YJELC-52897; No of Pages 5

Journal of Electrocardiology xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of Electrocardiology

journal homepage: www.jecgonline.com

Review

Cardiac arrhythmia detection using deep learning: A review

Saman Parvaneh a,⁎, Jonathan Rubin a, Saeed Babaeizadeh b, Minnan Xu-Wilson a
a
Philips Research North America, Cambridge, MA, USA
b
Advanced Algorithm Research Center, Philips Healthcare, Andover, MA, USA

a r t i c l e i n f o a b s t r a c t

Article history: Due to its simplicity and low cost, analyzing an electrocardiogram (ECG) is the most common technique for de-
Received 15 May 2019 tecting cardiac arrhythmia. The massive amount of ECG data collected every day, in home and hospital, may pre-
revised 18 July 2019 clude data review by human operators/technicians. Therefore, several methods are proposed for either fully
accepted 8 August 2019
automatic arrhythmia detection or event selection for further verification by human experts. Traditional machine
Available online xxxx
learning approaches have made significant progress in the past years. However, those methods rely on hand-
crafted feature extraction, which requires in-depth domain knowledge and preprocessing of the signal
(e.g., beat detection). This, plus the high variability in wave morphology among patients and the presence of
noise, make it challenging for computerized interpretation to achieve high accuracy. Recent advances in deep
learning make it possible to perform automatic high-level feature extraction and classification. Therefore, deep
learning approaches have gained interest in arrhythmia detection. In this work, we reviewed the recent advance-
ment of deep learning methods for automatic arrhythmia detection. We summarized existing literature from five
aspects: utilized dataset, application, type of input data, model architecture, and performance evaluation. We also
reported limitations of reviewed papers and potential future opportunities.
© 2019 Published by Elsevier Inc.

Introduction Traditional machine learning versus deep learning for arrhythmia

Cardiac arrhythmias affect more than four million people in US at a
total direct annual healthcare cost of up to $US67.4 billion, which is a
substantial economic burden [1]. Arrhythmia can lead to morbidity
and mortality. For instance, Atrial Fibrillation (AF), as the most common
arrhythmia, is associated with increased risk of stroke [2]. Analyzing an
ECG is the most common technique for detecting cardiac arrhythmia
due to its simplicity and low cost. The massive amount of ECG data col-
lected every day, in home and hospital, may preclude data review by
human operators/technicians [3]. Therefore, several methods are pro-
posed for either fully automatic arrhythmia detection or event selection
for further verification by human experts. Machine learning based
methods for ECG monitoring and interpretation range from traditional
machine learning to deep learning, and their combination [4]. In this
paper, we review the recent advancement of deep learning methods
for automatic arrhythmia detection.
⁎ Corresponding author at: Philips Research North America, 2 Canal Park, 3rd Floor,
Cambridge, MA, USA.
E-mail address: parvaneh@ieee.org (S. Parvaneh).
detection
In traditional machine learning (Fig. 1(a)), input data (e.g., ECG and
RR time series) is mapped to a set of features (e.g., lack of P wave, en-
tropy measures, Poincare section-based features, and heart rate vari-
ability) followed by a classifier (e.g., neural network, support vector
machine, and linear discriminant analysis). Features are often represen-
tations of cardiac arrhythmia and are identified by using knowledge
from experts. Traditional machine learning approaches have made sig-
nificant progress in the past years. They rely on hand-crafted feature ex-
traction, which requires in-depth domain knowledge and preprocessing
of the signal. Furthermore, the high variability in wave morphology
among patients and the presence of noise, make it challenging for com-
puterized interpretation to achieve high accuracy. The recent advance-
ment of deep learning makes it possible to perform automatic high-
level feature extraction and classification. Fig. 2(a) shows interest over
time in Google Trends for deep learning [5] where 100 is the peak pop-
ularity for deep learning and 0 means there was not enough data for the
search term “deep learning”. Furthermore, research trend for deep
learning is shown in Fig. 2(b) [6]. Both figures are created from January
2010 to December 2018. Fig. 2(a) shows the recent peak popularity for
deep learning and Fig. 2(b) confirms increased interest in using deep
learning in research.
In deep learning methods (Fig. 1(b)), raw or low-level processed
data is fed to a deep network architecture that is composed of multiple

https://doi.org/10.1016/j.jelectrocard.2019.08.004
0022-0736/© 2019 Published by Elsevier Inc.

