Opportunities and Challenges in Deep Learning Methods on Electrocardiogram Data: A Systematic Review

12/28/2019 ∙ by Shenda Hong, et al. ∙ 0

Objective: To conduct a systematic review of deep learning methods on Electrocardiogram (ECG) data from the perspective of model architecture and their application task. Methods: First, we extensively searched papers deploying deep learning (deep neural network networks) on ECG data that published between January 1st 2010 and September 30th 2019 from Google Scholar, PubMed and DBLP. Then we analyze them in three aspects including task, model and data. Finally, we conclude unresolved challenges and problems that existing models can not handle well. Results: The total number of papers is 124, among them 97 papers are published after in recent two years. Almost all kinds of common deep learning architectures have been used in ECG analytics tasks like disease detection/classification, annotation/localization, sleep staging, biometric human identification, denoising and so on. Conclusion: The number of works about deep learning on ECG data is growing explosively in recent years. Indeed, these works have achieve a far more better performance in terms of accuracy. However, there are some new challenges and problems like interpretability, scalability, efficiency, which need to be addressed and paid more attention. Moreover, it is also worth to investigate by discovering new interesting applications from both the dataset view and the method view. Significance: This paper summarizes existing deep learning methods on modeling ECG data from multiple views, while also point out existing challenges and problems, while can become potential research direction in the future.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

Electrocardiogram (ECG/EKG) is the most commonly used non-invasive diagnostic tool that records physiological activities of heart over a period of time. ECG can help diagnose many cardiovascular abnormalities such as premature contractions of atria (PAC) or ventricles (PVC), atrial fibrillation (AF), myocardial infarction (MI), and congestive heart failure (CHF). In recent years, we have witnessed a rapid development of portable ECG monitor in medical area such as Holter [98] , and wearble devices in healthcare area such as Apple Watch. Consequently the amount of ECG data grows rapidly so that human cardiologistis are inadequate for analyzing them. Thus how to analyze ECG data automatically and accurately has became an hot research topic for many years. Moreover, many new emerging applications such as biometric human identification, sleep staging can also be implemented based on ECG data.

Fig. 1: Comparison between traditional methods and deep learning methods.

Traditionally, automatic ECG analysis relies on diagnostic golden rules. As shown in the top of Figure 1, it is a two-stages method which firstly required extracting expert features

from raw ECG data, then deployed decision rules or other machine learning methods for the final results. The feature extraction step is mostly done by human experts. Hence, it is labor intensive and limited

[114, 118, 40].

Recently, deep learning methods have achieved promising results on many data analytic applications such as speech recognition, image classification, computer vision, and natural language processing

[67]. The main difference between deep learning methods and traditional methods is that they don’t require an explicit feature extraction step by human experts, as shown in the bottom of Figure 1. This feature extraction step is done automatically and implicitly by deep learning models, due to their powerful data learning abilities and flexible processing architectures. Some studies have experimentally shown that deep learning features are more informative than expert features on ECG data [51, 53]. The performance of deep learning methods are also better than traditional methods on many ECG analysis tasks such as disease detection [23], sleep staging [33] and so on.

Although some papers have reviewed deep learning methods on cardiovascular images [82], or more general Cardiology [61], there are no systematic reviews focusing on deep learning methods which we consider them as a promising way to mine the ECG data. Thus, we feel it’s amenable and necessary to conduct a systematic review of existing deep learning methods on ECG data from the perspective of model architecture and their application task. Challenges and problems of current research status are discussed, which, we believe, will give some inspiration and insights for the future work.

Ii Method

Ii-a Literature Search and Selection

In order to conduct a comprehensive review, we searched papers that deployed deep learning methods (deep neural network networks) on ECG data from Google Scholar, PubMed and DBLP from January 1st 2010 to September 30th 2019. The searched keywords are a combination of “deep learning”, “deep neural network”, “deep neural networks” and “Electrocardiogram”, “ECG”, “EKG”. We listed the search results of paper source as follows:

  • Medical Information and Biomedical Engineering (MI & BME): Circulation, Journal of the American College of Cardiology (JACC), Nature Medicine, Nature Biomedical Engineering, American Medical Informatics Association (AMIA), Journal of American Medical Informatics Association (JAMIA), Journal of American Biomedical Informatics (JBI), Transactions on Biomedical Engineering (TBME), Biomedical Signal Procession and Control, Computing in Cardiology (CinC), Physiological Measurement (PMEA), Computers in Biology and Medicine, IEEE Journal of Biomedical and Health Informatics

  • Artificial Intelligence and Data Mining (AI & DM): Nature Machine Intelligence, ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), AAAI Conference on Artificial Intelligence (AAAI), International Joint Conference on Artificial Intelligence (IJCAI), Neural Information Processing Systems (NIPS), International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), IEEE Transactions on Neural Networks and Learning Systems (TNNLS), IEEE Transactions on Knowledge and Data Engineering (TKDE), IEEE Transactions on Cybernetics, Neurocomputing, Knowledge Based System, Expert Systems with Applications.

  • Interdisciplinary: Nature Scientific Reports, PLoS One, IEEE Access.

The framework of literature search and selection is shown in Figure 2. It is a two-stages method including search and selection. To avoid missing papers that not explicitly mentioned their method in title, or named their method as some other names without above keywords. We expand our search field to “anywhere in the article”, then manually filter out papers that don’t use deep learning from them. Notice that a large majority of unrelated papers would more or less mention above keywords in their introduction part, related work part and so on. Thus, we result in the final list of deep learning papers (124 papers) on ECG would be much less than initial search (727 papers).

Fig. 2: Framework of literature search and selection.

Ii-B Literature Analysis

For each paper, we analyze them in the following three aspects:

  • Task: the targeted application tasks include: (1) disease detection, e.g. specific diseases like atrial fibrillation, myocardial infarction, congestive heart failure, ST elevation, or general diagnostic arrhythmia, (2) annotation or localization, e.g. QRS complex annotation, P wave annotation, localizing the origin of ventricular activation, (3) sleep staging, (4) biometric human identification, (5) denoising, and (6) the Others.

  • Model

    : deep model architectures include (1) convolutional neural networks (CNN), (2) recurrent neural networks (RNN), (3) combination of CNN and RNN (CRNN), (4) autoencoders (AE), (5) generative adversarial networks (GAN), (6) fully connected neural networks (FC) and others. Moreover, we also identify (7) bu whether they include traditional expert features or integrate expert knowledge in building the deep model.

  • Data: the statistics of data includes (1) the size of dataset, both number of samples and duration of each sample, (2) sampling frequency, and (3) number of channels (number of electrode leads).

We summarize all papers from the perspective of Models and Tasks, as shown in Figure 3. Finally, we conclude challenges and problems that existing models can not handle well.

Fig. 3: An overview of the analysis from aspects of task, model and data.

Iii Results

We include 124 papers in analysis, from aspects of task, model and data. An overview of the analysis is shown in Figure 3. An overall statistics of these papers is shown in Figure 4. Notice that 97 papers (about 78%) are published after 2018, 65 papers come from medical information and biomedical engineering community, while only 20 papers (about 16%) come from artificial intelligence and data mining community. In the rest of Results section, we will discuss more details from aspects of task, model and data.

Fig. 4: Overall statistics of all 124 papers.
Disease Detection Anotation or Localization Sleep Staging Biometric Human Identification Denoising Others
CNN [31, 6, 46, 5, 143, 64, 152, 154, 76, 55, 49, 141, 48, 25, 16, 123, 138, 78, 72, 72, 47, 101, 120, 97, 65, 51, 146, 4, 139, 112, 79, 34, 53, 102, 122, 145, 105, 121, 62, 74, 153, 14, 1, 140, 85, 77, 106, 37, 117, 57, 91] [150, 43, 15, 160, 71] [128, 89] [157, 44, 22, 158] [20, 159] [92, 7, 113]
RNN [129, 48, 25, 75, 88, 18, 116, 126, 151, 19] [8, 42, 103] [137] N.A. [70] N.A.
CRNN [127, 104, 131, 10, 52, 83, 163, 136, 107, 135, 125, 99, 29, 3, 132, 161, 96, 119] N.A. N.A. N.A. N.A. [39, 162, 27, 108]
AE [142, 129, 156, 28] [42, 41] [73] N.A. [20, 144, 70] [35]
GAN [132] N.A. N.A. N.A. [130] [162, 69]
With Expert Features [55, 51, 4, 139, 88, 34, 53, 126, 62, 37] N.A. N.A. [44] [159] N.A.
FC & Others [63, 87, 30, 90, 100, 147, 148, 32] [115] N.A. [68] N.A. [155]
TABLE I: Summary of papers from the perspective of models and tasks.

Iii-a Task

Iii-A1 Disease Detection

The goal of developing a deep learning model for disease detection is to map the input ECG/EKG data to the output disease target via multiple layers of neural networks. For instance, convolutional neural networks is applied on ECG/EKG signals for automated detection of cardiac arrhythmias (e.g., atrial flutter, supraventricular tachyarrhythmia, T Ventricular trigeminy, etc.) [152, 62, 46]. The atrial fibrillation (AF) detection can be regarded as a special case of cardiac arrhythima detection task, where all non-AF rhythms are grouped together [126, 161, 52]. In [101, 143], a deep learning technique is introduced to monitor ST change in ECG/EKG data.

