Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy on Jan 19, 2014
Content may be subject to copyright.
Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy on Jan 19, 2014
Content may be subject to copyright.
Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy
Content may be subject to copyright.
Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy
Content may be subject to copyright.
Obstructive Sleep Apnea Detection Using
SVM-Based Classification of ECG Signal Features
Laiali Almazaydeh, Khaled Elleithy, Miad Faezipour
Department of Computer Science and Engineering
University of Bridgeport
Bridgeport, CT 06604, USA
{lalmazay, elleithy, mfaezipo}@bridgeport.edu
Abstract— Sleep apnea is the instance when one either has
pauses of breathing in their sleep, or has very low breath while
asleep. This pause in breathing can range in frequency and
duration. Obstructive sleep apnea (OSA) is the common form
of sleep apnea, which is currently tested through
polysomnography (PSG) at sleep labs. PSG is both expensive
and inconvenient as an expert human observer is required to
work over night. New techniques for sleep apnea classification
are being developed by bioengineers for most comfortable and
timely detection. In this paper, an automated classification
algorithm is presented which processes short duration epochs
of the electrocardiogram (ECG) data. The classification
technique is based on support vector machines (SVM) and has
been trained and tested on sleep apnea recordings from
subjects with and without OSA. The results show that our
automated classification system can recognize epochs of sleep
disorders with a high degree of accuracy, approximately
96.5%. Moreover, the system we developed can be used as a
basis for future development of a tool for OSA screening.
Keywords: Sleep apnea, PSG, ECG, RR interval, feature
extraction, SVM.
I. INTRODUCTION
A. Background
Over the average lifespan, humans sleep for about 1/3 of
their lives. A sleeping disorder is when one cannot sleep,
causing the body to lose function. Just as the body’s benefits
of rest can range from physical to emotional and
psychological effects, lack of sleep can damage the body
physically, emotionally and psychologically. Till date, 84
kinds of sleep disorders have been discovered, including the
most common ones such as insomnia, sleep apnea,
narcolepsy and restless leg syndrome [1].
Sleep Apnea (SA) is the instance when one either has
pauses of breathing in their sleep, or has very low breath
while asleep. This pause in breathing is known as an apnea,
and can range in frequency and duration. The lack of
breathing during sleep is known as a hypopnea [2]. Sleep
apnea is classified into two different types. The first type is
Obstructive Sleep Apnea (OSA), which is more common,
occurring in 2% to 4% of middle-aged adults and 1% to 3%
of preschool children [3], and is generally caused by a
collapse of the upper respiratory airway. The second one is
Central Sleep Apnea (CSA), which is caused by an absent or
inhibited respiratory drive. Most cases of CSA are mixed,
meaning that it is often along with OSA cases, and the CSA
only cases appear exceedingly rarely [4]. Despite how
common it is, most cases go undiagnosed and can be
attributed to 70 billion dollars loss, 11.1 billion in damages
and 980 deaths each year [5].
Most sleep apnea cases go undiagnosed because of the
inconvenience, expenses and unavailability of testing. The
traditional testing process includes a polysomnography
(PSG), which is a standard procedure for all sleep disorder
diagnosis. It records the breath airflow, respiratory
movement, oxygen saturation, body position,
electroencephalogram (EEG), electrooculogram (EOG),
electromyogram (EMG), and electrocardiogram (ECG) [6].
B. Contribution and Paper Organization
It is clear that the mere dependency on PSG needs to be
taken away from the laboratory for simpler detection and
faster treatment of sleep apnea. Instead, automated, at-home
devices that patients can simply use while asleep seem to be
very attractive and highly on-demand. We propose a novel
methodology in this paper that combines most effective RR-
interval based features of the ECG signal based on the ones
suggested by Chazal et al., and Yilmaz et al. This work
relies on SVM for classification. Performance assessment of
the combination of these two approaches is done by
measuring the classification performance in determining the
presence of apnea for different epoch lengths.