Please cite this article as: S. Parvaneh, J. Rubin, S. Babaeizadeh, et al., Cardiac arrhythmia detection using deep learning: A review, Journal of
Electrocardiology, https://doi.org/10.1016/j.jelectrocard.2019.08.004
2 S. Parvaneh et al. / Journal of Electrocardiology xxx (xxxx) xxx

Fig. 1. (a) traditional machine learning and (b) deep neural network for classification.

hidden layers. Compared to traditional machine learning, feature ex-
traction/selection and classification is automatically performed within
the deep network. Deep learning based algorithms led to outstanding
performance on various benchmark datasets [7–9]. Especially,
convolutional neural networks (CNNs) have been used successfully for
solving complex image analysis tasks (both non-medical [10] and med-
ical [11] tasks). Of note, deep neural networks in recent years have been
developed or adapted for analysis of 1-D biosignals (time-series)
[7,8,12].
Well known deep learning architectures [13] include:
1- Convolutional Neural Network (CNN): CNNs use single or multiple
convolution and pooling layers followed by fully connected layer
for classification.
2- Recurrent Neural Network (RNN): RNNs perform the same task for
every element of an input sequence while output is dependent on
the previous computation. This network architecture has a memory
that captures information about data history, and what has been es-
timated so far.
3- Auto-encoder: An auto-encoder is a neural network that is trained to
attempt to copy its input to its output [13].

Fig. 2. (a) Interest over time for deep learning, and (b) research trend for deep learning between January 2010 and December 2018.

