Automatic Classification of Sleep Stages from Single ... - IEEE Xplore

IEEE INDICON 2015 1570202319 1   2   3   4   5   6   7   8   9   10   11   12   13   14   15   16   17   18   19   20   21   22   23   24   25   26   27   28   29   30   31   32   33   34   35   36   37   38   39   40   41   42   43   44   45   46   47   48   49   50   51   52   53   54   55   56   57   60   61   62   63   64   65  

Automatic Classification of Sleep Stages from Single-Channel Electroencephalogram Ahnaf Rashik Hassan∗ , Syed Khairul Bashar and Mohammed Imamul Hassan Bhuiyan Department of Electrical and Electronic Engineering Bangladesh University of Engineering and Technology, Dhaka, Bangladesh Email: ∗ [email protected] single channel EEG? Currently, PSG examination requires the patient to undergo overnight sleep in a specially equipped sleep laboratory. As a result, the sleep quality of the subject can be reduced due to the unfamiliar environment. A portable sleep quality monitoring system integrating an embedded system for EEG acquisition and the automatic sleep staging method with reduced channel requirement is essential with a view to making a sleep quality evaluation device feasible [3]. Thus an automatic sleep scoring algorithm can ensure wearability and portability of sleep quality evaluation at home. Excessive wire connections for PSG is often a problem that causes sleep disturbance. Automatic sleep scoring method based on single channel EEG can also reduce sleep disturbance caused by recording wires [4]. Furthermore, manual sleep scoring requires trained sleep technicians to apply visual pattern recognition to the signals. The classification of 8-hour (whole night) recording requires approximately 2-4 hours [5]. Since a large number of epochs have to be screened visually, manual sleep scoring is onerous for sleep technicians. Moreover, visual sleep scoring also makes the whole process subject to human error. An automatic sleep scoring system can not only assist the clinical staff but also speed up the diagnosis of various sleep disorders. Various single or multichannel based methods for automated sleep scoring have been reported in the literature. Berthomier et al. [6] presented a fuzzy logic based iterative method to classify sleep states from single channel EEG data. Karkovska et al. [7] extracted many features such as average amplitude, variance, spectral powers, coherence, fractal exponent etc. from data collected from six EEG channels, two EOG channels and one EMG channel and classified using quadratic discriminant analysis. Liang et al. [4] used multiscale entropy and autoregressive model parameters as features and linear discriminant analysis as classifier for single-channel automatic sleep scoring. Ronzhina et al. [8] proposed a power spectrum density and artificial neural network based method that used single channel EEG signal. Imtiaz [9] utilized spectral edge frequency, absolute and relative power of the signal for REM sleep detection from single channel EEG. Koch et al. [10] put forward a Latent Dirichlet Allocation topic model based method that uses four-channel EEG and EOG signals for sleep staging. Hassan et al. [11] utilized statistical moments and complete ensemble empirical mode decomposition with adap-

Abstract—A portable and wearable yet low-power sleep monitoring system necessitates an automatic sleep scoring algorithm with the use of minimum number of recording channels. Computer-aided sleep staging is also important to eradicate the onus of sleep scorers of analyzing an enormous volume of data. The existing works on sleep scoring are either multichannel based or yield poor performance. Therefore, an automatic sleep scoring algorithm based on single channel EEG signals is yet to emerge. In this work, we utilize spectral features to extract discriminatory information from EEG signal segments. We then perform statistical analyses to find out the efficacy and the discriminatory capability of the selected features for various sleep states. Afterwards, we employ Adaptive Boosting to perform classification. The experimental outcomes perspicuously manifest that the proposed scheme is superior to state-of-the-art ones in accuracy.

