Automatic Sleep Stage Classification - IEEE Xplore

7 downloads 140505 Views 1MB Size Report
Email: *[email protected]. Abstract-Automated sleep stage classification is essential for alleviating the burden of physicians since a large volume of ...
Proceedings of International Conference on Electrical Infonnation and Communication Technology (EICT 20 I 5)

Automatic Sleep Stage Classification Ahnaf Rashik Hassan* and Mohammed Imamul Hassan Bhuiyan Department of Electrical and Electronic Engineering Bangladesh University of Engineering and Technology, Dhaka, Bangladesh Email: *[email protected]

Abstract-Automated sleep stage classification is essential for alleviating the burden of physicians since a large volume of data

Segment the

have to be analyzed per examination. Most of the existing works

EEG sig nal into

IMFsof

in the literature are multichannel based or yield poor classifi­

30s epo chs

CEEMDAN

Calculate the

cation performance. A single-channel based computerized sleep

Comp ute Statistical Features from the IMFs

staging scheme that gives good performance is yet to emerge. In this work, we introduce a novel noise assisted decomposition scheme to perform automatic sleep stage classification from single channel EEG signals. At first, we decompose the EEG signal seg­ ments into mode functions using Complete Ensemble Empirical

Classify Using AdaBoost

Mode Decomposition with Adaptive Noise (CEEMDAN). Various statistical moment based features are then computed from these

Perfrom Statistical Analysis

mode functions. The effectiveness of statistical moment based features is validated by statistical analysis. In this work, we also introduce Adaptive Boosting for sleep stage classification. Experimental outcomes manifest that the computerized sleep

Fig. l. A schematic outline of the proposed computerized sleep staging method.

staging scheme propounded herein outperforms the state-of-the­ art ones in various cases of interest. Index Terms-EEG, Sleep Scoring, CEEMDAN, AdaBoost.

I.

INTRODUCTION

Sleep staging is traditionally performed by expert scorers based on visual observation of Polysomnographic (PSG) sig­ nals which is time-consuming, burdensome, error-prone, and subjective. Automated sleep staging therefore, can eradicate the burden of the physicians, expedite sleep disorder diagnosis, and benefit sleep research. Again, the push toward a wearable, portable yet low-power sleep quality monitoring device neces­ sitates the use of minimum number of channels and reduced computational cost. Most of the existing works, on account of being multichannel or multiple physiological signal based, are unsuitable for device implementation. As a result, automatic sleep scoring based on single channel Electroencephalogram (EEG) is gaining interest in sleep research community. Various single-channel, multichannel and multiple phys­ iological signal based method have been proposed in the literature for automated sleep staging. Long et al. [1] computed various respiratory amplitude, depth and volume based features from respiratory effort signals and used linear discriminant classifier to perform sleep classification. Zhu et al. [2] gener­ ated difference visibility graph (VG) and horizontal VG from single channel EEG signal and extracted nine features from them to classify using support vector machine. Huang et al. [3] used two channel of forehead EEG signals, extracted spectral features by short-time Fourier transform from the signals and performed classification using relevance vector machine. Koch et al. [4] proposed a Latent Dirichlet Allocation topic

978-1-4673-9257-0/15/$3\.00 ©2015 IEEE

211

model based method using four-channel multiple physiological signals (EEG and EOG) for sleep staging. Lajnef et al. [5] employed various features such as linear prediction error energy, variance, skewness, kurtosis, permutation entropy and multi-class support vector machine to perform automatic sleep scoring based on multichannel EEG, EOG and EMG signals. Kayikcioglu et al. [6] proposed an AR coefficient-based fea­ ture extraction scheme and utilized partial least squares (PLS) algorithm to classify sleep stages. In this work, we propound a single channel EEG based sleep scoring algorithm. The outline of our algorithm is presented in Fig. 1. After decomposing the EEG signals into 30s epochs, we further decompose the EEG signal segments into intrinsic mode functions using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). We then extract various statistical features from the intrinsic mode functions. The efficacy of the proposed feature extraction scheme is investigated using statistical analysis. Afterwards, sleep stage classification is perfonned using an eminent en­ semble learning based classification model, namely- Adaptive Boosting (AdaBoost). II.

MATERIALS AND METHODS

A. Experimental Data The data-set used for evaluation of the proposed scheme is Physionet's Sleep-EDF data-set which is publicly available [7]. The recordings have been obtained from Caucasian males and females (21 - 35 years old) without any medication. There are eight recordings in two subsets (marked as sc* and st*). The first four recordings (sc4002eO, sc4012eO, sc4102eO, sc4112eO) have been acquired in 1989 from ambulatory

