Content uploaded by Shervin Shirmohammadi
Author content
All content in this area was uploaded by Shervin Shirmohammadi on Aug 21, 2018
Content may be subject to copyright.
Classifying Measured Electrocardiogram Signal
Quality Using Deep Belief Networks
Bahareh Taji1, Adrian D. C. Chan2, Shervin Shirmohammadi1
1 School of Electrical Engineering and Computer Science (EECS), University of Ottawa, Canada
2 Department of Systems and Computer Engineering, Carleton University, Canada
btaji016@uottawa.ca, adcchan@sce.carleton.ca, shervin@eecs.uottawa.ca
Abstract—There is a current trend towards wearable
electrocardiogram (ECG) measurement systems, which enables
measurement while the subject performs their normal activities
of daily living (e.g., walking, driving, eating). This type of
measurement is susceptible to higher levels of contaminants,
compared to bedside measurements, due to subject movement
and a measurement environment that is not well-controlled.
Contaminants in the measured signal (e.g., motion artifact,
power line interference) can cause incorrect interpretations,
including misdiagnoses. Therefore, prior to ECG interpretations,
it is important to have an algorithm capable of automatically
classifying the measured ECG based on their signal quality. ECG
of low signal quality can undergo additional pre-processing to
mitigate the contaminants or the signal can be discarded. This
can reduce misdiagnoses, including false-alarms which is a top
medical technology hazard. In this paper, we propose an
algorithm based on Deep Belief Networks (DBN) which can
differentiate between noisy and clean signal measurements. Our
algorithm is designed based on a three layer Restricted
Boltzmann Machine (RBM) in which the first two RBMs are
trained to extract the features and apply them to the third layer
of RBM to classify the data. Results, using the MIT-BIH
Arrhythmia database, demonstrate that our algorithm can
successfully recognize a noisy ECG signal from a clean signal,
with a classification accuracy between 75% and 99.5%
depending on the level of contaminants. Our algorithm also
correctly identifies clean arrhythmic signals and does not
misidentify them as noisy. The proposed algorithm is applicable
to any ECG measurement systems including wearable and
bedside.
Keywords— ECG Signal; noise; classification; Deep Belief
Network; SNR
I. INTRODUCTION
Cardiovascular disease is the global leading cause of death
and, in the United States, approximately 86 million people are
living with some sort of cardiac disease [1]. The prevalence of
cardiovascular disease and complications leads to a high cost
burden on the public health sector, directly and indirectly; it
also motivates research within this area. The
electrocardiogram (ECG), which is the recording of the
electrical activity of the heart, is an important tool for
investigating the cardiovascular system. Advances in
technology have enabled new monitoring systems, including
wearables which enable continuous monitoring during normal
activities of daily living [2]. Significant strides have also been
achieved in computer-aided diagnosis (CAD) for cardiac
arrhythmias; however, there still exists some fundamental
challenges. One of the most problematic issues are
contaminants in the measured ECG signal.
The utility of the ECG signal depends on signal quality of
the measurement. ECG contaminated by various sources of
noise and artifacts (e.g., baseline wandering, power line
interference, and motion artifact) can lead to misdiagnosis of
various cardiac arrhythmias in the form of false positives and
false negatives [3] [4]. While mitigation strategies can be used
to remove or reduce contaminants in the measured signal,
applying such strategies on a clean signal can actually reduce
the signal quality [5].
Heart arrhythmia diagnosis from ECG is currently being
done clinically using some bedside monitoring or wearable
devices, with an alarm activated in case of any arrhythmia
occurrence. Wearable ECG measurement systems enable
continuous monitoring, including during a subject’s normal
activities of daily living, including walking, driving, eating,
sitting down, and standing up. The measured signal in
wearable systems is especially prone to contamination by
motion artifact. It is possible that an alarm is falsely activated
due to noise in the measured ECG noisy signal. The
Emergency Care Research Institute (ECRI) listed alarm
hazards, including false alarms, as the top patient safety
hazard from 2012 to 2015 [6] [7] [8] [9]. False alarm reduces
the perceived validity and importance of alarms, which can
lead to alarm fatigue resulting in alarm desensitization and
missed alarms.
