Content uploaded by Khaled Elleithy
Author content
All content in this area was uploaded by Khaled Elleithy
Content may be subject to copyright.
Apnea Detection Based on
Respiratory Signal Classification
Laiali Almazaydeh, Khaled Elleithy, Miad Faezipour and Ahmad Abushakra
Department of Computer Science and Engineering
University of Bridgeport
Bridgeport, CT 06604, USA
{lalmazay, elleithy, mfaezipo, aabushak}@bridgeport.edu
Abstract— Obstructive sleep apnea (OSA) is the most common
form of different types of sleep-related breathing disorders. It is
characterized by repetitive cessations of respiratory flow during
sleep, which occurs due to a collapse of the upper respiratory
airway. OSA is majorly undiagnosed due to the inconvenient
Polysomnography (PSG) testing procedure at sleep labs. This
paper introduces an automated approach towards identifying the
presence of sleep apnea based on the acoustic signal of
respiration. The characterization of breathing sound was carried
by Voice Activity Detection (VAD) algorithm, which is used to
measure the energy of the acoustic respiratory signal during
breath and breath hold. The performance of our classification
algorithm is tested on real respiratory signals and the
experimental results show that the VAD is useful as a predictive
tool for the segmentation of breath into sound and silence
segments. Moreover, the system we developed can be used as a
basis for future development of a tool for OSA screening.
Keywords: sleep apnea, OSA, PSG, VAD, respiration signal.
I. INTRODUCTION
Sleep apnea (SA) in the form of Obstructive sleep apnea
(OSA) is becoming the most common respiratory disorder
during sleep, which is characterized by cessations of airflow to
the lungs. These cessations in breathing must last more than 10
seconds to be considered an apnea event. Apnea events may
occur 5 to 30 times an hour and may occur up to four hundred
times per night in those with severe SA [1].
The most frequent night symptoms of SA can include
snoring, nocturnal arousals, sweating and restless sleep.
Moreover, like all sleeping disorders, symptoms of sleep apnea
do not occur just during the night. Daytime symptoms can
range from morning headaches, depression, impaired
concentration and excessive sleepiness which cause mortality
from traffic and industrial accidents. However, these symptoms
are not definitive to detect SA syndrome [2] [3].
In fact, SA is not a problem to be taken lightly, since it is
associated with a major risk factor of health implications and
increased cardiovascular disease and sudden death. It has been
linked to irritability, depression, sexual dysfunction, high blood
pressure (hypertension), learning and memory difficulties, in
addition to stroke and heart attack [2] [3]. Several treatment
options for OSA patients include weight loss, positional
therapy, oral appliances, surgical procedures and continuous
positive airway pressure (CPAP). CPAP is a common and
effective treatment especially for patients with moderate to
severe OSA. CPAP devices are masks worn during sleep that
improves oxygen saturation and reduces sleep fragmentation
[4].
Statistics show that around 100 million people worldwide,
and in the US from 18 to 50 million people, are suspected to
have OSA. This is while more than 80% of which remain
undiagnosed [5]. The trouble of having examinations
discourages patients prone to OSA undergo at the overnight
clinical research through polysomnographic data.
Polysomnography (PSG) is a complicated procedure and
certain way of assessing the OSA problem. Complete PSG
includes the monitoring of the breath airflow, respiratory
movement, oxygen saturation (SpO
2
), body position,
electroencephalography (EEG), electromyography (EMG),
electrooculography (EOG), and electrocardiography (ECG) [6].
However, PSG has received many criticisms from some
researchers. This is due to several reasons, including first, the
inconvenience since it requires the patient to be connected to
numerous sensors and to stay in hospital for one night. Second,
it is expensive. The average cost for a PSG is $2,625 due to the
need for the study to take place in a specially equipped medical
facility, in addition to the requirement of having a sleep lab
staff overnight, trained in ‘scoring’ the resultant measurements
manually. Third, a long wait list of up to 6 months is caused by
limited availability of PSG [7].
According to the American Academy of Sleep Medicine
(AASM), the Apnea-Hypopnea Index (AHI) is used to describe
the number of complete and partial apnea events per hour of
sleep and it is calculated to assess OSA syndrome severity.
OSA severity is usually determined as follows: AHI 5-15
indicates mild, 15-30 indicates moderate and over 30 indicates
severe OSA syndrome. Therefore, patients are diagnosed with
OSA if they have five or more apnea events per hour of sleep
during a full night sleep period [8].
