Content uploaded by Shadnaz Asgari
Author content
All content in this area was uploaded by Shadnaz Asgari on Nov 12, 2021
Content may be subject to copyright.
Automatic detection of atrial fibrillation using stationary wavelet
transform and support vector machine
Shadnaz Asgari
a,
n
, Alireza Mehrnia
b
, Maryam Moussavi
c
a
Department of Computer Engineering and Computer Science, California State University, Long Beach,1250 Bellflower Boulevard-MS 8302, #Long Beach, CA
90840, USA
b
Department of Electrical Engineering, University of California, Los Angeles, 56-125B Engineering IV Building, Box 951594, Los Angeles, CA 90095, CA
c
Department of Electrical Engineering, California State University, Long Beach, 1250 Bellflower Boulevard, Long Beach, CA 90840, CA
article info
Article history:
Received 19 August 2014
Accepted 6 March 2015
Keywords:
Atrial fibrillation
Support vector machine
Wavelet transform
Cardiac arrhythmia
Log-energy entropy
ROC curve analysis
abstract
Background: Atrial fibrillation (AF) is the most common cardiac arrhythmia, and a major public health
burden associated with significant morbidity and mortality. Automatic detection of AF could substan-
tially help in early diagnosis, management and consequently prevention of the complications associated
with chronic AF. In this paper, we propose a novel method for automatic AF detection.
Method: Stationary wavelet transform and support vector machine have been employed to detect AF
episodes. The proposed method eliminates the need for P-peak or R-Peak detection (a pre-processing
step required by many existing algorithms), and hence its performance (sensitivity, specificity) does not
depend on the performance of beat detection. The proposed method has been compared with those of
the existing methods in terms of various measures including performance, transition time (detection
delay associated with transitioning from a non-AF to AF episode), and computation time (using MIT-BIH
Atrial Fibrillation database).
Results: Results of a stratified 2-fold cross-validation reveals that the area under the Receiver Operative
Characteristics (ROC) curve of the proposed method is 99.5%. Moreover, the method maintains its high
accuracy regardless of the choice of the parameters' values and even for data segments as short as 10 s.
Using the optimal values of the parameters, the method achieves sensitivity and specificity of 97.0% and
97.1%, respectively.
Discussion: The proposed AF detection method has high sensitivity and specificity, and holds several
interesting properties which make it a suitable choice for practical applications.
&2015 Elsevier Ltd. All rights reserved.
1. Introduction
Atrial fibrillation (AF) is the most common cardiac arrhythmia
with an estimated prevalence of 1% corresponding to 2.3 million
patients in the US and 4.5 million in the European Union [1].This
prevalence is strongly associated with age, such that over 17% of
people aged 85 or above are AF patients [2].Itisexpectedthatthe
number of AF patients in U.S. will increase by 2.5-fold to more than
5.6 million people by 2050 [3].Thisincreasenotonlyreflects the
growing population of elderly individuals, but also improved
survival of people with predisposing conditions to AF (e.g. ischemic
heart disease) [4]. The presence of AF is associated with a 5-fold
increased risk of stroke [5] and about 2-fold increased risk of death,
independent of other risk factors [6]. Furthermore, due to high rates
of hospitalization and considerable health resource utilization, the
economic and clinical burden of AF is substantial and will continue
to increase in the future [7].
AF is usually characterized by rapidity and irregularity of
ventricular contraction [8]. The chronic forms of AF can be parox-
ysmal AF (more than one episode with spontaneous termination
within seven days); persistent AF (not self-terminated, or lasted
more than seven days) and permanent AF (not terminated, or
terminated but relapsed) [9]. Silent or asymptomatic AF can also
occur in any of these temporal forms, carrying a similar prognosis to
symptomatic AF [10]. Chronic AF adversely affects the blood flow
dynamics and can result in the stroke. Early and accurate detection
of AF and its management (e.g. anticoagulation, antiarrhythmic
therapy and radiofrequency ablation) could substantially help in
prevention of the complications associated with the chronic AF.
Currently, the diagnosis of AF mainly rests on the presence of
related symptoms (e.g. shortness of breath and fatigue), followed
by an electrocardiogram (ECG) study to verify the diagnosis.
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/cbm
Computers in Biology and Medicine
http://dx.doi.org/10.1016/j.compbiomed.2015.03.005
0010-4825/&2015 Elsevier Ltd. All rights reserved.
n
Corresponding author. Tel.: þ1 562 985 5467; fax: þ1 562 985 7823.
E-mail address: Shadnaz.Asgari@csulb.edu (S. Asgari).
Computers in Biology and Medicine 60 (2015) 132–142
During the ECG study, a trained clinician/technician visually
inspects the ECG signal collected over short period of time
(e.g. less than 48 h) to identify the AF associated chaotic patterns
or abnormal changes in the waveform morphology.
Early and accurate diagnosis of AF is a challenging task. As a
result, a majority of the early stage and easily treatable AF cases
are not diagnosed in time and evolves into chronic debilitating AF
with high cardiac related complications. Several factors hinder the
accurate and early detection of AF:
Many of the AF patients do not experience symptoms. Hence, a
considerable number of AF cases are undetected or diagnosed
fortuitously when the subject is being evaluated for other
cardiac complications. Moreover, the studies have shown that
a good correlation does not always exist between symptoms
and episodes of AF [11].
ECG study can miss detection of AF in patients with paroxysmal
intermittent AF who are not experiencing AF episodes during
the study.
Visual inspection of hours of collected ECG is time consuming.
Accurate interpretation of ECG study requires extensive train-
ing and experience. Hence, the reliability of diagnosis is
dependent on the level of training or experience of the
clinician. In fact, studies have shown that many primary care
practitioners are not able to detect AF (on an ECG) with
sufficient accuracy to guide therapy [12].
A possible solution to this problem is automatic AF detection
using ambulatory monitoring. This approach could eliminate the
need for visual inspection, and contribute to early and accurate
detection of AF by allowing for an in-depth analysis of ECG and
identification of abnormal patterns.
Over the last two decades, several algorithms have been devel-
oped for automatic AF detection. Many of these algorithms either
rely on the absence of P-waves (replaced by rapid oscillations or
fibrillatory waves) [13–16] or R–R irregularities [17–21], or combi-
nation of both characteristics to detect AF episodes [22–24].As
P-waves are prone to contamination with motion and noise
artifacts, the AF detection algorithms solely based on the absence
of P-waves perform poorly in the presence of noise [25]. Therefore,
employing heart rate variability for AF detection has become a
preferred approach in recent years. But many of these R–R interval
based methods compare the density histogram of R–R intervals for
a segment of data with previously compiled histograms of R–R
intervals during AF using the Kolmogorov–Smirnov test [18,24,26].
Since these algorithms require storage of large amount of histogram
data, they may not be suitable for implementation in an ambulatory
monitoring device with limited memory and processing power.
Other R–R algorithms need various parameters' tuning to guarantee
the high accuracy of AF detection [19–21]. Furthermore, most of the
existing methods do not perform well with short data segments
(less than one minute) [25].Thisdeficiency may result in missing
short-duration AF episodes (prevalent in Paroxysmal or early stage
AF), less accurate calculation of AF burden (a measure of the percent
of time a patient spends in AF), and a lower speed of AF detection.
