Noninvasive Automatic Sleep Apnea Classification System

Document Sample
Noninvasive Automatic Sleep Apnea Classification System Powered By Docstoc
					            Noninvasive Automatic Sleep Apnea Classification System
        Tarik AL-ANI a   Daniel NOVAK b            Frédéric LOFASO
                                                                    c,d
                       a                         b               c
        Yskandar HAMAM Paula Tamara POZZO MENDOZA Daniel ISABEY
                         Lenka LHOTSKA b           Redouane FODIL
                                                                   c




        a                         b                                          c
          ESIEE - A²SI,            Department of Cybernetics,                 INSERM U492, Créteil,
                                                                             d
        Noisy-Le-Grand,           Czech Technical University in Prague,       Hôpital R. Poincaré, Garches
        France                    Czech Republic                             France
        t.alani@esiee.fr


Abstract
Sleep Apnea Syndrome (SAS) is a very common sleep          way occludes (either partially or fully) but the effort to
disorder. SAS is considered as clinically relevant when    continue breathing is persistent. The primary causes of
the breath stops during more than 10 seconds and oc-       upper airway obstruction are lack of muscle tone dur-
curs more than five times per sleep hour. In this paper,   ing sleep, excess tissue in the upper airway, and anat-
we present a noninvasive automatic approach to sleep       omic abnormalities in the upper airway and jaw.
apnea classification. Only noninvasive records of the
respiratory and cardiac activities (Nasal Airway Flow      Central Sleep Apnea (CSA): It affects only 5-10% of
(NAF) and Pulse Transit Time (PTT)) issued by the          the sleep apnea population. It is a disorder of the cen-
technique of PolySomnoGraphy (PSG) are considered          tral motoneural system. CSA occurs when both airflow
for the detection of the different sleep apnea syn-        and respiratory effort cease. This cessation of breathing
dromes: obstructive, central and hypopnea. Experi-         results from a loss of the autonomic drive to breathe. It
mental results using clinical data are presented.          is when the brain seems to forget to tell the breathing
                                                           muscles to work. With central apnea, the breathing
Keywords: Sleep Apnea, Pulse Transit Time, Artificial      airways stay open while the chest muscles and dia-
Neural Network.                                            phragm stop working. When the amount of oxygen
                                                           starts to fall in the body, a brain alarm goes off. Then
                                                           the person wakes up and starts breathing.
1.    Introduction
                                                           Mixed Sleep Apneas (MSA): It occurs when there is a
Sleep Apnea Syndrome (SAS) is a very common sleep          temporary cessation in breathing followed by obstruc-
disorder. SAS is considered as clinically relevant when    tive ventilatory efforts. It is a CSA followed by an
the breath stops during more than 10 seconds and oc-       OSA.
curs more than five times per sleep hour. These non
breathing episodes may sometimes occur more than           The treatment of SAS depends on the cause of the ap-
300 times a night. Health studies affirm that more than    nea. That’s why we need to make a classification.
30 of these non breathing episodes per night should be     Nowadays the sleep apneas are classified manually by
considered abnormal. There exist two kinds of apneic       the expert physician thanks to the nocturnal polysom-
events that may cause insufficient pulmonary ventila-      nographic monitoring that simultaneously records sev-
tion during sleep, Apnea and Hypopnea. Apnea is de-        eral vital signals during the entire sleeping process
fined as the total absence of airflow, followed by the     (Nasal AirFlow (NAF), ElectroCardioGram (ECG),
reduction of oxygen levels in arterial blood. The term     ElectroEncephaloGram (EEG), ElectroMyoGram
hypopnea is used when the breath doesn’t stop but de-      (EMG), Esophageal Pressure (Pes), gastric pressure
crease over 50% of its normal value , followed by the      (Pga), Oxygen Saturation (OS),…) [1, 3]. A sleep ap-
reduction of oxygen levels in arterial blood. The SAS      nea diagnosis is a very time consuming, expensive and
is present mainly in adults and in 11% of children es-     tedious task consisting of expert visual evaluation all
pecially in the male population. [1, 2]. Different form    10 minutes pieces of approximately 8 hour recording
of apnea/hypopnea may be distinguished: obstructive,       with a setting of many channels.
and mixed.                                                 In a previous work [3], it was demonstrated that sleep
Obstructive Sleep Apnea (OSA): The most common             apnea classification may be done automatically using
type of sleep apnea. OSA occurs when the upper air         three simultaneous records of NAF, Peso and Pgas.
The current techniques of investigating patients with
suspected sleep disordered breathing are inadequate.