I. I NTRODUCTION Sleep is a rapidly reversible state that is characterized by a loss of consciousness and reduced responsiveness to external stimuli. Sleep related ailments deteriorate the quality of lives of humans- the commonest ones being circadian rhythm disorder, narcolepsy, hypersomnia, insomnia, and parasomnias. Sleep related complaints are second only to complaints of pain as a cause to seek medical attention. On the other hand, the function of mammalian sleep is mostly unknown. So research based on analysis of sleep data is of paramount importance not only for the diagnosis of sleep disorders but also to augment our understanding about sleep. Sleep scoring is the first step of sleep data analysis. Traditionally, all night polysomnographic (PSG) recordings are visually scored by experts based on Rechtschaffen and Kales’s (R&K) recommendations [1] or a new guideline developed by the American Academy of Sleep Medicine (AASM) [2]. Our study uses six sleep stages in accordance with R&K standard: Awake (AWA), Non-Rapid Eye Movement stages 1-4 (S1-S4) and Rapid Eye Movement (REM). The 5-state stages of sleep combine S3 and S4 of 6-state as Slow Wave Sleep (SWS) and the 4-state stages combine S1 and S2 of 5-state. Albeit most of prior studies’ attempt to work on multichannel EEG signal, in this work we purport to solve the problem based on single channel EEG. At this point, one might wonder-why do we need a computerized automatic sleep scoring system in the first place? Why must we work on

c 2015 IEEE 978-1-4673-6540-6/15/$31.00

1

tive noise for feature extraction and performed classification using bagging. In this work, we extract various spectral features from the EEG signal segments. After performing statistical analysis, we classify various sleep stages using boosted decision trees. The whole procedure is depicted in Fig. 1. The rest of the paper is organized as follows: Section II describes the experimental data used in this work. We elucidate the feature extraction part of our work in Section III. Section IV then explicates the statistical analyses that are performed in this study. Afterwards, we describe our classifier in Section V. Section VI presents the performance of our algorithm and compares it with that of the state-of-the art methods. Finally, Section VII expounds how this work can be extended further and concludes the article. II. E XPERIMENTAL DATA D ESCRIPTION The data used for evaluation of the proposed scheme are taken from Physionet Data Bank’s [12] Sleep-EDF Database [13] [14] which is publicly available and widely used in the literature. The recordings were obtained from Caucasian males and females (21 - 35 years old) without any medication. The first four recordings (sc4002e0, sc4012e0, sc4102e0, sc4112e0) were obtained in 1989 from ambulatory healthy volunteers during 24 hours in their normal daily life. The last four data recordings (st7022j0, st7052j0, st7121j0, st7132j0) were obtained in 1994 from subjects who had mild difficulty falling asleep but were otherwise healthy. They contain horizontal EOG, Fpz-Cz and Pz-Oz EEG data, each sampled at 100 Hz. EEG signal from Pz-Oz channel yields better classification performance than that of the Fpz-Cz channel [4] [6] [8] [11]. So in our study Pz-Oz channel EEG signal is used. Expert scoring of the EEG data is also obtained from [13]. Each 30s of EEG data was scored in accordance with the R&K recommendations [1]. So the interval of each epoch in this study is defined as 30s or (30 × 100 =)3000 data points. Each epoch was scored by expert scorers in one of the eight classes: AWA, S1, S2, S3, S4, REM, MVT (Movement Time) and ‘Unscored’. The entire data-set is divided into two halvesthe odd numbered epochs are chosen as the training data and the rest of them are used as test data. Table I summarizes the number of epochs of different classes that are used in this work.

Fig. 1. Schematic outline of the proposed framework.

Fig. 1. Spectral features have widely been used in the literature for audio signal analysis [15] [16]. Recently, these features have emerged as useful means of analyzing and classifying physiological signals of nonlinear or non-stationary nature such as EEG [17] [18] [19]. In essence, spectral features attempt to capture classification related information from the magnitude spectrum of the signal. All these factors motivated the use of spectral features in this work. Let us assume x(n) is an N point epoch where n = 0, 1, 2, ..., N − 1 and X(m) is the Discrete Fourier Transform (DFT) of x(n) where m = 0, 1, 2, ..., N − 1. The magnitude spectrum of x(n) is the absolute value of its DFT, |X(m)|. 1) Spectral Roll-off: Spectral Roll-off (SR) is the frequency sample mc below which c% of the coefficients of the magnitude spectrum of an epoch are concentrated. Mathematically, it can be expressed as: N −1 c X |X(m)| = |X(m)| 100 m=0 m=0 mc X