Lastly, a majority of existing approaches for AF detection require P
or R peak detection as a pre-processing step and consequently their
performance will degrade if the related peaks are missed or
erroneously detected.
Aimed at addressing the existing issues, the current work
proposes a novel method for automatic AF detection using sta-
tionary wavelet transform (SWT) and support vector machine
(SVM). As opposed to traditional methods, the proposed method
does not require P or R peak detection pre-processing step. Only
few parameters are involved in the development of the proposed
method and the method maintains its high accuracy regardless of
the choice of the parameters' values. The proposed method also
performs well even for short data segments of 10 s. In what follows,
we first describe (Section 2) our proposed method in detail. Then in
Section 3, we present the result of applying the proposed method
on MIT-BIH Atrial fibrillation database. These results are discussed
and compared with those of the existing methods in Section 4.
Conclusions are presented in Section 5.
2. Materials and methods
2.1. Automatic AF detection using wavelet transform and support
vector machine
The proposed method consists of three major steps: pre-
processing; feature extraction; and AF classification. Fig. 1 illus-
trates a block diagram of various steps of this method. In the
following sub-sections, we will describe each step in detail.
2.1.1. Pre-processing
The incoming ECG signal is divided every Tseconds into
segments of length T. An elliptical band-pass filter with pass-
band of 0.5–50 Hz (effective filter order of 10) is applied to each
data segment to remove noise and baseline wander. After filtering
in the forward direction, the filtered sequence was then reversed
and run back through the filter to obtain a zero-phase distortion.
Then a wavelet transform is performed on the filtered data
segments.
Wavelet transform has proved to be a useful tool for denoising,
delineation and compression of signals. Using wavelet transform,
one can observe a signal at different scales where each scale
emphasizes on various signal properties and characteristics [27].
Wavelet transform allows for the analysis of transients, aperiodi-
city and other non-stationary signal features where, through the
Fig. 1. Modular framework of the proposed method for AF detection.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142 133
interrogation of the transform, subtle changes in signal morphol-
ogy may be highlighted over the scales of interest [28].
In the current work, we employ wavelet transform which
allows for arbitrarily high resolution of the high frequency signal
components (in contrast to short term Fourier transform) and then
extract features that would enable us to detect AF episodes.
Among several wavelet transform techniques, we choose station-
ary wavelet transform (SWT). SWT is time-invariant and at each
decomposition level, its wavelet coefficients carry the same
number of samples (temporal information) as that of the original
signal. SWT overcomes the problem of repeatability and robust-
ness of the analysis which exists with a discrete wavelet trans-
form. We leverage these remarkable properties of SWT, and
employ it for the purpose of AF detection.
It is known that to implement SWT with L-levels on a signal,
the length of signal should be a multiple of 2
L
. In order to meet
this data length requirement, we zero-pad the filtered data
segment (Let us call the zero-padded signal xðnÞwhere
n¼1; :::; N). Fig. 2 presents a tree-structure diagram of an L-level
SWT ðℓ¼1;2; :::; LÞ.
In this figure, GðzÞand HðzÞare high-pass and low-pass filters
respectively, designed based on the wavelet basis function, and
XðzÞis the zero-padded signal in z-domain. Note that at decom-
position level ℓ, the impulse responses of high-pass filter Gðz
2
ℓ1
Þ
and low-pass filter Hðz
2
ℓ1
Þare up-sampled versions (by a factor of
2) of impulse responses of high-pass filter Gðz
2
ℓ2
Þand low-pass
filter Hðz
2
ℓ2
Þat previous level ðℓ1Þ. The detail coefficients d
ℓ
ðnÞ
and coarse coefficients c
ℓ
ðnÞin time domain can be recursively
obtained as
D
ℓ
ðzÞ¼Gðz
2
ℓ1
Þ:D
ℓ1
ðzÞ)d
ℓ
ðnÞ¼P
m
gðnÞ:d
ℓ1
ðn2
ℓ1
mÞ;C
ℓ
ðzÞ
¼Hðz
2
ℓ1
Þ:C
ℓ1
ðzÞ)c
ℓ
ðnÞ¼X
m
hðnÞ:c
ℓ1
ðn2
ℓ1
mÞ:
n¼1; :::; Nð1Þ
Hence, at the end of pre-processing step, a total of 2 Ltime
series are generated (Ldetail coefficients and Lcoarse coefficients),
where each coefficient time series has the same time resolution as
the original signal xðnÞ. Note that the optimal value for the length
of the data segment (T) is not a trivial choice. The parameter
tuning to maximize the performance of the proposed method will
be discussed later in Section 2.3.
2.1.2. Feature extraction
Following the calculation of wavelet coefficients, at this step, two
features are extracted for each coefficient: Peak-to-average power
ratio; and log-energy entropy. These features are later used in
combination with a support vector machine classifier to detect AF.
2.1.2.1. Peak-to-average power ratio. The power spectrum of the
wavelet coefficients at each level provides spectral information of
ECG signal for different scales [29], and hence it can be used to
analyze atrial activity. In fact, power spectrogram of detail
coefficient at ℓth level ðS
D
ℓ
ðfÞÞ and that of coarse coefficient
ðS
C
ℓ
ðfÞÞ can be expressed as
S
D
ℓ
ðfÞ¼ΕfjD
ℓ
ðfÞj
2
g;
S
C
ℓ
ðfÞ¼ΕfjC
ℓ
ðfÞj
2
g;ð2Þ
where Εdenotes expectation operation, and D
ℓ
ðfÞand C
ℓ
ðfÞare
Fourier transforms of detail coefficient ðD
ℓ
ðfÞ¼D
ℓ
ðzÞj
z¼e
j2
π
ft
Þand
coarse coefficient ðC
ℓ
ðfÞ¼C
ℓ
ðzÞj
z¼e
j2
π
ft
Þ, respectively.
It is known that atrial activity usually occurs in the frequency
range of F¼[4–9 Hz]. Therefore during an AF episode, one can
expect to have higher power concentration in this frequency range
[25]. Peak-to-average power ratio has been shown to be effective
in capturing the power distribution profile in the aforementioned
frequency range [29], and is obtained according to the following
formula:
ρ
D
ℓ
¼
max S
D
ℓ
ðfÞ
fAF
R
fAF
S
D
ℓ
ðfÞ
;
ρ
C
ℓ
¼
max S
C
ℓ
ðfÞ
fAF
R
fAF
S
C
ℓ
ðfÞ
;
ð3Þ
where ρ
D
ℓ
and ρ
C
ℓ
are Peak-to-average power ratio of detail and
coarse coefficients at level ℓ, respectively. By calculating ρ
D
ℓ
and ρ
C
ℓ
for all levels ℓ¼f1;2; :::; Lg, we obtain 2 Lpower spectrum
related features for each data segment.