The Obstructive apnea episodes are not usually diffi-
cult to detect even when only a basic measure of respi-
ratory effort such as thoracic and abdominal movement
is used. On the other hand, correctly identifying ob-
structive hypopneas and episodes of upper airway re-
sistance needs a sensitive measure of airflow and inspi-        Figure 1. A polysomnograph example.
ratory effort. The measurement of swings in pleural             From the clinical point of view, episodes of sleep ap-
pressure by esophageal manometry is the current gold            nea are often detected first using only the respiration
standard techniques for detecting changes in respira-           signals, such as nasal airflow and abdominal and tho-
tory effort. However, the placement of an esophageal            racic movements. The nasal airflow measures the Ap-
catheter is often uncomfortable and unacceptable, it            nea/Hypopnea Index (AHI), which indicates the num-
may modify the upper airway dynamics, and some be-              ber of apnea/hypopnea during one sleeping hour. The
lieve that it contributes to the sleep disturbance during       abdominal and thoracic signals show the difference be-
the sleep study. Furthermore, this technique is avail-          tween OSA and CSA.
able in only a proportion of sleep laboratories and, if
performed, adds significantly to the cost of the sleep          When the type of sleep apnea is not able to be deter-
study. For all these reasons, other new techniques for          mined from the PSG signals, the esophageal pressure
detecting and classifying sleep apneas and other                signal has to be measured. This signal shows the dia-
breathing disorders are developed using mainly the              phragm activity, but to record it requires the applica-
ECG [4] or Pulse Transit Time (PPT) [5, 13]. In this            tion of an invasive method, which increase the risk of
preliminary work, we are presently investigating the            arousal due to patient discomfort.
use of the neural networks based on simultaneous NAF            The blood oxygen saturation (SaO2) signal measured
and PPT records as noninvasive signals, to classify and         by the oxymetry allows distinguishing between apnea
detect automatically the different types of apnea.              and hypopnea.
In this paper, we present a noninvasive automatic ap-           The EEG signals aim is to evaluate arousals. Further-
proach to sleep apnea classification. Only noninvasive          more, the physiologic range varies during sleep. For
records of the NAF and PTT are considered for the               example, during the low stable sleep phase the thoracic
classification of the different sleep apnea syndromes.          movements increase, the abdominal movements de-
Experimental results using clinical data are presented.         crease, and the esophageal pressure increases 50% in
The structure of this paper is as follows. In Section 2,        comparison to the situation of waking state.
we introduce the most important diagnostic tool in
sleep apnea diagnosis. In section 3, the state of art is        The PSG is uncomfortable for the patient and involves
presented. In Section 4, we introduce a noninvasive             a considerable capital investment for the healthcare
method for measuring respiratory effort. In Section 5, a        system in equipment, bed space and specialized techni-
review of our system is presented. In Section 6, we             cal support. Interpretation of the test data is also com-
give some results. The conclusions and perspectives             plex and time consuming and consequently the overall
are presented in Section 7.                                     cost of performing a PSG is estimated to be around
                                                                1000-2000 dollars [6].
                                                                Each PSG has its own signal analysis software. Most
2.     The polysomnography (PSG)                                of them use time domain algorithms, which examine
The most important diagnostic tool in any medical condi-        the amplitude and time, with a precision of about 80 to
tion is for the physician to take the time to obtain a good     90%. Even if it is quite efficient and allows a good di-
history and physical examination. A chest x-ray along           agnosis in most of the case it is still difficult to detect
with laboratory tests are usually performed to evaluate         the patient’s type of sleep apnea with good precision.
other possible contributing factors, such as diabetes or        Even experts do not always agree on the PSG results
hypothyroidism. The definitive diagnostic exam is a poly-       interpretation.
somnograph where the patient stays in a sleep laboratory
                                                                Another difficulty is the dependence between the sig-
or even at home overnight while measurements of his
                                                                nals interpretation and the sleep stage and/or the envi-
brain activity, respiratory activity, oxygen levels, and car-
                                                                ronment. These are the most important reasons why
diac activity are performed. A sleep apnea recording is a
                                                                automatic detection is needed.
very time consuming task consisting of expert visual
evaluation all 10 minutes pieces of approximately 8 hour
recording with a setting of many channels: NAF, Peso,
Pgas, ECG, EEG and other signals, Figure 1.
3.    State-of-the-art                                     the interval between two R waves. The Discrete Fou-
                                                           rier Transform of fragments from RR interval series is
To try to solve the PSG interpretation problems, many      computed for a spectral analysis. Pattern searching is
different approaches have been suggested. These ap-        used to identify the sleep apnea events by means of ‘U’
proaches are based on the analysis of different PSG        shaped pattern. This study is mainly based on the real
signals with the help of the most common kinds of arti-    necessity of finding an inexpensive SAS diagnosis.
ficial intelligence (AI) methods. In this section, a non   This method still shows some problems such as, possi-
exhaustive overview of the state-of-the-art is pre-        ble misleading in RR power spectral analysis and erro-
sented.                                                    neous results due to the ‘U’ pattern search. However
                                                           the SAS diagnosis based on the ECG analysis is feasi-
3.1   Thoracic and abdominal exploration                   ble.
The aim of this research is to make a classification of
Obstructive Sleep Apnea and Central Sleep Apnea us-        3.5       Expert system
ing only the abdominal and thoracic movement signals       This research focuses on the SAS detection and inter-
[7]. Each SAS type present different abdominal and         pretation [11]. It detects and interprets with the help of
thoracic movement signals. These differences help to       an expert system. This intelligent system is divided
realize a SAS classification. This method also presents    into many modules, each with its own function. It is a
the advantage to detect the different level of obstruc-    gain of time and of money for the hospitals but there
tion, which allows an evaluation of the respiratory ef-    are still some situations which are not detected. Fur-
fort. The main problem in this method is the noise,        thermore, the experts and the system do not always
which cannot be eliminated with the standard methods;      agree (on event detection or on time interval).
