery [4]. In addition to basic traditional diagnosis and monitoring, ECG is currently employed in telemedicine and home-care monitoring. Home-care monitoring plays a role in patients who suffered from cardiovascular disease, providing timely diagnosis with low price [5]. The telemedicine framework developed by Costa and Oliveira [6] reached 2600 examinations/month in their first 6 months and the most remote client is 85 km away.
Therefore, fast and accurate ECG diagnosis is necessary clinically. However, ECG signal has characteristics of high noise and high complexity, making it time-consuming and labor-intensive to identify certain diseases. Another problem is the individual variability [7], for example, every slight movement produces a baseline drift which is a low-frequency interference for ECG signal and the magnitude of electric potential measured changes with the placement of electrodes [8]. Furthermore, signal interpretation is a tiresome and complicated task, so there is probability of subjective uncertainty and human error in the process of analysis even for experts trained for years. Therefore, it is of great value to attach importance to the research and development of computer-aided method. Computer-supported analysis is able to analyze ECG signal in a more accurate and faster way without difference caused by inter-operators and operator-specific.
Computer-aided ECG interpretation system was first developed in 1960s [9]. The fully automatic system of the traditional intelligent algorithm for ECG interpretations includes three main steps: data preprocessing, feature extraction and classification. Data is denoised and padded or cut to make them into signal segments of the same length in the preprocessing stage. The feature extraction phase is the key part for ECG signal classification. Features can be extracted from the morphology of the ECG signal in the time and frequency domain or directly from the heart rhythm. Finally, the signals are classified into different types of heartbeat or disease according to the features extracted. The ultimate goal is to design algorithms that with high accuracy, efficiency and robustness and are able to reduce doctors’ burden.
Deep learning, as a computer-aided method with strong ability to feature extraction, managed to achieve high accuracy in ECG signal classification [10]. Deep learning is achieved by building hierarchical artificial neural networks [11]. The simple non-linear modules of each layer allow deep learning with great advantage in processing complex non-linear signals such as ECG signal. Passing through each layer, more abstractive and high-level features can be gained [12], which contributes to the high classification accuracy. As a result, compared with conventional machine learning, deep learning has better ability to representation learning of intricate data with large samples. LeCun et al. [13] concluded that deep learning discovers intricate structure in large datasets by using the backpropagation algorithm to indicate how a model should change its internal weight values that are used to compute the representation in each layer. This enables deep learning to have fine fault tolerance and prevent the misjudgment caused by overfitting. Deep learning can automatically complete the task of feature extraction and classification by imitating the general-propose learning of human brain while these need human engineers to design in the past. In this way can it learn the implicit knowledge that was only mastered by experts in the past, which means human burden can be greatly reduced. What is more, the progress of Center Processing Unit (CPU) and Graphic Processing Unit (GPU) performance reduces the training and execution time dramatically [14]. This allows deep learning to have large sets of data trained and to apply more complexed algorithms, giving it greater development potential.
Started from the use of greedy layer-wise pretraining, stacked auto-encoders (SAE) and deep belief network (DBN) serve as early typical methods in deep learning field. Following this, convolutional neural network (CNN) becomes the most popular algorithm with the success in visual recognition and are extended to various fields. Recurrent neural network (RNN) is recursive networks popular for its outstanding performance in time series data processing, which is quite suitable for ECG signal classification task. Long short-term memory (LSTM) outperforms traditional RNN in long-term dependence, becoming a more common method. A number of deep learning methods have been applied to feature extraction and classification in ECG interpretation. SAE is an unsupervised way to extract features by encoding and decoding the input ECG segments. DBN can either works as SAE unsupervised or serve as a classifier in supervised manner. Both of these two methods often fine-tuned with active learning so that the important data can be focused. Since ECG is 1-D signal, it is input directly into 1-D CNN or transformed into image and processed by 2-D CNN. As for RNN, ECG is usually processed as time-series signal. There are networks combining CNN and RNN to learn both space and time information, which is proved to be efficient in ECG classification tasks. The concept of SAE is executed with LSTM and CNN by changing auto-encoder into LSTM and convolutional layer while remaining the procedure of encoding and decoding of the network. Most of deep learning methods have achieved accuracy which is higher than manual classification with little or no human assistance, reducing the burden on professionals efficiently. Compared with traditional methods, deep learning can process raw ECG signals directly without the requirement of preliminary feature extraction, allowing higher efficiency and simpler steps for usage. However, there is limitation on dealing with imbalanced input due to deep learning’s strong dependence on input, while some kinds of heartbeat samples exist less in reality. Also, the high complexity and large amount of calculation makes the application in wearable devices still a difficulty. Moreover, ECG collected from reality always accompanied with noise and the preprocess of denoising requires much calculation resources. Given these existing problems, more robust models with less parameters should be constructed. In addition to the theoretical progress, the application of methods in reality should receive more attention. There have been some reviews focusing on ECG signal and deep learning. Murat et al. [15] only focused on heartbeat classification task. Hong et al. [16] went through deep learning’s application in all field of ECG signal more than disease diagnosis. Faust’s review contains researches on a range of physiological signals where only a small space is used to introduce ECG [17]. Here we can see that no comprehensive review of deep learning’s application in ECG diagnosis has been carried out, while this direction has practical significance and development potential. Hence, we adopt the position that a review in studies of deep learning method applied in ECG diagnosis is necessary. In this paper, state-of-the-art studies are reviewed in a systematic way and their characteristics are highlighted. Meanwhile, an overview of deep learning on ECG diagnosis is illuminated by pointing out problems and potential development.