2.1.2.2. Log-energy entropy. Entropy is a measure of the degree of
disorder, uncertainty or randomness of a signal. Wavelet entropy
provides useful information about the underlying dynamic system
(associated with a signal) over different frequency bands. For
example, entropy of a mono-frequency signal is very low, given
that the wavelet coefficient at the decomposition level containing
the representative frequency will contain most of the signal
energy. In contrast, a random white noise has high entropy, as
there is a significant contribution of energy from wavelet
coefficients at various decomposition levels [30].
During AF, electrical discharges conducted from the atrium into
the ventricles are irregular and disorganized. Consequently, one
can expect that the ECG signal demonstrates more degree of
randomness and a higher entropy during an AF episode [30].
Leveraging this note, in order to reveal underlying dynamic
process relevant to AF pathology, we calculate the wavelet entropy
of the signal. Various wavelet entropy measures have been
defined. In this paper, we use a simple non-normalized log-energy
entropy [31] (which is conveniently included in Matlab Wavelet
Packet and quickly executable) defined as
E
D
ℓ
¼X
N
n¼1
log ðd
ℓ
ðnÞÞ
2
;
E
C
ℓ
¼X
N
n¼1
log ðc
ℓ
ðnÞÞ
2
;
ð4Þ
where E
D
ℓ
and E
C
ℓ
are the log-energy entropy of detail coefficient
and coarse coefficient at ℓth level, respectively. Calculation of log-
energy entropy for all wavelet coefficients results in obtaining 2
Ltime related features for each data segment. Hence, at the end of
this step, a vector of 4 Lfeatures are extracted for each data
segment as
Φ¼ρ
C
1
::: ρ
C
L
ρ
D
1
::: ρ
D
L
E
C
1
::: E
C
L
E
D
1
::: E
D
L
hi
T
:ð5Þ
Fig. 2. Tree-structure of an L-level stationary wavelet transform ðℓ¼1;2;…;LÞ.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142134
2.1.3. AF classification
Our goal is to classify each data segment into AF or non-AF
category using 4 Lextracted features. For this purpose, we
employ support vector machine (SVM) as our classification algo-
rithm. SVM is a non-parametric binary classifier which has shown
promising results in various medical diagnostics [32–35].
In a conventional problem of binary classification, a data point is
viewed as a p-dimensional vector which belongs to one the two
possible categories. An SVM classifies these data points by finding the
best (p1)-dimensional hyperplane that separates all data points of
one class from those of the other class. The separating hyperplane
has the largest margin between the two classes (Margin is the
maximal width of the slab parallel to the hyperplane that has no
interior data points). SVM then classifies new samples based on
which side of the hyperplane they fall into and how far the samples
are from the hyperplane (SVM score)[36]. One of the major advan-
tages of SVM is that in addition to linear classification, it can
efficiently perform a non-linear classification [37].Usingakernel
trick, samples are mapped into higher-dimensional feature space
where a linear classifier can separate the two classes with the largest
margin among the samples. Mapping the separating hyperplane back
to the original space results in having a non-linear classifier.
For our AF detection, we use a common kernel choice, Gaussian
Radial Basis Function kernel
KðΦ
i
;Φ
j
Þ¼e
‖
Φ
i
Φ
j‖2
2
2
σ
2
;ð6Þ
with scaling factor σ¼1. Φ
i
and Φ
j
are features of two data
segments and ‖Φ
i
Φ
j
‖
2
2
is the squares Euclidean distance
between the two feature vectors. The separating hyperplane for
our SVM classifier is obtained using the conventional “Least-
Squares”method.
2.2. Patient data
To evaluate the performance of the proposed method, we
employ MIT-BIH Atrial fibrillation (MIT-BIH AFIB) dataset [38],
the most popular and most frequently used publicly available
dataset for AF detection. This dataset contains 23 annotated
records of ECG signal from atrial fibrillation patients (mostly
paroxysmal) with sampling frequency of 250 Hz and 12-bit resolu-
tion over a range of 710 mV. Each record is about 10 h-long and
the whole dataset includes slightly less than 234 h of data. The
dataset has 605 annotated episodes: 291 atrial fibrillation episodes
with average time duration of 115 s, 14 atrial flutter episodes with
average time duration of 419 s, 12 episodes of junctional rhythm
with average time duration of 27 s and 288 episodes of all other
rhythms with average time duration of 174 s. Fig. 3 presents
examples of ECG signal during four annotated episodes.
2.3. Data analysis and validation protocol
Daubechies 5 is an orthogonal wavelet resembling the ECG
waveform in morphology [29]. Hence, we choose the Daubechies
5 wavelet as the mother wavelet to perform wavelet analysis.
Given the sampling frequency of 250 Hz and the frequency range
of atrial activities (4–9 Hz) [39],aL¼6 level wavelet transform is
implemented. A periodogram spectral estimator with a Hamming
window is used to obtain the power spectrum of each wavelet
coefficient.
The original beat-to-beat annotations of data are converted to a
T-second resolution (Tis the duration of each data segment) by
using a minimum percentage parameter P[19,20]: A data segment
is classified as a true AF only if the percentage of annotated AF beats
in that data segment is more than P. A two-fold stratified cross-
validation on the feature vectors (extracted from the dataset) is
employed to train and test the classifier. At the testing phase of each
fold, SVM's scores for the testing samples (distance of the testing
samples to the separating hyperplane) are obtained and compared
with a pre-set threshold ςto classify corresponding testing samples.
The classification results (from each fold) is then compared to the
ground truth (converted annotation) and number of True Positive
(TP), False Negative (FN), True Negative (TN) and False Positive (FP)
cases are calculated. Finally, the results from the two folds are
combined to calculate True Positive Rate (TPR) as TPR ¼TP=TPþFN,
FalsePositiveRate(FPR)asFPR ¼FP=FPþTN and Accuracy (ACC) as
ACC ¼TPþTN=TPþTNþFP þFN.
In order to obtain the optimal values of the parameters of the
proposed method (T,Pand ς), an exhaustive standard Receiver
Operating Characteristic (ROC) Curve analysis is implemented as
the following: Parameter T(length of the data segment in seconds)
is varied over a reasonable range of 10–120 s with incremental
steps of 5 s. Parameter P(percentage threshold for annotation of
AF data segments) is varied from 0% to 100% with incremental
steps of 10%. Then for each specific value of Tand P, an ROC curve
is derived by varying the score threshold ςfrom 1.5 to 1.5 with
incremental steps of 0.01. The ROC with the highest Area Under
the Curve (AUC) is used to obtain the optimal values of parameters
0 5 10 15
−200
−100
0
100
200
Time (seconds)
ECG Signal
0 5 10 15
−600
−400
−200
0
200
400
Time (seconds)
ECG Signal
0 5 10 15
−200
0
200
400
Time (seconds)
ECG Signal
0 5 10 15
−200
0
200
400
600
Time (seconds)
ECG Signal
Fig. 3. Examples of ECG signal during four annotated episodes. (A) Atrial fibrillation; (B) atrial flutter; (C) junctional rhythm; and (D) other rhythm.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142 135
Tand P. Furthermore, the score threshold ςcorresponding to the
knee point of this ROC curve (closest point to ideal performance of
FPR¼0 and TPR¼1) is considered as the optimal value for score
threshold.