moreover the Piecewise Linear Approximation (PLA)
method used in this approach introduces a small error.     3.6       Artificial Neural Network and NAF signal

3.2   Pressure and airflow exploration                     This investigation is based on the artificial neural net-
                                                           work (ANN) for the classification between apnea and
This study focuses on the classification between OSA,      hypopnea (without taking into account if it is OSA or
CSA, and the obstruction quantification, with the help     CSA) [12]. It was used the NAF signal and creation of
of the airway impedance [8]. It uses the Forced Oscilla-   four ANN. One of the created ANN showed good re-
tion Technique, which is an application of an oscilla-     sults, the robustness of the system could be improved
tory pressure signal to the respiratory system. By using   by including other signals, such as oxygen saturation,
the nasal pressure and the airflow signals the imped-      heart rhythm, etc. Today it is still difficult the inclusion
ance is determined. This impedance showed higher           of other signals, due to these signals depend on other
values in OSA cases. This method allows avoiding an        phenomena.
analysis with the esophageal pressure, due to this it
avoids the use of an invasive act.                         There are many other approaches and systems, which
                                                           are capable to detect and classify SAS. Each approach
3.3   Nonlinear multivariable model                        uses different types of signals.

This investigation uses an algorithm approach, which       The classification of sleep apnea syndrome today is not
is based on the model and analysis of physiological        perfect. On one hand the way to diagnose SAS can be
data. The recorded data is composed of blood oxygen        improved in order to be more comfortable for the pa-
saturation, heart rate and respiration signals. The mod-   tient and on the other, actual systems do not correctly
els estimated from the data can distinguish between        distinguish the different kinds of SAS. These are the
two different patterns (normal breathing and apnea).       main reasons that researchers are actively working to
The estimated models are nonlinear, autoregressive,        create a new system to detect and classify the different
moving average and with exogenous inputs (NAR-             types of sleep apnea syndrome.
MAX) [9]. It was also proved that, for sleep apnea         Obtaining non-linear analytical models for the different
syndrome data, this method is better than a nonlinear      Sleep Apnea Syndromes is difficult problem [3]. A
monovariable model. A linear model clearly performs        good approach to be applied could be a method with
worse than the nonlinear multivariable model.              the help of a expert system, due to their large impact in
                                                           automated diagnostic could be the expert systems. The
3.4   Electrocardiogram analysis                           main reasons for not choosing this approach are :
This research focuses on the obstructive sleep apnea             •    The necessity of doctor’s support that are ex-
detection based on ECG signal [10]. The analysis is                   perts on the SAS. In general, it is difficult to
based on the detection of respiratory disturbance re-                 find an expert doctor on the SAS field which
flected in the ECG. The R wave detection from the                     would have enough time to help in the project;
QRS complex allows obtaining RR interval, which is
     •    The difficulty of building and maintaining        where
          large rule bases;
                                                                     PEP = pre-ejection (systolic) period.
     •    The difficulty to act in real time.               The instrumentation used for this technique is only a
The approach presented in this paper is based on Arti-      plethysmograph and ECG electrodes. The plethys-
ficial Neural Network. The main reason to choose this       mograph sensor is put on the finger, to have a better
approach is because it has showed good results in           measure of PTTreal and on the ear lobe, to study the
physiologic applications and in SAS applying different      PEP. The pulse wave signals used are from the finger.
kinds of other signals than the PTT signal (which is the    The PTT signal provides a good measure of respiratory
main one applied in our work) [12]. It should be noted      effort, quantification of the obstruction, and therefore a
that in the ANN it is important to have enough good         classification of the type of apnea/hypopnea. There is
data; because these data are needed in the learning         no respiratory effort in central apnea. If there is an in-
process.                                                    crease of PTT oscillations then there is an OSA or up-
                                                            per airway resistance, because the respiratory effort in-
4.       Pulse Transit Time                                 creases. If there is a decrease of PTT oscillations then a
                                                            CSA occurs. Signal for different breathing events is
The measurement of swings in pleural pressure, for de-      displayed in the following Figures 3a-d.
tecting changes in respiratory effort is usually assessed
by measuring esophageal pressure (Peso), trough an
esophageal balloon catheter. This technique has several
disadvantages; it causes some discomfort, due to the
placement of the esophageal catheter, can lead to frag-
mentary sleep, and may modify the upper airway dy-
namics.
A new noninvasive method for measuring respiratory
effort has been proposed [5, 13]. It is based on the es-
timation of the Pulse Transit Time (PTT) signal, which      Figure 2 . Relation between PTT and intra-mask pressure
has been demonstrated that its oscillations yield a valid
measure of inspiratory effort, Figure 2. The aim of this
work is to find a new technique for detecting and clas-
sifying automatically sleep apnea syndrome with the
help of the PTT, to validate the use of PTT as a method
to automatic diagnosis of sleep apnea syndrome.