In [85] and [131], convolutional neural networks was applied to automate the detection of the Myocardial Infarction and Congestive Heart Failure respectively. Some studies include data from multiple modalities (e.g., ECG, transthoracic echocardiogram [5]), and support both binary classification (e.g.,patients with paroxysmal atrial fibrillation or Healthy [106]), multi-class classification (e.g., detection of acute cognitive stress detection [47], decompensation of patients detection [149]) and multi-task classification (e.g., detection of prevalent hypertension, sleep apnea, and diabetes [127]).

Iii-A2 Annotation or Localization

Annotation and localization of specific waves (e.g., QRS, P-wave) in ECG/EKG signals is of great importance to cardiologists in order to help them diagnosing cardiac disease such as atrial fibrillation. Most studies which applied deep learning method to ECG/EKG annotation or localization focus on using deep learning method for the QRS detection, P-wave detection, localization of ventricular tachycardia, localization of origins of premature ventricular, as well as for the discovering of the patterns of ECG/EKG.

Autoencoder (AE) [41] and recurrent neural network (RNN) [42] are employed to automatically detecting the exit of ventricular tachycardia from 12-lead electrocardiogram. [15], [160] and [71] focused instead on the annotation of fetal QRS complex (detecting the Q-wave, R-wave and S-wave and calculating the heart rate), which is critical to determine various arrhythmias, from the MIT-BIH arrhythmia dataset. Some studies use the QT database (QTDB) on PhysioNet to explored ways of other types of annotating ECG/EKG waves, including P-wave [103, 8], T-wave annotation [8], etc.

Iii-A3 Sleep Staging

Understanding sleep is critical for the whole healthcare system, as sleep is a key ingredient to our well-being. Sleep disorder may lead to catastrophes in personal medicine or public health [89]. In [73]

, sparse autoencoder (SAE) and Hidden Markov model (HMM) are combined together to detect the obstructive sleep apnea (OSA) using the PhysioNet challenge 2000 dataset. Convolutional Neural Network is used to detect the OSA instead of the SAE-HMM method

[89].

To automatically identify the sleep stage, LSTM network is employed for [137] from a multi-channel physiological signals dataset (EEG, EOG, and EMG signals) which is collected from the sleep disorders diagnosis center of Xijing Hospital, Fourth Military Medical University. In [128], convolutional neural network (CNN) based deep learning architecture is employed for multi-class classification of obstructive sleep apnea and hypopnea (OSAH), which is the most common sleep-related breathing disorder, using single-lead electrocardiogram (ECG) recordings.

Iii-A4 Human Identification

With the rapid development of information technology, body sensor networks are reshaping people’s daily lives, especially in smart health applications. Biometric-based human identification (ID) is a promising technology for automatic and accurate individual recognition using various body sensor data, such as heart rate, temperature and activity. For example, [158] proposes a novel ECG biometric authentication system that incorporates generalized S-transformation (GST) and CNN techniques. Zhang et al. built a multiresolution convolutional neural network (MCNN) based biometric human identification system [157] and evaluated their system using eight dataset from PhysioNet [36] including CEBSDB, WECG, FANTASIA, NSRDB, STDB, MITDB, AFDB, VFDB and showed better performance using CNN than existing systems.

A secure multimodal biometric system that uses convolution neural network (CNN) and Q-Gaussian multi support vector machine (QG-MSVM) based on a different level fusion is employed

[44]. Moreover, PTB Diagnostic dataset (containing 549 15-channels ECG records from 290 subjects) [12] and CYBHi dataset (containing 65 subjects with an average age between 21.64 and 40.56 years) [24] are used to evaluate the proposed method. [22] present an ECG biometric authentication method based on parallel multi-scale one-dimensional residual network, which can improve the generalization ability of the method on different ECG signals sampled in the different environment for the matching task. In [68]

, the features are extracted by the principal component analysis network (PCANet) and then using a robust EigenECG network (REECGNet) based on time-frequency representations of ECGs for personal identification.

Iii-A5 Denoising

The ECG signals acquisition process is often accompanied by a large amount of noise, which will seriously affect the doctor’s diagnosis of patients, especially in the telemedicine environment. [159] addresses this issue by proposing a noise rejection method based on the combination of modified frequency slice wavelet transform (MFSWT) and convolutional neural network (CNN). Generative adversarial method is employed for ECG signals de-noising and evaluates the quality of de-noised signals against SVM algorithm. [130]

Xiong et al. [144] and Chiang et al. [20] propose a denoising auto-encoders (DAEs) and Fully convolutional network (FCN) based DAE method respctively for ECG signal denoising which are evaluated on ECG signals from the bench-marker MIT-BIH Arrhythmia Database and the noises come from the MIT-BIH noise stress test database. In [70], bidirectional recurrent denoising auto-encoder (BRDAE) is used, and the evaluation results on MIMIC-III database indicate that it promises values beyond traditional denoising method by providing PPG feature accentuation for pulse waveform analysis.

Iii-A6 Others

The convolutional neural network (CNN) used in the other types of the studies are mainly including emotion detection [113], and drug assessment [7], data compression [155]

, etc. In addition to the CNN, Long Short-Term Memory (LSTM) is used to learn long-term dependencies. For instance,

[27]

use the combination of CNN and LSTM to detect pulse and classify an organized ECG into pulsed rhythm (PR) or pulseless electrical activity (PEA). In

[39] and [108], the combination of CNN and LSTM method is employed, respectively, to predict needs for urgent revascularization and classify the driver stress level.

Besides combination of CNN and LSTM, [162] formulated Electrocardiogram generation as a sequential data generation problem using generative adversarial network (GAN). V-lead ECG signals is synthesized from limb leads using an R-peak aligned generative adversarial network (GAN) to [69]. In the reviewed articles, some method study the use of Residual Network (ResNet) to deal the raw ECG/EKG waveform data to assess the risk of future cardiac disease. [92] Risk scores derived from an ECG/EKG measurement do not depend on patient history and can be more generally applicable to new patients, patients with missing health information, or patients from clinically underserved populations. Gogna et al. [35] propose a semi-supervised stacked label consistent autoencoder (AE) for reconstruction and analysis of biomedical signals.

Iii-B Model

Iii-B1 Cnn

Convolutional neural network (CNN) is a class of deep neural networks widely applied for image classification, natural language processing and signal analysis. It can automatically extract hierarchical pattern in data using stacked learnable small filters or kernels which means it requires little pre-processing compared to hand-engineered features. A typical CNN is composed of convolutional layers followed by batch normalization layer, non-linear activation layer, dropout layer, pooling layer in the first few stages and classification layers like fully connected layers in the subsequent stages as done in 

[46]. In some works, support vector machine, boosting classifier tree and RNN can also be the alternatives for fully connected layers to summarize the global feature from CNNs. CNNs have proved to achieve superior performance and compute fast due to its shared-weights architecture and ability of parallelization.

Two types of CNN are commonly used for ECG classification, namely, 1-D CNN and 2-D CNN. In detail, 1-D CNN works by applying the kernel along temporal dimension on the raw ECG data, while 2-D CNN is usualy done on the transformed ECG data such as distance distribution matrix by entroy calculation [76], gray-level co-occurrence matrix (GLCM) [123] or combined feature like morphology, RR intervals and beat-to-beat correlation [72]

. But there is some contradition when it comes to multi-head ECG or ECG time-frequency spectrograms extracted by wavelet transform, fast Fourier transform (FFT) and short-term Fourier transform (STFT). Some work like 

[55] directly applied 2D-CNN on it but the problem is the different frequency resolutions, meaning that most signal characteristics are reflected by intra-component patterns, not inter-component behaviors. To handle this problem, shared 1-D CNN is used in [85] and multi-scale 1-D CNN is similarily used in [157] for biometric human identification.

In addition, 2-D CNN can be done on an one-head ECG signal treated as a image. In this case, pre-trained ResNet, DenseNet and Inception-Net on ImageNet can be fine-tuned on a ECG dataset for heart beat problems detection 

[138] or examining ST changes [132]. Particularly for Localization task such as QRS detection, fully convolutional network (FCN) with larger kernel size can be applied on a ECG image to squeeze the image’ height size to 1 and keep the output length same as input to get the QRS window label [71].

Some advanced techniques like atrous spatial pyramid pooling (ASPP) module is used to exploit multi-scale features from ECG. Moreover, active learning 

[141], data augmentation [132, 161] can be incorporated into the CNN framework to tackle the imbalance problem and further improve the accuracy.

Iii-B2 Rnn

Recurrent Neural Network(RNN) is a type of neural network naturally designed to model sequential data, such as time series, event sequences and natural language text. It works in the way where the output from previous step are fed as input to the current step. By iteratively updating the hidden state and memory, it is capable to remember information in sequence order.

In particular for ECG data, RNN is a preferred choice for both capturing the temporal dependency and handling varied length input. GRU/LSTM, Bidirectional-LSTM (BiLSTM) are commonly used RNN variants which tackles one critical problem called vanishing gradient caused by vanilla RNN. In [75], attention mechanism is combined with BiLSTM to provide performance gain and interpretability by visualizing the attention weight.

