20th International Conference on Information Fusion
Xi'an, China - July 10-13, 2017
Atrial Fibrillation Detection by Multi-scale
Convolutional Neural Networks
Zhenjie Yao∗ , Zhiyong Zhu∗ , Yixin Chen†∗
∗ Beijing advanced innovation center for future internet technology, Beijing University of Technology, Beijing, China
Email: yaozhenjie@gmail.com,zhiyong.zhu@bjut.edu.cn
† Department of Computer Science and Engineering, Washington University in St.Louis, USA
Email: chen@cse.wustl.edu
Abstract—Atrial Fibrillation (AF) is the most common chronic
arrhythmia. Effective detection of the AF would avoid serious
consequences like stroke. Conventional AF detection methods
need heuristic or hand-craft feature extraction. In this paper,
A deep neural network named multi-scale convolutional neural
networks (MCNN) based AF detector is proposed. Instant heart
rate sequence is extracted from ECG signal, then an end-to-
end MCNN detects AF with the instant heart rate sequence as
input and detection result as output. The algorithm was tested
on both public and private datasets. On the public dataset,
with the sensitivity achieved being 0.9822, the corresponding
specificity is 0.9811, and the overall accuracy is 0.9818. The
area under an ROC curve is as high as 0.9962, compared to
the AUC of the best conventional method is 0.9947. Comparison
shows that the MCNN based AF detector give superior accuracy
than conventional methods. Test on private dataset also shows
significant improvement.
I. I NTRODUCTION
Atrial Fibrillation (AF) is the most common chronic ar-
rhythmia. According to Framingham heart study, lifetime risk
of AF is about 25% [1]. When AF occurs, the frequency
of atrial activation is of 300 to 600 times per minute, heart
rate can reach 100 to 160 beats per minute. Furthermore,
the heart beat rhythm disorders. During AF, Atrium failed
to contract effectively leads poor cardiac ejection, and thus
easy to form thrombosis. The awful outcome of AF is stroke,
which can result in death, paralysis, limb movement or sensory
disturbances. The risk of stroke in patients with AF is 5 times
more than others [2][3].
Generally, AF is significant different from normal heart
rhythm on ECG signal [4]. First, P-wave is absent, and
there are continuous, irregular f-wave. At the same time, the
heart rhythm disorders, which result in irregularity of RR (R
peak to R peak) intervals. It is possible to detect most AF
automatically from ECG. Effective detection of AF leads to
timely intervention, which would avoid disease progression to
serious consequences.
In the last few decades, many researchers tried to detect
AF automatically. Lots of effective works has been done.
According to the methods adopted, AF detection can be
divided into the following categories. The first category is
RR interval based algorithm. R waves are identified from the
ECG signal, then RR interval is derived. We can detect AF
by identify the irregularity of heat pace. Moody and Mark
modeled the RR interval as Movkov process, likelihood of
978-0-9964-5270-0©2017 ISIF
observing pairs of RR intervals was calculated to identify AF
[5]. Logan et. al. detected AF according to the variance of RR
intervals [6]. Linker determined an instantaneous heart rate
for each of the heart beat intervals, and determined a non-
linear value that represents the variability of the instantaneous
heart rates [7]. Tatento and Glass detected AF by Kolmogorov
Smirnov test of RR intervals [8]. Cerutti et.al. found sig-
nificant difference between AF and normal sinus rhythm, in
parameters extracted through a time-variant identification of an
autoregressive model and non-linear measurements based on
the corrected conditional entropy [9]. Sakar et. al. proposed
a series of features derived from the Lorenz distribution of
RR intervals to detect AF [10]. Lake and Moorman took
advantage of sample entropy of RR intervals to detect AF [11].
Later, a series of works were based on sample entropy and its
variants. Colloca et. al. extracted features mentioned in [7],
[10] and [11] etc., and recognized AF by SVM [12]. Lee et. al.
introduced time-varying coherence functions, which describes
the similarity between adjacent segments, to distinguish AF
from normal sinus rhythm. Zhou et. al. calculated sample
entropy in symbolic dynamic space to identify AF, which
showed outstanding detection performance [13]. In 2015, they
improved the method and got the best detection performance
until now [14]. Petrenase et. al. used fusion of irregularity
of RR intervals and bigeminy suppression for AF detection,
which result in a very low complexity and high efficient
AF detector, only 8 beats is needed for detection [15]. The
second category of AF detector is waveform analysis based
method and hybrid method. ECG waveform analysis need to
determine the location of every parts of ECG, or the frequency
constituents. As for hybrid method, it means taking both RR
interval and waveform into consideration. Slocum detected
AF by waveform analysis, distribution of power and P-wave
absence were identified as the evidence of AF [16]. Schmidt
et. al. combine irregularity of RR intervals and absence of
P-wave for AF detection [17]. Babaeizadeh et. al. calculated
a P-R interval variability, a P wave morphology similarity
measure and a R-R Markov score for AF detection [18].
Couceiro detected AF based on the analysis of the three
main physiological characteristics of AF, P wave absence,
heart rate irregularity and atrial activity [19]. Larburu gave a
good survey on AF detection method before 2010. They also
compared these methods systematically, which show that R-R