Please cite this article as: S. Parvaneh, J. Rubin, S. Babaeizadeh, et al., Cardiac arrhythmia detection using deep learning: A review, Journal of
Electrocardiology, https://doi.org/10.1016/j.jelectrocard.2019.08.004
 S. Parvaneh et al. / Journal of Electrocardiology xxx (xxxx) xxx
4- Deep Belief Network (DBN): DBNs are a composition of unsuper-
vised networks such as restricted Boltzmann machines (RBMs) or
auto-encoders where each sub-network's hidden layer serves as
the visible layer for the next [13].
CNNs and RNNs are supervised machine learning and auto-encoders
and DBNs are categorized as unsupervised machine leaning.
Deep learning for arrhythmia detection
Paper selection
In this section, papers using deep learning for arrhythmia detection
are reviewed. Searches were performed in the following databases:
PubMed, Embase, and Google Scholar. The latest literature searches
were completed in April 2019. The following general search term
were compiled for each database: (“deep learning” OR “deep neural net-
work” OR “convolutional neural network” OR “cnn” OR “recurrent neu-
ral network” OR “rnn” OR “long short term memory” OR “lstm” OR
“autoencoder” OR “deep belief network” OR “dbn”) AND (“ecg” OR
“electrocardiography” OR “electrocardiogram” OR “arrhythmia” OR
“cardiac arrhythmia”).
Only peer-reviewed papers published from January 12,011 to April
30 2019 were considered for this review and pre-prints were excluded.
After removing duplicate records and non-English papers, only deep
learning papers for arrhythmia detection were selected by screening
title, key words, and abstract. At the end 52 papers were selected for
this review.
Trend of deep learning research for arrhythmia detection
As shown in Fig. 3, there is an increasing trend in the deep learning
research for arrhythmia detection. This trend is potentially due to
a) popularity of deep learning and its successful application in other do-
mains (e.g., image classification and speech recognition), b) availability
of open source machine learning platforms such as TensorFlow,
c) availability of public datasets that allow training of deep learning
models, d) access to affordable processing infrastructure (e.g. Graphics
Processing Unit and cloud computing), and e) better adaptation of
deep learning techniques for time-series (e.g., biosignals).
The majority of deep learning based algorithms are developed for
image (2-D data) classification. A recently developed approach is to
convert 1-D data (e.g., ECG) to a 2-D representation
(e.g., spectrogram) and use a network structure without any modifica-
tion. An alternative is to adapt the network to better suit 1-D signals,
for example, adapting the kernel size in CNNs.
Fig. 3. Trend of deep learning research for arrhythmia detection.
Please cite this article as: S. Parvaneh, J. Rubin, S. Babaeizadeh, et al., Cardiac arrhythmia detection using deep learning: A review, Journal of
Electrocardiology, https://doi.org/10.1016/j.jelectrocard.2019.08.004
3
Utilized dataset and main goal
Only a limited number of research papers used private datasets and
the majority of selected papers used the following public datasets:
1- MIT-BIH Arrhythmia (MITDB) [14,15]: this database includes 48
half-hour excerpts of two-channel ambulatory ECG recordings
(sampling frequency: 360 samples per second) that was annotated
by two or more cardiologists.
2- Creighton University Ventricular Tachyarrhythmia (CUDB) [14,16]:
eight-minute ECG recordings of 35 patients who experienced epi-
sodes of sustained ventricular tachycardia, ventricular flutter, and
ventricular fibrillation are provided in this dataset (sampling fre-
quency: 250 samples per second).
3- MIT-BIH Atrial Fibrillation (AFDB) [14,17]: this database contains 25
long-term two-channel ECG recordings (10 h) of patients with atrial
fibrillation (sampling frequency: 250 samples per second).
4- PhysioNet/CinC Challenge 2017 [18,19]: single-channel short ECG
recording and associate human annotations for 8528 and 3658
human subjects in public training set and blind test set is available
for this dataset (sampling frequency: 300 samples per second).
The main goal of papers that used MITDB, CUDB, and AFDB was dis-
crimination between a) N, S, V, F, and Q classes of ECG heartbeats
(please see Table 1 for the definition of different classes) or b) normal
sinus rhythm (NSR), Atrial Fibrillation (AF), Atrial Flutter, and Ventricu-
lar Fibrillation. The main goal of papers that used PhysioNet/CinC Chal-
lenge 2017 was also discrimination between NSR, AF, other rhythm, and
noise.
Of note, research papers that had access to heartbeat annotations
were focused on automatic heartbeat detection (e.g., marking a heart-
beat as normal or premature) and other papers with access to
recording-based annotations were focused on arrhythmia detection in
a record (e.g., AF or normal recording in 10 s ECG recording).
Input data
Input data to the deep neural networks in the selected papers were:
1- ECG waveform: total duration of data, specific duration of data (e.g.
15 s), or specific number of beat(s) from digitized ECG were used
in majority of selected papers (96.1%). Only one paper used images
of ECG strips as input to deep neural network. Also, time-frequency
representation of ECG was used in some papers to convert 1D signal
to 2D representation for using deep neural networks that were de-