The PTT signal is a method to measure the variations
in blood pressure. It is the time needed for the arterial
pulse pressure wave to travel from the aortic valve to      Figure 3 a. PTT during normal breathing.
the periphery, generally the ear or the finger. This time
is estimated as the delay between the R wave in the
ECG and the arrival of the pulse wave at the periphery
as determined by pulse oxymetry (about 200-250 ms).
We measure one value of PTT by heart beat.
There is a link between the PTT and the esophageal
pressure (Pes). If the esophageal pressure increases,
then the amplitude of the PTT oscillations falls. A de-
                                                            Figure 3 b. PTT during OSA.
crease of the esophageal pressure corresponds to an in-
crease of the blood pressure (BP), which is directly due
to arousals and not to hypoxemia.
             ↓Pes ↑ PTT         ↑BP     arousal
Actually the R wave does not correspond to the open-
ing of the aortic valve but to the beginning of contrac-
tion of the left ventricle. Therefore the PTT measured
corresponds to the sum of the real PTT and the contrac-
tion time of the left ventricle:                            Figure 3 c. PTT during CSA.
                PTTmeasured = PTTreal +PEP,
                                                            Figure 5 c. NAF during a hypopnea event: the airflow
                                                            decreases
Figure 3 d. PTT during MSA.
In this work the PTT was estimated as the interval be-
tween the ECG R-wave and the point at which the
                                                            5.      System overview
pulse wave at the finger reached 50% amplitude, Fig-        To carry out the automatic diagnostic system for sleep
ure 4.                                                      apnea classification, it was needed to realize four main
                      PTT = t2-t1,                          steps, shown in Figure 6.

where
t1: the point where ECG R wave occurs,
t2: the point at which the pulse wave (PW) reaches 50%
amplitude (this percentage could change in each study.
some of the more recent studies chose a percentage of       Figure 6. System overview.
25%).
                                                            All these steps are described in this section. The feature
                                                            extraction and selection approach is one of the most
                                                            important in the automatic diagnostic system, because
                                                            most of the neural network accuracy depends on it. The
                                                            ANN training is the set of examples used for learning.
                                                            Enough data is necessary in the training to fit the pa-
                                                            rameters of the ANN to be efficient. The ANN testing
                                                            is the set of examples used to assess the performance of
                                                            the ANN.
Figure 4. PTT estimation.
It should be emphasized that the PTT signal is not able     5.1     Data acquisition
to make a differentiation between apnea and hypopnea        The data files used for training and testing the neural
[13]. To be able to distinguish between these two ap-       networks are given by a sleep laboratory at the hospital
neic events we need to use another signal in addition. It   Raymond Poincaré in France. These data files are all in
has been decided to use the NAF, which allows a good        European Data Format (EDF) and consist of three pa-
distinction between apnea and hypopnea, Figures 5 a-c.      tients PSG record. Each record measured the different
                                                            vital parameters: EOG, EEG, ECG, abdominal and tho-
                                                            racic movements, EMG, NAF, phonograph, pulse and
                                                            SaO2. A more detail description for each patient is pre-
                                                            sented in [20]. For reading the data, it was necessary to
                                                            use other toolboxes: EEGLAB toolbox [14] and the
                                                            EDF toolbox [15].

                                                            5.1.1    Pre-processing of the data
Figure 5 a. NAF during normal breathing.
                                                            The ECG baseline is the interference that appears due
                                                            to different reasons, such as: patient movement, breath-
                                                            ing, physical exercise, etc. The baseline wandering can
                                                            make the inspection of the ECG difficult; therefore it is
                                                            very important to reduce as much as possible its effect.
                                                            The method used to remove baseline in this work is
                                                            based on wavelet transform, which removes the low
                                                            frequency artifacts.
Figure 5 b. NAF during an apnea event: the airflow          The wavelet approximation of signal was applied, as
ceases.                                                     the tool to obtain a good approximation of the ECG
baseline, because both are low frequency signals. To        be generated and the feature selection stage, which de-
obtain a good result, the level of such approximation       cides the best amount of features to be used. . In this
must be defined. The best level depends on the ampli-       work two signals are selected for features extraction:
tude and main spectrum distribution of the baseline in-     PTT and NAF signals.
terference. It was made an automatic method to get the
                                                            In order to avoid having too many inputs in the neural
best level [16], which is based on measures of the re-
                                                            network, several windows of the signal were taken.
sulting signal variance and on spectrum energy disper-
                                                            According to the definition of sleep apnea syndrome,
sion of the wavelet approximation, Figure7.
                                                            the length of this window should be at least 10 seconds
                                                            in order to detect apnea within a single time window.
                                                            But in order to assure that some events are not missed,
                                                            it is better to take a window longer than 10 seconds.
                                                            Since signals may contain short apnea or hypopnea
                                                            episodes (lasting only several seconds) which are not
                                                            pathological. Therefore it was decided to use a window
                                                            of 14 seconds length. The use of a longer window
                                                            would unnecessarily increase the complexity of the
Figure 7. Original ECG signal (blue solid) and corrected    system.