However, new simplified methods for diagnosis and
screening of OSA are needed, in order to have a major benefit
of the treatment on OSA outcomes. In this work, we develop an
efficient algorithm for automatic classification of respiratory
signal to detect abnormalities in breathing or breathing
cessations. We use Voice Activity Detection (VAD) to classify
respiratory signals into normal respiration and sleep apnea. The
detection process would check for apnea attack for a time
period of fifteen seconds or more.
In the following sections, we glance at a variety of sleep
apnea detection methods. Section III, gives a general
description of VAD. In section IV, the methodology of our
proposed system is described. Section V demonstrates the
results of our system. Then, we conclude our paper in section
VI, and highlight some directions for future research.
II. RELATED WORK
Over the past few years most of the related research has
focused on presenting methods for the automatic processing of
different statistical features of different signals such as thorax
and abdomen effort signals, nasal air flow, oxygen saturation,
electrical activity of the heart (ECG), and electrical activity of
the brain (EEG) for the detection of SA.
In our previous published research, we developed a Neural
Network (NN) as a predictive tool for OSA using SpO2 signal
and evaluated its effectiveness [9]. In addition, in [10] we
further developed a model based on a linear kernel Support
Vector Machines (SVMs) using a selective set of RR-interval
features from short duration epochs of the ECG signal. The
results show that our automated classification system can
recognize epochs of SA with a high degree of accuracy,
approximately 96.5%
As OSA is generally caused by a blocked of the airflow
airway and it is characterized by repetitive episodes of
breathing cessation, the respiratory signal recording analysis
during sleep becomes very valuable in order to estimate
respiratory flow and distinguish the changes in the breathing
pattern of the patient. Then, to provide additional and
complementary information, other biological signal data
measurements such as ECG and SpO2 could be bridged to
analyze sleep data, as clinical experience indicates that an
apneic event is frequently accompanied by a fall in the blood
oxygen saturation (SpO2) [11], and cyclic variations in the
duration of a heartbeat (ECG); this consists of bradycardia
during apnea followed by tachycardia upon its cessation [12].
Recently, based on the SpO2 and tracheal breathing sound
recording analysis during sleep, the study in [13] reports a new
fully automatic technology for OSA detection. Different
parameters were investigated to distinguish the breathing level
during each individual apnea event. Therefore, in the first step,
the drops (more than 4%) and rises of the SpO2 signal were
marked, then, the total energy of the tracheal sound segments
within the periods between a drop and the following rise in the
SpO2 were found. After collection of data, each parameter was
then fuzzified with a sigmoid function and the fuzzy outputs
were added together to classify the sound signals. The results
show high sensitivity and specificity values of more than 90%
in differentiating normal respiration from disordered breathing
in patients.
Several studies for the non-invasive sensing of the
respiration rate are based either on the measurement of the
nasal airflow or on the evaluation of the respiration movements
(abdominal or chest respiration signals). Using mean absolute
amplitude analysis of the combination of the thoracic and the
abdominal signals in [14] showed that both signals are able to
indicate the occurrence of SA events. The analysis
identification results achieved a sensitivity ranged from 70.29-
86.25% and the specificity values ranged from 74.82 to
90.09%.
It has been reported that snoring is a common finding in
people with OSA. OSA is generally caused by a blocked of the
airflow airway. Therefore, the snoring must be due to the
vibration of soft tissues when the airflow stimulates the ill
structure in the upper airway during sleep [15]. Of all methods
for diagnosing OSA, the formants estimation method is most
widely used. The formants information contains the essential
acoustic properties of the upper airway. It has been discovered
by studies that there is a correlation between the state of the
upper airway and the first formant frequency. A narrower upper
airway is usually led to a higher first formant frequency.
Andrew et al. [16] and [17] proposed fixed formant frequency
thresholds to detect the hypopneic snores which must be higher
than that of the typical ones.
Various portable monitor devices already exist in the
market. SleepStrip
TM
is one of the carriages available in home
sleep test diagnostic devices. This device has to be worn for a
minimum of five hours of sleep, and the actual device is placed
on the individual’s face where the two flow sensors (oral and
nasal thermistors) are placed just below the nose and above the
upper lip to capture the breath of the patient. For all samples
combined, sensitivity and specificity values ranged from 80-
86% and 57-86% respectively [18].
Even though most of the related studies yielded promising
initial results, more improvement is needed, as it either requires
physical attachment to a user or may be unreliable. Therefore,
since apnea is a condition where patient pauses breathing; this
can be of great concern for detecting breathing through the
sound signal where an appropriate alarm can be released upon
its cessation.