signed for image classification.
2- RR Intervals: papers in this group used RR time series extracted from
ECG.
Utilized deep network architecture for arrhythmia detection
As shown in Fig. 4, CNN is the most common technique used for
arrhythmia detection (23 papers). Furthermore, combination of
CNN with RNN or feature-based approach is used in 10 papers. In
papers that used deep unsupervised techniques (e.g., auto-
encoders and DBNs), features learned by deep unsupervised tech-
niques were used by a classifier (e.g., Softmax classification) for ar-
rhythmia detection.
Performance evaluation
Papers that used MITDB, CUDB, and AFDB used two approaches for
performance evaluation:
a) Splitting data into train and test sets: train data was used for training
a model and test data was used for model evaluation.
b) Using k-fold cross validation: In the kth run, k-1 folds are used for
training a model and one fold is used for model evaluation.
4 S. Parvaneh et al. / Journal of Electrocardiology xxx (xxxx) xxx
Table 1
A summary table of ECG beats categorized as per ANSI/AAMI EC57; 2012 standard [20].
N S
Normal Atrial premature
Left bundle branch block Aberrant atrial premature
Right bundle branch block Nodal (junctional) premature
Atrial escape Supra ventricular premature
Nodal (junctional) escape
Sensitivity, specificity, precision, recall, and F1-score were used as
performance metrics by papers in this group. Of note, the random crea-
tion of train/test sets or k-folds that was performed independently by
each team a) makes it harder to compare results between studies that
used the same dataset and had the same goal, and b) can increase risk
of overfitting to the training dataset.
Papers that used PhysioNet/CinC Challenge 2017, trained models on
the available training set and evaluated performance on the blind test
set and utilized the average F1 score as a performance metric. Perfor-
mance evaluation on the blind test set in papers of this group
a) enables comparison between different papers, and b) allows assess-
ment of potential overfitting to the training data set.
Limitations of reviewed studies and potential future opportunities
The applicability of the majority of reviewed studies for arrhythmia
classification is limited as:
1- In reviewed papers, except ones that used PhysioNet/CinC Challenge
2017 database, only clean data (data with minimal measurement
imprecision) was used for ECG classification and presence of noisy
data/segments in real ECG recording scenario was not considered.
2- Computational efficiency of algorithms (e.g., training time, evalua-
tion time for each record, and number of hyper-parameters in the
deep neural network) was not reported.
3- Classification algorithms were mainly trained and tested on a spe-
cific dataset captured with a specific device and generalizability of
developed algorithms was not evaluated.
4- Reviewed paper did not explore interpretability of trained models.
Therefore, potential opportunities for future studies are:
1- Benchmarking performance of different algorithms on the same
dataset (preferably a large dataset recorded from multiple sites
with different equipment and real noise/interference) to improve
comparison between different algorithms and evaluate generaliz-
ability of developed deep models.
2- Reporting computational efficiency/running time of algorithms and
potentially integrating them in scoring different algorithms in
Fig. 4. Utilized deep network architecture for arrhythmia detection.
Please cite this article as: S. Parvaneh, J. Rubin, S. Babaeizadeh, et al., Cardiac arrhythmia detection using deep learning: A review, Journal of
Electrocardiology, https://doi.org/10.1016/j.jelectrocard.2019.08.004
V F Q
Premature ventricular contraction Fusion of ventricular and normal Paced
Ventricular escape Fusion of paced and normal
Unclassified
addition to reported evaluation metrics. Reported computational ef-
ficiency/running time could be used to assess feasibility of integra-
tion developed model in resource limited devices (e.g., wearable
devices).
3- Improving interpretability of developed deep learning models.
Learnings from PhysioNet/CinC Challenge 2017
As highlighted in Performance evaluation section, using the same
training and test set in PhysioNet/CinC Challenge 2017 enables compar-
ison between different algorithms that are developed based on an avail-
able dataset. In this section, the main leanings from reviewed papers
based on this challenge are summarized.
Top scores in this challenge varied from traditional machine learning
to deep learning [18]. Therefore, there is no superior algorithm for clas-
sification of NSR/AF/O/noise. The performance of traditional machine
learning techniques used in this challenge was higher by ~3% compared
to top scores achieved by deep learning methods. The data size for the
challenge may not have been large enough to unlock the power of
deep learning in this challenge. However, this point needs to be vali-
dated in the future studies. Scores for deep learning methods range
from low to high score that indicates that deep learning methods
could get promising results if they are used properly. In this challenge,
performance of an ensembling method was about 5% more than the in-
dividual top score [18] that suggest that a collection of different algo-