ECG signal (red hashed).
To be able to eliminate or reduce the baseline wander-      5.2.1   NAF extraction
ing, the approximations must have a narrow spectrum,        The first approach applied for NAF signal extraction
because such interferences are usually almost pure si-      was the decreasing of sampling rate from 200 Hz to 20
nusoids. Moreover, the variance of the resulting signal     Hz, due to practical considerations of having less input
should be as low as possible, because the approxima-        in the network. This change of sampling rate was done
tions must no have high frequency components such as        using a simple moving average filter. These signals are
peaks following R peaks, the final signal must be flat.     then re-sampled at 2 Hz using the antialiasing filter. At
The QRS detection by Tompkins and Pan, improved by          the end of all the re-sampling the selected 14 second
Fokapu and Girard algorithm [17] was chosen because         window consisted only of 28 sample points for the
it showed the best detection accuracy, Figure 8.            NAF instead of 2,800. It has been noticed that the de-
                                                            creasing of sampling rate was not a suitable approach
                                                            to be applied, because the data was re-sampled too
                                                            much. Because of this re-sampling some of the infor-
                                                            mation was lost. Therefore a new approach was needed
                                                            for the features extraction of this signal. The most im-
                                                            portant information in the NAF signal is the amplitude.
                                                            This information can be used to determine if normal
                                                            breathing, apnea, or hypopnea are present. The method
                                                            used for calculating the amplitude, is detailed below,
                                                            Figures 9 a-c:
Figure 8. ECG-R peak detection using the Tompkis and
                                                                 • Calculate all the local maximum and local
Pan, improved by Fokapu.                                             minimum amplitudes of the 14 second window
                                                                     from the appearance of the first R peak in the
5.2    Feature extraction                                            ECG. The calculation starts at the first occur-
                                                                     rence of the ECG-R peak in the 14 second
The pattern classification problem is generating opti-               window in order to correlate with the PTT sig-
mal decision boundaries or decision procedures to                    nal. This part was difficult to implement due
separate the data into pattern classes based on the fea-             to high noise presence in the NAF signal. To
ture vectors. For efficient pattern classification, meas-            be able to get rid of the misclassified local
urements that could lead to disjoint sets of features                minimums and maximums in the signal due to
vectors are desired. This point underlines the impor-                noise, two conditions were created in the algo-
tance of the preprocessing and features extraction pro-              rithm: Firstly, the next value between two
cedures. This section aim is to choose features, which               maximums or two minimums must be at least
will be part of the feature vectors. It is necessary to              200 samples far away from each other. Sec-
perform a feature selection in order to have only the                ondly, the interval between one maximum and
most relevant features.                                              one minimum must be at least 150 samples far
There are two stages that must to be fulfill, the feature            away form each other. With the help of these
extraction stage, which decides how the features will
        two conditions reasonable results were ob-               •    Extract the amplitude of the signal and the pe-
        tained.                                                       riod where the amplitude is constant. If the
                                                                      amplitude changes, two new values are ob-
    •   Calculate the Instantaneous Respiration Am-
                                                                      tained: the new amplitude and the period dur-
        plitude (IRA), which is the amplitude between
                                                                      ing this amplitude occur.
        each minimum and each maximum, and the
                                                                 •    The number of points that are computed using
        Instantaneous Respiration Interval (IRI),
                                                                      segmentation procedure is not the same for
        which is the interval between each minimum
                                                                      each 14 second window. To avoid this prob-
        and each one maximum. These calculations
        were made because they showed very good re-                   lem, the average of the PTT segmentations and
        sults in a previous study [18].                               their intervals are taken to be part of the neural
                                                                      network input.
    •   The averages of these two results were taken
                                                           Note that a third approach based on the PTT signal fea-
        to form part of the neural network input.
                                                           ture extraction that correlates with the NAF signal was
                                                           suggested [19]. This approach calculates the PTT am-
                                                           plitude and duration of each breathing cycle.




Figure 9 a. Maximums and minimums for normal breath-
ing event in a NAF signal.
                                                           Figure 10 a. PTT segmentation for a normal breathing
                                                           event.




Figure 9 b. Maximums and minimums for apneic event
in a NAF signal.
                                                           Figure 10 b. PTT segmentation for apneic event.




Figure 9 c. Maximums and minimums for hypopneic            Figure 10 c. PTT segmentation for hypopneic event.
event in a NAF signal.
                                                           5.3       Training and testing sets creation
5.2.2   PTT extraction
                                                           The whole training is composed of noisy and non noisy
The first approach used was based on re-sampling the       data. In this work the noisy data is consider to be the
data in the same way done as for the NAF signal. For       one that is in REM sleep stage, because during this pe-
the PTT signal case the re-sampling was performed          riod the PTT signal is difficult to understand. Two
first from 200 Hz to 22.22 Hz and at last to 2.469 Hz.     Training sets were created for this work, each of them
It was obtained between 22 and 30 sample points. It        containing 45 windows, 15 windows for each event
should be noted that the PTT signal is a stair function    (i.e. normal breathing, apnea, hypopnea) these data
and not a continuous function, i.e. the numbers of sam-    were taken from the PSG records of three patients. The
ple points are not the same for each part of the signal.   testing set is a set of examples used only to assess the
This method is not suitable for the PTT signal, because    performance (generalization) of a fully-specified clas-
this reduction does not keep all the information in the    sifier. For testing the network performance, 36 win-
signal. Because of this a new approach was suggested.      dows were taken, 12 windows for each event case (i.e.