To further investigate the efficacy of the chosen classification
algorithm (SVM), we compare the ROC curve of AF detection using
SVM classifier with those of three other popular classifiers—viz.,
Naïve Bayesian, Linear discriminant Analysis (LDA), and Logistic
Regression. For this purpose, the ROC curves of other classifiers are
obtained similar to that of SVM by varying either the score thresh-
old (logistic regression), or a threshold on the AF posteriori (Naïve
Bayesian, LDA). Finally, we compare the performance of our
proposed method with those of the existing AF detection methods
in terms of various measures including accuracy, transition time,
and computation time.
3. Results
Fig. 4 presents a typical ECG waveform prior to and during an AF
(extracted from 6th AF episode of file 4048 in MIT-BIH AFIB
database), along with its detail wavelet coefficients at levels 4–6
(d
4
ðnÞ;d
5
ðnÞand d
6
ðnÞ). This example demonstrates signature
characteristics of an AF episode, namely irregular R–Rintervaland
absence of P wave. It also shows the efficacy of wavelet transform to
capture the atrial activity, especially at detail coefficients.
Fig. 5 presents two examples of the application of the proposed
method (with T¼10 and P¼50%), on segments of ECG signal. The
ECG signal of Fig. 5(A) contains the 8th AF episode of file 4936,
while the ECG signal of Fig. 5(E) contains the 27th normal sinus
rhythm (NSR) episode of file 8219. From Fig. 5(D), we observe that
by using a score threshold of ς¼0, our proposed method is able to
correctly identify the entire AF episode (a true positive case).
Similarly, as displayed in Fig. 5(H), our proposed method is able to
correctly identify the entire non-AF episode (a true negative case).
The application of the proposed method resulted in few cases of
false negative and false positive in detection of AF episodes. Fig. 6
presents two examples of those cases (a false negative and a false
positive case). The ECG signal of Fig. 6(A) contains the 2th AF episode
of file 4746, while the ECG signal of Fig. 6(E) contains the 26th NSR
episode of file 4936. As Fig. 6(D) illustrates, with T¼10 s, our proposed
method has missed the detection of the short AF episode (a false
negative case). Note that in both of the two 10-s data segments which
partially contain the AF episode (corresponding to time between 30
and 50), only less than P¼50%of the segment holds the AF pattern.
This could have contributed to missing the detection of the related AF
episode. On the other hand, Fig. 6(H) shows that although our
proposed method has been able to identify the NSR episode, but its
estimated duration is less than the duration of the annotated episode.
This is due to the fact that calculated SVM score for the 4th 10 s data
segment (corresponding to time between 30 and 40) is slightly above
the threshold ς¼0, and hence, this data segment has been falsely
detected as AF (false positive case).
Fig. 7 presents two more examples of false positive cases. The
ECG signal of Fig. 7(A) (extracted from file 6426) contains an atrial
flutter episode with duration of approximately 118 s. As Fig. 7
(D) illustrates, few data segments along the corresponding atrial
flutter episode have been wrongly detected as atrial fibrillation. The
ECG signal of Fig. 7(E) contains a junction rhythm episode of file
7879. Similarly (as Fig. 7(H) demonstrates), the proposed method
has falsely labeled the junction rhythm episode as an AF episode. In
both examples, the irregularity associated with a non-AF episode
(i.e., atrial flutter, and junctional rhythm) has resulted in a false
positive case. To evaluate the prevalence of the false positive (and
false negative cases), we employed the ROC curve analysis (Fig. 8).
Note that although Fig. 8 depicts the results of parameter value
optimization only for few values of Tand P, we performed the
0 5 10 15 20 25 30
−1
0
1
Time (second)
Original Signal
0 5 10 15 20 25 30
−2
0
2
Time (second)
d4
0 5 10 15 20 25 30
−2
0
2
Time (second)
d5
0 5 10 15 20 25 30
−2
0
2
Time (second)
d6
Beginning of an AF episode
Fig. 4. An example of ECG signal prior to and during AF (extracted from 6th AF episode from file 4048 of MIT-BIH AFIB database) and its detail wavelet coefficients.
(A) Original signal; (B) detail coefficient at level 4; (C) detail coefficient at level 5; and (D) detail coefficient at level 6.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142136
010 20 30 40 50 60 70 80 90 100 110 120 130 140 150150
−1
0
1
Time (second)
Original Signal
010 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Time (second)
True Annotation
010 20 30 40 50 60 70 80 90 100 110 120 130 140 150
−1
0
1
Time (second)
SVM Score
010 20 30 40 50 60 70 80 90 100 110 120 130 140 150
Time (second)
Detection Result
010 20 30 40 50 60 70 80 90 100
−1
0
1
Time (second)
Original Signal
010 20 30 40 50 60 70 80 90 100
Time (second)
True Annotation
010 20 30 40 50 60 70 80 90 100
−1
0
1
Time (second)
SVM Score
010 20 30 40 50 60 70 80 90 100
Time (second)
Detection Result
non−AF
AF
Transition Delay
non−AF
AF
AF
non−AF
non−AF
AF
Fig. 5. Two examples of application of the proposed method: Left side plots are related to an ECG data segment containing an AF episode. (A) Original signal; (B) true
annotation; (C) calculated score (output of SVM classifier); and (D) detection result (AF versus non-AF). Right side plots are related to an ECG data segment containing a NSR
episode. (E) Original signal; (F) true annotation; (G) calculated score (output of SVM classifier); and (H) detection result (AF versus non-AF).
010 20 30 40 50 60 70 80
Time (second)
True Annotation
010 20 30 40 50 60 70 80
−1
0
1
Time (second)
SVM Score
010 20 30 40 50 60 70 80
Time (second)
Detection Result
010 20 30 40 50 60 70 80 90 100
−1
0
1
Time (second)
Original Signal
010 20 30 40 50 60 70 80 90 100
Time (second)
True Annotation
010 20 30 40 50 60 70 80 90 100
−1
0
1
Time (second)
SVM Score
010 20 30 40 50 60 70 80 90 100
Time (second)
Detection Result
010 20 30 40 50 60 70 80
−1
0
1
Time (second)
Original Signal
AF
AF AF
non−AF
non−AF
non−AF
non−AF
AF
Fig. 6. Examples of false negative and false positive cases in AF detection using the proposed method: Left side plots are related to an ECG data segment containing an AF
episode. (A) Original signal; (B) true annotation; (C) calculated score (output of SVM classifier); and (D) detection result (AF versus non-AF). Right side plots are related to an
ECG data segment containing a NSR episode. (E) Original signal; (F) true annotation; (G) calculated score (output of SVM classifier); and (H) detection result (AF versus non-
AF).