The rest of this paper is organized as follows. In Section
II, we glance at a variety of sleep apnea detection methods.
Section III contains an overview of the system, and details
on the analysis methodology of the paper. We describe the
steps to determine RR-interval and features extraction for
different epoch lengths in the same Section. In Section IV,
we present the results of our system, and then we provide a
comparison with other SA detection works. Finally, Section
V concludes this paper regarding the potential usefulness of
our system, and highlights some directions for future
research.
II. RELATED WORK
Several methods have been suggested for identification
of sleep apnea over the past few years. Statistical features of
different signals such as nasal air flow, the thorax and
abdomen effort signals, acoustic speech signal, oxygen
saturation, electrical activity of the brain (EEG), and
electrical activity of the heart (ECG) are commonly used in
the detection.
Ng et al. [7] showed that thoracic and the abdominal
signals were good parameters for the identification of the
CONFIDENTIAL. Limited circulation. For review only.
Preprint submitted to 34th Annual International IEEE EMBS Conference.
Received March 29, 2012.
occurrence of sleep apnea. Using the mean of absolute
amplitudes of the thoracic and the abdominal signals, they
have achieved a good performance with a receiver operating
characteristic value higher than 80%.
Depending on the hypothesis that speech signal
properties of OSA patients will be different than those not
having OSA, Goldshtein et al. [8] developed a gaussian
mixture model-based system to classify between the OSA
and non-OSA groups. They achieved a specificity and
sensitivity of 83% and 79% for the male OSA and 86% and
84% for the female OSA patients, respectively. Their
technique relied on vocal tract length and linear prediction
coefficients features.
The study in [9] assessed analysis of a comprehensive
feature set based on blood oxygen saturation (SaO
2
) from
nocturnal oximetry in order to evaluate sleep quality. The
three features of SaO
2
signal which are delta index, central
tendency measure and oxygen desaturation index are
evaluated. Central tendency measure accuracy was higher
than those provided by delta index and oxygen desaturation
index. With central tendency measure the sensitivity was
90.1% and the specificity was 82.9%.
The relationship between periodic changes in the
oxygen saturation (SaO2) profile and in the EEG pattern due
to apnea events during the night was investigated in [10].
The spectral analysis of these two signals achieved 91%
sensitivity, 83.3% specificity and 88.5% accuracy in OSA
diagnosis.
In [11], the authors analyze various feature sets and a
combination of classifiers based on the arterial oxygen
saturation signal measured by pulse oximetry (SpO
2
) and the
ECG in order to evaluate sleep quality and detect apnea.
With selected features of the SpO2 and ECG signals, the
Bagging with REP Tree classifier achieved sensitivity of
79.75%, specificity of 85.89% and overall accuracy of
84.40%.
Wavelet transforms and an artificial neural network
(ANN) algorithm were applied to the EEG signal in [12] to
find a solution to the problem of identifying sleep apnea
episodes. The system's identification results achieved a
sensitivity of approximately 69.64% and a specificity of
approximately 44.44%.
Many studies show that detection of obstructive sleep
apnea can be performed through heart rate variability (HRV)
and the ECG signal.
Quiceno-Manrique et al. [13] proposed a simple
diagnostic tool for OSA with a high accuracy (up to 92.67%)
using time-frequency distributions and dynamic features in
ECG signal. Moreover, based on spectral components of
heart rate variability, frequency analysis was performed in
[14] using Fourier and Wavelet Transformation with
appropriate application of the Hilbert Transform, where the
sensitivity was 90.8%. In addition, in [15] a bivariate
autoregressive model was used to evaluate beat-by-beat
power spectral density of HRV and R peak area, where the
classification results showed accuracy higher than 85%. The
technique in this work also relies on features of the ECG
signal.
III. METHODOLOGY
In this work, we focus on the ECG signal features to
detect sleep apnea. The block diagram of the overall
methodology used in this study is shown in Figure 1.
