Sleep-Wake Classification Based on Electrocardiogram

records
Hayet Werteni#1, Slim Yacoub*2, Noureddine Ellouze#3
#
Signal, Image and pattern recognition research unit
Dept. of Genie Electrique, ENIT
BP 37, 1002, The Belvédère, Tunisia
1
[email protected]
3
[email protected]
*
INSAT, Dept. of Physique and Instrumentation,
BP 676,1080, Centre Urbain, Tunisia
2
[email protected]

Abstract—This paper presents application of an automatic showed significant differences in parameters of heart rate
classification system on MIT-BIH Polysomnographic Database. variability during all stages of sleep. Recently, new
This system aimed to score sleep-wake stages using only the approaches have emerged [6], [2], [7] and [8] which classified
electrocardiogram (ECG) records. The feature extraction stage sleep stages based only on the RR interval, which used in the
described in this paper was performed using methods of Heart extraction phase detrended fluctuation analysis (DFA), the
Rate Variability analysis (HRV) and Detrended fluctuation analysis of heart rate variability analysis (HRV) and
analysis (DFA). Those features will be evaluated using a neural windowed detrended fluctuation analysis (PDFA).
network algorithm called Extreme Learning Machine (ELM).
In this paper, we present a method of sleep-wake stages
We make a comparative study of our algorithm using two
classifiers back propagation neural network (BPNN) and support scoring using ECG signal. The feature extraction was
vector machine (SVM). performed using the methods of Heart Rate Variability
analysis and detrended fluctuation analysis (DFA). These
Keywords— ECG, Heart rate variability, Detrended fluctuation methods were tested on the MIT/BIH Polysomnographic
analysis, sleep stages, Extreme learning machine. Database (MITBPD) using the Extreme Learning Machine
(ELM). The proposed method was compared using the back
I. INTRODUCTION propagation neural network (BPNN), or support vector
Recently, the analysis of the sleep becomes an important machine (SVM) classifiers.
issue by reason the considerable number of pathologies .
associated with sleep periods. Therefore, a good analysis of
II. METHODS
sleep was presented with a classification of the night stages.
For this reason, a polysomnography (PSG) was done; which is In this study, ECG epochs were classified in one of two
a reference test for quantifying and qualifying sleep disorders. categories “Awake” and “Sleep” .The classification system
It is presented by the recording of several physiological used in this paper consists of three steps: Firstly a
variables during the night's sleep [1], in order to classify the preprocessing is applied on the ECG signal in order to build
different phases of sleep. These variables was: the electro- the RR series. Secondly, the feature extraction is used by the
encephalogram (EEG), electro-oculogram (EOG), electro- Heart Rate Variability analysis (HRV) and Detrended
myogram (EMG), respiratory effort, air flow entering through fluctuation analysis (DFA) methods. Finally, we adopt the
nose and mouth, the saturation of oxygen in hemoglobin, the simple learning algorithm for a single-hidden layer
electro-cardiogram (ECG) and effort of the thoracic and feedforward neural network called Extreme Learning Machine
abdominal muscles. ECG recording is one of the simple (ELM) classifier for the classification phase. Finally, the
technologies in sleep disorders detection. Cyclic variations in performance of ELM method was compared with the back
RR intervals of ECG signals have been reported to be propagation neural network (BPNN)-based or the support
associated with sleep apnea and different sleep stages [2]. vector machine (SVM)
Various studies have confirmed that several methods could
A. Preprocessing
possibly recognize the stages of sleep from the ECG signal.