This second approach is PTT segmentation, described        normal breathing, apnea, and hypopnea) were taken
below, Figures 10 a-c:                                     from the PSG records of three patients. It should be
                                                           noted that the selected test record sections contain no
                                                           training patterns. Three testing set were created each
containing one class of desired outputs. For more de-         network had a problem differentiating between normal
tails, see [20].                                              breathing and hypopnea.
                                                              Due to this, the average of NAF amplitude in these
5.4    Network parameters
                                                              cases is almost the same as the one in a normal breath-
The network outputs were chosen to be binary values,          ing event or a hypopnea event.
coding the respiration patterns corresponding to the ac-
                                                              The reason to choose this element to be part of the neu-
tual input training pattern.
                                                              ral network input is following: when the window con-
[1; 0; 0]: Normal breathing event                             tains the beginning or the end of an apneic event, it can
                                                              be taken the first or last point from a normal breathing,
[0; 1; 0]: Apneic event (OSA, CSA, MA)
                                                              Figure 12.
[0; 0; 1]: Hypopnea
It should be noted that a hidden layer should be needed       Table 1. Neural Network Input specification.
to add for further analysis such as classification be-
tween OSA, CSA or MA.                                            Neural Net-     First     Second     Third    Fourth
                                                                 work Input      Net-       Net-      Net-      Net-
The feature extraction and selection helped us to build                          work       work      work      work
the neural network input. Four simple different neural
networks were created using three structures, Figure            Average of        Yes        Yes       Yes      Yes
                                                                NAF ampli-
11. Table 4 describes neural networks inputs for the
                                                                tude times
first neural network. The input vector of the second
                                                                total maxi-
neural network is composed of two elements. The third           mum peaks
and the fourth neural networks have three elements in
their input vector.                                              Average of       Yes        No        No        No
                                                                 each NAF
                                                                 amplitude
                                                                  duration
                                                                 Average of       No         Yes       No        No
                                                                   PTT for
                                                                 each ampli-
Figure 11. The architecture of the first neural network: an
                                                                  tude seg-
array of one row and four columns, where: LW {1, 1} =
                                                                  mentation
First layer weight coefficient, b {1} = First layer thresh-
old, LW {1, 2} = Second layer weight coefficient, b {2} =        Average of       No         No        No        No
Second layer threshold.                                           PTT dura-
                                                                 tion of each
                                                                  segmenta-
6.     Results                                                       tion
The feature extraction and selection helped us to build          Average of       Yes        No        Yes      Yes
these neural network inputs. The testing set used was            PTT ampli-
the same for all the neural networks in order to com-           tude for each
pare the results between them.                                    breathing
                                                                    cycle
One of the reasons could be that one of the elements of
the neural network input, listed in Table 1, is the aver-        Average of       Yes        No        Yes      Yes
age of NAF amplitude multiply by the total amount of             each PTT
maximum peaks in 14 second window.                               amplitude
                                                                  duration
Table 2. First Neural Network results.
                                                              In order to avoid the misclassification between apnea
 Classifica-    Normal       Apnea       Hypopnea             and normal breathing the amplitude average was mul-
    tion       breathing                                      tiplied by the amount of maximum peaks presented on
 Accuracy        75 %       83.33 %        75 %               the window, so this assures us to have a good differen-
                                                              tiation between the normal breathing and apnea. How-
Table 2 show that, for the first network, some (three)        ever; this multiplication could lead to a misclassifica-
normal breathings events were misclassified as hy-            tion between apnea and hypopnea. Another way of
popneic events. It can be also noted that the neural          solving these misclassifications is modifying the win-
dow start time. From the left picture of Figure 12, it     To make a better comparison among all of the neural
could be seen that the amplitude average multiplied by     networks their diagnostic accuracy [21] is calculated
                                                           for each of them: Accuracy = Se*P(A) + Sp*P(N),
                                                           where
                                                           Se is the sensibility = (number of true positive) / (num-
                                                           ber of subjects with this disease)
                                                           Sp is the specificity = (number of true negative) /
                                                           (number of subjects without the disease)
                                                           P(A) is the fraction of subjects with the disease, P(N) is
                                                           the fraction of normal subjects.