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142 137
actual parameter optimization on the entire range of each para-
meter, as described in Section 2.3. It can be noticed that the ROC
curves of both subplots (Fig. 8(A,B)) are very closely spaced. In fact,
our results showed that the AUC of all four ROC curves of Fig. 8
(A) and all 5 ROC curves of Fig. 8(B) are above 0.992. This indicates
that our proposed method is able to maintain its high accuracy
regardless of the choice of the parameter values Tand P. Never-
theless, the highest AUC of 0.995 is obtained when T¼30 and
020 40 60 80 100 120 140 160
−1
0
1
Time (second)
Original Signal
020 40 60 80 100 120 140 160
non−AF
AF
Time (second)
True Annotation
020 40 60 80 100 120 140 160
−1
0
1
Time (second)
SVM Score
020 40 60 80 100 120 140 160
non−AF
AF
Time (second)
Detection Result
0 5 10 15 20 25 30 35 40 45 50
−1
0
1
Time (second)
Original Signal
0 5 10 15 20 25 30 35 40 45 50
non−AF
AF
Time (second)
True Annotation
0 5 10 15 20 25 30 35 40 45
−1
0
1
Time (second)
SVM Score
0 5 10 15 20 25 30 35 40 45 50
non−AF
AF
Time (second)
Detection Result
Fig. 7. Examples of false positive cases in AF detection using the proposed method: Left side plots are related to an ECG data segment containing an atrial flutter episode.
(A) Original signal; (B) true annotation; (C) calculated score (output of SVM classifier); and (D) detection Result (AF versus non-AF). Right side plots are related to an ECG data
segment containing an episode with junctional rhythm. (E) Original signal; (F) true annotation; (G) calculated score (output of SVM classifier); and (H) detection result (AF
versus non-AF).
0 0.02 0.04 0.06 0.08 0.1
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
FPR (1−specificity)
TPR (sensitivity)
T=15
T=30
T=45
T=60
0 0.02 0.04 0.06 0.08 0.1
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
FPR (1−specificity)
TPR (sensitivity)
P=10%
P=30%
P=50%
P=70%
T=90%
Knee Point
Fig. 8. Results of parameter value optimization using ROC curve analysis. (A) P¼0:7, varying score threshold ς, varying Tas 15, 30, 45 and 60 s; and (B) T¼30, varying score
threshold ς, varying Pas 10%, 30%, 50%, 70% and 90%.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142138
P¼50%. Moreover, the score threshold ςcorresponding to the
knee point of the ROC curve with T¼30 and P¼50%(closest
point to the ideal performance of FPR¼0 and TPR ¼1) is
ς¼0:05.
Fig. 9(A) presents results of AF detection (with T¼30 and
P¼50%), obtained from applying various classification algorithms
on the extracted 28-dimensional feature space. Note that similar
to Fig. 8, all the ROC curves are calculated by using a two-fold
stratified cross validation as described in Section 2.3. We observe
that Naïve Bayesian classifier with AUC of 0.915 has the lowest
accuracy. LDA and Logistic regression perform slightly better with
AUC of 0.937 and 0.949, respectively, while our proposed SVM
classifier achieves the best performance with AUC of 0.995.
Fig. 9(B) compares the result of AF detection using our proposed
method with that of Weng et al. which have used a similar SWT
approach [29]. We observe that AUC of Weng method is 0.89, and
hence our proposed method outperforms Weng method by 10%.
Furthermore, at sensitivity of 100%, the specificity of the proposed
method is 86.3%, while the specificity of the Weng method is only 2%.
Finally, Table 1 summarizes the result of the performance of our
proposed method on MIT_BIH AFIB database (two-fold stratified cross-
validation) using the optimal values of the parameters P¼50%an
ς¼0:05. With optimal data length of T¼30, our method achieves
sensitivity of 97%, specificity of 97.1% and an accuracy of 97.1%.
To complete our study, we also analyzed the transition delay
and computation time of our proposed method. Note that in all
segment-based AF detection methods, the transition from a non-
AF to an AF episode is detected with a delay (i.e. transition delay).
For example, the transition delay in the Fig. 5 is 6.8 s. In general,
transition delay is a function of the length of the data segment ðTÞ,
and a longer data segment will induce a larger transition delay.
Our analysis showed that the median transition delay of our
proposed method for T¼10 is 9.8 s.
The computation time of our proposed method implemented in
Matlab 2012a (on an Intel
s
core™CPU@ 3.5 GHz, 32 GB RAM,
64 bit OS) is approximately 35 ms and 40 ms, with T¼10 and
T¼30, respectively. Given this short computation time, our pro-
posed algorithm is easily realizable in real-time for automatic AF
detection.
4. Discussion
Automatic and accurate detection of AF using ambulatory ECG
monitoring could lead to earlier diagnosis, and consequently
provides more opportunities for effective management or treat-
ment of the condition, and avoiding the complications associated
with chronic AF.
Over the last two decades, several algorithms have been devel-
oped for automatic AF detection. These algorithms either rely on the
absence of P-waves or R–R irregularities, or combination of both
characteristics to detect AF episodes. Table 2 lists some of the most-
cited and recently proposed AF detection algorithms. Among the
listed methods, algorithm #1 is a P-wave-based method, algorithms
#2–#6areR–R interval-based algorithms, and algorithms #7–#9
use the combination of these characteristics to detect AF. Note that
all of these algorithms require a pre-processing step where the P
and/or R peaks of ECG signals need to be located. As a result, their
performances depend on the accuracy of peak detection step and
they do not perform well in the case of missing peaks or erro-
neously detected peaks.
The current work introduces a novel AF detection method which
does not require any P and/or R peak detection step. Our method
leverages on the high time-frequency resolution of stationary wavelet
transform, and captures atrial activity by calculating peak-to-average
power spectrum ratio over different frequency bands. Furthermore, we
extract the log-energy entropy of the wavelet coefficients, to enhance
the performance of the method. The extracted features are then
classified into AF (and non-AF) cases using a support vector machine.
4.1. Number of parameters and the choice of parameters values
The proposed algorithm has only few parameters to optimize:
segment length T, minimum AF percentage P,andSVMscore
threshold ς. Interestingly, the results of our parameter optimization
(Fig. 8) indicated that the method maintains its high accuracy
regardless of the choice of Tand Pvalues (less than 0.3% change
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR (1−Specificity)
TPR (Sensitivity)
00.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
FPR (1−Specificity)
TPR (Sensitivity)
Naive Bayesian
LDA
Logistic Regression
SVM
Weng Method
Proposed Method
Fig. 9. Comparison of AF detection results in terms of ROC curves. (A) Using various classification algorithms; and (B) proposed method versus that of Weng method [29].
Table 1
Result of the performance of our proposed method on MIT_BIH AFIB database
(two-fold stratified cross-validation) using the optimal values of the parameters
(P¼50%and ς¼0:05).