All these studies was based on analysis of RR interval The feature extraction phase is based on accurate QRS
variability, which present the variation in time period between complex detection. For this reason we develop a new
consecutive heart beats and gives a measure of the modulation algorithm for R wave's locations using the multiscale wavelet
of sympathetic and parasympathetic nervous system. During analysis, based on singularity detection of local maxima of the
sleep, the study of this variability can show the modification wavelet coefficients [9].We used the QRS annotations to build
of the cardiovascular system regulation based on sleep cycles the RR intervals, which defined as the interval from the peak
[3]. Several studies [4], [5] were investigated in blood of one QRS complex to the peak of the next.
pressure changes of breathing with the sleep depth, and
B. Feature extraction  TOTPWR: Total spectral power of all RR intervals
There are ten features used in this paper extracted for every up to 0.04 Hz
30-s, using the Heart Rate Variability and Detrended  VLF: Total spectral power of all RR intervals
fluctuations analysis methods. between 0.003 and 0.04 Hz
 LF: Total spectral power of all RR intervals between
1) Heart Rate Variability (HRV): The HRV is a method 0.04 and 0.15 Hz.
used to interpret the oscillation between consecutive heart
 HF: Total spectral power of all RR intervals between
beats and oscillations in consecutive instantaneous heart rates.
0.15 and 0.4 Hz
The features that are extracted by this method can be divided
 LF/HF: Ratio of low to high frequency power
into two categories: time domain measures and frequency
domain measures [10]. Due to short epoch duration, only short 2) Detrended fluctuations analysis (DFA): In this paper we
term time-domain and frequency domain HRV features are present the DFA method for quantifying the correlation
used in this study [2]. property in non-stationary time series. First, the RR interval
time series is integrated [12]:
As time domain, we calculated the following measures:
 k
y  k     RR  i   RR 
RR mean: Average of all RR intervals (3)
 SDNN: Standard deviation of all RR intervals
i 1
 RMSSD: Square root of the mean of the squares of
differences between adjacent RR intervals k  1,..., N
Before the extraction of the frequency measures, a
preprocessing was carried out on the RR interval based on the Where RR is the average RR .Then, the integrated series
fact that the spectral analysis cannot be affected only on y  k  is divided into segments of equal length n. With in each
regularly sampled signals [3]. This preprocessing is presented
by resampling the data at a frequency of 1 Hz using an segment, a least squares line is fitted into the data [14]. Let
interpolation method developed by Berger et al [11]. Thus, the yn  k  denote these regression lines. Next the integrated
y  k  is detrended by subtracting the local trend within
frequency measures were extracted from the power spectral
density of the RR series by the Autoregressive Model (AR). series
The AR has fine properties to deal with non-stationary time each segment and the root-mean-square fluctuation of this
series [7]. For this reason, the RR series is modeled by the integrated and detrended time series is calculated by [14]:
autoregressive model (AR) which is presented as follows:
p
x   ap(i ) x  n  i   e  n  (1) 1 N
F  n    y  k   yn  k  
2 (4)
i 1
N k 1
Where is the filter order, is the sample of
intervals, are the coefficients of the model, x
F  n Characterizes the fluctuations of the indefinite integral
denotes the prediction output. is the prediction error, of the series RR i  , Corrected by her trend. F  n is
which present a white noise which has variance  . The
2
computed for all time-scales n. Typically, F  n increases
estimation of coefficient of the AR model was determined by
the Levinson-Durbin algorithm of order 13. The selection of with n, the “box-size”. If log  F  increases linearly
this order can be verified by “arfit toolbox” [12], [13]. This
selection is a compromise, because the proper order may vary. with log  n  , then the slope of the line relating F  n and n
This is caused by short interval length and by the natural in a log-log scale gives the scaling exponent, which can be
diversity in heart rates [13]. used to discriminate between sleep stages [15]. The DFA
measures consist of:
The coefficients are used to estimate the power  : Computed from the basic boxel size n = 6 to
spectral density of the RR intervals, according to the boxel size n = 16[16].
following expression:  : Computed from basic boxel size n = 6 to boxel
size of the total number of samples in the analyzed
2
Par  f   2 (2)
interval [16].
p
1   ap(i )e  j 2 fi C. ELM Classification
i 1 The classification is performed using a recent neural
network algorithm: the Extreme Learning Machine (ELM).