To explain the importance of deep learning in the field of ECG analysis, this paper is organized as follows. The basic knowledge of ECG signal and typical cardiovascular diseases is represented in Section 2 and the common-used databases are introduced in Section 3. In Section 4, the theoretical background of deep learning in is introduced and the relevant researches are presented in Section 5. To make it clear, we present the studies according to 4 classic deep learning architectures: SAE, DBN, CNN and RNN. Section 6 discussed the limitation and future opportunities of deep learning in ECG diagnosis, key issues of deep learning and other network architectures that are promising in ECG field. Finally, a brief conclusion is drawn in Section 7.
Section snippets
ECG signal
The heart contracts in a rhythmical manner with regular excitement of myocardium, pumping blood throughout the body. In the process myocardium contraction, slight current is generated by heart and conducted to body surface, causing potential changes in each part of the body.
For ECG diagnosis, it is 12 leads that utilized most widely, in which 10 electrodes are applied and one of them serves as a reference to others (usually is the right leg) [18]. The configuration of 12 leads can be seen in
Database
A few studies run their networks on datasets of more than one database to prove validity of the methods, which also make it easier to compare experiment results horizontally. Most of the studies we reviewed used the following databases. The specific parameters of commonly used databases are listed in Table 1.
MIT-BIH Database is the most widely used database. It consists of 9 sub-databases of different diseases, among which the most popular one is the MIT-BIH Arrhythmia Database [21]. A
Theoretical background of deep learning
Deep learning is a subordinate branch of machine learning. It can learn and construct intrinsic features from neurons in numerous hidden layers of a neuron network. Deep learning network is developed based on neural network. Warren McCulloch and Walter Pitts came up with the concept of artificial neural network and the mathematical model of artificial neuron, thus opening the era of artificial neural network research. Following this, Frank Rosenblatt proposed perceptron [22], which is the
Researches
In this section, deep learning methods of classifying ECGs are introduced according to their algorithm. In each part, a table is listed to show the research work’s application, Deep Learning (DL) algorithm, the database used for the study, and the study results. The studies requiring more explanation are further explained following the table.
Discussion
We have analyze literature reports that apply deep learning to ECG diagnosis systematically. Therefore, in this section, current limitations and future potential directions of both deep learning in ECG diagnosis and deep learning method itself will be discussed.
Conclusion
In this paper, existing deep learning studies on ECG diagnosis classification tasks were reviewed and summarized systematically. We reviewed relevant studies from perspective of data, basic algorithm and models. State-of-the-art studies were reviewed according to the deep learning algorithm they used and their distinguishing features were highlighted with discussion. We found that deep learning has achieved good performance in various cardiovascular diseases, though there is still limitation
CRediT authorship contribution statement
Xinwen Liu: Literature review, Figures and tables, Writing - original draft, Writing - review & editing. Huan Wang: Writing - review & editing. Zongjin Li: Basic knowledge summary. Lang Qin: Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the Innovation Fund of Glasgow College, University of Electronic Science and Technology of China and in part by the Sichuan Science and Technology Program under Grant 2020108.
References (118)
- et al.
Telecardiology through ubiquitous internet services
Int. J. Med. Inform.
(2012) - et al.
Application of deep learning techniques for heartbeats detection using ECG signals-analysis and review
Comput. Biol. Med.
(2020) - et al.
Opportunities and challenges of deep learning methods for electrocardiogram data: A systematic review
Comput. Biol. Med.
(2020) - et al.
Expected energy-based restricted Boltzmann machine for classification
Neural Netw.
(2015) - et al.
Deep learning approach for active classification of electrocardiogram signals
Inform. Sci.
(2016) - et al.
Automated diagnosis of arrhythmia using combination of CNN and LSTM techniques with variable length heart beats
Comput. Biol. Med.