To address the above problems, signal quality verification
should be performed prior to CAD. If ECG signal of poor
quality is detected, the monitoring system must be able to act
appropriately, which could include applying mitigation
strategies or discarding the poor quality signal.
Simultaneously, it important that the system does not
mistakenly consider arrhythmic signals as noisy signals, as it
could result in a false negative (e.g., ECG signals being
discarded and a missed true alarm). As ECG contaminants can
present themselves in various forms and mixtures, biomedical
signal quality analysis can be challenging.
Machine learning algorithms can provide one approach to
confront this challenge. In this paper, we design, implement,
train and test a machine learning algorithm, specifically a deep
belief networks (DBN) algorithm, which can discriminate
between clean and noisy ECG measurements. Our algorithm,
the block diagram of which is demonstrated in Fig. 1, works
as follows. First, data is applied to the DBN. Then, two layers
of Restricted Boltzmann Machine (RBM) are dedicated to
feature extraction, the output of which is then applied to
another layer of RBM which is specifically designed to
classify the data.
Generative
RBMs
(RBM 1&2)
Test
Data
Extracted
features
Classifier
(Discriminative
RBM (RBM3))
Training Da ta
with labels Labels
DBN
Fig. 1 Block diagram of proposed method
The rest of the paper is organized as follows. Section II
covers the related work and the background to DBNs, while
section III, introduces our DBN energy-based model as well as
the datasets we have used in our study. Section IV discusses
the results and finally the paper concludes in section V while
presenting our future activities in this topic.
II. BACKGROUND AND RELATED WORK
A. Machine learning in bio-signal classification
There are many machine learning algorithms used for a
wide range of purposes, from image recognition to financial
applications. In the context of bio-signals, machine learning
algorithms are used for applications, such as medical
diagnoses [10] [11] [12], prosthetic control [13] and human
identification [14]. Machine learning algorithms have also
been employed in biomedical signal quality analysis [3] [15]
[16] [17] [18]. In [3], [15], [16] and [18] support vector
machine (SVM) is the machine learning algorithm performing
the task of classifying bio-signals based on their quality. In
[17], SVM and multi-layer perceptron (MLP) artificial neural
network are compared in terms of their performance in
classifying ECG signals based on their quality.
Among various machine learning algorithms, deep learning
is set of algorithms that has recently been garnering attention
because of its ability to perform a massive amount of
computational tasks with excellent accuracy. Deep belief
networks (DBN) are deep learning algorithms which are
effective, fast, and can perform feature extraction by
themselves. Hence, by applying DBN, the time consuming
and expensive step of the classification process, feature
extraction, will be automatically done. Some researchers are
creating algorithms and conducting tests in order to apply
DBNs in the context of bio-signal classification, e.g. in [19]
DBN is applied to extract features of six ECG signal types and
these features are fed into SVM which is in charge of ECG
classification based on their type. However, applying DBN as
a classifier for signal quality assessment, as we have done in
this work, is a fresh area of research.
B. Deep Belief Networks
Deep architectures are multiple processing layers
composed of multiple linear as well as non-linear
transformations [20]. Deep architectures have been shown to
outperform other classifiers in image and audio data
processing [21] [22] [23]. There is much evidence that show
deep learning models have greater or similar performance as
state-of-the-art methods in some areas, especially when it
comes to big and complex data [24]. Deep architectures are
arguably suitable for the complicated ECG data discussed in
this paper.
Among all deep architectures, we choose DBN because of
its strength in feature extraction and classification. DBNs can
be used for feature extraction in an unsupervised method.
Moreover, they are an appropriate tool for classification and in
our case, they will be used as supervised methods, taking the
labels of the training set.
DBNs were first proposed by Hinton [25] and are deep
architectures composed of a stack of Restricted Boltzmann
Machines (RBM), which utilize training data to create an
influential generative model. RBMs are essentially neural
networks having one hidden and one visible layer; they are a
special type of Boltzmann Machine with no connection
between hidden units and no connection between visible units.