Figure 12. Ending of Apnea and Beginning of Apnea.         The true positive is the situation when the test is posi-
the amount of peaks can still give a result similar for    tive for a subject with the disease and the true negative
the hypopnea case. It may also appear a misclassifica-     represents the case when the test is negative for a sub-
tion between normal breathing and hypopnea, because        ject who doesn’t have the disease. The results are
sometimes the 14 second window contains more               summarized in Table 6.
maximum peaks for the hypopnea event than the nor-
mal breathing event. The replacement of this element       Table 6. Diagnostic accuracy calculation for each neural
to the average of NAF amplitude would not help to          network.
solve those misclassifications. Instead, it showed worst
results.                                                      CALCULATIONS          First   Second    Third   Fourth
                                                                                    Net-     Net-     Net-     Net-
From Table 3, it can be seen that some (three) apneic                               work     work     work     work
events were misclassified as hypopneic events. One of
the reasons of these misclassifications was mentioned
                                                                 True positive       22       22       22       23
above. Table 4 shows some (two) hypopneic events
misclassification. The third neural network was created            Amount
to show that the window start time affects in the mis-          True negative         9       12       12       12
classifications. Table 5 shows a small (one) misclassi-            amount
fication between an apneic event and a hypopneic
event and yields better results than the Table 4. This        Sensibility in per-   91.7     91.7     91.7     95.8
proofs that the window start time affects the neural               centage
networks inputs.
                                                              Specificity in per-    75      100       100     100
                                                                  centage
Table 3. Second Neural Network results
                                                              Fraction of breath-   0.667   0.667     0.667   0.667
  Classifica-      Normal        Apnea     Hypopnea           ing event with the
     tion         breathing                                          SAS
                                                              Fraction of normal    0.333   0.333     0.333   0.333
  Accuracy         100 %         75 %       91.67 %
                                                               breathing events
Table 4. Third Neural Network Results.                        Diagnostic Accu-
                                                              racy in percentage
                                                                                    86.10   94.50     94.50   99.97
  Classifica-      Normal       Apnea     Hypopnea
     tion         breathing
   Accuracy        100 %        100 %      66.67 %         To calculate the diagnostic accuracy, the results were
                                                           divided into two sets: the normal breathing and the
                                                           SAS which includes both apnea and hypopnea.
Table 5. Forth Neural Network Results.
                                                           In this experiment we used 36 windows: 24 of them
                                                           containing the disease SAS, and the rest normal breath-
 Classifica-     Normal       Apnea      Hypopnea
                                                           ing
    tion        breathing
                                                           The accuracy in Tables 2-5 was calculated for each
 Accuracy        100 %        91.66 %     100 %
                                                           type of event by the following formula:
                                                           (Amount of correct detected window/Total amount of
                                                           windows)*100
The fourth neural network presents the best diagnostic     De Physiologie et d’Exploration Fonctionnelle at Hôpi-
accuracy of 99.97%, followed by the second and third       tal Raymond Poincaré, France in particular Dr. MA
ones with an accuracy of 94.50%. The first neural net-     Duera Salva for providing the data used in this work
work has an accuracy of 86.10%. These proves that the      and their help.
window start time affects the neural network accuracy,
                                                           The work of Tamara Pozzo and Lenka Lhotska has
the extraction segmentation approach for the PTT sig-
                                                           been supported by the research program "Information
nal is better than the amplitude extraction of each
                                                           Society" under Grant No. 1ET101210512 "Intelligent
breathing cycle approach for the PTT.
                                                           methods for evaluation of long-term EEG recordings".

7.    Conclusion                                           References
In this preliminary work, our objective was just to ex-
                                                           [1] AASM. "Sleep-Related Breathing Disorders in Adults:
amine the possibility to use the NAF and PTT records       Recommendations for Syndrome Definition and Meas-
for automatic sleep apnea classification.                  urement Techniques in Clinical Research". The Report of
We manage to make a differentiation between normal         an American Academy of Sleep Medicine Task Force,
breathing, apnea, and hypopnea. Four neural networks       SLEEP, Vol. 22, No. 5, (1999).
were created, based on different feature extraction and    [2] Eliot S. Kaltz, Janita Lutz, Cheryl Black, and Carole
selection. The best diagnostic accuracy was obtained       L. Marcus, "Pulse Transit Time as a measure of arousal
by the fourth neural networks, showing a diagnostic        and respiratory effort in children with sleep-disorder
accuracy of 99.97%.                                        breathing", in Pediatric research, April 1, 2003; Vol. 53,
                                                           No.4, Pages 580-588.
The first neural network showed worst results than the
                                                           [3] T. Al-Ani, Y. Hamam, R. Fodil, F. Lofaso and
third neural network. This proves that the extraction by   D.Isabey, "Using Hidden Markov Models For Sleep Dis-
segmentation approach for the PTT signal is better than    orderd Breathing Identification". Simulation Practice and
the amplitude extraction [19] of each breathing cycle      Theory, Special Issue, Vol. 12, Issue 2, Elsevier, Mai
for the PTT.                                               2004.