Generally, the value of c is 85 or 95. In this study, we use c =95. Spectral roll-off actually measures where most of the spectral energy is concentrated. It is a measure of skewness of the spectral shape of the signal. 2) Spectral Centroid: Spectral Centroid (SC) is defined as the frequency-weighted sum of the magnitude spectrum of the signal normalized by its unweighted sum. SC can be written as: PN −1

III. S PECTRAL F EATURE E XTRACTION Once we have generated the epochs, the first step is to extract various spectral features from them as we can see from

TABLE I E XPERIMENTAL DATA D ESCRIPTION AWA

S1

S2

S3

S4

REM

Total

Train Epochs

4027

302

1810

336

313

804

7592

Test Epochs

4028

302

1811

336

314

805

7596

m|X(m)| m=0 |X(m)|

m=0

SC = PN −1

SC evinces where the ‘center of mass’ of the spectrum is.

2

3) Spectral Spread: Spectral Spread (SS), sometimes also referred to as instantaneous bandwidth, describes the concentration of the magnitude spectrum around the spectral centroid [16]. It can be interpreted as the standard deviation of the magnitude spectrum around the spectral centroid. Mathematically, SS can be defined as: PN −1 2

x 10

18 16 14 12 10

4) Spectral Flatness: Spectral flatness (SF ) is a measure of the noisiness of the magnitude spectrum. It is the ratio of the geometric mean to the arithmetic mean of the magnitude spectrum of the signal. SF can be expressed mathematically as: 1 QN −1 N SF =

8 6 4

m=0 |X(m)| 1 PN −1 m=0 |X(m)| N

2 0 S1

Flat spectra correspond to noise or impulse-like signals. Hence high flatness values evince noisiness. Whereas low flatness values generally evince the presence of harmonic components. 5) Spectral Slope: Spectral slope (SSl) is a measure of the slope of the spectral shape [15]. It is calculated using a linear approximation of the magnitude spectrum. Specifically, a linear regression approach is used. In the presented form, the linear function is modeled from the magnitude spectrum. SSl is then estimated using the following equation [17]: PN −1 PN −1 PN −1 N m=0 fm |X(m)| − m=0 fm . m=0 |X(m)| SSl = PN −1 2 PN −1 N m=0 fm − ( m=0 |X(m)|)2

m=1

S3−S4

AWA

REM

AWA

REM

Spectral Flatness 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

.(|X(m)| − |X(0)|) m PN −1 m=1 |X(m)|

0 S1

S2

S3−S4

(b)

IV. F EATURE S ELECTION Statistical hypothesis testing must be performed for the feature selection stage of any classification problem. This ensures that the chosen set of features has discriminatory capability among various classes [20]. Hypothesis testing also helps us ascertain whether the discriminatory capability of the selected set of features is statistically significant or not [21]. To assess whether the values of the features in the six classes

P -VALUES OF THE

S2

(a)

m denotes the frequency of the m-th component where fm = N of Fourier transform and m = 0, 1, ..., N − 1 6) Spectral Decrease: Spectral decrease (SD) estimates the steepness of the decrease of the spectral envelope over frequency [16]. SSl is defined by [15] as: PN −1 1 SD =

Spectral Spread

20

m=0 (m − SC) |X(m)| PN −1 m=0 |X(m)|

SS =

5

Fig. 2. Box-whisker plots of (a)SS (b)SF manifest the superior discriminatory capability and efficacy of the selected features.

differ significantly, we perform a non-parametric KruskalWallis one-way analysis of variance [22] [23]. Unlike its parametric equivalent which is known as one-way analysis of variance (ANOVA), Kruskal-Wallis test does not require an assumption of normal distribution of the data samples [24]. The tests are carried out in MATLAB’s Statistics Toolbox. The test is performed at 95% confidence level. Hence, a difference is statistically significant if p < α(= 0.05). Table II gives the result of the hypothesis testing and we can see that all of the IMFs features give lower p-values indicating