Iii-B3 Crnn

CRNN, as the name suggests, is composed of CNN and RNN modules introduced above. It’s a preferred choice to handle long ECG sigal with varied sequence length and multi-channel input. 1-D CNN [48] or 2-D CNN [119] is applied on segments of ECG to extract the local feature followed by BiLSTM to summarize the multiple segments feature long time dimension as the global feature for classification.

DeepHeart [10] follows the CRNN framework for cardiovascular risk prediction, and utilized an auto-encoder model(see Sec. III-B4) to initialize model weights which achieved better performance. In order to provide interpretability for diagnosis decision, recently a model called MINA [52]

incorporates CRNN with multi-level attention model where beat, rhythm and frequency-level domain knowledge is extracted and different attentions are learned.

Iii-B4 Ae

Autoencoder(AE) is a type of neural network composed of an encoder module and a decoder module used to learn hidden embedding in an unsupervised manner. The goal of the AE is to learn a reduced dimension representation by the encoder while the decoder trys to generate from the representation as close as the original data. There are three commonly used varients of AE, namely, Denoising autoencoder (DAE), Sparse autoencoder (SAE) and Contractive autoencoder (CAE). DAEs take a partially corrupted input and are trained to recover the original undistorted input, while SAE and CAE utilize different regularization method like KL-divergence and Frobenius norm of the Jacobian matrix respectively to learn more robust hidden representation for classification.

Stacked DAE [142], SAE [156] and CAE [144] are widely used for ECG denoising purpose, since ECG signals are prone to be contaminated by various kinds of noise, such as baseline wander, electrode contact noise and motion artifacts, which may lead to wrong interpretation. In practise, the crux of the matter comes to the choice of the encoder and decoder module. As readly introduced in Sec. III-B1III-B2III-B3, CNN, RNN and CRNN would be pairwise combined. [20] utilized the fully convolutional networks as the encoder and decoder while [66] used BiLSTM. To further improve the classification performance, [35] simultaneously carried out the reconstruction and classification procedure.

Iii-B5 Gan

Generative adversarial network (GAN) is a class of neural network framework invented by Ian Goodfellow et. al [38]. This generative model consists of two models: a generative model that captures the data distribution of the training dataset from a latent representation, and a discriminative model

that distinguishes the probability that a sample produced by the generator from the true data distribution. These two models are trained iteratively to conduct a minimax game.

GAN applications have increased rapidly, especially in areas like image generation [13] and language generation [80]. Recently, it has been applied to tackle the imbalanced-data challenge remaining in ECG data. To name a few, [132] proposed an abnormality detection model for ECG signals based a CRNN framework and using GAN composed of multiple 1-D CNN to do data augmentation which shows high performance for class-imbalanced dataset. [138] utilized GAN to de-noising the ECG and [162] proposed GAN composed of a BiLSTM(generator) and CNN(discriminator) to generate synthetic ECG data to tackle the problem that a large volume of labeled clinical data is required to train a deep learning model.

Iii-B6 With Expert Features

According to ENCASE [51]

, the expert features can be divided into three categories: 1) Statistical features includes count, mean, maximum, minimum and so on. 2) Signal procession features which transform ECG data from time domain to frequency domain including Fast Fourier Transform (FFT), Wavelet Transform (WT), Shannon entropy and so on. 3) Medical Features based on medical domain knowledge, for example, features based on P,Q,R,S and T waves, sample entropy, coefficient of variation and density histograms (CDF) and so on.

All the methods mentioned above could benefit a lot from the expert features although extra efforts are needed to extract them compared to the raw morphological features as the input to Deep Neural Networks (DNN). Nowadays, ensemble methods [51, 55, 139] take as the input of the expert features combined with raw morphological features and incorporate the random foresets, boosting tree models with DNN are the state-of-the-art method as far as we know.

Iii-B7 FC & Others

Some works also rely on the fully connected neural networks (FC) to do disease detection and classification [30, 63] especially for extremelly short ECG like only 10-RR-interval [63]. Other works borrow ideas from image vision area. To name a few, U-net [111], stemming from the fully convolutional network (FCN) and consisting of a contracting path and an expansive path is widely used in image segmentation task. [100] proposed a modified U-net to handle varied length ECG for classification and R-peak detection. PCANet [17], a image classification model working with the help of cascaded principal component analysis (PCA), binary hashing, and block-wise histograms, is modified for biometric human identification task in nonstationary ECG noise environment [68]. But its effect remains to be evaluated with the state-of-the-arts baselines introduced in above sections.

Iii-C Data

The datasets can be categorized into two groups: (1) collected from medical devices, and (2) collected from healthcare devices. The biggest difference is that medical devices data have more channels than healthcare devices data, so that it is more informative. However, medical devices data is also much harder to collect, so that the amount of healthcare devices data is much more. Moreover, most of the works (98 out of 124) used open source datasets, which makes it easier for follow-up works and reproducibility.

Concretely, three most frequently used open source datasets are as follows:

  • MIT-BIH Arrhythmia Database111https://physionet.org/content/mitdb/1.0.0/ [93] (MITDB, 39 papers) consists of 48 half hour ECG records from 47 subjects at Boston’s Beth Israel Hospital. Each ECG data is an 11-bit resolution over a 10 mV range with a sampling frequency of 360 Hz. This dataset is fully annotated with both beat level diagnosis and rhythm level diagnosis.

  • PhysioNet Computing in Cardiology Challenge 2017222https://www.physionet.org/content/challenge-2017/1.0.0/ [23] (28 papers) contains 8,528 de-identified ECG recordings lasting from 9s to just over 60s and sampled at 300Hz by the AliveCor healthcare device. Among them, 5154 recordings are normal, 717 recordings are AF, 2557 recordings are others and 46 recordings are noise.

  • PTB Diagnostic ECG Database333https://physionet.org/content/ptbdb/1.0.0/ [12] (PTBDB, 8 papers) contains 549 15-channels ECG records from 290 subjects. The sampling rates is available at up to 10 KHz. Among these subjects, 216 of them have 8 types of heart disease patients, and 52 of them are healthy control, while 22 is unknown.

Others including: MIT-BIH Atrial Fibrillation Database [95] (AFDB, 5 papers), MIT-BIH Normal Sinus Rhythm Database (NSRDB, 4 papers), 2018 China Physiological Signal Challenge 444http://2018.icbeb.org/Challenge.html [84] (CPSC, 3 papers), QT Database [66] (QTDB, 4 papers), St Petersburg INCART 12-lead Arrhythmia Database (INCARTDB, 3 papers), MIT-BIH Malignant Ventricular Ectopy Database (VFDB, 3 papers), CU Ventricular Tachyarrhythmia Database (CUDB, 3 papers), Medical Information Mart for Intensive Care [60] (MIMIC-III, 2 papers), Congestive Heart Failure RR Interval Database (CHF2DB, 2 papers), AHA Database Sample Excluded Record (AHADB, 2 papers), MIT-BIH Noise Stress Test Database [94] (NSTDB, 2 papers), ECG-ID Database [86] (ECGIDDB, 2 papers), BIDMC Congestive Heart Failure Database [9] (BIDMC, 1 paper), MIT-BIH ST Change Database [2] (STDB, 1 paper), ECG Effects of Ranolazine, Dofetilide, Verapamil, and Quinidine [59] (ECGRDVQ, 1 paper), ECG Effects of Dofetilide, Moxifloxacin, Dofetilide+Mexiletine, Dofetilide+Lidocaine and Moxifloxacin+Diltiazem [58] (ECGDMMLD, 1 paper), Long Term ST Database [56] (LTSTDB, 1 paper), European ST-T Database [124] (EDB, 1 paper), Noninvasive Fetal ECG - The PhysioNet Computing in Cardiology Challenge 2013 (1 paper).

Above ECG dataset without providing link can be found in PhysioNet website555https://physionet.org [36], or the alternative archived PhysioNet website666https://archive.physionet.org.

Iv Discussion of Opportunities and Challenges

In this section, we will discuss current challenges and problems of deep learning on ECG works. In the meantime, potential opportunities are also identified along with these challenges and problems. Again, we will discuss from the perspective of model, data and task.

From the perspective of model, following three challenges and problems need to be considered:

  • Interpretability. Deep learning models are often regarded as black box model, because they usually have many model parameters, or complex model architecture, that is hard for human to understand the reason why a certain result is given by the model. This challenge is much severer in medical domain tasks, since diagnosis without any explanation is not acceptable by medical experts. To handle with this, two directions are worth studying. The first one is how to interpret a complex deep learning model by a relatively simple model. For example, one can first build a black box deep learning model for a specific task, then build a separate interpretable simple model which is according with deep learning model’s prediction, and interpret the prediction based on the simple model [109, 110]

    . The second one is how to directly build an interpretable deep model. For example, when designing deep model architecture, one can borrow neuron connection ideas from tree based model

    [133], or adding attention mechanism on hidden layers [21, 52], which can be better understood by human.