interval based methods outperform waveform based method
and hybrid methods comprehensively [20]. This conclusion
is reasonable. R peak is most prominent feature, it is robust
to noise. However, waveform features such as f-wave and P-
wave is not as robust as R peak, they are likely to be impacted
by noise and other disturbances. So we tend to take the RR
interval based AF detection method.
Recently, deep learning has got great success in image
recognition and speech recognition, which made the system
performance improve significantly [21][22]. A deep neural
network is capable of extracting features from the original
input data, it is robust to variances like translation, rotation,
and noise. The MCNN (Multi-scale Convolutional Neural
Networks) is a deep neural network that designed for time
series classification. It showed impressive advantages com-
pared to other methods in time series classification [23].
Inspired by these works, we proposed an MCNN based RR
interval analysis method to detect AF from ECG automatically.
Experimental results show that the proposed method is capable
of detecting AF effectively.
The remainder of this paper is organized as follows. Section
II presents our MCNN based RR interval analysis algorithm
for AF detection in detail. Section III covers the experimental
results on both public and private datasets and its discussion.
Finally, in Section IV, the conclusion of this study is drawn
and further work is prospected.
II. MCNN FOR AF DETECTION
The detection method is straightforward. Instant heart rate is
derived from ECG signal by preprocessing and transformation.
Instant heart rate is feed into the MCNN segmentation by
segmentation for AF detection. The MCNN is an end-to-end
classifier with instant heart rate as input and detection result
as output.
A. Preprocessing
First, the R locations should be extracted from ECG signal.
You can skip this step if you can read R location from the
corresponding annotations. Most public ECG datasets have
corresponding annotations by clinicians. RR intervals is easily
derived from R locations. Then Instant Heart Rate (IHR) is
calculated by:
IHRi = 60 ∗ f /RRI i (1)
in which IHRi is the ith IHR, f is the sample rate of ECG
signal, RRI i is the ith RR interval. For example, if f is 250
Hz, the ith RRI is 200, then the IHR is 75 beat per minute.
After the preprocessing and transformation, the ECG signal
is transformed into an IHR sequence. Considering most AF
detection method need 128 beats, in this AF detection appli-
cation, we take 127 IHR samples into consideration. For each
IHR, The forward 63 IHRs, itself and the backward 63 IHRs
format an input IHR subsequence of the MCNN. The MCNN
is a binary classifier, whose output is whether it is AF or not.
B. MCNN
MCNN is an end-to-end deep neural network model, as
other deep neural network, it incorporates feature extraction
and classification in a single framework. The learnable con-
volutional layers automatically extracts features at different
position and in different scales, leading to superior feature
representation and better accuracy performance on time series
classification.
Figure 1 presents the architecture of the MCNN. The IHR
subsequence with length of 127 is the input time series of
MCNN. As shown in the figure, the MCNN framework has
three sequential stages: transformation stage, local convolution
stage, and full convolution stage. Details of each stage is given
below:
• Transformation stage. Time series has its intrinsic
feature in different scale. Long term (also known as
low frequency) reflect the overall trends and short term
(also known as high frequency) indicate local changes.
Generally, High frequency tends to include noises and
perturbations. To capture information in different scales
and avoid the influence of noises and perturbations, the
transformation stage applies various transformations on
the input time series. In our implementation, there are 3
types of transformation.
The first type is identity mapping, which keep the input
IHR sequence unchanged, so the length of the identity
mapping output is 127.
The second is smoothing the original input by low pass
filter. A low frequency filter can reduce the variance of
time series, which reduce the influence of high-frequency
perturbations and random noises. In particular, we employ
moving average to smooth the IHR sequence. Mathemat-
ically, it is
(IHRi−l + IHRi−l+1 + ... + IHRi+l )
IHRSM
i = ,
2∗l+1
(2)
where 2 ∗ l + 1 is window length of the low frequency
filter. We take l = 2 and l = 5, corresponding output is
of length 123 and 117, respectively. In fact, the filter is
not limited to low pass filter. All kinds of filter maybe
useful.
The third is down-sampling transformations, which pick
one from every k (down-sample factor) samples in the
time domain. Mathematically, it is
n−1
IHRDS = IHR1+k∗i , i = 0, 1, ...,  , (3)
k
in which IHRDS is the down-sampled sequence, k is
the down-sample factor, IHRi is the ith sample in IHR
sequence, n is the length of the original input IHR
sequence. In the AF detection application, we get only
one down-sample sequence with k = 3, the length of the
output is 43.
Each transformation result is called a branch, as it is a
branch input to the following local convolutional stage.
 It is worth noting that transformation is not limited to the
mentioned 3 types, other transformations may also work.
• Local convolution stage. As described above, after the
transformation stage, we obtain multiple time series with
different lengths from a single input sequence. In the local