III. VOICE ACTIVITY DETECTION – THE PRINCIPLE
In this work, the principle of voice activity detection
(VAD) algorithm employed to detect the presence or absence
of apnea on real breathing signals is described.
Voice Activity Detector plays an important role in speech
processing techniques such as speech coding [19], speech
enhancement, and speech recognition [20]. Other examples
include cellular radio systems (GSM and CDMA based) [21],
hands-free telephony [22], VoIP applications and echo
cancellation.
VAD relies on measurement of features from speech which
yield highly in differentiating between voiced and unvoiced
segments, where the regions of voice information within a
given audio signal are referred to as ‘voice-active’ segments
and the pauses between talking are called ‘silence’ or ‘voice-
inactive’ segments. Therefore, the performance trade-offs of
VAD algorithm are made by maximizing the detection rate of
active speech while minimizing the false detection rate of
inactive segments [23].
The most important part in VAD classifier is feature
extraction, from which different regions in the audio signal
can be separated. Common features used in the VAD detection
process are cepstral coefficient [24], spectral entropy [19],
zero-crossing rate [20, 25], least square periodicity measure
[26], and average magnitude difference function [27]. Another
important and widely used feature in this regard is signal
energy, which is presented in this work, and compared with the
dynamically calculated threshold.
IV. APNEA DETECTION USING VAD BASED - ENERGY
The general VAD block diagram used in our methodology
is shown in figure 1.
The assumptions on the VAD algorithm used here is that
the speech is quasi-stationary and its spectral changes quickly
over short periods like 20-30ms, but the background noise is
relatively stationary and changes very slowly with time. In
addition, the energy of the active speech level is usually higher
than background noise energy [28].
In the first step, the respiratory signal is filtered to remove
the undesired low frequency components. Then, the power with
different window sizes of the Fast Fourier Transform (FFT) is
calculated for the filtered signal [29].
Let x(t) be the input signal samples, and X(n) be Fast
Fourier Transform (FFT) samples. The VAD algorithm begins
with the energy computation within the smallest integer range
of frequency values n
1
and n
2
:
(1)
The energy of the signal is computed in two window
frames; short window and long window for every window
number i:
E
short
(i) = φ
short
E
nergy
+
(1 - φ
short
) E
short
(i) (2)
E
long
(i) = φ
long
E
nergy
+ (1 - φ
long
) E
long
(i) (3)
The number of frames used is N/L, where N represents the
number of samples in the signal and L represents the window
size in frequency domain. The coefficients φ
short
and φ
long
refer
to the window-length factors, where φ
short
= 1/16, and
φ
long
=
1/128, and L= 528 were used in this study.
At this point, since VAD aims to differentiate voice and
silence, where silence is mostly referring to background noise,
the noise level at every frame needs to be computed. For this
purpose, a threshold THR value needs to be determined for
comparing the signal value against noise:
THR =
+ M (4)
In the above formulation, K
f
is the K-th frame and M is a
margin value that can be considered to separate voice and
silence in the event that noise level is flat.
The VAD technique eventually makes a decision by
comparing every frame of signal energy against the THR value.
It is important to note that transitional periods from active
voice to silence may also affect the decision. Based on the
above steps and discussions, the decision on the VAD identifier
(ID) values is made as follows:
Figure 1. A block diagram of VAD design.
1, if Energy > THR
VAD- ID = 1, if Energy ≤ THR and in transitional period (5)
0, if Energy ≤ THR and not in transitional period
The outcome of the VAD technique is the separated speech
and silence phases which can be fine-tuned for identifying
breath versus breathing cessations for apnea detection.
In our work, there is a second threshold (Tr) that would be
used to decide whether a silence phase corresponds to apnea or
not. According to the sleep apnea literature, a breathing
cessation (silence) of 15 seconds or more would be classified as
apnea, as shown below:
If (VAD-ID
j
== 0) and (T
VAD-IDj
≥ 15Sec) =>
VAD-ID
j
is a SA event period
(6)
In the above relationship, T
VAD-IDj
corresponds to the
duration of silence phase j detected by the VAD technique.
The testing procedure and experimental results to evaluate
the described VAD algorithm will be discussed through the
next section.
V. EXPERIMENTAL RESULTS
MATLAB environment was used to perform our
methodology on various samples of breathing signals during
breathing and breath hold in 50 normal people. The volunteers
were asked to breath 20 cycles. They were asked to hold their
breath before, during and after the 20 cycles. The breath
recording was done using the SONY VAIO VPCEB42FM
microphone (Realtek High Definition Audio [30] professional
microphone with the accompanying Audacity software).