rithms can be beneficial by combining complementary extracted
information [7].
Benefits and limitations of deep learning for arrhythmia detection
Using deep learning reduces pre-processing and feature engineering
efforts and may allow novel feature discovery by learning hidden pat-
terns that might not be clear using traditional machine learning
methods [4]. However, interpretation of deep learning models is chal-
lenging [4] and gaining benefit from deep learning models requires ac-
cess to large datasets. Therefore, future studies on improving
interpretability of developed deep learning models as well as identify-
ing the right size of training and test datasets for developing optimum
arrhythmia detection using deep learning are warranted.
Limitations of current review
This paper is a conceptual overview of deep learning-based algo-
rithms for arrhythmia detection and a systematic review is required to
highlight details of deep network architectures and mechanisms for
prevention of overfitting.
Conclusion
In this article, the main application, utilized dataset, deep network
architecture, and metrics used for performance evaluation of deep
learning-based algorithms for automated cardiac arrhythmia detection
were reviewed. Our review suggests that further efforts are required
to a) improve comparison between different algorithms (both perfor-
mance and computational efficiencies), b) evaluate generalizability of
proposed algorithms, and c) enhance interpretability of developed
deep learning-based models.
 S. Parvaneh et al. / Journal of Electrocardiology xxx (xxxx) xxx
References
[1] Tang DH, Gilligan AM, Romero K. Economic burden and disparities in healthcare re-
source use among adult patients with cardiac arrhythmia. Appl Health Econ Health
Policy 2014;12:59–71.
[2] Khoo CW, Krishnamoorthy S, Lim HS, Lip GY. Atrial fibrillation, arrhythmia burden
and thrombogenesis. Int J Cardiol 2012;157:318–23.
[3] Hannun AY, Rajpurkar P, Haghpanahi M, Tison GH, Bourn C, Turakhia MP, et al.
Cardiologist-level arrhythmia detection and classification in ambulatory electrocar-
diograms using a deep neural network. Nat Med 2019;25:65–9 2019/01/01.
[4] Parvaneh S, Rubin J. Electrocardiogram monitoring and interpretation: from tradi-
tional machine learning to deep learning, and their combination. Computing 2018;
45:1.
[5] Interest over time for deep learning. Available https://trends.google.com/trends/
explore?date=2012-01-01%202019-04-11&geo=US&q=%2Fm%2F0h1fn8h.
[6] Research trend for deep learning. Available https://app.dimensions.ai/analytics/
publication/viz/overview-publications?search_text=%22deep%20learning%
22&search_type=kws&search_field=text_search.
[7] Parvaneh S, Rubin J, Rahman A, Conroy B, Babaeizadeh S. Analyzing single-lead short
ECG recordings using dense convolutional neural networks and feature-based post-
processing to detect atrial fibrillation. Physiol Meas 2018;39:084003.
[8] Potes C, Parvaneh S, Rahman A, Conroy B. Classifier ensemble for detection of abnor-
mal heart sounds; 2016.
[9] Rubin J, Parvaneh S, Rahman A, Conroy B, Babaeizadeh S. "Densely connected
convolutional networks and signal quality analysis to detect atrial fibrillation
using short single-lead ECG recordings," presented at the Computing in Cardiology
Rennes-France; 2017.
Please cite this article as: S. Parvaneh, J. Rubin, S. Babaeizadeh, et al., Cardiac arrhythmia detection using deep learning: A review, Journal of
Electrocardiology, https://doi.org/10.1016/j.jelectrocard.2019.08.004
5
[10] Taigman Y, Yang M, Ranzato MA, Wolf L. Deepface: closing the gap to human-level
performance in face verification. Proceedings of the IEEE conference on computer vi-
sion and pattern recognition; 2014. p. 1701–8.
[11] Rubin J, Sanghavi D, Zhao C, Lee K, Qadir A, Xu-Wilson M. Large scale automated
reading of frontal and lateral chest x-rays using dual convolutional neural networks
arXiv preprint arXiv:804.07839 ; 2018.
[12] Rubin J, Parvaneh S, Rahman A, Conroy B, Babaeizadeh S. Densely connected
convolutional networks for detection of atrial fibrillation from short single-lead
ECG recordings. J Electrocardiol 2018;51(6):S18–21.
[13] Goodfellow I, Bengio Y, Courville A. Deep learning. MIT press; 2016.
[14] Goldberger A, Amaral L, Glass L, Hausdorff J, Ivanov P, Mark R, et al. PhysioBank,
PhysioToolkit, and PhysioNet: components of a new research resource for complex
physiologic signals. Circulation 2000;101:e215.
[15] Moody GB, Mark RG. The impact of the MIT-BIH arrhythmia database. IEEE Eng Med
Biol Mag 2001;20:45–50.
[16] Nolle F, Badura F, Catlett J, Bowser R, Sketch M. CREI-GARD, a new concept in com-
puterized arrhythmia monitoring systems. Computers in Cardiology 1986;13:515–8.
[17] Moody G. A new method for detecting atrial fibrillation using RR intervals. Com-
puters in Cardiology 1983:227–30.
[18] Clifford G, Liu C, Moody B, Silva I, Li Q, Johnson A, et al. AF classification from a short
single lead ECG recording: the PhysioNet computing in cardiology challenge 2017
presented at the Computing in Cardiology Rennes-France ; 2017.
[19] Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al.
Physiobank, physiotoolkit, and physionet. Circulation 2000;101:e215–20.
[20] Acharya UR, Oh SL, Hagiwara Y, Tan JH, Adam M, Gertych A, et al. A deep
convolutional neural network model to classify heartbeats. Comput Biol Med
2017;89:389–96.