From the spectrum of RR series presented for each sleep stage,
The ELM resolves difficulties of a neural network through a
we selected the following spectral indexes:
fast learning speed and high performance [17]. This algorithm
makes up for the defaults of gradient-based learning algorithm The data files contain the header file, the QRS annotations
by analytically calculating the optimal weights of the Single file and the sleep stages file. The header file is a text file
Hidden Layer Feedforward Neural Network (SLFN). In the containing the information related to the record such as, the
beginning, weights between the input layer and the hidden types of the signals, calibration constants, the length of
layer are randomly selected to make the SLFN into a linear recordings, and the anthropometric data of the subject. The
equation, and after we can get the optimal value for weights QRS annotations give the time occurrence of each QRS
between the output layer and the hidden layer by calculating complex, whereas sleep stages file contains the class label of
the linear matrix equation [17]. each epoch, i.e., wake, REM sleep, stages 1, 2, 3 or 4 of the
The ELM with a single hidden layer can be summarized as NREM sleep, epoch with MT (movement time) annotation
follows [18]: was not included in this work.
Given a training set:
B. Sleep stage classification
In our work, all records of the MITBP database are used,
by eliminating for every signal some periods of 30 seconds
Where is a training sample and is the corresponding which are not exploitable. We used 2 classes of sleep stage as
target value, the activation function and the number of shown in Table I, where: Wake (W) is a period when no sleep
hidden neurons , perform the following steps. stages can be classified. Sleep (S) period includes sleep stage
1 or 2 or 3 or 4 or REM sleep. The classification of these two
Step1: Assign arbitrary input weight and bias , classes of sleep was carried out by the ELM classifier by
fixing the number of input neurons to 10, which the
corresponding number of features extracted. The output is set
Step2: Calculate the output matrix at the hidden layer at the number of classes that is equal to 2 to classify 'W /S'.
The number of hidden neurons is set at 300 after a study of
H  g (w.x  b) (5) several configurations, this is the learning process. The
activation function chosen is the sigmoid function. The
Step 3: Calculate the output weight β classification of sleep-wake stages was performed with two
scenarios: subject- specific system and subject-independent
  H †T (6)
system. A subject specific classification is trained by
Where: selecting epochs from 1/3 of the night’s sleep, remaining 2/3
of the night’s sleep was used to test the system [8]. A subject-
 t1T  independent classification was constructed using training
  epochs drawn from all subjects. For this reason, we construct
  the training set by 1/3 of the database and 2/3 for testing. In
T   Table II and Table III, we present the results of independent
  and specific classification. We note from these results that the
  rate of classification is different from wake to sleep class. In
t T  the Wake class, classification rate is always less than a sleep
 N class. A subject specific classifier was trained and tested, and
yielded an estimated classification accuracy of 83.11 %.When
H † is the Moore-Penrose generalized inverse of the a subject-dependent classifier was trained and tested, the
matrix H . estimated classification accuracy dropped to 69.13%. The
performance of the classifier on the subject independent task
III. RESULTS AND DISCUSSION is poor, compared to specific classification. The results show
can be obtained to classify sleep-wake stages when we use the
methods of HRV and DFA. In fact the HRV features are the
A. Database markers of autonomic nervous system which changes during
The approach presented in this paper is tested using ECG the stages of sleep. Also, DFA is a successful method for the
registrations from the MIT-BIH polysomnographic (MITBP) classification of sleep stages [15].
database. This database is a collection of recordings of
multiple physiologic signals during sleep with sampling rate
250Hz. The database contains over 80 hours' worth of four-, TABLE I NUMBER OF EPOCHS W/S FOR TRAINING AND TESTING DATA IN
RECORDS OF THE MITBPD
six-, and seven-channel polysomnographic recordings, each
with specifically an ECG signal annotated beat-by-beat. The Wake Sleep
16 patients (16 males aged 32 to 56) in this database suffer Training Data 1937 4736
from a condition known as sleep apnea in which they are Testing Data 968 2368
repeatedly awoken during the night due to their breathing
being interrupted [19].