Fig. 2 shows how RBM and general Boltzmann Machines are
different from each other. In DBNs, the hidden layer of the nth
RBM is the visible layer of the (n+1)th RBM. Obviously, this
is not the case for the RBM on the top of the stack. A regular
stack of RBMs and a stack of RBMs forming a DBN are
illustrated in Fig. 3. According to this, DBNs are in fact multi-
layer neural networks having several layers of RBMs [19]. If a
regular neural network has many hidden layers, it could easily
be trapped in local minima; moreover, the training process
takes a long time. These facts lead to a non-efficient
performance of artificial neural network. To overcome such
difficulties, DBN is an appropriate alternative. To train DBNs,
a layer-wise greedy algorithm is used. RBMs are trained layer
by layer, and then a fine-tuning process is performed to adjust
all parameters of DBN [25].
Fig. 2 General Boltzmann Machine (left) vs. Restricted
Boltzmann Machine (right)
Fig. 3 General Stack of RBMs (left) vs. its corresponding DBN (right)
III. PROPOSED METHOD
A. Architecture of the DBN
DBNs are a probabilistic generation model composed of
stacked RBMs. Each RBM is composed of a set of visible
units and a set of hidden units where
and h are the number of visible and hidden units
respectively. Visible and hidden units are interconnected with
symmetric weights.
A general Boltzmann Machine is a network with stochastic
binary units and its energy function of joint distribution is
given as [26]:
(1)
where h is the set of hidden units and v is the set of visible
units of the Boltzmann Machine. W is the symmetric weight
between hidden and visible units, L is the weight between
visible and visible units while J is the weight between hidden
and hidden units. In the above function, bias is not considered
for simplicity
In RBMs, visible units are disjoint (L = 0) and hidden units
are also disjoint (J = 0). Therefore; the following equation
represents the energy of joint distribution between hidden and
visible units in an RBM with respect to bias [26]:
(2)
where and ℎ are the number of visible and hidden units
respectively, is the symmetric interaction term between
the ith visible unit and the jth hidden unit, and and are
bias scalars of visible and hidden units, respectively.
RBM assigns a probability value with an energy function to
each state in the visible and hidden units. The probability of
joint distribution for visible and hidden units can be
determined as follows [26]:
(3)
where Z is computed from the following equation:
(4)
The network assigns a marginal probability (P) to a visible
unit:
(5)
By regulating the weights and biases, the network
maximizes the probability assigned to the training data and
lowers the probability of other data. By maximizing the
following objective function, we can obtain the optimal set of
weights that can do so. This function is indeed “Average log-
probability” of the data [26]:
(6)
where m represents the number of training data samples. The
ultimate goal in this step is to increase the model probability
for the set of training data. Thus, partial derivative of the
previous function with respect to a weight is required which is
as follows [26]:
(7)
where Xil is the ith unit of lth data instance. The left side of
the above equation can be calculated to its exact value,
whereas the right hand side showing the expectation under the
model distribution is the refractor. That leads to create other
methods to estimate this partial derivative. The derivative of
the log probability of a training vector with respect to a weight
can be computed as follows [26]:
(8)
where angle brackets denote the expectations under the
distribution specified by the subscript that follows [19].
According to this, the learning rule for executing stochastic
steepest ascent in the log probability of the training data is:
(9)
where ϵ is the learning rate.
As we mentioned before, in RBMs, hidden units are disjoint
and therefore independent. hj (binary state of hidden unit j),
given a randomly selected training data v, is set to 1 with
probability
(10)
where σ(x) is the logistic sigmoid function 1/(1 + exp(-x)) and
bj is the bias scalar for the hidden unit. According to this,
<vihj>data can easily be calculated.
Similarly, because visible units in RBM are disjoint, vi
(binary state of each visible unit i), given a hidden vector, is
set to 1 where its probability is:
(11)
where σ(x) is the logistic sigmoid function 1/(1 + exp(-x)) and
ai is the bias scalar for the visible unit.
Computing < > can be performed by using
sequential Gibbs sampling in which each iteration includes
updating all hidden units and also using (10) at the same time.