convolution stage, we use several convolutional layers to
extract the features for each branch independently. All
the local convolution filters are with the same window
size 8. It is worth noting that with the same window size,
convolution on shorter signal means larger local receptive
field, while on longer signal means smaller local receptive
field. Same filter length on different branch means feature
extraction across different scales.
As conventional deep neural networks, all the outputs
will pass through a nonlinear down sampling, name max
pooling, which can reduce the size of the data, and make
the neural network robust to spatial shifting and rotation.
Instead of adopting a constant pooling factor, the pooling
size for each branch is different. In our implementation,
no matter how long is the branch, the length of its output
is 12. For example, if the length of a branch is 123, we
select pooling factor as 10, and the length of outcome
sequence is 12.
• Full convolution stage. After extracting feature maps
from multiple branches, all these features are concatenat-
ed into one sequence, and fed into a full convolutional
layers (each followed by max pooling), as well as a fully
connected layer followed by a softmax transformation.
The output of MCNN will be the predicted distribution
of each possible label for the input time series. we
could derive final result by simply taking the label with
maximal possibility.
As shown in Figure 1, the MCNN is an end-to-end system.
No heuristic or hand-craft feature extraction is needed. When
labeled training dataset is available, all the parameters could
be trained jointly through back propagation. In the application
of AF detection, the MCNN served as a binary classifier. AF
is positive example, and other heart rhythms (including normal
rhythm and other arrythmia) are negative examples.
III. E XPERIMENTAL R ESULTS
Three datasets are adopted to evaluate the MCNN AF
detector, including the Long-Term AF Database (LTAFDB),
the MIT-BIH AF Database (AFDB), and the Holter ECG
dataset (HEDB). LTAFDB, and AFDB are public datasets,
could be accessible from [24]. HEDB is a private dataset,
which is constitute of 499 Holter ECG signals, each is about 24
hours long. The summary of all the datasets are listed in Table
I. Informations include the sample frequency, total number of
beats, number of AF beats, number of records and average
duration of each records. For comparison, we follow the way
of dataset division in [14].the LTAFDB database is used as the
training set to determine the model parameters, while AFDB
and HEDB are used as the testing sets. Besides of this test,
we also evaluate the proposed algorithm on HEDB with half
data as training set and the other half as testing set.
The sensitivity, specificity and accuracy are calculated to
measure the performance of the proposed method for AF
detection. They are defined as:
sensitivity = T P/(T P + F N ), (4)
specif icity = T N/(T N + F P ), (5)
Accuracy = (T P + T N )/(T P + F N + T N + F P ), (6)
in which TP (True Positive) is the number of AF beats that
are recognized as AF; FN (False Negative) is the number of
AF beats that are recognized as not AF; TN (True Negative)
is number of not AF beats that are recognized as not AF
beats; FP (False Positive) is number of not AF beats that are
recognized as AF beats. In general, a good detection method
will maximize the all the 3 performance metrics.
Table II lists the tests results on AFDB together with
four existing methods that tested on the same dataset, with
LTAFDB as training dataset. To our knowledge, the four
methods give better results in the existing methods. Zhou
et. al. [14] give the best detection performance compared to
other existing works, the accuracy is 97.99%. The proposed
MCNN based AF detection method improved the detection
performance significantly, the accuracy is 98.18%, while the
sensitivity is 98.22% and the specificity is 98.11%. Despite the
accuracy only increased 0.19%, but we essentially reduced the
error of about 10%. So we conclude that, on AFDB dataset,
the MCNN based AF detection outperforms the best of the
existing methods.
Since symbolic dynamics and Shannon ENtropy (SEN) [14]
shows best detection performance compared to other existing
works, we realized the method for extensive comparison.
In statistics, a receiver operating characteristic (ROC) curve
is a graphical plot that illustrates the performance of a de-
tection system as its discrimination threshold is varied. The
curve is created by plotting the sensitivity (also known as
Detection Rate, DR) against the False Alarm Rate (FPR) at
various threshold settings. FAR is defined as:
F AR = 1 − specif icity. (7)
Figure 2 shows the ROC curves obtained by MCNN and
SEN on AFDB dataset, respectively. Most parts of the two
curves are similar. To enlarge the detail of the ROC curves,
only the top-left part is cropped. Both curves are interpolation
results of 21 actual test points. Each test point contains a
pair of DR and FAR values. The test points of MCNN are
obtained by modifying the threshold on the output of softmax
layer, each actual test point is corresponding to a threshold.
Conventional softmax layer is corresponding to threshold of
0.5. As for SEN, tuning the threshold of the final entropy leads
to the ROC curve. As shown in the figure, when the FAR is
smaller than 0.02, corresponding to the very left part of the
figure, the two curves are so similar that they almost coincide
with each other. In the other parts of the curves (FAR¿0.02),
MCNN outperform SEN consistently.
 Fig. 1. Overall architecture of MCNN [23].