A. Subjects
The database of ECG signals used is available from the
PhysioNet web site [16]. The Apnea-ECG Database contains
70 recordings, containing a single ECG signal varying in
length from slightly less than 7 hours to nearly 10 hours
each. The sampling frequency used for ECG acquisition was
100 Hz, with 16-bit resolution, and one sample bit
representing 5µV. The standard sleep laboratory ECG
electrode positions were used (modified lead V2) [6].
Figure 1. Schematic diagram of the system.
B. ECG
ECG is considered as one of the most efficient features to
detect sleep disorders. Cyclic variations in the duration of a
heartbeat, also known as RR intervals (time interval from
one R wave to next R wave) of ECG have been reported to
be associated with sleep apnea episodes. This consists of
bradycardia during apnea followed by tachycardia upon its
cessation [6]. RR-interval is defined as the time interval
between two consecutive R peaks. The RR interval time
series generated for each ECG beat can be written as follows
[17]:
)1(1,...,2,1),()1()( niiririrr
Several researches have been conducted to recognize sleep
apnea using the features derived from the RR interval such as
median, mean, inter-quartile range (IQR), and the standard
deviation of the change in RR intervals [6][17][18].
C. Data Preparation
To select the data, we chose the ECG records which
have continuous apnea data for a certain period of time,
followed by a regular (normal) data representation for a
period of time, or vice versa. The data preparation is used
for training the SVM classifier (see subsection III.G).
The next step in our procedure after data selection is data
partitioning. In our work, three cases of partitioning were
analyzed, as follows:
ECG signal
Data Preparation
RR Interval Detection
Features Extraction
Data Randomization
SVM
Performance Evaluation
CONFIDENTIAL. Limited circulation. For review only.
Preprint submitted to 34th Annual International IEEE EMBS Conference.
Received March 29, 2012.
Case 1. The apnea and regular data are partitioned into
10 second pieces.
Case 2. The apnea and regular data are partitioned into
15 second pieces.
Case 3. The apnea and regular data are partitioned into
epochs of 30 second pieces.
Since apnea is defined as a pause in breathing, and can
last from a few seconds to minutes (almost >=10 sec); we
investigate the three above cases to determine the best
accuracy that can be achieved.
Figure 2. Detection of R- Peak.
D. RR Interval Detection
We need to distinguish the R waves from the other
waves of the ECG signal. Therefore, we developed the
following two conditions, in which R-peak was detected. An
R peak will be identified if both conditions 1 and 2 are
satisfied:
1) It has to be a local maximum, which is detected by a
local max function within a window of 150ms.
2) The local max peaks must be at least 2 standard
deviation above the mean.
Once the R-peak was determined, RR intervals were
computed. The RR interval is the peak to peak time period
from two continuous peak signals as shown in Equation 1.
Figure 2 shows the detection of R-peaks.
E. Features Extraction
Our technique relies on an effective combination of
ECG signal features which is a novel hybrid of features
extracted from [6] and [19]. The following ECG features
which are most effective for apnea detection are calculated:
Mean epoch and recording RR-interval.
Standard deviation of the epoch and recording RR-
interval.
The NN50 measure (variant 1), defined as the number
of pairs of adjacent RR- intervals where the first RR-
interval exceeds the second RR- interval by more than
50 ms.
The NN50 measure (variant 2), defined as the number
of pairs of adjacent RR-intervals where the second RR-
interval exceeds the first RR interval by more than 50
ms.
Two pNN50 measures, defined as each NN50 measure
divided by the total number of RR-intervals.
The SDSD measures, defined as the standard deviation
of the differences between adjacent RR- intervals.
The RMSSD measures, defined as the square root of the
mean of the sum of the squares of differences between
adjacent RR- intervals.
Median of RR-intervals.
Inter-quartile range, defined as difference between 75
th
and 25
th
percentiles of the RR-interval value
distribution.
Mean absolute deviation values, defined as mean of
absolute values obtained by the subtraction of the mean
RR-interval values from all the RR-interval values in an
epoch.