This process is followed by updating all visible units using
(11) [19]. This computation is time consuming. A faster
algorithm, Contrastive Divergence (CD), was proposed by
Hinton [27]. In this faster method utilizing the training data,
we initialize the visible units. Then using (10), the binary
values of the hidden units are computed followed by re-
computing the vi values using (11). Now the probability of the
hidden units can be calculated and eventually using these
values and also the values of the visible units, < > is
obtained. Now, DBN is constructed by combining several
RBMs from bottom to top. Each RBMs’ hidden layer is the
next RBMs’ visible layer in the DBNs. A greedy layer-wise
training algorithm is applied to train the DBN [25].
B. Dataset
ECG measurements used in our study are from the MIT-
BIH Arrhythmia Database [28], which includes different types
of arrhythmias. Each measurement recording is approximately
30 minutes in length and has a sampling frequency of 360 Hz.
There are 20 measurement recordings that are contaminant-
free in their channel one.
We contaminate 8 other recordings from the same
database, which are noisy in very limited periods of time (<=7
seconds) during the entire recording, with a calibrated amount
of motion artifact from the MIT-BIH Noise Stress Test
Database (NSTDB) [28]. Motion artifact noise is scaled
appropriately to obtain five desired signal to noise ratios
(SNR): -10, -5, 0, 5 and 10 db. Noise injection is performed
following the guidelines of MIT-BIH NSTDB [28].
To obtain a contaminated ECG signal of a certain SNR, we
should utilize (12) [29]:
(12)
where a is the scaling factor of noise, S is the power of signal
and N is the power of noise signal. This SNR definition is
different from a traditional definition of SNR and desired for
ECG [29]. ECG signal power (S) is computed by measuring
the peak-to-peak amplitude of 300 normal QRS complexes
from channel 1 of each recording. The top and bottom 5% of
the measurements are discarded and the average QRS
amplitude of the recording is determined using the remaining
90%. The signal power for the channel of the recording is
estimated by dividing the average QRS peak-to peak
amplitude by 8.
The noise signal we apply to contaminate the ECG signals
is the “em” measurement, available in the MIT-BIH Noise
Stress Test Database (NSTDB) [28] containing electrode
motion artifacts. ‘em’ is approximately half an hour in
duration and has a sampling frequency of 360 Hz. Noise
recording is done by using standard ECG recorders, leads and
electrodes on volunteers who were physically active. The
electrodes were connected to their limbs in a position where
their ECG was not visible. As a result, the recorded noise
signal is actual noise [28]. To calculate the power of the noise
(N), 300 first seconds of ‘em’ are broken down to one second
windows and the mean amplitude and the root mean squared
difference between the signal and this mean amplitude are
calculated. The largest and the smallest 5% of each
measurement are discarded and the RMS noise amplitude is
estimated by the mean of the remaining 90% of the
measurements. This estimate is squared to find the noise
power.
The contaminated signal is constructed by estimating S and
N, then computing the scaling factor a from (12), and finally
using (13):
(13)
where is the desired noisy signal, is the
original clean signal and is the noise signal.
Applying the above estimations and calculations, we obtain
the noisy signals at five levels of SNR. The original signals we
chose to contaminate may contain some limited period of
being noisy which is not taken into account by the calibration.
Each 30 minute recording is divided into 5 second
segments. Therefore, we have 7200 clean ECG segments and
2880 segments, which are contaminated with calibrated noise.
A total of six different datasets were constructed. In the first
five datasets, each dataset contains noisy signals from only
one level of SNR. In the last dataset, the dataset contains noisy
signals randomly taken from contaminated signals having
SNR = 0, 5 and 10 db for training, and for test set, noisy data
is randomly taken from contaminated signals of all levels of
SNRs including -10, -5, 0, 5 and 10db. We perform the
classification for all the datasets separately and compare the
results.
C. Training and testing sets
Clean data segments are labeled with a 0 and noisy data
segments are labeled with a 1. One part (75%) of our datasets
are labeled and used to train the RBMs. Performance of the
DBN is evaluated on the remaining part (25%) of the dataset
that is separate from the training data; i.e., there is no overlap
between recordings used for training and those applied for
testing. A 4-fold cross validation is performed for all tests.