  • Efficiency. Since deep models are much more complex, it’s hard to deploy big models to portable healthcare devices, which is a huge obstacle to applying deep learning model on real world applications. In this situation, a promising research direction is model compression technique. For example, knowledge distillation is commonly used to transform a big and powerful model to a simple model with minor accuracy decrease [50]. Also, we can use the quantization, weight sharing, and careful coding of network weights [45] to compress a big model.

  • Integration with expert features. Most of existing deep learning models are trained end-to-end, so it is hard to integrate with existing expert features or expert knowledge once the model finished training. And, there are two research lines to tackle aforementioned problems. The first one is use domain expert knowledge to design deep neural network architecture [54, 52]

    . The second one is to regard deep models as feature extractors, coined as deep features

    [51, 53], and explicitly extract latent embeddings from deep learning models. Then one can easily combine expert features with deep features, and build traditional machine learning methods on them.

Then, from the perspective of model, following two challenges and problems need to be considered:

  • Imbalanced labels. The ECG disease labels are very likely to be a biased distribution, since many severe diseases happen rarely, but they are actually much more important. It is hard to train an effective deep learning method with a large amount of model parameters on such a small dataset. There are two ways to handle with this problem. The first one is data augmentation such as data preprocessing using side-and-cut technique, or generating more training dataset using generative models like variational autoencoders (VAE) [26] or GAN [38]

    . The second one is to design new loss functions like Focal loss

    [81], or other model training schema like few shot learning [134].

  • Multi-modal data. Currently, most works only consider ECG for analysis. With the development of medical devices and healthcare devices, many other vital signs like temperature, respiratory rate, blood pressure can also be collected along with ECG. However, these data are not always synchronized in time line, and their sampling frequencies are also varied, so they can be regarded as multi-modal data. It is a potential opportunity to study on how to design a model which is capable of utilizing these multi-modal data simultaneously to improve task performance compared with the model trained on any individual data.

Finally, there remains a lot of interesting and innovative application tasks to be studied, including: mental stress measurement (the electrode is equipped in steering wheel for driver), emotion detection, analyze mammalian ECG [11], create risk scores of out-hospital patients via healthcare devices and so on.

V Conclusion

In this paper, we systematically reviewed existing deep learning (deep neural network) methods on Electrocardiogram (ECG) data, from the perspective of model, data and task. We found that deep learning methods can also achieve better performance than traditional methods on ECG modeling. However, we also found that these deep learning still face up with some unresolved challenges and problems. This paper can provide an systematic overview of how deep learning methods can be used to solve real applications for engineers, while also point our some potential research opportunities in the future for researchers.

Acknowledgement

This work was in part supported by the National Science Foundation award IIS-1418511, CCF-1533768 and IIS-1838042, the National Institute of Health award NIH R01 1R01NS107291-01 and R56HL138415.