TABLE I
S UMMARY OF DATASETS

Name Frequency(Hz) # of beats # of AF beats # of records Average duration per record

LTAFDB 128 8996056 5326145 84 24h
AFDB 250 1221574 519687 25 10h
HEDB 250 54074562 28170333 499 24h

TABLE II
C OMPARISON OF DIFFERENT METHODS ON AFDB

Year Sensitivity(%) Specificity(%) Accuracy(%)

Lake et. al. [11] 2011 91 94 N/A
Lee et. al. [25] 2013 98.22 97.68 97.91
Petrenase et. al. [15] 2015 97.1 98.3 N/A
Zhou et. al. [14] 2015 97.83 98.44 97.99
Proposed 2016 98.22 98.11 98.18

The AUC is a measurement of the detection accuracy, which
is defined as the ratio of the area under the ROC curve to
the area of the whole figure. The larger the AUC is, the
better performance the detector achieves. For the AF detection
problem, the AUC of MCNN is 0.9962, while the AUC of SEN
is 0.9947, which implies that the proposed MCNN based AF
detection method is more effective than SEN based method.
Table III lists the test results on HEDB. As HEDB is private
dataset, no test result exists in other literatures. As shown
in the table, while the sensitivity is 95.87%, the specificity
is 93.90%, the accuracy is 94.93%. The proposed MCNN
based AF detection method failed to improve the detection
performance, the accuracy is 94.32%, while the sensitivity is
95.57% and the specificity is 92.89%.
We infer that the failure on HEDB is caused by differ-
ence between the training dataset (LTAFDB) and test dataset
(HEDB). Both LTAFDB and AFDB are from Physionet, all the
ECGs are collected from Americans; but HEDB are collected
in China. Because ethnic characteristics, there are significant
difference between the datasets. We check the mean heart rate
of all the 3 dataset to valid the guess. The mean heart rate
of LTAFDB is 91.36 beats per minute. The mean heart rate
of AFDB is 88.07 beats per minute. The mean heart rate of
HEDB is 83.09 beats per minute. We see that compared to
 Fig. 2. ROC curves on AFDB.

TABLE III
C OMPARISON OF DIFFERENT METHODS ON HEDB

Year Sensitivity(%) Specificity(%) Accuracy(%)

Zhou et. al. [14] 2015 95.87 93.90 94.93
Proposed 2016 95.57 92.89 94.32

TABLE IV
C OMPARISON OF DIFFERENT METHODS ON HEDB

Year Sensitivity(%) Specificity(%) Accuracy(%)