D. Applying the DBN as a classifier
In this study, we designed a Deep Belief Network with
three RBMs. The first two RBMs are generative RBMs which
do not need labels, and the last RBM is a discriminative one
which uses data with their labels and can classify data. Each
RBM has 1000 hidden units and 1000 visible units. Benefiting
from the Contrastive Divergence (CD) method [27], RBMs
are trained through a 200 epochs attempt. For fine tuning, the
Back-Propagation (BP) method [30] is employed.
Accuracy, Precision (also known as Positive Predictive
Value), Recall (also known as Sensitivity) and Specificity
(also known as True Negative Rate) are defined as follows:
(14)
(15)
(16)
(17)
We use these four parameters to evaluate the performance
of the algorithm. For our study, a true positive (TP) is defined
as a signal which is noisy and the algorithm correctly
recognizes it as a noisy one, a false positive (FP) is defined as
a signal which is clean but the system takes it as a noisy
signal, and a false negative (FN) is defined as a signal which
is noisy but the system recognizes it as a clean signal. We also
report results for noise-free segments corresponding to an
arrhythmic event, separately. There are 2,220 noise-free
arrhythmic segments. Ideally, these segments should be
identified as clean; however the arrhythmia could be
mistakenly classified as noise.
IV. RESULTS
The results of all tests are presented in Table 1. As it shows,
Precision and Recall rates of classifying the ECG
measurements fluctuate with the SNR. Distinguishing noisy
and clean signals seems to be more difficult for high SNR; this
would be expected because noisy signals have little noise and
would closely resemble clean signals. All three performance
measures are in the high 90s for SNR that are 0 dB or lower.
Table 1 Accuracy, Precision, Recall and Specificity
SNR
-10dB
-5dB
0dB
5dB
10dB
Random
Accuracy (%)
99.5
96.7
97.8
96.8
75.0
97.2
Precision (%)
97.7
98.9
92.2
75.43
53.4
98.4
Recall (%)
100.0
99.9
99.4
96.4
73.2
98.2
Specificity (%)
100.0
99.9
99.4
96.4
73.1
98.2
Results confirm that our DBN is able to identify noisy
signals. There is low probability that the system identifies a
noisy ECG segment as a clean one. Such an error in
classifying ECG measurements can potentially be hazardous
because it is not a valid signal to be interpreted and can cause
a critical misperception of ones’ heart performance. Fig. 4
demonstrates a five-second ECG segment with calibrated
amount of motion artifact noise. For high signal to noise ratios
(SNR>=5), the level of noise is quite low and visually these
segments would be relatively difficult to identify as noisy
segments. In such situations, although the DBN may falsely
classify the signal as “clean”, the likelihood of a false alarm is
low given that the noise is quite modest.
It is also important that the algorithm not identify
arrhythmic signals as noisy signals, which could result in a
false negative. Therefore, we also investigated if our algorithm
can recognize the arrhythmic signals, in the clean part of our
test set, as clean ones. Table 2 shows the number of
arrhythmic ECG segments and the type of arrhythmia in the
test set and the number of these segments that were classified
as clean. Results suggest that the DBN classifier does not get
confused and can identify clean segments regardless of being
arrhythmic or not. Results are better with lower SNR because
noisy signals are easier to discern from clean signals given the
higher level of noise.
Fig. 4 Five-second ECG segment with calibrated amounts of motion artifact
noise
Table 2 Number of arrhythmias available in each classified as clean set vs.
number of them in clean test set
Arrhythmia
type
Left
bundle
branch
block
Aberrated
atrial
premature
Premature
ventricular
contraction
Fusion of
ventricular
and normal
Clean test signals
1254
109
845
12
signals
Classified as clean
(SNR = -10dB)
1253
109
843
12
signals
Classified as clean
(SNR = -5dB)
1254
109
843
10
signals
Classified as clean
(SNR = 0dB)
1997
109
834
11
signals
Classified as clean
(SNR = 5 dB)
1000
105
780
11
signals
Classified as clean
(SNR = 10dB)
976
105
682
10
signals
Classified as clean
(Random SNR)
1189
109
793
12
In total, results show much promise in using our DBN as
an effective classifier of ECG measurements into noisy and
clean signals. Such a classifier can be used in a preprocessing
step that can help lower the rate of false alarms, and at the
same time being intelligent enough not to gate true alarms.