References

  • [1] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, M. Adam, A. Gertych, and R. San Tan (2017) A deep convolutional neural network model to classify heartbeats. Computers in biology and medicine 89, pp. 389–396. Cited by: TABLE I.
  • [2] P. Albrecht (1983) ST segment characterization for long term automated ecg analysis. Ph.D. Thesis, Massachusetts Institute of Technology, Department of Electrical Engineering …. Cited by: §III-C.
  • [3] R. S. Andersen, A. Peimankar, and S. Puthusserypady (2019) A deep learning approach for real-time detection of atrial fibrillation. Expert Systems with Applications 115, pp. 465–473. Cited by: TABLE I.
  • [4] F. Andreotti, O. Carr, M. A. Pimentel, A. Mahdi, and M. De Vos (2017) Comparing feature-based classifiers and convolutional neural networks to detect arrhythmia from short segments of ecg. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [5] Z. I. Attia, S. Kapa, F. Lopez-Jimenez, P. M. McKie, D. J. Ladewig, G. Satam, P. A. Pellikka, M. Enriquez-Sarano, P. A. Noseworthy, T. M. Munger, et al. (2019) Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nature medicine 25 (1), pp. 70. Cited by: §III-A1, TABLE I.
  • [6] Z. I. Attia, S. Kapa, F. Lopez-Jimenez, S. J. Asirvatham, P. A. Friedman, and P. A. Noseworthy (2018) Electrocardiographic screening for atrial fibrillation while in sinus rhythm using deep learning. Circulation 138 (Suppl_1), pp. A15871–A15871. Cited by: TABLE I.
  • [7] Z. I. Attia, A. Sugrue, S. J. Asirvatham, M. J. Ackerman, S. Kapa, P. A. Friedman, and P. A. Noseworthy (2018) Noninvasive assessment of dofetilide plasma concentration using a deep learning (neural network) analysis of the surface electrocardiogram: a proof of concept study. PloS one 13 (8), pp. e0201059. Cited by: §III-A6, TABLE I.
  • [8] M. Attin, A. Cogliati, and Z. Duan (2017) Annotating ecg signals with deep neural networks. Circulation 136 (suppl_1), pp. A19056–A19056. Cited by: §III-A2, TABLE I.
  • [9] D. S. Baim, W. S. Colucci, E. S. Monrad, H. S. Smith, R. F. Wright, A. Lanoue, D. F. Gauthier, B. J. Ransil, W. Grossman, and E. Braunwald (1986) Survival of patients with severe congestive heart failure treated with oral milrinone. Journal of the American College of Cardiology 7 (3), pp. 661–670. Cited by: §III-C.
  • [10] B. Ballinger, J. Hsieh, A. Singh, N. Sohoni, J. Wang, G. H. Tison, G. M. Marcus, J. M. Sanchez, C. Maguire, J. E. Olgin, et al. (2018) DeepHeart: semi-supervised sequence learning for cardiovascular risk prediction. In Thirty-Second AAAI Conference on Artificial Intelligence, Cited by: §III-B3, TABLE I.
  • [11] J. A. Behar, A. A. Rosenberg, I. Weiser-Bitoun, O. Shemla, A. Alexandrovich, E. Konyukhov, and Y. Yaniv (2018) PhysioZoo: a novel open access platform for heart rate variability analysis of mammalian electrocardiographic data. Frontiers in physiology 9, pp. 1390. Cited by: §IV.
  • [12] R. Bousseljot, D. Kreiseler, and A. Schnabel (1995) Nutzung der ekg-signaldatenbank cardiodat der ptb über das internet. Biomedizinische Technik/Biomedical Engineering 40 (s1), pp. 317–318. Cited by: 3rd item, §III-A4.
  • [13] A. Brock, J. Donahue, and K. Simonyan (2018) Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096. Cited by: §III-B5.
  • [14] W. Cai, Y. Chen, J. Guo, B. Han, Y. Shi, L. Ji, J. Wang, G. Zhang, and J. Luo (2019) Accurate detection of atrial fibrillation from 12-lead ecg using deep neural network. Computers in Biology and Medicine, pp. 103378. Cited by: TABLE I.
  • [15] J. Camps, B. Rodríguez, and A. Mincholé (2018) Deep learning based qrs multilead delineator in electrocardiogram signals. In 2018 Computing in Cardiology Conference (CinC), Vol. 45, pp. 1–4. Cited by: §III-A2, TABLE I.
  • [16] X. Cao, B. Yao, and B. Chen (2019) Atrial fibrillation detection using an improved multi-scale decomposition enhanced residual convolutional neural network. IEEE Access 7, pp. 89152–89161. Cited by: TABLE I.
  • [17] T. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma (2015) PCANet: a simple deep learning baseline for image classification?. IEEE transactions on image processing 24 (12), pp. 5017–5032. Cited by: §III-B7.
  • [18] Y. Chang, S. Wu, L. Tseng, H. Chao, and C. Ko (2018) AF detection by exploiting the spectral and temporal characteristics of ecg signals with the lstm model. In 2018 Computing in Cardiology Conference (CinC), Vol. 45, pp. 1–4. Cited by: TABLE I.
  • [19] S. Chauhan, L. Vig, and S. Ahmad (2019) ECG anomaly class identification using lstm and error profile modeling. Computers in biology and medicine 109, pp. 14–21. Cited by: TABLE I.
  • [20] H. Chiang, Y. Hsieh, S. Fu, K. Hung, Y. Tsao, and S. Chien (2019) Noise reduction in ecg signals using fully convolutional denoising autoencoders. IEEE Access 7, pp. 60806–60813. Cited by: §III-A5, §III-B4, TABLE I.
  • [21] E. Choi, M. T. Bahadori, J. Sun, J. Kulas, A. Schuetz, and W. F. Stewart (2016) RETAIN: an interpretable predictive model for healthcare using reverse time attention mechanism. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pp. 3504–3512. Cited by: 1st item.
  • [22] Y. Chu, H. Shen, and K. Huang (2019) ECG authentication method based on parallel multi-scale one-dimensional residual network with center and margin loss. IEEE Access 7, pp. 51598–51607. Cited by: §III-A4, TABLE I.
  • [23] G. D. Clifford, C. Liu, B. Moody, H. L. Li-wei, I. Silva, Q. Li, A. Johnson, and R. G. Mark (2017) AF classification from a short single lead ecg recording: the physionet/computing in cardiology challenge 2017. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: §I, 2nd item.
  • [24] H. P. Da Silva, A. Lourenço, A. Fred, N. Raposo, and M. Aires-de-Sousa (2014) Check your biosignals here: a new dataset for off-the-person ecg biometrics. Computer methods and programs in biomedicine 113 (2), pp. 503–514. Cited by: §III-A4.
  • [25] H. Dang, M. Sun, G. Zhang, X. Qi, X. Zhou, and Q. Chang (2019) A novel deep arrhythmia-diagnosis network for atrial fibrillation classification using electrocardiogram signals. IEEE Access. Cited by: TABLE I.
  • [26] C. Doersch (2016) Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908. Cited by: 1st item.
  • [27] A. Elola, E. Aramendi, U. Irusta, A. Picón, E. Alonso, P. Owens, and A. Idris (2018) Deep learning for pulse detection in out-of-hospital cardiac arrest using the ecg. In 2018 Computing in Cardiology Conference (CinC), Vol. 45, pp. 1–4. Cited by: §III-A6, TABLE I.
  • [28] J. Farhadi, G. Attarodi, N. J. Dabanloo, M. Mohandespoor, and M. Eslamizadeh (2018) Classification of atrial fibrillation using stacked auto encoders neural networks. In 2018 Computing in Cardiology Conference (CinC), Vol. 45, pp. 1–3. Cited by: TABLE I.
  • [29] O. Faust, A. Shenfield, M. Kareem, T. R. San, H. Fujita, and U. R. Acharya (2018) Automated detection of atrial fibrillation using long short-term memory network with rr interval signals. Computers in biology and medicine 102, pp. 327–335. Cited by: TABLE I.
  • [30] N. Feng, S. Xu, Y. Liang, and K. Liu (2019) A probabilistic process neural network and its application in ecg classification. IEEE Access 7, pp. 50431–50439. Cited by: §III-B7, TABLE I.
  • [31] M. Gadaleta, M. Rossi, S. R. Steinhubl, and G. Quer (2018) Deep learning to detect atrial fibrillation from short noisy ecg segments measured with wireless sensors. Circulation 138 (Suppl_1), pp. A16177–A16177. Cited by: TABLE I.
  • [32] C. D. Galloway, A. V. Valys, F. L. Petterson, V. P. Gundotra, D. L. Treiman, D. E. Albert, J. J. Dillon, Z. I. Attia, and P. A. Friedman (2018) Non-invasive detection of hyperkalemia with a smartphone electrocardiogram and artificial intelligence. Journal of the American College of Cardiology 71 (11 Supplement), pp. A272. Cited by: TABLE I.
  • [33] M. M. Ghassemi, B. E. Moody, L. H. Lehman, C. Song, Q. Li, H. Sun, R. G. Mark, M. B. Westover, and G. D. Clifford (2018) You snooze, you win: the physionet/computing in cardiology challenge 2018. In 2018 Computing in Cardiology Conference (CinC), Vol. 45, pp. 1–4. Cited by: §I.
  • [34] S. Ghiasi, M. Abdollahpur, N. Madani, K. Kiani, and A. Ghaffari (2017) Atrial fibrillation detection using feature based algorithm and deep convolutional neural network. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [35] A. Gogna, A. Majumdar, and R. Ward (2016) Semi-supervised stacked label consistent autoencoder for reconstruction and analysis of biomedical signals. IEEE Transactions on Biomedical Engineering 64 (9), pp. 2196–2205. Cited by: §III-A6, §III-B4, TABLE I.
  • [36] A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. Peng, and H. E. Stanley (2000) PhysioBank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. Circulation 101 (23), pp. e215–e220. Cited by: §III-A4, §III-C.
  • [37] Z. Golrizkhatami and A. Acan (2018) ECG classification using three-level fusion of different feature descriptors. Expert Systems with Applications 114, pp. 54–64. Cited by: TABLE I.
  • [38] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio (2014) Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680. Cited by: §III-B5, 1st item.
  • [39] S. Goto, M. Kimura, Y. Katsumata, S. Goto, T. Kamatani, G. Ichihara, S. Ko, J. Sasaki, K. Fukuda, and M. Sano (2019) Artificial intelligence to predict needs for urgent revascularization from 12-leads electrocardiography in emergency patients. PloS one 14 (1), pp. e0210103. Cited by: §III-A6, TABLE I.
  • [40] M. E. Guglin and D. Thatai (2006) Common errors in computer electrocardiogram interpretation. International journal of cardiology 106 (2), pp. 232–237. Cited by: §I.
  • [41] P. K. Gyawali, S. Chen, H. Liu, B. M. Horacek, J. L. Sapp, and L. Wang (2017) Automatic coordinate prediction of the exit of ventricular tachycardia from 12-lead electrocardiogram. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: §III-A2, TABLE I.
  • [42] P. K. Gyawali, B. M. Horacek, J. L. Sapp, and L. Wang (2019) Sequential factorized autoencoder for localizing the origin of ventricular activation from 12-lead electrocardiograms. IEEE Transactions on Biomedical Engineering. Cited by: §III-A2, TABLE I.
  • [43] A. Habib, C. Karmakar, and J. Yearwood (2019) Impact of ecg dataset diversity on generalization of cnn model for detecting qrs complex. IEEE access 7, pp. 93275–93285. Cited by: TABLE I.
  • [44] M. Hammad, Y. Liu, and K. Wang (2018) Multimodal biometric authentication systems using convolution neural network based on different level fusion of ecg and fingerprint. IEEE Access 7, pp. 26527–26542. Cited by: §III-A4, TABLE I.
  • [45] S. Han, H. Mao, and W. J. Dally (2016) Deep compression: compressing deep neural network with pruning, trained quantization and huffman coding. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, External Links: Link Cited by: 2nd item.
  • [46] A. Y. Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourn, M. P. Turakhia, and A. Y. Ng (2019) Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature medicine 25 (1), pp. 65. Cited by: §III-A1, §III-B1, TABLE I.
  • [47] J. He, K. Li, X. Liao, P. Zhang, and N. Jiang (2019) Real-time detection of acute cognitive stress using a convolutional neural network from electrocardiographic signal. IEEE Access 7, pp. 42710–42717. Cited by: §III-A1, TABLE I.
  • [48] R. He, Y. Liu, K. Wang, N. Zhao, Y. Yuan, Q. Li, and H. Zhang (2019) Automatic cardiac arrhythmia classification using combination of deep residual network and bidirectional lstm. IEEE Access 7, pp. 102119–102135. Cited by: §III-B3, TABLE I.
  • [49] W. He, G. Wang, J. Hu, C. Li, B. Guo, and F. Li (2019) Simultaneous human health monitoring and time-frequency sparse representation using eeg and ecg signals. IEEE Access 7, pp. 85985–85994. Cited by: TABLE I.
  • [50] G. E. Hinton, O. Vinyals, and J. Dean (2015) Distilling the knowledge in a neural network. CoRR abs/1503.02531. External Links: Link, 1503.02531 Cited by: 2nd item.
  • [51] S. Hong, M. Wu, Y. Zhou, Q. Wang, J. Shang, H. Li, and J. Xie (2017) ENCASE: an ensemble classifier for ecg classification using expert features and deep neural networks. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: §I, §III-B6, §III-B6, TABLE I, 3rd item.
  • [52] S. Hong, C. Xiao, T. Ma, H. Li, and J. Sun (2019) MINA: multilevel knowledge-guided attention for modeling electrocardiography signals. In Tweenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI), pp. 5888–5894. Cited by: §III-A1, §III-B3, TABLE I, 1st item, 3rd item.
  • [53] S. Hong, Y. Zhou, M. Wu, J. Shang, Q. Wang, H. Li, and J. Xie (2019) Combining deep neural networks and engineered features for cardiac arrhythmia detection from ecg recordings. Physiological measurement 40 (5), pp. 054009. Cited by: §I, TABLE I, 3rd item.
  • [54] Z. Hu, X. Ma, Z. Liu, E. H. Hovy, and E. P. Xing (2016) Harnessing deep neural networks with logic rules. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers, External Links: Link Cited by: 3rd item.
  • [55] J. Huang, B. Chen, B. Yao, and W. He (2019) ECG arrhythmia classification using stft-based spectrogram and convolutional neural network. IEEE Access 7, pp. 92871–92880. Cited by: §III-B1, §III-B6, TABLE I.
  • [56] F. Jager, A. Taddei, G. B. Moody, M. Emdin, G. Antolič, R. Dorn, A. Smrdel, C. Marchesi, and R. G. Mark (2003) Long-term st database: a reference for the development and evaluation of automated ischaemia detectors and for the study of the dynamics of myocardial ischaemia. Medical and Biological Engineering and Computing 41 (2), pp. 172–182. Cited by: §III-C.
  • [57] J. Jiang, H. Zhang, D. Pi, and C. Dai (2019) A novel multi-module neural network system for imbalanced heartbeats classification. Expert Systems with Applications: X 1, pp. 100003. Cited by: TABLE I.
  • [58] L. Johannesen, J. Vicente, J. W. Mason, C. Erato, C. Sanabria, K. Waite-Labott, M. Hong, J. Lin, P. Guo, A. Mutlib, et al. (2016) Late sodium current block for drug-induced long qt syndrome: results from a prospective clinical trial. Clinical Pharmacology & Therapeutics 99 (2), pp. 214–223. Cited by: §III-C.