Zhou et. al. [14] 2015 94.98 95.20 95.09
Proposed 2016 96.32 95.58 95.98

the difference between LTAFDB and AFDB, the difference
between LTAFDB and HEDB is much more significant. The
difference between training dataset and test dataset is the main
cause of the failure on HEDB.
Furthermore, we divide the HEDB into two datasets. It is
worth noting that, the separation is person-independent. The
ECG of each person will only appear in one subdataset. Cross
validation is taken to evaluate the MCNN based AF detector,
the two subdataset take turns to serve as training dataset. For
the motivation of justice, the parameters of SEN are also
optimized for HEDB. Results are shown in Table IV. For
the optimized SEN method, while the sensitivity is 94.98%
and the specificity is 95.20%, the accuracy is 95.09%. Only
a little improvement is achieved. The proposed MCNN based
AF detection method improve the detection performance, the
accuracy is 95.98%, while the sensitivity is 96.32% and the
specificity is 95.58%. We see that while both training and
testing dataset are from HEDB dataset, The MCNN based AF
detection outperforms SEN based method.
IV. C ONCLUSIONS AND F URTHER WORKS
This paper investigated AF detector based on MCNN. First,
instant heart rate sequence is extracted from ECG signal.
Then an end-to-end MCNN was trained to differentiate the
sequence is AF or not. The test results on both public and
private dataset show that the proposed method is effective.
On the public dataset, with the sensitivity achieved being
0.9822, the corresponding specificity is 0.9811, and the overall
accuracy is 0.9818. The area under an ROC curve is as high
as 0.9962, compared to AUC of 0.9947 for the best result
until now. As for the private dataset, the significant difference
between training dataset and test dataset is the main cause of
the failure on HEDB. If take half of the dataset as training
dataset and half as test dataset, the MCNN based AF detector
outperform traditional SEN based AF detector. Comparison
show that the MCNN based AF detector give higher accuracy
than conventional methods.
The testing results on HEDB show that the deep learning
method is more sensitive to the difference between training
and testing datasets. Such difference usually exists in real life
scenarios. Transfer learning is branch of machine learning that
specially developed for solving this problem. We infer that
transfer deep learning model could accomplish the work.
Although the performance of the proposed method is good,
the AF detection problem is far from being solved. There are
still quite a lot of problems worth exploring. According to the
literatures and our experience, bigeminy, trigeminy and sinus
arrhythmia are easy to be mistaken for AF, which would lead
to false alarm. It is hard to distinguish AF and part of these
arrhythmias. More attention should be paid to these problems.
Another concern is that compared to conventional AF
detection method, deep learning is computationally intensive.
It is difficult to calculate in real-time, especially for wearable
devices with limited power supply and limited computation
capacity. Generally, 128 heartbeats usually last 1 minute and
a half, most ECG last only 10 seconds. We would try to detect
AF in 10 seconds in our further work.
MCNN show power on AF detection, deeper model may
improve the result. Besides modifying the MCNN, another
important kind of deep learning model should also work on AF
detection. It is Recurrent Neural Networks (RNN), which is
designed for time sequence classification. Deep learning is the
most promising direction for AF detection, more investigations
are needed.
R EFERENCES
[1] D. M. Lloyd-Jones, T. J. Wang, E. P. Leip, M. G. Larson, D. Levy, R. S.
Vasan, R. B. DAgostino, J. M. Massaro, A. Beiser, P. A. Wolf et al.,
“Lifetime risk for development of atrial fibrillation the framingham heart
study,” Circulation, vol. 110, no. 9, pp. 1042–1046, 2004.
[2] T. J. Wang, M. G. Larson, D. Levy, R. S. Vasan, E. P. Leip, P. A. Wolf,
R. B. DAgostino, J. M. Murabito, W. B. Kannel, and E. J. Benjamin,
“Temporal relations of atrial fibrillation and congestive heart failure
and their joint influence on mortality the framingham heart study,”
Circulation, vol. 107, no. 23, pp. 2920–2925, 2003.
[3] L. Mainardi, L. Sornmo, and S. Cerutti, Understanding atrial fib-