So far in existing work, DBNs has mostly been used for
two dimensional data such as hand written digits and rarely
used for one dimensional data. We have shown that our DBN
has encouraging results in classifying one dimensional signals
which are fed into it as time series. Moreover, feature
extraction is not an issue anymore and we do not spend time
and cost on it when applying DBN as the classifier.
V. CONCLUSION AND FUTURE WORK
Results in this study suggest that DBN is a profitable and
powerful machine learning technique for separating clean and
noisy ECG segments prior to any analyses for diagnoses. It
can perform the task with a high precision and recall rates.
The DBN can also appropriately discriminate arrhythmic ECG
signals and noisy ECG signals, even if the signal is only
slightly noisy (i.e., relatively high SNR).
Such algorithms can be added to an ECG monitoring device
as a preprocessing step to validate adequate signal quality
prior to analyses for diagnoses. This can help reduce the false
alarm rate by ensuring that only clean signals are being
processed. Another advantage of having such systems is that
the process of making decision of keeping the signal and
passing it through for further analysis can be performed in a
real-time manner because signals are fed to the system as time
series.
In future work, we intend to define a Signal Quality Index
(SQI) applicable to all types of contaminating noises and
classify ECG signals according to the SQI using DBNs. Also,
we will apply DBN to ECG signals aiming to differentiate
arrhythmias for diagnostic purposes. Furthermore, we will
apply a new clinically recorded ECG signal dataset to the
algorithm. For the future, to elevate the rate of precision and
recall, a new learning method called “Persistent Contrastive
Divergence (PCD)” [31] will be implemented, as well.
REFERENCES
[1]
"www.heart.org," 2015. [Online]. Available:
https://www.heart.org/idc/groups/ahamahpublic/@wcm/@sop/@smd/d
ocuments/downloadable/ucm_470704.pdf.
[2]
C. De Capua, A. Meduri and R. Morello, "A smart ECG measurement
system based on web-service-oriented architecture for telemedicine
applications," IEEE Transactions on Instrumentation and
Measurement, vol. 59, no. 10, pp. 2530-2538, 2010.
[3]
J. Behar, J. Oster, Q. Li and G. D. Clifford, "ECG signal quality during
arrhythmia and its application to false alarm reduction," IEEE
Transactions on Biomedical Engineering, vol. 60, no. 6, pp. 1660-
1666, 2013.
[4]
M. Abdelazez, P. X. Quesnel, A. D. Chan and H. Yang, "Signal
Quality Analysis of Ambulatory Electrocardiograms to Gate False
Myocardial Ischemia Alarms," in Press in IEEE Transactions on
Biomedical Engineering, 2016.
[5]
J. G. Webster, Medical Instrumentation Application and Design, 4th
ed. Wiley Global Education, 2009.
[6]
E. Institute, "Top 10 Health Technology Hazards for 2015," Nov 2014.
[7]
E. Institute, "Top 10 Health Technology Hazards for 2014," Health
devices, vol. 42, no. 11, Nov 2013.
[8]
E. Institute, "Top 10 Health Technology Hazards for 2013," Health
Devices, vol. 41, no. 11, Nov 2012.
[9]
E. Institute, "Top 10 Health Technology Hazards for 2012," Health
Devices, vol. 40, no. 11, Nov 2011.
[10]
K. R. Foster, R. Koprowski and J. D. Skufca, "Machine learning,
medical diagnosis, and biomedical engineering research-
commentary.," Biomedical engineering online, vol. 13, no. 1, 2014.
[11]
K. Polat and S. Güneş, "Detection of ECG Arrhythmia using a
differential expert system approach based on principal component
analysis and least square support vector machine," Applied
Mathematics and Computation, vol. 18, no. 1, pp. 898-906, 2007.
[12]
S. Mitra, M. Mitra and B. Chaudhuri, "A rough-set-based inference
engine for ECG classification," IEEE Transactions on instrumentation
and measurement, vol. 55, no. 6, pp. 2198-2206, 2006.
[13]
K. Englehart, B. Hudgins and A. D. Chan, "Continuous multifunction
myoelectric control using pattern recognition," Technology and
Disability, vol. 15, no. 2, pp. 95-103, 2003.