  • [59] L. Johannesen, J. Vicente, J. Mason, C. Sanabria, K. Waite-Labott, M. Hong, P. Guo, J. Lin, J. S. Sørensen, L. Galeotti, et al. (2014) Differentiating drug-induced multichannel block on the electrocardiogram: randomized study of dofetilide, quinidine, ranolazine, and verapamil. Clinical Pharmacology & Therapeutics 96 (5), pp. 549–558. Cited by: §III-C.
  • [60] A. E. Johnson, T. J. Pollard, L. Shen, H. L. Li-wei, M. Feng, M. Ghassemi, B. Moody, P. Szolovits, L. A. Celi, and R. G. Mark (2016) MIMIC-iii, a freely accessible critical care database. Scientific data 3, pp. 160035. Cited by: §III-C.
  • [61] K. W. Johnson, J. T. Soto, B. S. Glicksberg, K. Shameer, R. Miotto, M. Ali, E. Ashley, and J. T. Dudley (2018) Artificial intelligence in cardiology. Journal of the American College of Cardiology 71 (23), pp. 2668–2679. Cited by: §I.
  • [62] R. Kamaleswaran, R. Mahajan, and O. Akbilgic (2018) A robust deep convolutional neural network for the classification of abnormal cardiac rhythm using single lead electrocardiograms of variable length. Physiological measurement 39 (3), pp. 035006. Cited by: §III-A1, TABLE I.
  • [63] T. Kimura, S. Takatsuki, H. Yamashita, H. Miyama, T. Fujisawa, K. Nakajima, Y. Katsumata, T. Nishiyama, Y. Aizawa, and K. Fukuda (2018) A 10-rr-interval-based rhythm classifier using a deep neural network. Circulation 138 (Suppl_1), pp. A12693–A12693. Cited by: §III-B7, TABLE I.
  • [64] S. Kiranyaz, T. Ince, and M. Gabbouj (2015) Real-time patient-specific ecg classification by 1-d convolutional neural networks. IEEE Transactions on Biomedical Engineering 63 (3), pp. 664–675. Cited by: TABLE I.
  • [65] S. Kiranyaz, T. Ince, and M. Gabbouj (2017) Personalized monitoring and advance warning system for cardiac arrhythmias. Scientific reports 7 (1), pp. 9270. Cited by: TABLE I.
  • [66] P. Laguna, R. G. Mark, A. Goldberg, and G. B. Moody (1997) A database for evaluation of algorithms for measurement of qt and other waveform intervals in the ecg. In Computers in cardiology 1997, pp. 673–676. Cited by: §III-B4, §III-C.
  • [67] Y. LeCun, Y. Bengio, and G. E. Hinton (2015) Deep learning. Nature 521 (7553), pp. 436–444. External Links: Link, Document Cited by: §I.
  • [68] J. Lee and K. Kwak (2019) Personal identification using a robust eigen ecg network based on time-frequency representations of ecg signals. IEEE Access 7, pp. 48392–48404. Cited by: §III-A4, §III-B7, TABLE I.
  • [69] J. Lee, K. Oh, B. Kim, and S. K. Yoo (2019) Synthesis of electrocardiogram v lead signals from limb lead measurement using r peak aligned generative adversarial network. IEEE journal of biomedical and health informatics. Cited by: §III-A6, TABLE I.
  • [70] J. Lee, S. Sun, S. Yang, J. Sohn, J. Park, S. Lee, and H. C. Kim (2018) Bidirectional recurrent auto-encoder for photoplethysmogram denoising. IEEE journal of biomedical and health informatics. Cited by: §III-A5, TABLE I.
  • [71] J. S. Lee, S. J. Lee, M. Choi, M. Seo, and S. W. Kim (2019) QRS detection method based on fully convolutional networks for capacitive electrocardiogram. Expert Systems with Applications 134, pp. 66–78. Cited by: §III-A2, §III-B1, TABLE I.
  • [72] F. Li, Y. Xu, Z. Chen, and Z. Liu (2019) Automated heartbeat classification using 3-d inputs based on convolutional neural network with multi-fields of view. IEEE Access 7, pp. 76295–76304. Cited by: §III-B1, TABLE I.
  • [73] K. Li, W. Pan, Y. Li, Q. Jiang, and G. Liu (2018) A method to detect sleep apnea based on deep neural network and hidden markov model using single-lead ecg signal. Neurocomputing 294, pp. 94–101. Cited by: §III-A3, TABLE I.
  • [74] Q. Li, C. Liu, Q. Li, S. P. Shashikumar, S. Nemati, Z. Shen, and G. D. Clifford (2019) Ventricular ectopic beat detection using a wavelet transform and a convolutional neural network. Physiological measurement 40 (5), pp. 055002. Cited by: TABLE I.
  • [75] R. Li, X. Zhang, H. Dai, B. Zhou, and Z. Wang (2019) Interpretability analysis of heartbeat classification based on heartbeat activity’s global sequence features and bilstm-attention neural network. IEEE Access 7, pp. 109870–109883. Cited by: §III-B2, TABLE I.
  • [76] Y. Li, Y. Zhang, L. Zhao, Y. Zhang, C. Liu, L. Zhang, L. Zhang, Z. Li, B. Wang, E. Ng, et al. (2018) Combining convolutional neural network and distance distribution matrix for identification of congestive heart failure. IEEE Access 6, pp. 39734–39744. Cited by: §III-B1, TABLE I.
  • [77] Y. Li, Y. Pang, J. Wang, and X. Li (2018) Patient-specific ecg classification by deeper cnn from generic to dedicated. Neurocomputing 314, pp. 336–346. Cited by: TABLE I.
  • [78] Z. Li, X. Feng, Z. Wu, C. Yang, B. Bai, and Q. Yang (2019) Classification of atrial fibrillation recurrence based on a convolution neural network with svm architecture. IEEE Access 7, pp. 77849–77856. Cited by: TABLE I.
  • [79] M. Limam and F. Precioso (2017) Atrial fibrillation detection and ecg classification based on convolutional recurrent neural network. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [80] K. Lin, D. Li, X. He, Z. Zhang, and M. Sun (2017) Adversarial ranking for language generation. In Advances in Neural Information Processing Systems, pp. 3155–3165. Cited by: §III-B5.
  • [81] T. Lin, P. Goyal, R. B. Girshick, K. He, and P. Dollár (2017) Focal loss for dense object detection. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, pp. 2999–3007. External Links: Link, Document Cited by: 1st item.
  • [82] G. Litjens, F. Ciompi, J. M. Wolterink, B. D. de Vos, T. Leiner, J. Teuwen, and I. Išgum (2019) State-of-the-art deep learning in cardiovascular image analysis. JACC: Cardiovascular Imaging 12 (8), pp. 1549–1565. Cited by: §I.
  • [83] F. Liu, X. Zhou, J. Cao, Z. Wang, H. Wang, and Y. Zhang (2019) A lstm and cnn based assemble neural network framework for arrhythmias classification. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1303–1307. Cited by: TABLE I.
  • [84] F. Liu, C. Liu, L. Zhao, X. Zhang, X. Wu, X. Xu, Y. Liu, C. Ma, S. Wei, Z. He, et al. (2018) An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detection. Journal of Medical Imaging and Health Informatics 8 (7), pp. 1368–1373. Cited by: §III-C.
  • [85] W. Liu, M. Zhang, Y. Zhang, Y. Liao, Q. Huang, S. Chang, H. Wang, and J. He (2017) Real-time multilead convolutional neural network for myocardial infarction detection. IEEE journal of biomedical and health informatics 22 (5), pp. 1434–1444. Cited by: §III-A1, §III-B1, TABLE I.
  • [86] T. S. Lugovaya (2005) Biometric human identification based on ecg. PhysioNet. Cited by: §III-C.
  • [87] J. Maidens and N. B. Slamon (2018) Artificial intelligence detects pediatric heart murmurs with cardiologist-level accuracy. Circulation 138 (Suppl_1), pp. A12591–A12591. Cited by: TABLE I.
  • [88] V. Maknickas and A. Maknickas (2017) Atrial fibrillation classification using qrs complex features and lstm. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [89] J. Malik, Y. Lo, and H. Wu (2018) Sleep-wake classification via quantifying heart rate variability by convolutional neural network. Physiological measurement 39 (8), pp. 085004. Cited by: §III-A3, TABLE I.
  • [90] S. M. Mathews, C. Kambhamettu, and K. E. Barner (2018) A novel application of deep learning for single-lead ecg classification. Computers in biology and medicine 99, pp. 53–62. Cited by: TABLE I.
  • [91] S. Mehta, F. Fernandez, C. Villagran, S. Niklitschek, A. Frauenfelder, F. Nola, M. R. Ceschim, C. Matheus, C. Chaves, S. Quintero, et al. (2019) APPLICATION of artificial intelligence to detect st elevation mi with a single lead ekg. Journal of the American College of Cardiology 73 (9 Supplement 1), pp. 1328. Cited by: TABLE I.
  • [92] A. C. Miller, Z. Obermeyer, and S. Mullainathan (2019) A comparison of patient history-and ekg-based cardiac risk scores. AMIA Summits on Translational Science Proceedings 2019, pp. 82. Cited by: §III-A6, TABLE I.
  • [93] G. B. Moody and R. G. Mark (2001) The impact of the mit-bih arrhythmia database. IEEE Engineering in Medicine and Biology Magazine 20 (3), pp. 45–50. Cited by: 1st item.
  • [94] G. B. Moody, W. Muldrow, and R. G. Mark (1984) A noise stress test for arrhythmia detectors. Computers in cardiology 11 (3), pp. 381–384. Cited by: §III-C.
  • [95] G. Moody (1983) A new method for detecting atrial fibrillation using rr intervals. Computers in Cardiology, pp. 227–230. Cited by: §III-C.
  • [96] S. Mousavi and F. Afghah (2019) Inter-and intra-patient ecg heartbeat classification for arrhythmia detection: a sequence to sequence deep learning approach. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1308–1312. Cited by: TABLE I.
  • [97] M. T. Nguyen, B. Van Nguyen, and K. Kim (2018) Deep feature learning for sudden cardiac arrest detection in automated external defibrillators. Scientific reports 8 (1), pp. 17196. Cited by: TABLE I.
  • [98] G. Nikolic, R. Bishop, and J. Singh (1982) Sudden death recorded during holter monitoring.. Circulation 66 (1), pp. 218–225. Cited by: §I.
  • [99] S. L. Oh, E. Y. Ng, R. San Tan, and U. R. Acharya (2018) Automated diagnosis of arrhythmia using combination of cnn and lstm techniques with variable length heart beats. Computers in biology and medicine 102, pp. 278–287. Cited by: TABLE I.
  • [100] S. L. Oh, E. Y. Ng, R. San Tan, and U. R. Acharya (2019) Automated beat-wise arrhythmia diagnosis using modified u-net on extended electrocardiographic recordings with heterogeneous arrhythmia types. Computers in biology and medicine 105, pp. 92–101. Cited by: §III-B7, TABLE I.
  • [101] Y. Park, I. D. Yun, and S. Kang (2019) Preprocessing method for performance enhancement in cnn-based stemi detection from 12-lead ecg. IEEE Access 7, pp. 99964–99977. Cited by: §III-A1, TABLE I.
  • [102] S. Parvaneh, J. Rubin, A. Rahman, B. Conroy, and S. Babaeizadeh (2018) Analyzing single-lead short ecg recordings using dense convolutional neural networks and feature-based post-processing to detect atrial fibrillation. Physiological measurement 39 (8), pp. 084003. Cited by: TABLE I.
  • [103] A. Peimankar and S. Puthusserypady (2019) An ensemble of deep recurrent neural networks for p-wave detection in electrocardiogram. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1284–1288. Cited by: §III-A2, TABLE I.
  • [104] A. Picon, U. Irusta, A. Álvarez-Gila, E. Aramendi, F. Alonso-Atienza, C. Figuera, U. Ayala, E. Garrote, L. Wik, J. Kramer-Johansen, et al. (2019) Mixed convolutional and long short-term memory network for the detection of lethal ventricular arrhythmia. PloS one 14 (5), pp. e0216756. Cited by: TABLE I.
  • [105] F. Plesinger, P. Nejedly, I. Viscor, J. Halamek, and P. Jurak (2018) Parallel use of a convolutional neural network and bagged tree ensemble for the classification of holter ecg. Physiological measurement 39 (9), pp. 094002. Cited by: TABLE I.
  • [106] B. Pourbabaee, M. J. Roshtkhari, and K. Khorasani (2017) Deep convolutional neural networks and learning ecg features for screening paroxysmal atrial fibrillation patients. IEEE Transactions on Systems, Man, and Cybernetics: Systems 48 (12), pp. 2095–2104. Cited by: §III-A1, TABLE I.
  • [107] D. Rajan, D. Beymer, and G. Narayan (2018) Generalization studies of neural network models for cardiac disease detection using limited channel ecg. In 2018 Computing in Cardiology Conference (CinC), Vol. 45, pp. 1–4. Cited by: TABLE I.
  • [108] M. N. Rastgoo, B. Nakisa, F. Maire, A. Rakotonirainy, and V. Chandran (2019) Automatic driver stress level classification using multimodal deep learning. Expert Systems with Applications 138, pp. 112793. Cited by: §III-A6, TABLE I.
  • [109] M. T. Ribeiro, S. Singh, and C. Guestrin (2016) ”Why should I trust you?”: explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, pp. 1135–1144. External Links: Link, Document Cited by: 1st item.
  • [110] M. T. Ribeiro, S. Singh, and C. Guestrin (2018) Anchors: high-precision model-agnostic explanations. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pp. 1527–1535. External Links: Link Cited by: 1st item.
  • [111] O. Ronneberger, P. Fischer, and T. Brox (2015) U-net: convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pp. 234–241. Cited by: §III-B7.
  • [112] J. Rubin, S. Parvaneh, A. Rahman, B. Conroy, and S. Babaeizadeh (2017) Densely connected convolutional networks and signal quality analysis to detect atrial fibrillation using short single-lead ecg recordings. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [113] L. Santamaria-Granados, M. Munoz-Organero, G. Ramirez-Gonzalez, E. Abdulhay, and N. Arunkumar (2018) Using deep convolutional neural network for emotion detection on a physiological signals dataset (amigos). IEEE Access 7, pp. 57–67. Cited by: §III-A6, TABLE I.
  • [114] J. Schläpfer and H. J. Wellens (2017) Computer-interpreted electrocardiograms: benefits and limitations. Journal of the American College of Cardiology 70 (9), pp. 1183–1192. Cited by: §I.
  • [115] M. Schram, J. B. Shreibati, M. Bos, C. Galloway, A. Valys, F. Petterson, V. Gundotra, D. Albert, Z. I. Attia, P. Noseworthy, et al. (2019) PREDICTION of the heart rate corrected qt interval (qtc) from a novel, multilead smartphone-enabled ecg using a deep neural network. Journal of the American College of Cardiology 73 (9 Supplement 1), pp. 368. Cited by: TABLE I.
  • [116] P. Schwab, G. C. Scebba, J. Zhang, M. Delai, and W. Karlen (2017) Beat by beat: classifying cardiac arrhythmias with recurrent neural networks. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [117] A. Sellami and H. Hwang (2019) A robust deep convolutional neural network with batch-weighted loss for heartbeat classification. Expert Systems with Applications 122, pp. 75–84. Cited by: TABLE I.
  • [118] A. P. Shah and S. A. Rubin (2007) Errors in the computerized electrocardiogram interpretation of cardiac rhythm. Journal of electrocardiology 40 (5), pp. 385–390. Cited by: §I.
  • [119] S. P. Shashikumar, A. J. Shah, G. D. Clifford, and S. Nemati (2018)