The human respiratory signal is given to the classification
system as the input, and the coding is developed in such way
that it calculates the fundamental feature of the respiratory
signal, which is the energy. The threshold is then applied to the
extracted energy feature and the binary decision is made.
VAD=1 is declared if the energy feature exceeds the threshold.
Otherwise, VAD=0 is for no breath or when silence (cessations
of breathing) is present.
Figure 2 shows the results obtained from the segmentation
technique of the input signal which splits the acoustic signal of
respiration into silence and voiced phases. The start point and
end point of a respiratory signal which contains breathing
Processing
Thresholds
Computation
VAD
decision
correction
VAD
decision
Features
Extraction
Input
Signal
VAD output
phases is determined in this work. Hence, the apnea events that
are silence phases lasting 15s or longer can be detected.
Figure 2. Segmentation acoustic signal of breath using VAD.
VI. CONCLUSIONS AND FUTURE WORK
This work sought to determine the effectiveness of VAD
based - energy in distinguishing the apnea in breathing signal.
The provided respiratory signal is classified successfully
with the help of the formulated algorithm with more than 97%
accuracy.
In order to detect sleep apnea in real time, the proposed
algorithm could be improved and adjusted by adding
calibration procedures to run on an FPGA [31]. After the
successful design and implementation of the OSA system, it is
planned to be experimentally tested in order to evaluate its
accuracy and practicality. The tests will take place in a local
hospital for a set of patients who have symptoms of OSA. In
addition, a different set of other subjects without SA symptoms
will test the system to verify its false positive accuracy.
REFERENCES
[1] Sleep Disorders Guide. www.sleepdisorderguide.com.
[2] N. Israel, A. Tarasiuk, and Y. Zigel, “Nocturnal Sound Analysis for the
Diagnosis of Obstructive Sleep Apnea,” in Proceedings of the 32
nd
IEEE
International Conference on Engineering in Medicine and Biology
Society (EMBS 2010), pp. 6146-6149, Sep. 2010.
[3] A. Yilmaz, and T. Dundar, “Home Recording for Pre-Phase Sleep Apnea
Diagnosis by Holter Recorder Using MMC Memory,” in Proceedings of
the 2010 IEEE International Conference on Virtual Environments
Human-Computer Interfaces and Measurement Systems (VECIMS), pp.
126-129, Sep. 2010.
[4] “Choosing a CPAP,” American Sleep Apnea Association,
www.sleepapnea.org/resources/pubs/cpap/htm.
[5] SleepMedInc. www.sleepmed.md.
[6] D. Baraglia et al., “Automated Sleep Scoring and Sleep Apnea Detection
in Children,” in Proceedings of SPIE 6039, 2005.
[7] “Sleep Study Cost”, Shop & Compare Healthcare Facilities and Cost
with NewChoiceHealth.com. Web. 03 Dec. 2010.
www.newchoicehealth.com/SleepStudyCost.
[8] P. Chazal, C. Heneghan, and W. Mcnicholas, “ Multimodal Detection of
Sleep Apnoea Using Electrocardiogram and Oximetry Signals,”
Philosophical Transactions of The Royal Society A: Mathematical,
Physical and Engineering Sciences, vol. 367, no. 1887, pp. 369-389,
2009.
[9] L. Almazaydeh, M. Faezipour, and K. Elleithy, “A Neural Network
System For Detection Of Obstructive Sleep Apnea Through SpO
2
Signal
Features,” (IJACSA) International Journal of Advanced Computer
Science and Applications, vol. 3, no. 5, pp. 7-11, Jun. 2012.
[10] L. Almazaydeh, K. Elleithy, and M. Faezipour, “Detection of obstructive
sleep apnea through ECG signal features,” in Proceedings of the IEEE
International Conference on Electro Information Technology (IEEE
eit2012), pp. 1-6, May. 2012.
[11] M. Canosa, E. Hernandez, and V. Moret, “Intelligent Diagnosis of Sleep
Apnea Syndrome,” In IEEE Engineering in Medicine and Biology
Magazine., vol. 23, no. 2, pp. 72–81, 2004.
[12] P. Chazal, T. Penzel, and C. Heneghan, “Automated Detection of
Obstructive Sleep Apnoea at Different Time Scales Using the
Electrocardiogram,” Institute of Physics Publishing., vol. 25, no. 4, pp.
967–983, Aug. 2004.
[13] A. Yadollahi, and Z. Moussavi, “Acoustic Obstructive Sleep Apnea
Detection”, In Proceedings of IEEE Conference on Engineering in
Medicine and Biology Society (EMBC 2009), pp. 7110-7113, Sep. 2009.