 TABLE II. RESULTS FOR SUBJECT-SPECIFIC SYSTEM classification is tested on two types of give; the subject-
Classification rate (%) specific and the subject independent. The performance of the
Test Train classifier on the subject independent task was poor, compared
Wake(W) 69.13 98.91 to specific classification, this is logical since similarities exist
Sleep(S) 83.11 99.17 in records of the same person. Subject-independant scheme
TABLE III. RESULTS FOR SUBJECT-DEPENDENT SYSTEM
gives less good results because of big differences between
persons' data.
Classification rate (%)
Test Train
Wake(W) 50.59 97.53
Sleep(S) 68.99 98.47 REFERENCES
[1] D. Stéphanie, “ Classification automatique en stades du sommeil :
extraction de caractéristiques et comparaison des principaux
C. Comparison with other classifiers classificateurs,” thesis, Académie universitaire Wallonie-
Bruxelles .2007.
We make a comparative study of ELM algorithm with the [2] A. Viola, “ La Variabilité Cardiaque au Cours des Cycles de Sommeil
back propagation neural network (BPNN) [20] and the support chez l'Homme,” Thèse, l'Université de STRASBOURG .2004.
vector machine (SVM) [21]. BPNN is a learning algorithm for [3] M.Sami, M. Isa, I. Wasito,M. Aniti, “ Kernel Dimensionality
Reduction on Sleep Stage Classification using ECG Signal,” IJCSI
the multi-layer neural network, this performance are based on: International Journal of Computer Science Issues, Vol. 8, Issue 4, No 2,
the learning rate, the momentum, the number of training July .2011.
iterations and hidden neurons, and the number of hidden [4] V.Vanoli, P.B, Adamson,B. Lin, G.D,Pinna R. Lazzara, W.C. Orr,
layers used. In this paper, the BPNN are part of these “Heart Rate Variability During Specific Sleep Stages Circulation,”
91,pp.1918–1922, 1995.
parameters: the number of hidden layer is equal to that in the [5] M. Malik. Chairman of Writing Committee of Task Force of the
ELM case as 1, and the learning rate of 0.01, the momentum European Society of Cardiology and the North American Society of
constant of 0.4, iteration number 400, and a number of hidden Pacing Electrophysiology. Heart Rate Variability: Standards of
neurons 500 after a study of several configurations. Measurement, Physiological Interpretation, and Clinical Use,
Circulation, 93, pp.1043—1065.1996.
For the SVM, we used the LibSVM toolbox [22]. Thus using [6] M.Adnane, Z. Jiang, “Automatic sleep-wake stages classifier based on
SVM, there are these parameters to be determined: the kernel ECG,” ICROS-SICE International Joint Conference .2009.
parameter,  and the constant of regularization C . In this [7] MO. Mendez, M. Matteucci, S. Cerutti, F. Aletti, A.M.Bianch, “Sleep
Staging Classification Based on HRV: Time-Variant Analysis,” 31st
paper, we use the following couples C ,  [23]: Annual International Conference of the IEEE EMBS Minneapolis,
Minnesota, USA, September 2-6.2009.
C   2 0 , 21 ,..., 216  and [8] S.Redmond, C. Heneghan,“ Electrocardiogram- BasedAutomatic Sleep
Staging in Sleep Disordered Breathing,” Computers in Cardiology
  2( 15) , 2 14 ,..., 2 2 
30:pp.609—612.2003.
[9] H.Werteni, S.Yacoub, N. Ellouze,“ Multiscale products for the
detection of an electrocardiogramme R wave :MIT-BIH
The results of this comparison are summarized in TableIV and polysomnographic,”6th International Conference: Sciences of
TableV. We conclude that the performance of the proposed Electronic Technologies of Information and Telecommunications. 2012.
algorithm showed the best performance, the rate of [10] American Academy of Sleep Medicine Task Force. Sleep-related
classification with SVM was slightly lower, but the BPNN breathing disorders in adults: Recommendations for syndrome
definition and measurement techniques in clinical research Sleep.
algorithm were low compared to our proposed algorithm. vol.22,pp. 667—689.1999.