healthy volunteers during 24 hours in their normal daily life. The last four data recordings (st7022jO, st7052jO, st7121jO, st7132jO) have been obtained in 1994 from subjects who had mild difficulty falling asleep but were otherwise healthy, during a night in the hospital, using a miniature telemetry system [8]. The recordings contain horizontal EOG, Fpz-Cz and Pz-Oz EEG data, each sampled at 100 Hz. It has been suggested in various studies [9] [10] [11] [2] that EEG signal from Pz-Oz channel yields better classification performance than that of the Fpz-Cz channel. So, in the present work Pz-Oz channel EEG signal is used. The sc* recordings also contain the submental-EMG envelope, oro-nasal airflow, rectal body temperature and an event marker, all sampled at sampling rate of I Hz. The st* recordings contain submental EMG sampled at sampling rate of 100 Hz and an event marker sampled at sampling rate of 1 Hz. Expert scoring of the EEG data is obtained from the same source. Each 30s of EEG data has been scored in accordance with the R&K recommendations [12]. The duration of each epoch is 30s or 3000 data points. Each epoch was scored by expert scorers in one of the eight classes: AWA, Sl , S2, S3, S4, REM, MVT (Movement Time) and 'Unscored' . Table I summarizes the number of epochs of different classes that are used in this work. In total, 15,188 EEG epochs are used in this study. This study uses six sleep stages in accordance with R&K standard: AWA, SI-S4 and REM. The 5-state stages of sleep combine S3 and S4 of 6-state as Slow Wave Sleep (SW S) and the 4-state stages combine SI and S2 of 5-state. The 3-state and 2-state stages of sleep include: AWA, NREM, REM and AWA, Sleep (SI-S4 and REM) respectively.

co

500



-5000

IMF

3

t ��1/1-'�'''.-' ' '''':-' ...,-:.....,,,.. _. ._.-: -_+�'_:' --,

o

5

10

15

20

25

. ..,,,1

... ._. -

30

time (5) Fig. 2. I!vI F3 of CEEMDAN of various sleep stages. Note the variations of the IMFs in different sleep states.

3) Obtain the envelope of local maxima Vmax and that of local minima Vmin using cubic spline interpolation. 4) Generate the local mean curve m by generating the upper and lower envelopes: m=

Vmax + Vmin

(I)

2

5) Compute h2 by subtracting the local mean curve from

hI:

(2) B. Empirical Mode Decomposition

6) Repeat steps 2)-5) until the difference between hk +1 and hk (SD(k)) defined as follows reaches a predefined value E.

Empirical Mode Decomposition (EMD) aims to generate highly localized time-frequency estimation of a signal in a data-driven fashion by decomposing it into a finite sum of intrinsic mode functions (IMF) or modes. Each mode must satisfy two conditions [13]:

2

hk+l - hk l1 SD(k) = Il 2 IIhk l1

I) Set hI = X. 2) Identify the local maxima and minima of hI.

(3)

Thus, the input signal can be decomposed into L IMFs until the residue becomes a monotonic function such that further extraction of an IMF is not possible. The input X can be reconstructed from all the IMFs as: X=

TABLE I

E

where 11. 11 is the Euclidian L2-norm. 7) Set Cl = hk as the first mode. 8) Find the residue, Tl = X - Cl. Steps I)-7) are known as sifting. 9) Substitute X in 1) with T. Repeat steps 1)-7) to find the rest of the IMFs C2, C3, .... , CL.

1) The number of extrema and the number of zero crossings must be the same or differ at most by one. 2) At any point the mean value of the envelope defined by the local maxima and the envelope defined by the local minima is zero. EMD iteratively decomposes an N-point EEG epoch X into amplitude and frequency modulated oscillatory IMFs using the following steps:



216

[20] S. Bashar, A. Hassan, and M. Bhuiyan, "Identification of motor imagery movements from eeg signals using dual tree complex wavelet transform," in Advances in Computing, Communications and Informatics (ICACC1), 2015 International Conference on, Aug 2015, pp. 290-296. [21] A. R. Hassan, "A comparative study of various classifiers for auto­ mated sleep apnea screening based on single-lead electrocardiogram," in Electrical and Electronic Engineering (ICEEE), 2015 International Conference on, Nov. 2015, pp. 1-4. [22] Y Freund and R. E. Schapire, "A decision-theoretic generalization of on-line learning and an application to boosting," Journal of Computer and System Sciences, vol. 55, no. 1, pp. 119 - 139, 1997. [23] L. Herrera, A. Mora, C. Fernandes, D. Migotina, A. Guillen, and A. Rosa, "Symbolic representation of the eeg for sleep stage classi­ fication," in Intelligent Systems Design and Applications (lSDA), 2011 11th International Conference on, Nov 2011, pp. 253-258. [24] c.-S. Huang, c.-L. Lin, L.-w. Ko, S.-Y Liu, T.-P. Sua, and c.-T. Lin, "A hierarchical classification system for sleep stage scoring via forehead eeg signals," in Computational 1ntelligence, Cognitive Algorithms, Mind, and Brain (CCMB), 2013 IEEE Symposium on, April 2013, pp. 1-5. [25] Y-L. Hsu, Y-T. Yang, J.-S. Wang, and c.-Y Hsu, "Automatic sleep stage recurrent neural classifier using energy features of eeg signals," Neurocomputing, vol. 104, no. 0, pp. 105 - 114, 2013. [26] A. R. Hassan, S. K. Bashar, and M. 1. H. Bhuiyan, "Automatic classi­ fication of sleep stages from single-channel electroencephalogram," in India Conference (INDICaN), 2015 Annual IEEE, Dec 2015, pp. 1-6. [27] A. Hassan, S. Bashar, and M. Bhuiyan, "On the classification of sleep states by means of statistical and spectral features from single channel electroencephalogram," in Advances in Computing, Communications and Informatics (ICACCI), 2015 International Conference on, Aug 2015, pp. 2238-2243.