 This preliminary study of automatic diagnosis system      [4] Z Shinar, A Baharav, S Akselrod, “Obstructive Sleep
was not designed for real online diagnosis, due to its     Apnea Detection Based on Electrocardiogram Analysis”,
drawbacks such as the binary output, and the window        IEEE Transactions on Computers in Cardiology, 2000;
start time dependence. Some implementations and            Vol. 27, Pages 757-760.
modifications must be done before obtaining a real         [5] Robin P S. et al. "Pulse transit time: an appraisal of
automatic diagnosis system. The binary output should       potential clinical applications". THORAX, vol. 54, No 5,
be changed, in order to be understandable for the doc-     (1999) 452-457.
tors. Furthermore, the type of window selection should      [6] John Ronald, Kenneth Delaive, Les Roos, Jure Man-
be changed, and finally more data must be collected.       freda, and Meir H Kryger, “Obstructive Sleep Apnea pa-
For further studies, more data should be added and it      titents use more health care resources ten years prior to
will be necessary to add more than one hidden layer to     diagnosis”, Sleep Research Online, 1998; Vol.1, No.1,
the neural networks.                                       Pages 71-74.
The feature extraction and selection of the PTT signal      [7] Péter Várady, Szabolcs Bongár, Zoltán Benyó, “De-
and the NAF signal was the most difficult part to im-      tection of Airway Obstructions and Sleep Apnea by ana-
                                                           lyzing the Phase Relation of Respiration Movements Sig-
plement in the artificial neural network design. We be-
                                                           nals”, IEEE Transactions on Instrumentation and Meas-
lieved that further improvements can be achieved in
                                                           urement, February 2003; Vol.52, No.1, Pages 2-6.
order to get better input data for neural network. This
part is one of the most important, because most of the     [8] Fu-Chung Yen, Khosrow Behbehani, Edgar A. Lucas,
neural network accuracy depends on the step of feature     John R. Burk, John R. Axe, “A Noninvasive Technique
                                                           for Detecting Obstructive and Central Sleep Apnea”,
extraction and selection.
                                                           IEEE Transactions on Biomedical Engineering, December
The authors are presently investigating the use of Hid-    1997, Vol. 44, No. 12, Pages 1262-1268.
den Markov Models (HMMs) and make some com-                [9] Luis Antonio Aguirre, Vinícius C. Barros, Álvaro V.P
parisons between the approaches presented in this pre-     Souza, “Nonlinear multivariable modeling and analysis of
sent work and in [3].                                      sleep apnea time series”, Computers in Biology and
                                                           Medicine, May 1999; Vol. 29, No. 3, Pages 207-228.
                                                           [10] Z Shinar, A Baharav, S Akselrod, “Obstructive Sleep
8.    Acknowledgements                                     Apnea Detection Based on Electrocardiogram Analysis”,
The authors would like to thanks Daniel Coquelle,          IEEE Transactions on Computers in Cardiology, 2000;
manager engineer in MEDATEC France, the Service            Vol. 27, Pages 757-760.
[11] M. Cabrero-Canosa, M. Castro-Pereiro, M. Grana-
Ramos, E. Hernandez-Pereira, V. Moret-Bonillo, M. Mar-
tin-Agana, H. Verea-Hernando, “An intelligent system for
the detection and interpretation of sleep apneas”, Expert
Systems with Applications, May 2003; Vol. 24, No.4,
Pages 335-349.
[12] Péter Várady, Tamas Micsik, Sandor Benedek,
Zoltán Benyó, “A Novel Method for the Detection of Ap-
nea and Hypopnea Events in Respiratory Signals”, IEEE
Transactions on Biomedical Engineering, September
2002; Vol. 49, No. 9, Pages 936-942.
[13] Eliot S. Kaltz, Janita Lutz, Cheryl Black, and Carole
L. Marcus, “Pulse Transit Time as a measure of arousal
and respiratory effort in children with sleep-disorder
breathing”, in Pediatric research, April 1, 2003; Vol. 53,
No.4, Pages 580-588.
[14] http://www.sccn.ucsd.edu/eeglab/
[15] http://www.hsr.nl/edf/companies/companies.html
[16] Novak D., Cuesta Frau D., “Baseline Wandering
Removal Using Wavelets”, 17th international EURASIP
conference, BIOSIGNAL, Brno,Czech Republic, 2000.
[17] Jiapu Pan, Willis J. Tompkins, “A Real-Time QRS
Detection Algorithm”, IEEE Transactions on Biomedical
Engineering, March 1985; Vol.32, No.3, Pages 230-237.
[18] Z Shinar, A Baharav, S Akselrod, “Obstructive Sleep
Apnea Detection Based on Electrocardiogram Analysis”,
IEEE Transactions on Computers in Cardiology, 2000;
Vol. 27, Pages 757-760.
[19] J. Pagani, MP Villa, G Calcagnini, E Lombardozzi, F
Censi, S. Poli, P Bartolini, V Barbaro, R Ronchetti, “De-
tection of Central and Obstructive Sleep Apnea in Chil-
dren using Pulse Transit Time”, Computers in Cardiol-
ogy, September 2002; Vol. 29, Pages 529-532.
[20] Paula Tamara Pozzo Mendoza, , “Automated diagno-
sis system for sleep apnea classification”, Master Thesis,
Czech Technical University in Prague, Faculty of Electri-
cal Engineering, Department of Cybernetics, Prague
2005.
[21] Rangaraj M. Rangayyan, “Biomedical Signal Analy-
sis: A Case-Study Approach”, IEEE Press Series on Bio-
medical Engineering, January 2002; 552 Pages.