[14] A. Ng, J. Chung, M. Gohel, et al., “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.
[15] Y. Zhao, H. Zhang, W. Liu, and S. Ding, “A Snoring Detector for
OSAHS Based on Patient’s Individual Personality,” In 3rd International
Conference in Awareness Science and Technology (iCAST)., pp. 24-27,
2011.
[16] K. Andrew, S. Tong, et al. “Could Formant Frequencies of Snore Signals
Be an Alternative Means for the Diagnosis of Obstructive Sleep
Apnea?”, Sleep Medicine, vol. 9, pp. 894-898, Dec. 2008.
[17] K. Andrew, T. Koh, E. Baey, et al., “Speech-Like Analysis of Snore
Signals for the Detection of Obstructive Sleep Apnea”, In International
Conference on Biomedical and Pharmaceutical Engineering 2006
(ICBPE 2006), pp. 99-103, Dec. 2006.
[18] T. Shochat, N. Hadas, M. Kerkhofs, et al., “The SleepStrip
TM
: An
Apnoea Screener for the Early Detection of Sleep Apnoea Syndrome,”
European Respiratory Journal, vol. 19, pp. 121-126, 2002.
[19] S. McClellan, J. Gibson,”Spectral entropy: An alternative indicator for
rate allocation,” In IEEE International Conference on Acoustics, Speech,
Signal Processing, pp. 201-204, Apr. 1994.
[20] B. Atal, L. Rabiner, “A Pattern Recognition Approach to Voiced-
Unvoiced-Silence Classification with Applications to Speech
Recognition,” IEEE Transactions on Acoustics, Speech, Signal
Processing, vol. 24, no. 3, pp. 201-212, June 1976.
[21] ETSI TS 126 094 V3.0.0 (2000-01), 3G TS 26.094 version 3.0.0 Release
1999, Universal Mobile Telecommunications System (UMTS);
Mandatory Speech Codec speech processing functions AMR speech
codec; Voice Activity Detector (VAD), 2000.
[22] A. Benyassine, E. Shlomot, and H. Su, “ITU-T recommendation G.729
annex B: A silence compression scheme for use with G.729 optimized
for V.70 digital simultaneous voice and data application,” IEEE
Communications Magazine, vol. 35, no. 9, 1997.
[23] E. Verteletskaya, K. Sakhnov, “Voice Activity Detection for Speech
Enhancement Applications,” ACTA Polytechnica, vol. 50, no. 4, pp. 100-
105, 2010.
[24] J. Haigh, and J. Mason, “Robust Voice Activity Detection Using
Cepstral Features,” In Proceedings of IEEE Region 10 Conference on
Computer, Communication, Control and Power Engineering
(TENCON’93), vol. 3, pp. 321-324, 1993, Beijing.
[25] A. Sangwan, C. MC, H. Jamadagni, et al., “VAD Techniques for Real-
Time Speech Transmission on the Internet,” In Proceedings of 5
th
IEEE
International Conference on High-Speed Networks and Multimedia
Communications, pp. 46-50, 2002.
[26] S. Tanyer, and H. Ozer, “Voice Activity Detection in Nonstationary
Gaussian Noise,” In Proceedings of International Conference on Signal
Processing ( ICSP’98), pp. 1620-1623, 1998.
[27] M. Orlandi, A. Santarelli, and D. Falavigna, “Maximum Likelihood
Endpoint Detection with Time Domain Features,” In Proceedings of 8
th
European Conference on Speech Communication and Technology
(eurospeech 2003), pp. 1757-1760, 2003, Geneva.
[28] K. Sakhnov, E. Verteletskay, and B. Simak, “Approach for Energy-
Based Voice Detector with Adaptive Scaling Factor,” IAENG
International Journal of Computer Science”, vol. 36, no. 4, pp. 394,
2009.
[29] A. Abushakra and M. Faezipour, “Acoustic Signal Classification of
Breathing Movements to Virtually Aid Breath Regulation,” IEEE
Journal of Biomedical and Health Informatics, vol. 17, no. 2, pp. 493-
500, March 2013.
[30] VAIO® Laptop ComputersVPCEB42FM/BJ: Available from:
http://www.docs.sony.com/release/VPCEB3_Series.pdf., 2011.
[31] B. Marinkovic, M. Gillette, and T. Ning, “FPGA Implementation of
Respiration Signal Classification Using a Soft-Core Processor,” In
Proceedings of the IEEE 31
st
Annual Northeast Bioengineering
Conference, pp. 54-55, April 2005.