[14]
L. Biel, O. Pettersson, L. Philipson and P. Wide, "ECG analysis: a new
approach in human identification," IEEE Transactions on
Instrumentation and Measurement, vol. 50, no. 3, pp. 808-812, 2001.
[15]
G. D. Fraser, A. D. Chan, J. R. Green and D. T. MacIsaac, "Automated
Biosignal Quality Analysis for Electromyography using a One-Class
Support Vector Machine," IEEE Transactions on Instrumentation and
Measurement, vol. 63, no. 12, pp. 2919-2930, 2014.
[16]
P. McCool, G. D. Fraser, A. D. Chan, L. Petropoulakis and J. J.
Soraghan, "Identification of Contaminant Type in Surface
Electromyography (EMG) Signals," IEEE Transactions on Neural
Systems and Rehabilitation Engineering, vol. 22, no. 4, pp. 774-783,
2014.
[17]
G. D. Clifford, j. Behar, Q. Li and I. Rezek, "Signal quality indices and
data fusion for determining clinical acceptability of
electrocardiograms," Physiological measurement, vol. 33, no. 9, p.
1419, 2012.
[18]
J. Kužílek, M. Huptych, V. Chudáček, J. Spilka and L. Lhotská, "Data
driven approach to ECG signal quality assessment using multistep
SVM classification," Computing in Cardiology, pp. 453-455, 2011.
[19]
M. Huanhuan and Z. Yue, "Classification of Electrocardiogram Signals
with Deep Belief Networks," in IEEE 17th International Conference
on Computational Science and Engineering (CSE), 2014.
[20]
Y. Bengio, A. Courville and P. Vincent, "Representation Learning: A
Review and New Perspectives," IEEE transactions on pattern analysis
and machine intelligence, vol. 35, no. 8, pp. 1798-1828, 2013.
[21]
H. Lee, R. Grosse, R. Ranganath and A. Y. Ng, "Convolutional deep
belief networks for scalable unsupervised learning of hierarchical
representations," in Proceedings of the 26th annual international
conference on machine learning, 2009.
[22]
H. Lee, P. Pham, Y. Largman and A. Y. Ng, "Unsupervised feature
learning for audio classification using convolutional deep belief
networks," Advances in neural information processing systems, pp.
1069-1104, 2009.
[23]
D. Erhan, Y. Bengio, A. Courville, P. A. Manzagol, P. Vincent and S.
Bengio, "Why Does Unsupervised Pre-training Help Deep Learning,"
Journal of Machine Learning Research, vol. 11, pp. 625-660, 2010.
[24]
Y. Lv, Y. Duan, W. Kang, Z. Li and F. Y. Wang, "Traffic Flow
Prediction with Big Data: A Deep Learning Approach," IEEE
Transactions on Intelligent Transportation Systems, vol. 16, no. 2, pp.
865-873, 2015.
[25]
G. E. Hinton, S. Osindero and Y. W. Teh, "A fast learning algorithm
for deep belief nets," Neural computation, vol. 18, no. 7, pp. 1527-
1554, 2006.
[26]
G. E. Hinton, "A practical guide to training restricted Boltzmann
machines," Neural Networks: Tricks of the Trade, vol. 9, no. 1, pp.
599-619, 2012.
[27]
M. A. Carreira-Perpinan and G. E. Hinton, "On contrastive divergence
learning," Artificial Intelligence and Statistics, vol. 10, pp. 33-40,
2005.
[28]
A. L. G. e. al, "PhysioBank, PhysioToolkit, and PhysioNet
Components of a New Research Resource for Complex Physiologic
Signals," Circulation, vol. 101, no. 23, p. e215–e220, Jun 2000.
[29]
"physionet," [Online]. Available:
https://www.physionet.org/physiotools/wag/nst-1.htm.
[30]
R. Hecht-Nielsen, "Theory of the backpropagation neural network," in
International Joint Conference , 1989.
[31]
T. Tieleman, "Training restricted Boltzmann machines using
approximations to the likelihood gradient," in Proceedings of the 25th
international conference on Machine learning, New York, 2008.