rillation: the signal processing contribution. Morgan & Claypool
Publishers, 2008.
[4] S. Petrutiu, J. Ng, G. M. Nijm, H. Al-Angari, S. Swiryn, and A. V.
Sahakian, “Atrial fibrillation and waveform characterization,” IEEE
engineering in medicine and biology magazine, vol. 25, no. 6, pp. 24–30,
2006.
[5] G. B. Moody and R. G. Mark, “A new method for detecting atrial
fibrillation using rr intervals,” Computers in Cardiology, vol. 10, no. 1,
pp. 227–230, 1983.
[6] B. Logan and J. Healey, “Robust detection of atrial fibrillation for a long
term telemonitoring system,” in Computers in Cardiology, 2005. IEEE,
2005, pp. 619–622.
[7] D. T. Linker, “Long-term monitoring for detection of atrial fibrillation,”
Dec. 8 2009, uS Patent 7,630,756.
[8] K. Tateno and L. Glass, “Automatic detection of atrial fibrillation using
the coefficient of variation and density histograms of rr and δrr intervals,”
Medical and Biological Engineering and Computing, vol. 39, no. 6, pp.
664–671, 2001.
[9] S. Cerutti, L. Mainardi, A. Porta, and A. Bianchi, “Analysis of the
dynamics of rr interval series for the detection of atrial fibrillation
episodes,” in Computers in Cardiology 1997. IEEE, 1997, pp. 77–
80.
[10] S. Sarkar, D. Ritscher, and R. Mehra, “A detector for a chronic
implantable atrial tachyarrhythmia monitor,” IEEE Transactions on
Biomedical Engineering, vol. 55, no. 3, pp. 1219–1224, 2008.
[11] D. E. Lake and J. R. Moorman, “Accurate estimation of entropy in very
short physiological time series: the problem of atrial fibrillation detection
in implanted ventricular devices,” Am J Physiol Heart Circ Physiol, vol.
300, pp. H319–H325, 2011.
[12] R. Colloca, A. E. Johnson, L. Mainardi, and G. D. Clifford, “A support
vector machine approach for reliable detection of atrial fibrillation
events,” in Computing in Cardiology 2013. IEEE, 2013, pp. 1047–
1050.
[13] X. Zhou, H. Ding, B. Ung, E. Pickwell-MacPherson, and Y. Zhang,
“Automatic online detection of atrial fibrillation based on symbolic
dynamics and shannon entropy,” Biomedical engineering online, vol. 13,
no. 1, p. 1, 2014.
[14] X. Zhou, H. Ding, W. Wu, and Y. Zhang, “A real-time atrial fibrillation
detection algorithm based on the instantaneous state of heart rate,” PloS
one, vol. 10, no. 9, p. e0136544, 2015.
[15] A. Petrėnas, V. Marozas, and L. Sörnmo, “Low-complexity detection
of atrial fibrillation in continuous long-term monitoring,” Computers in
biology and medicine, vol. 65, pp. 184–191, 2015.
[16] J. Slocum, A. Sahakian, and S. Swiryn, “Diagnosis of atrial fibrilla-
tion from surface electrocardiograms based on computer-detected atrial
activity,” Journal of electrocardiology, vol. 25, no. 1, pp. 1–8, 1992.
[17] R. Schmidt, M. Harris, D. Novac, and M. Perkhun, “Atrial fibrillation
detection,” WO Patent, vol. 2, 2008.
[18] S. Babaeizadeh, R. E. Gregg, E. D. Helfenbein, J. M. Lindauer, and
S. H. Zhou, “Improvements in atrial fibrillation detection for real-time
monitoring,” Journal of electrocardiology, vol. 42, no. 6, pp. 522–526,
2009.
[19] R. Couceiro, P. Carvalho, J. Henriques, M. Antunes, M. Harris, and
J. Habetha, “Detection of atrial fibrillation using model-based ecg
analysis,” in Pattern Recognition, 2008. ICPR 2008. 19th International
Conference on. IEEE, 2008, pp. 1–5.
[20] N. Larburu, T. Lopetegi, and I. Romero, “Comparative study of algo-
rithms for atrial fibrillation detection,” in 2011 Computing in Cardiology.
IEEE, 2011, pp. 265–268.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural infor-
mation processing systems, 2012, pp. 1097–1105.
[22] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly,
A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath et al., “Deep neural
networks for acoustic modeling in speech recognition: The shared views
of four research groups,” IEEE Signal Processing Magazine, vol. 29,
no. 6, pp. 82–97, 2012.
[23] Z. Cui, W. Chen, and Y. Chen, “Multi-scale convolutional neural net-
works for time series classification,” arXiv preprint arXiv:1603.06995,
2016.
[24] “The physionet ecg database,” http://physionet.org/physiobank/
database/.
[25] J. Lee, Y. Nam, D. D. McManus, and K. H. Chon, “Time-varying
coherence function for atrial fibrillation detection,” IEEE Transactions
on Biomedical Engineering, vol. 60, no. 10, pp. 2783–2793, 2013.