    Detection of paroxysmal atrial fibrillation using attention-based bidirectional recurrent neural networks

    .
    In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 715–723. Cited by: §III-B3, TABLE I.
  • [120] Y. Shen, M. Voisin, A. Aliamiri, A. Avati, A. Hannun, and A. Ng (2019) Ambulatory atrial fibrillation monitoring using wearable photoplethysmography with deep learning. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 1909–1916. Cited by: TABLE I.
  • [121] P. Sodmann, M. Vollmer, N. Nath, and L. Kaderali (2018) A convolutional neural network for ecg annotation as the basis for classification of cardiac rhythms. Physiological measurement 39 (10), pp. 104005. Cited by: TABLE I.
  • [122] N. Strodthoff and C. Strodthoff (2018) Detecting and interpreting myocardial infarction using fully convolutional neural networks. Physiological measurement. Cited by: TABLE I.
  • [123] W. Sun, N. Zeng, and Y. He (2019) Morphological arrhythmia automated diagnosis method using gray-level co-occurrence matrix enhanced convolutional neural network. IEEE Access. Cited by: §III-B1, TABLE I.
  • [124] A. Taddei, G. Distante, M. Emdin, P. Pisani, G. Moody, C. Zeelenberg, and C. Marchesi (1992) The european st-t database: standard for evaluating systems for the analysis of st-t changes in ambulatory electrocardiography. European heart journal 13 (9), pp. 1164–1172. Cited by: §III-C.
  • [125] J. H. Tan, Y. Hagiwara, W. Pang, I. Lim, S. L. Oh, M. Adam, R. San Tan, M. Chen, and U. R. Acharya (2018) Application of stacked convolutional and long short-term memory network for accurate identification of cad ecg signals. Computers in biology and medicine 94, pp. 19–26. Cited by: TABLE I.
  • [126] T. Teijeiro, C. A. García, D. Castro, and P. Félix (2018) Abductive reasoning as a basis to reproduce expert criteria in ecg atrial fibrillation identification. Physiological measurement 39 (8), pp. 084006. Cited by: §III-A1, TABLE I.
  • [127] G. H. Tison, A. C. Singh, D. A. Ohashi, J. T. Hsieh, B. M. Ballinger, J. E. Olgin, G. M. Marcus, and M. J. Pletcher (2017) Cardiovascular risk stratification using off-the-shelf wearables and a multi-task deep learning algorithm. Circulation 136 (suppl_1), pp. A21042–A21042. Cited by: §III-A1, TABLE I.
  • [128] E. Urtnasan, J. Park, and K. Lee (2018) Multiclass classification of obstructive sleep apnea/hypopnea based on a convolutional neural network from a single-lead electrocardiogram. Physiological measurement 39 (6), pp. 065003. Cited by: §III-A3, TABLE I.
  • [129] E. K. Wang, X. Zhang, and L. Pan (2019) Automatic classification of cad ecg signals with sdae and bidirectional long short-term term network. IEEE Access. Cited by: TABLE I.
  • [130] J. Wang, R. Li, R. Li, K. Li, H. Zeng, G. Xie, and L. Liu (2019) Adversarial de-noising of electrocardiogram. Neurocomputing 349, pp. 212–224. Cited by: §III-A5, TABLE I.
  • [131] L. Wang, W. Zhou, Q. Chang, J. Chen, and X. Zhou (2019) Deep ensemble detection of congestive heart failure using short-term rr intervals. IEEE Access. Cited by: §III-A1, TABLE I.
  • [132] P. Wang, B. Hou, S. Shao, and R. Yan (2019) ECG arrhythmias detection using auxiliary classifier generative adversarial network and residual network. IEEE Access 7, pp. 100910–100922. Cited by: §III-B1, §III-B1, §III-B5, TABLE I.
  • [133] S. Wang, C. C. Aggarwal, and H. Liu (2017)