(2018) - et al.
A novel application of deep learning for single-lead ECG classification
Comput. Biol. Med.
(2018) - et al.
Patient-specific ECG classification by deeper CNN from generic to dedicated
Neurocomputing
(2018) - et al.
ECG arrhythmia classification using stft-based spectrogram and convolutional neural network
IEEE Access
(2019) - et al.
Multiple-feature-branch convolutional neural network for myocardial infarction diagnosis using electrocardiogram
Biomed. Signal Process. Control
(2018)
Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals
Inform. Sci.
Automated detection of coronary artery disease using different durations of ECG segments with convolutional neural network
Knowl.-Based Syst.
Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network
Inform. Sci.
Arrhythmia detection using deep convolutional neural network with long duration ECG signals
Comput. Biol. Med.
Classification of myocardial infarction with multi-lead ECG signals and deep CNN
Pattern Recognit. Lett.
Electrocardiogram heartbeat classification based on a deep convolutional neural network and focal loss
Comput. Biol. Med.
Application of stacked convolutional and long short-term memory network for accurate identification of CAD ECG signals
Comput. Biol. Med.
Comprehensive electrocardiographic diagnosis based on deep learning
Artif. Intell. Med.
A deep learning approach for real-time detection of atrial fibrillation
Expert Syst. Appl.
Automated detection of atrial fibrillation using long short-term memory network with RR interval signals
Comput. Biol. Med.
A novel wavelet sequence based on deep bidirectional LSTM network model for ECG signal classification
Comput. Biol. Med.
Heart disease and stroke statistics-2019 update: A report from the American heart association
Circulation
Cardiovascular risks associated with incident and prevalent periodontal disease
J. Clin. Periodontol.
Insights into the importance of the electrocardiogram in patients with acute heart failure
Eur. J. Heart Fail.
Coronarv and precoronarv care
Am. J. Phys. Med.
Integrated heart failure telemonitoring system for homecare
Cardiol. J.
Weighted conditional random fields for supervised interpatient heartbeat classification
IEEE Trans. Biomed. Eng.
Basic ECG theory, recordings, and interpretation
Automatic screening of normal and abnormal electrocardiograms by means of a digital electronic computer
Proc. Soc. Exp. Biol. Med.
Deep learning
Deep learning for corpus callosum segmentation in brain magnetic resonance images
Deep learning
Nature
Towards ultra-high performance and energy efficiency of deep learning systems: An algorithm-hardware co-optimization framework
Application of deep learning techniques for heartbeats detection using ECG signals- analysis and review
Comput. Biol. Med.
Deep learning for healthcare applications based on physiological signals: A review
Comput. Methods Programs Biomed.
A review of signal processing techniques for electrocardiogram signal quality assessment
IEEE Rev. Biomed. Eng.
Electrocardiographic systems with reduced numbers of leads synthesis of the 12-lead ECG
IEEE Rev. Biomed. Eng.
A deep convolutional neural network model to classify heartbeats
Comput. Biol. Med.
Physiobank, physiotoolkit, and physionet : Components of a new research resource for complex physiologic signals
Circulation
The Perceptron - A Perceiving and Recognizing AutomatonCornell Aeronautical Laboratory Report, Vol. 85
Perceptrons, vol. 84
Learning representations by back propagating errors
Nature
Reducing the dimensionality of data with neural networks
Science
A review of semantic segmentation using deep neural networks
Int. J. Multimed. Inf. Retr.
A survey on deep learning techniques for image and video semantic segmentation
Appl. Soft Comput.
Object detection with deep learning: A review
Analytics of deep neural network-based background subtraction
J. Imaging
Deep neural network concepts for background subtraction: A systematic review and comparative evaluation
Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion
J. Mach. Learn. Res.
Cited by (149)
Hot rolled prognostic approach based on hybrid Bayesian progressive layered extraction multi-task learning
2024, Expert Systems with ApplicationsA novel hybrid GRU-CNN and residual bias (RB) based RB-GRU-CNN models for prediction of PTB Diagnostic ECG time series data
2024, Biomedical Signal Processing and ControlUniversal 12-lead ECG representation for signal denoising and cardiovascular disease detection by fusing generative and contrastive learning
2024, Biomedical Signal Processing and ControlDeep learning for ECG classification: A comparative study of 1D and 2D representations and multimodal fusion approaches
2024, Biomedical Signal Processing and ControlECG-based data-driven solutions for diagnosis and prognosis of cardiovascular diseases: A systematic review
2024, Computers in Biology and Medicine
- ☆
This work was supported by the Innovation Fund of Glasgow College,University of Electronic Science and Technology of China and in part by the Sichuan Science and Technology Program under Grant 2020108.