Segment length T(s) Sensitivity (%) Specificity (%) Accuracy (%)
10 96.6 96.3 96.4
15 97.0 96.8 96.9
30 97.0 97.1 97.1
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142 139
in AUC). In contrast to our method, many of the existing algorithms
need to tune at least 5 parameters to guarantee a high performance
[19–21]. Moreover, the accuracy of these methods is much more
sensitive to the choice of those parameters. For example, in Dash
[20] changing the segment length from 128 beats to 64 beats
decreases TPR by more than 10% (from 96% to approximately 85%),
whereas in Lee method the TPR decreases by 5%, if one changes the
segment length from 64 beats to 12 beats. Note that in general,
there is a tradeoff between accuracy and speed of diagnosis. Shorter
data segments will result in having a higher speed, but lower
accuracy in AF detection. A measure that can substantiate one AF
detection method from another is the degree of this trade-off. For
our proposed method, the extent of this trade-off is low, because
our algorithm is able to maintain its high accuracy (96.4%) even
with a data segment as short as 10 s.
4.2. Comparison of performance of the proposed method using
various Classifiers
Among various classifiers we evaluated for AF detection, Naïve
Bayesian demonstrated the lowest performance (Fig. 9(A)). This
observation is consistent with the fact that underlying assumption
of Naïve Bayesian classifier, i.e. all features are independent, does
not hold for our feature space. The performance of LDA and
Logistic regression were close (around 94%). But our SVM classifier
had the highest performance with an AUC of 99.5%.
4.3. Comparison of performance of the proposed method with Weng
[29] method
Similar to our approach, Weng et al. has employed stationary
wavelet transform to detect AF episodes [29]. In their paper, the
authors compared their proposed SWT-based method with the
frequency domain QRST-cancellation based method which obtains
the dominant frequency components in the residual signal after
QRST cancellation using Short-time Fourier transform. Their results
showed that Weng method has a superior performance (in terms of
AUC) to the frequency domain QRST-cancellation method by 14%.
Given the efficacy of Weng method, we compared the performance
of our method with that of Weng (Fig. 9(B)). The results showed
that our proposed method outperformed Weng method by 10% in
AUC (and consequently, it outperformed the frequency domain
QRST-cancellation method by 24%). Note that although our
proposed method is similar to Weng method in terms of employing
stationary wavelet transform for AF detection, there exist two major
differences between the two methods:(1) Our proposed method
uses entropy measure as additional feature (in addition to peak-to-
average-ratio) to classify AF data segments. This additional feature
selection contributed to enhancement of the performance of AF
detection; (2) While our proposed method employs SVM with
Gaussian radial basis function kernel for classification (a non-
linear classification), Weng method implements Perceptron classi-
fier, following a Fisher Discriminant Ratio analysis of the features
(to reduce dimensionality of the feature space). A linear classifier
such as Perceptron may not be the best option for AF classification,
as the peak-to-average features may not be well separated linearly.
Moreover, feature dimensionality reduction using Fisher Discrimi-
nant Ratio has the underlying assumption of independency and
Gaussian distribution for features; both assumptions may not hold
for the chosen feature space.
4.4. Comparison of performance of the proposed method with the
existing Methods
Comparison of our algorithm with the existing methods listed in
Table 2 shows that the proposed method has a superior perfor-
mance (in terms of both sensitivity and specificity) to Slocum [14],
Taten o [18],Dash[20], Couceiro [22] and Babaeizadeh [23] algo-
rithms. While the specificity of our proposed method is 1% lower
than that of Huang [26], its sensitivity is about 1% higher. The
sensitivity and specificity of our algorithm is slightly lower than
Sarkar [19] (by 0.5% and 1.9%), Jiang [24] (by 1.1% and 0.4%), and Lee
[21] (by 1.2% and 0.6%), respectively. Although the last four afore-
mentioned methods have shown a slightly better performance in
AF detection, the proposed method holds some other properties
which may make it a better choice for practical applications:
Huang [26] and Jiang [24] method need to implement the
computationally expensive Kolmogorov–Smirnov test on the
histograms of R–R intervals to detect AF. The comprehensive
storage capacity requirement of these two methods may limit
their applications in mobile monitoring devices with limited
memory and processing power.
Sarkar [19] and Jiang [24] method need a long data segment
(2 min for Sarkar and 50 beats for Jiang) to guarantee a high
accuracy on AF detection. Our proposed method is able to
Table 2
Comparison of the performances of recent AF detection algorithms on MIT-BIH AFIB database. The bold percentages inside the parentheses are the results re-evaluated by
Larbruru et al. [25].
#Algorithm name Sensitivity
(%)
Specificity
(%)
Weakness points (relative to the proposed method)
1 Slocum et al. [14]
a
(62.8) (77.5) Needs beat detection, has lower performance
2 Tateno 2001 [18]
b
94.4
(91.2)
97.2
(96.1)
Needs beat detection, has lower performance, has larger memory consumption
3 Sarkar et al. [19]
b
97.5 99 Needs beat detection, has larger memory consumption, needs longer data segment to guarantee high performance
4 Dash et al. [20]
b
94.4 95.1 Needs beat detection, has lower performance, has larger transition delay, has higher sensitivity to the selection of
parameter values
5 Huang et al. [26]
b
96.1 98.1 Needs beat detection, has larger memory consumption, has larger transition delay
6 Lee et al. [21]
b
98.2 97.7 Needs beat detection, has higher sensitivity to the selection of parameter values
7 Couceiro et al.
[22]
c
93.8
(96.6)
96.1
(82.7)
Needs beat detection, has lower performance
8 Babaeizadeh et al.
[23]
c
92
(87.3)
(95.5) Needs beat detection, has lower performance
9 Jiang et al. [24]
c
98.2 97.5 Needs beat detection, has larger memory consumption, needs longer data segment to guarantee high performance,
has lower AF episode detection rate
10 proposed method 97.0 97.1
a
Based on the absence of P-wave.
b
Based on R–R irregularity.
c
Based on combination of R–R irregularity and absence of P-wave.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142140
achieve almost the same level of performance with short data
segment of 10 s, and hence has a better speed of diagnosis.
The long data segment requirement by the existing methods
also results in missing detection of some of the short duration
AF episodes. For example, while Jiang [24] method was only
able to detect 260 out of 291 AF episodes of MIT-BIH AFIB
database, our proposed method detected 271 episodes includ-
ing short AF episodes.
Another advantage of our proposed method is that it does not
require pre-processing step of peak detection. The accuracy of
AF detection for other methods such as Dash [20] and Lee [21]
depends on the accuracy of beat detection. Therefore, if the R
peak detection/annotation is not correct, the algorithm would
not be able to correctly identify the AF episodes. In fact, due to
this reason, the authors in Dash [20] and Lee [21] had to
exclude 2 data files (files 4936 and 5091) from MIT-BIH AFIB
database while evaluating their proposed method. These two
files had wrong R peak annotations and with those wrongly
located R peaks, their method could have not been able to
perform well. In contrast, our proposed method does not rely
on R peak locations for AF detection, so we did not need to
exclude those files in our analysis.
A summary of weakness points of existing algorithms relative
to the proposed method are presented in Table 2.
4.5. Comparison of transition time and computation time of the proposed
method with the existing Methods
The median transition delay of our proposed method was 9.8 s.
This result outperforms those of two recently published algo-
rithms: 18 beats in Dash [20] and 70 beats Huang [26]. In Lee [40],
the authors were able to decrease the transition delay to only
6 beats by using a data segment of 12 beats at the cost of having a
lower accuracy rate of 92%. However, our proposed method is able
to achieve the accuracy of 96.4% with transition delay of 9.8 s.