TABLE IV RESULTS FOR SUBJECT-SPECIFIC SYSTEM [11] RD.Berger, S.Akselrod, D.Gordon,“An efficient algorithm for spectral
analysis of the heart rate variability,” IEEE Trans Biomed Eng,
Testing ELM BPNN SVM vol.33,pp. 220—222.1986.
accuracy (%) [12] T. Schneider, A. Neumaier ,“Algorithm. arfit—a matlab package for
Wake(W) 69.13 64.99 69.66 the estimation of parameters and eigenmodes of multivariate
autoregressive models,” ACM Transactions on Mathematical Software,
Sleep(S) 83.11 80.83 83.06
vol.27,pp.58—65.
[13] T.Al-ani, R. Kazbunda, D.Nov´ak ,“ Automatic sleep scoring based
only on electrocardiogram records,” ISBN 978-3-901608-32-2. Proc.
TABLE V RESULTS FOR SUBJECT- DEPENDENT SYSTEM EUROSIM .2007.
[14] .A. Absil, “ Analyse non linéaire de signaux cardiologiques en vue
Testing ELM BPNN SVM
d’applications cliniques,” Travail de fin d’ études, niversité é de
accuracy (%) Liège 1998.
Wake(W) 50.59 45.78 49.22 [15] T. Penzel, J.W. kantelhardt, L.Grote, J.H. Peter, A. Bunde,
Sleep(S) 68.99 58.88 68.25 “ Comparison of detrended fluctuation analysis and spectral analysis
for heart rate variability in sleep and sleep apnea, ” IEEE Trans on
IV. CONCLUSIONS Biomedical Engineering, 50(10),pp.1143–1151.2003.
In this work, we studied the feasibility of sleep-wake [16] T.Al-ani R . Kazbunda , D.Nov´ak , “Automatic sleep scoring based
only on electrocardiogram records,”ISBN 978-3-901608-32-2,Proc.
classification using only the electrocardiogram signal. The EUROSIM. 2007.
features extraction is realized using HRV and DFA methods. [17] K.Jinkwon, S.Hang, S. Kwangsoo, L. Myoungho ,“Robust algorithm
We evaluated the performance of these features using three for arrhythmia classification in ECG using extreme learning machine,”
different classifiers: ELM, BPNN and SVM. This approach of
 BioMedical Engineering OnLine 8--31 doi, 10.1186/1475-925X-8-31.
2009.
[18] C.Kwak and O.Wook Kwon , “Cardiac Disorder Classification Based
on Extreme Learning Machine,” World Academy of Science,
Engineering and Technology 48 .2008.
[19] A.Goldberger,L. Amaral, L. Glass, M. Hausdorff,P.CH. Ivanov,
R.G.Mark, J.E, Mietus, G.B, Moody, C-K, Peng ,H.E, Stanley
PhysioBank, PhysioToolkit, and PhysioNet: Components of a New
Research Resource for Complex Physiologic Signals. Circulation, Vol.
101, No. 23, pp. 215–220.2000.
[20] RO, Duda ,PE. Hart , DG. Stork, Pattern classification 2nd edition,
New York. Wiley.2001.
[21] H.Wei-Chih, C. Chih-Chung, L. Chih-Jen, “A Practical Guide to
Support Vector Classifcation National,” Taiwan University, Taipei
106, Taiwan.2009.
[22] (2013) http://www.csie.ntu.edu.tw/~cjlin/libsvm/
[23] S. Sameh , Z. Lachiri, “Multiclass support vector machines for
environmental sounds classification in visual domain based on log-
Gabor filters,” Springer International Journal of Speech
Technology ,ISSN 1381-2416.2012.