Deep learning in ECG diagnosis: A review

Abstract

Cardiovascular disease (CVD) is a general term for a series of heart or blood vessels abnormality that serves as a global leading reason for death. The earlier the abnormal heart rhythm is discovered, the less severe the sequela and the faster the recovery. Electrocardiogram (ECG), as a main way to detect the electrical activity of heart, is a very important harmless means of predicting and diagnosing CVDs. However, ECG signal has characteristics of complex and high chaos, making it time-consuming and exhausting to interpret ECG signal even for experts. Hence, computer-aided methods are required to relief human burden and reduce errors caused by tiredness, inter- and intra-difference. Deep learning shows outstanding performance on ECG classification studies recent few years. Its hierarchical architecture enables higher-level features obtained and its strong ability to feature extraction contributes to classification project. Latest studies can achieve higher accuracy and efficiency than manual classification by experts. In this paper, we review the existing studies of deep learning applied in ECG diagnosis according to four typical algorithms: stacked auto-encoders, deep belief network, convolutional neural network and recurrent neural network. We first introduced the mechanism, development and application of the algorithms. Then we review their applications in ECG diagnosis systematically, discussing their highlights and limitations. Our view about future potential development of deep learning in ECG diagnosis is stated in the final part of this paper.

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.