Finally, our results indicated that the computation time of the
proposed method for segment length of 30 s or less, is less than
40 ms. The most time-consuming part of our algorithm (more than
90% of running time) pertains to calculation of stationary wavelet
coefficients. Nevertheless, a computation time of 40 ms enables our
proposed method to be realizable in practical applications. In general,
the computational time of an AF detection method is highly
dependent on the software/hardware used for the algorithm imple-
mentation/evaluation, and the length of the data segment used [25].
Many of the existing algorithms have reported their computation
time for the segment length resulting in the highest accuracy of
detection. For example, for a data segment of 128 beats, the
computation times of Lee [21] and Dash [20] methods are less than
30 ms and 200 ms, respectively. It should be noted that since R peak
annotations have been already provided in MIT-BIH AFIB database,
the reported computation time of many existing algorithms do not
include the computation time required to obtain the location of the R
peaks. The average computation time of R peak detection using a
well-known algorithm such as that of Pan and Tompkin [41] for data
segment of 100 s is about 40 ms. Considering this additional compu-
tation time, our algorithm runs substantially faster than many
existing algorithms.
4.6. Potential limitation of the study
The conversion of the original beat-to-beat annotation to a
T-second resolution (by using a minimum percentage parameter
P) limited the duration of the shortest detectable AF episode to
TP=100 . For example, with T¼10 and P¼50%, the shortest
detectable AF will be 5 s. One solution to address this issue is to
decrease the value of Tto less than 10 s. However, such investiga-
tion requires a computer with a processing power higher than the
one used in this study (in order to train support vector machine
with the larger number of resulted data points).
In general, AF detection algorithms need near 100% sensitivity
to be clinically acceptable; otherwise there will be some FN cases
that cannot be correctly diagnosed and intervened in a timely
manner. Although our proposed method performed well in AF
detection (sensitivity of 97.0% and specificity of 97.1%), it demon-
strated a specificity of 86.3% at sensitivity of 100%. Thus, further
enhancement of the proposed method in terms of specificity may
be needed to make it a clinically acceptable method.
The proposed AF detection method applies support vector
machine on the ECG features (peak-to-average power ratio and
entropy) extracted over different frequency bands using a stationary
wavelet transform. The method does not dynamically choose the
most effective wavelet scale for the purpose of denoising. Hence,
the possibility of further enhancement of the method using a multi-
resolution analysis remains to be investigated.
The validation of the proposed method was conducted on
MIT_BIH atrial fibrillation dataset using two-fold cross validation.
Although this dataset is the most popular publicly available data
set for AF detection, it includes a limited group of patients. A larger
dataset with more diverse arrhythmia patterns could allow for a
more comprehensive analysis of the proposed method.
5. Conclusion
The current work proposes a novel AF detection method using
stationary wavelet transform and support vector machine. This
method eliminates the need for P and/or R peak detection; a pre-
processing step required by many of the existing algorithms. Compar-
ison of the performance of the proposed method (on MIT-BIH Atrial
Fibrillation database) with other major algorithms showed that our
method achieving the sensitivity and specificity of 97.0% and 97.1%,
respectively outperforms many of existing methods. Moreover, the
proposed method holds interesting properties (e.g. high performance
for data segments as short as 10 s, reasonably low computation and
transition times, high accuracy regardless of the choice of the
parameters of the method) which make it a suitable choice for various
practical applications.
Conflict of interest statement
None declared.
Acknowledgment
The authors would like to thank “Yong Bai”from UCLA for his
assistance with implementation of Multi-scale Entropy.
References
[1] M. Lainscak, N. Dagres, G.S. Filippatos, S.D. Anker, D.T. Kremastinos, Atrial
fibrillation in chronic non-cardiac disease: where do we stand? Int. J. Cardiol.
128 (2008) 311–315.
[2] J. Heeringa, D.A. van der Kuip, A. Hofman, J.A. Kors, G. van Herpen,
B.H. Stricker, T. Stijnen, G.Y. Lip, J.C. Witteman, Prevalence, incidence and
lifetime risk of atrial fibrillation: the Rotterdam study, Eur. Heart J. 27 (2006)
949–953.
[3] A.S. Go, E.M. Hylek, K.A. Phillips, Y. Chang, L.E. Henault, J.V. Selby, D.E. Singer,
Prevalence of diagnosed atrial fibrillation in adults: national implications for
rhythm management and stroke prevention: the Anticoagulation and risk
factors in atrial fibrillation (ATRIA) study, J. Am. Med. Assoc. 285 (2001)
2370–2375.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142 141
[4] C.D. Furberg, B.M. Psaty, T.A. Manolio, J.M. Gardin, V.E. Smith, P.M. Rautaharju,
Prevalence of atrial fibrillation in elderly subjects (the cardiovascular health
study), Am. J. Cardiol. 74 (1994) 236–241.
[5] P.A. Wolf, R.D. Abbott, W.B. Kannel, Atrial fibrillation as an independent risk
factor for stroke: the Framingham Study, Stroke J. Cereb. Circ. 22 (1991)
983–988.
[6] E.J. Benjamin, P.A. Wolf, R.B. D'Agostino, H. Silbershatz, W.B. Kannel, D. Levy,
Impact of atrial fibrillation on the risk of death: the Framingham heart study,
Circulation 98 (1998) 946–952.
[7] C.A. Sanoski, Clinical, economic, and quality of life impact of atrial fibrillation,
J. Manag. Care Pharm. 15 (2009) S4–S9.
[8] V. Markides, R.J. Schilling, Atrial fibrillation: classification, pathophysiology,
mechanisms and drug treatment, Heart (Br. Card. Soc.) 89 (2003) 939–943.
[9] K. Harris, D. Edwards, J. Mant, How can we best detect atrial fibrillation? J. R. Coll.
Physicians Edinb. 42 (Suppl. 18) (2012) S5–S22.
[10] A.J.Camm,P.Kirchhof,G.Y.Lip,U.Schotten,I.Savelieva,S.Ernst,I.C.Van
Gelder,N.Al-Attar,G.Hindricks,B.Prendergast,H.Heidbuchel,O.Alfieri,
A. Angelini, D. Atar, P. Colonna, R. De Caterina, J. De Sutter, A. Goette,
B. Gorenek, M. Heldal, S.H. Hohloser, P. Kolh, J.Y. Le Heuzey,
P. Ponikowski, F.H. Rutten, Guidelines for the management of atrial
fibrillation: the task force for the management of atrial fibrillation of
the european society of cardiology (ESC), Eur. Heart J. 31 (2010)
2369–2429.
[11] J.R. Mehall, R.M. Kohut Jr., E.W. Schneeberger, W.H. Merrill, R.K. Wolf, Absence
of correlation between symptoms and rhythm in symptomatic atrial fibrilla-
tion, Ann. Thorac. Surg. 83 (2007) 2118–2121.