The first seven features are proposed by Chazal et al.
[6], while the three latter feature are proposed by Yilmaz et
al.[19], who claimed that RR interval mean, standard
deviation, and range are sensitive to outliers, and thus
classification performance deteriorates when only these
features are included.
Our hybrid technique includes a combination of the most
effective set of RR-interval based features of the ECG signal
for classification. The classification results confirm the
improved accuracy compared to the two above techniques.
F. Support Vector Machines
We use Support Vector Machines (SVM) as a
classification (also known as supervised learning) method in
order to investigate apneaic epoch detection. In our
implementation, we use a linear kernel function to map the
training data into kernel space. In the optimization process,
we use a method called sequential minimal optimization to
find the separating hyperplane.
For data randomization, we separate the apnea and non
apnea data. We then separate training data and testing data,
with 80% for the training and 20% for the testing. After the
signals are separated, we perform the training for SVM.
IV. RESULTS
A. Performance Evaluation
We evaluated the effectiveness of our model on the
different records in the Apnea-ECG database. MATLAB
toolset was used for signal processing and classification.
Two statistical indicators, Sensitivity (Se) and
Specificity (Sp) in addition to the Accuracy (Acc) have been
used to evaluate the performance of our classification
system. Table I, II and III show the classification results for
the three cases mentioned in the data partitioning step. Our
model was based on a linear kernel SVM using various RR-
interval features of the ECG signal. The three cases used
here are:
seconds. The accuracy of our approach is 86.1%,
96.5%, and 95%, respectively. From Table II, SVM with
linear kernel using 15 second epochs shows the best
classification accuracy with high successful rate of correct
prediction.
Input\Output
Regular
Apnea
Regular
97.2%
2.78%
Apnea
25%
75%
CONFIDENTIAL. Limited circulation. For review only.
Preprint submitted to 34th Annual International IEEE EMBS Conference.
Received March 29, 2012.
Input\Output
Regular
Apnea
Regular
100%
0%
Apnea
7.1%
92.9%
Input\Output
Regular
Apnea
Regular
100%
0%
Apnea
10%
90%
B. Comparison with other techniques
We performed a comparison with other SA detection
works. Table IV represents comparative results. As can be
seen, our system has achieved a comparable or better
performance.
Q-Manrique et
al. [13]
V. CONCLUSIONS AND FUTURE WORKS
In this work, we studied the possibility of the detection of
sleep apnea or hypopnea events from the ECG signal
variation patterns during sleep. We further developed a
model using the ECG signal features and evaluated its
effectiveness. We evaluated our model on three different
epoch lengths. From the experimental results, we conclude
that SVM with linear kernel shows the best accuracy with 15
second epoch length.
As a future work, we plan to do performance
optimization for feature selection, and then incorporate this
work into a real- time monitoring system that acquires and
analyzes the ECG signal of subjects during sleep.
REFERENCES
[1] Sleep Disorder Overview. www.neurologychannel.com.
[2] S. Isa, M. Fanany, W. Jatmiko and A. Arymurthy, “Sleep Apnea
Detection from ECG Signal, Analysis on Optimal Features, Principal
Components, and Nonlinearity,” in Proceedings of the 5
th
IEEE
International Conference on Bioinformatics and Biomedical
Engineering (iCBBE), pp. 1-4, May 2011.
[3] Sleep Apnea: What Is Sleep Apnea? www.nhlbi.nih.gov.
[4] Apnea guide. www.apneaguide.com.
[5] Treating Sleep Apnea Could Cut Road Deaths.
www.americanvoiceinstitute.org.
[6] P. Chazal, T. Penzel and C. Heneghan,“Automated Detection of
Obstructive Sleep Apnoeaa at Different Time Scales Using the
Electrocardiogram,” Institute of Physics Publishing, vol. 25, no. 4, pp.
967-983, Aug. 2004.