    Using a random forest to inspire a neural network and improving on it

    .
    In Proceedings of the 2017 SIAM International Conference on Data Mining, Houston, Texas, USA, April 27-29, 2017., pp. 1–9. External Links: Link, Document Cited by: 1st item.
  • [134] Y. Wang and Q. Yao (2019) Few-shot learning: A survey. CoRR abs/1904.05046. External Links: Link, 1904.05046 Cited by: 1st item.
  • [135] P. A. Warrick and M. N. Homsi (2018) Ensembling convolutional and long short-term memory networks for electrocardiogram arrhythmia detection. Physiological measurement 39 (11), pp. 114002. Cited by: TABLE I.
  • [136] P. Warrick and M. N. Homsi (2017) Cardiac arrhythmia detection from ecg combining convolutional and long short-term memory networks. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [137] Y. Wei, X. Qi, H. Wang, Z. Liu, G. Wang, and X. Yan (2019) A multi-class automatic sleep staging method based on long short-term memory network using single-lead electrocardiogram signals. IEEE Access 7, pp. 85959–85970. Cited by: §III-A3, TABLE I.
  • [138] A. Wołk et al. (2019) Early and remote detection of possible heartbeat problems with convolutional neural networks and multipart interactive training. IEEE Access. Cited by: §III-B1, §III-B5, TABLE I.
  • [139] Y. Xia, N. Wulan, K. Wang, and H. Zhang (2017) Atrial fibrillation detection using stationary wavelet transform and deep learning. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: §III-B6, TABLE I.
  • [140] Y. Xia, N. Wulan, K. Wang, and H. Zhang (2018) Detecting atrial fibrillation by deep convolutional neural networks. Computers in biology and medicine 93, pp. 84–92. Cited by: TABLE I.
  • [141] Y. Xia and Y. Xie (2019) A novel wearable electrocardiogram classification system using convolutional neural networks and active learning. IEEE Access 7, pp. 7989–8001. Cited by: §III-B1, TABLE I.
  • [142] Y. Xia, H. Zhang, L. Xu, Z. Gao, H. Zhang, H. Liu, and S. Li (2018) An automatic cardiac arrhythmia classification system with wearable electrocardiogram. IEEE Access 6, pp. 16529–16538. Cited by: §III-B4, TABLE I.
  • [143] R. Xiao, Y. Xu, M. M. Pelter, D. W. Mortara, and X. Hu (2018) A deep learning approach to examine ischemic st changes in ambulatory ecg recordings. AMIA Summits on Translational Science Proceedings 2018, pp. 256. Cited by: §III-A1, TABLE I.
  • [144] P. Xiong, H. Wang, M. Liu, F. Lin, Z. Hou, and X. Liu (2016) A stacked contractive denoising auto-encoder for ecg signal denoising. Physiological measurement 37 (12), pp. 2214. Cited by: §III-A5, §III-B4, TABLE I.
  • [145] Z. Xiong, M. P. Nash, E. Cheng, V. V. Fedorov, M. K. Stiles, and J. Zhao (2018) ECG signal classification for the detection of cardiac arrhythmias using a convolutional recurrent neural network. Physiological measurement 39 (9), pp. 094006. Cited by: TABLE I.
  • [146] Z. Xiong, M. K. Stiles, and J. Zhao (2017) Robust ecg signal classification for detection of atrial fibrillation using a novel neural network. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.
  • [147] S. S. Xu, M. Mak, and C. Cheung (2018) Towards end-to-end ecg classification with raw signal extraction and deep neural networks. IEEE journal of biomedical and health informatics. Cited by: TABLE I.
  • [148] S. S. Xu, M. Mak, and C. Cheung (2019) I-vector based patient adaptation of deep neural networks for automatic heartbeat classification. IEEE journal of biomedical and health informatics. Cited by: TABLE I.
  • [149] Y. Xu, S. Biswal, S. R. Deshpande, K. O. Maher, and J. Sun (2018) Raim: recurrent attentive and intensive model of multimodal patient monitoring data. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 2565–2573. Cited by: §III-A1.
  • [150] T. Yang, L. Yu, Q. Jin, L. Wu, and B. He (2017) Localization of origins of premature ventricular contraction by means of convolutional neural network from 12-lead ecg. IEEE Transactions on Biomedical Engineering 65 (7), pp. 1662–1671. Cited by: TABLE I.
  • [151] Ö. Yildirim (2018) A novel wavelet sequence based on deep bidirectional lstm network model for ecg signal classification. Computers in biology and medicine 96, pp. 189–202. Cited by: TABLE I.
  • [152] W. Yin, X. Yang, L. Zhang, and E. Oki (2016) ECG monitoring system integrated with ir-uwb radar based on cnn. IEEE Access 4, pp. 6344–6351. Cited by: §III-A1, TABLE I.
  • [153] Ö. Yıldırım, P. Pławiak, R. Tan, and U. R. Acharya (2018) Arrhythmia detection using deep convolutional neural network with long duration ecg signals. Computers in biology and medicine 102, pp. 411–420. Cited by: TABLE I.
  • [154] X. Zhai and C. Tin (2018) Automated ecg classification using dual heartbeat coupling based on convolutional neural network. IEEE Access 6, pp. 27465–27472. Cited by: TABLE I.
  • [155] B. Zhang, J. Zhao, X. Chen, and J. Wu (2017) ECG data compression using a neural network model based on multi-objective optimization. PloS one 12 (10), pp. e0182500. Cited by: §III-A6, TABLE I.
  • [156] J. Zhang, F. Lin, P. Xiong, H. Du, H. Zhang, M. Liu, Z. Hou, and X. Liub (2019) Automated detection and localization of myocardial infarction with staked sparse autoencoder and treebagger. IEEE Access. Cited by: §III-B4, TABLE I.
  • [157] Q. Zhang, D. Zhou, and X. Zeng (2017) HeartID: a multiresolution convolutional neural network for ecg-based biometric human identification in smart health applications. Ieee Access 5, pp. 11805–11816. Cited by: §III-A4, §III-B1, TABLE I.
  • [158] Z. Zhao, Y. Zhang, Y. Deng, and X. Zhang (2018) ECG authentication system design incorporating a convolutional neural network and generalized s-transformation. Computers in biology and medicine 102, pp. 168–179. Cited by: §III-A4, TABLE I.
  • [159] Z. Zhao, C. Liu, Y. Li, Y. Li, J. Wang, B. Lin, and J. Li (2019) Noise rejection for wearable ecgs using modified frequency slice wavelet transform and convolutional neural networks. IEEE Access 7, pp. 34060–34067. Cited by: §III-A5, TABLE I.
  • [160] W. Zhong, L. Liao, X. Guo, and G. Wang (2018) A deep learning approach for fetal qrs complex detection. Physiological measurement 39 (4), pp. 045004. Cited by: §III-A2, TABLE I.
  • [161] Y. Zhou, S. Hong, J. Shang, M. Wu, Q. Wang, H. Li, and J. Xie (2019) K-margin-based residual-convolution-recurrent neural network for atrial fibrillation detection. IJCAI, pp. 6057–6063. Cited by: §III-A1, §III-B1, TABLE I.
  • [162] F. Zhu, F. Ye, Y. Fu, Q. Liu, and B. Shen (2019) Electrocardiogram generation with a bidirectional lstm-cnn generative adversarial network. Scientific reports 9 (1), pp. 6734. Cited by: §III-A6, §III-B5, TABLE I.
  • [163] M. Zihlmann, D. Perekrestenko, and M. Tschannen (2017) Convolutional recurrent neural networks for electrocardiogram classification. In 2017 Computing in Cardiology (CinC), pp. 1–4. Cited by: TABLE I.