[12] J. Mant, D.A. Fitzmaurice, F.D. Hobbs, S. Jowett, E.T. Murray, R. Holder,
M. Davies, G.Y. Lip, Accuracy of diagnosing atrial fibrillation on electrocardio-
gram by primary care practitioners and interpretative diagnostic software:
analysis of data from screening for atrial fibrillation in the elderly (SAFE) trial,
Br. Med. J. (Clin. Res. Ed.) 335 (2007) 380.
[13] K. Aytemir, S. Aksoyek, A. Yildirir, N. Ozer, A. Oto, Prediction of atrial
fibrillation recurrence after cardioversion by P wave signal-averaged electro-
cardiography, Int. J. Cardiol. 70 (1999) 15–21.
[14] J. Slocum, A. Sahakian, S. Swiryn, Diagnosis of atrial fibrillation from surface
electrocardiograms based on computer-detected atrial activity, J. Electrocar-
diol. 25 (1992) 1–8.
[15] L. Clavier, J.M. Boucher, R. Lepage, J.J. Blanc, J.C. Cornily, Automatic P-wave
analysis of patients prone to atrial fibrillation, Med. Biol. Eng. Comput. 40
(2002) 63–71.
[16] I. Dotsinsky, Atrial wave detection algorithm for discovery of some rhythm
abnormalities, Physiol. Meas. 28 (2007) 595–610.
[17] G.B. Moody, R.G. Mark, A new method for detecting atrial fibrillation using R–R
intervals, Comput. Cardiol. (1983) 227–230.
[18] K. Tateno, L. Glass, Automatic detection of atrial fibrillation using the
coefficient of variation and density histograms of RR and deltaRR intervals,
Med. Biol. Eng. Comput. 39 (2001) 664–671.
[19] S. Sarkar, D. Ritscher, R. Mehra, A detector for a chronic implantable atrial
tachyarrhythmia monitor, IEEE Trans. Bio-Med. Eng. 55 (2008) 1219–1224.
[20] S. Dash, K.H. Chon, S. Lu, E.A. Raeder, Automatic real time detection of atrial
fibrillation, Ann. Biomed. Eng. 37 (2009) 1701–1709 .
[21] J. Lee, Y. Nam, D.D. McManus, K.H. Chon, Time-varying coherence function for
atrial fibrillation detection, IEEE Trans. Bio-Med. Eng. 60 (2013) 2783–2793.
[22] R. Couceiro, P. Carvalho, J. Henriques, M. Antunes, M. Harris, J. Habetha,
Detection of atrial fibrillation using model-based ECG analysis, in: Proceedings
of the 19th International Conference on Pattern Recognition, 2008. ICPR 2008,
2008, pp. 1–5.
[23] S. Babaeizadeh, R.E. Gregg, E.D. Helfenbein, J.M. Lindauer, S.H. Zhou, Improve-
ments in atrial fibrillation detection for real-time monitoring, J. Electrocardiol.
42 (2009) 522–526.
[24] K. Jiang, C. Huang, S.M. Ye, H. Chen, High accuracy in automatic detection of
atrial fibrillation for Holter monitoring, J. Zhejiang Univ. Sci. B 13 (2012)
751–756.
[25] N. Larburu, T. Lopetegi, I. Romero, Comparative study of algorithms for atrial
fibrillation detection, Comput. Cardiol. (2011) 265–268.
[26] C. Huang, S. Ye, H. Chen, D. Li, F. He, Y. Tu, A novel method for detection of the
transition between atrial fibrillation and sinus rhythm, IEEE Trans. Bio-Med.
Eng. 58 (2011) 1113–1119.
[27] P.R. Gomes, F.O. Soares, J.H. Correia, C.S. Lima, ECG Data-Acquisition and
classification system by using wavelet-domain Hidden Markov Models, in:
Proceedings of the Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, 2010, 2010 pp. 4670–4673.
[28] P.S. Addison, Wavelet transforms and the ECG: a review, Physiol. Meas. 26
(2005) R155–R199.
[29] B. Weng, J.J. Wang, F. Michaud, M. Blanco-Velasco, Atrial fibrillation detection
using stationary wavelet transform analysis, in: Proceedings of the Annual
International Conference of the IEEE Engineering in Medicine and Biology
Society, 2008, 2008 pp. 1128–1131.
[30] R. Alcaraz, C. Vaya, R. Cervigon, C. Sanchez, J.J. Rieta, Wavelet sample entropy:
a new approach to predict termination of atrial fibrillation, Comput. Cardiol.
2006 (2006) 597–600.
[31] Q. Shaoyu, Z. Ping, Wavelet and wavelet packet transform analysis in the ECG
signals of atrial fibrillation, in: Proceedings of the Complex Medical Engineer-
ing, 2007, CME 2007, IEEE/ICME International Conference on, 2007, pp. 1766–
1769.
[32] S. Karatsiolis, C.N. Schizas, Region based support vector machine algorithm for
medical diagnosis on Pima Indian Diabetes dataset, in: Proceedings of the
2012 IEEE 12th International Conference on, Bioinformatics & Bioengineering
(BIBE), 2012, pp. 139–144 .
[33] I. El-Naqa, Y. Yang, M.N. Wernick, N.P. Galatsanos, R.M. Nishikawa, A support
vector machine approach for detection of microcalcifications, Med. Imaging
IEEE Trans. 21 (2002) 1552–1563.
[34] S. Gao, H. Li, Breast cancer diagnosis based on support vector machine, in:
Proceeedings of the Uncertainty Reasoning and Knowledge Engineering
(URKE), 2012 2nd International Conference on, IEEE, 2012, pp. 240–243.
[35] C. Campbell, Kernel methods: a survey of current techniques, Neurocomputing
48 (2002) 63–84.
[36] N. Cristianini, J. Shawe-Taylor, An Introduction to Support Vector Machines
and Other Kernel-Based Learning Methods, Cambridge university press,
Cambridge, United Kingdom, 2000.
[37] B.E. Boser, I.M. Guyon, V.N. Vapnik, A training algorithm for optimal margin
classifiers, in: Proceedings of the fifth annual workshop on Computational
learning theory, ACM, 1992, pp. 144–152.
[38] A.L. Goldberger, L.A. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark,
J.E. Mietus, G.B. Moody, C.K. Peng, H.E. Stanley, PhysioBank, PhysioToolkit, and
PhysioNet: components of a new research resource for complex physiologic
signals, Circulation 101 (2000) E215–E220.
[39] M. Stridh, L. Sornmo, C.J. Meurling, S.B. Olsson, Sequential characterization of
atrial tachyarrhythmias based on ECG time-frequency analysis, IEEE Trans.
Bio-Med. Eng. 51 (2004) 100–114.
[40] D.-S. Kim, J.-S. Lee, Propagation and attenuation characteristics of various
ground vibrations, Soil Dyn. Earthq. Eng. 19 (2000) 115–126.
[41] J. Pan, W.J. Tompkins, A real-time QRS detection algorithm, Biomed. Eng. IEEE
Trans. (1985) 230–236.
S. Asgari et al. / Computers in Biology and Medicine 60 (2015) 132–142142