[7] A. Ng, J. Chung, M. Gohel, W. Yu, K. Fan and T. Wong, “Evaluation
of the Performance of Using Mean Absolute Amplitude Analysis of
Thoracic and Abdominal Signals for Immediate Indication of Sleep
Apnoea Events,” Journal of Clinical Nursing, vol. 17, no. 17, pp.
2360-2366, Sep. 2008.
[8] E. Goldshtein, A. Tarasiuk and Y. Zigel, “Automatic Detection of
Obstructive Sleep Apnea Using Speech Signals,” in IEEE
Transactions on Biomedical Engineering, vol. 58, no. 5, pp. 1373-
1382, May. 2011.
[9] D. Alvarez, R. Hornero, D. Abasolo, F. Campo and C. Zamarron,
“Nonlinear Characteristics of Blood Oxygen Saturation from
Nocturnal Oximetry for Obstructive Sleep Apnoea Detection,”
Institute of Physics Publishing, vol. 27, no. 4, pp. 399-412, Apr. 2006.
[10] D. Avarez, R. Hornero, J. Marcos, F. Campo and M. Lopez, “Spectral
Analysis of Electroencephalogram and Oximetric Signals in
Obstructive Sleep Apnea Diagnosis,” in Proceedings of the 31
st
IEEE
International Conference on Engineering in Medicine and Biology
Society (EMBS 2009), pp. 400-403, Sep. 2009.
[11] B. Xie, H. Minn, “Real Time Sleep Apnea Detection by Classifier
Combination,” in IEEE Transactions on Information Technology in
Biomedicine (in Press), 2012.
[12] R. Lin, R. Lee, C. Tseng, H. Zhou, C. Chao, J. Jiang,“A New
Approach for Identifying Sleep Apnea Syndrome Using Wavelet
Transform and Neural Networks,” Biomedical Engineering:
Applications, Basis & Communications, vol. 18, no. 3, pp. 138-143,
2006.
[13] Q. Manrique, A. Hernandez, T. Gonzalez, F. Pallester and C.
Dominquez, “ Detection of Obstructive Sleep Apnea in ECG
Recordings Using Time-Frequency Distributions and Dynamic
Features,” in Proceedings of the IEEE International Conference on
Engineering in Medicine and Biology Society( EMBS 2009), pp.
5559-5562, Sep. 2009.
[14] M. Schrader, C. Zywietz, V. Einem, B. Widiger, G. Joseph,
“Detection of Sleep Apnea in Single Channel ECGs from the
PhysioNet Data Base,” Computers in Cardiology 2000, vol. 27, pp.
263-266, Sep. 2000.
[15] M. Mendez, D. Ruini, O. Villantieri, M. Matteucci, T. Penzel and A.
Bianchi, “Detection of Sleep Apnea from Surface ECG Based on
Features Extracted by an Autoregressive Model, ” in Proceedings of
the IEEE International Conference on Engineering in Medicine and
Biology Society (EMBS 2007), pp. 6105-6108, Aug. 2007.
[16] PhysioNet, www.physionet.org.
[17] S. Isa, M. Fanany, W. Jatmiko and A. Murini, “Feature and Model
Selection on Automatic Sleep Apnea Detection Using ECG,” in
International Conference on Computer Science and Information
Systems, ICACSIS 2010, pp. 357-362, 2010.
[18] P. Langley, E. Bowers and A. Murray, “Principal Component
Analysis as Tool for Analyzing Beat-to-Beat Changes in ECG
Features: Application To ECG-Derived Respiration,” in IEEE
Transactions on Biomedical Engineering, vol. 57, no. 4, pp. 821-
829, Apr. 2010.
[19] B. Yilmaz, M. Asyali, E. Arikan, S. Yektin and F. Ozgen, “Sleep
Stage and Obstructive Apneaic Epoch Classification Using Single-
Lead ECG,” in Biomedical Engineering Online, vol. 9, 2010.
CONFIDENTIAL. Limited circulation. For review only.
Preprint submitted to 34th Annual International IEEE EMBS Conference.
Received March 29, 2012.