TABLE II P ROPOSED F EATURES S HOW G OOD D ISCRIMINATORY C APABILITY

Features

SC

SR

SF

SS

SSl

SD

p-Values

0

0

0

0

2.6466e-23

6.9652e-245

3

Algorithm 1: AdaBoost.M2 Input: • Sequence of N instances S = [(xi , yi )] of C classes with labels yi ∈ ω, ω = {ω1 , ω2 , ...., ωC } and i = 1, 2, 3, ....., N . • Weak learning algorithm WeakLearn. • Integer T specifying the number of iterations. Let, B = {(i, y) : i = 1, 2, ..., N, y 6= yi } Initialize D1 (i, y) = 1/|B|f or(i, y) ∈ B for t = 1, 2, 3, ...., T do 1. Call WeakLearn, providing it with distribution Dt . 2. Get back a hypothesis ht : X × Y → [0, 1]. 3. Calculate the pseudo-error t of hT : 1X t = Dt (i, y)(1 − ht (xi , yi ) + ht (xi , y)) 2 4. Set βt = t /(1 − t ). 5. Update distribution Dt :

a pseudo-loss which itself is then required to have an error no larger than 0.5. This pseudo-loss measures the goodness of the weak hypotheses. Moreover, the probability of error for random guessing is much higher for multi-class problems for C-class classification problem) than that of binary ( C−1 C classification problems. Since the learners are weak, achieving an error of 0.5 becomes harder with the increase of classes. These two factors make AdaBoost.M2 a powerful tool for multi-class classification and motivated its use in the proposed framework. AdaBoost.M2 is expounded in Algorithm 1. In Step 1 and 2, the weak classifier denoted by WeakLearn is invoked with a distribution Dt and a hypothesis ht is obtained. We used decision trees as WeakLearn in the proposed method. The weak learner’s goal then is to minimize the pseudo-loss t , as defined in Step 3. The distribution function is updated as shown in Step 5. For a given instance x, the final hypothesis hf inal (x) outputs the label y that maximizes a weighted average of the weak hypothesis values ht (x, y).

Dt (i, y) 12 (1+ht (xi ,yi )−ht (xi ,y)) βt Zt P where is a Zt = Dt (i) normalization constant chosen so that Dt+1 becomes a distribution function. end Output: The final hypothesis: Dt+1 (i, y) =

hf inal (x) = arg max y∈Y

VI. E XPERIMENTAL R ESULTS AND D ISCUSSIONS We perform experiments to demonstrate the validity of the proposed method. All the simulations are done using MATLAB 2013a on a computer with Intel(R) Core(TM) i53470, 3.2 GHz CPU, 4 GB of RAM.

T X 1 (log ) × ht (x, y) β t t=1

A. Performance Comparison It is hard to compare various methods of automatic sleep scoring due to the differences of the EEG data and the number of channel used for sleep staging in prior studies. For meaningful comparison, results obtained from the same data-set have to be compared. Keeping that in mind, Table III lists accuracies of some existing methods that utilize Physionet’s Sleep-EDF data-set. All the accuracies reported in this subsection and in Table III are the best values for a given method. Our proposed method’s accuracy surpasses that of all the state-of-the-art methods for 4-stage to 6-stage sleep classification as we can see from Table III. Some results of prior studies that did not use Physionet’s Sleep-EDF data-set are also reported for the sake of comparison. Karkovska et al. [7] reported 81% agreement with the hypnograms of experts for 5-stage classification. Imtiaz [9] reported 80.60% sensitivity using Sleep-EDF. But no accuracy value was reported and hence it is not listed in Table III. Our algorithm yields better accuracy than both of the aforementioned works too for 5-class classification. Besides Sleep-EDF data-set, Berthomier et al. [6] also used their own EEG recordings that yielded 82.90% for 5-stage classification of sleep stages. So, accuracy obtained from both data-sets are comparable to that of the proposed framework. More over, our experiments are conducted using more number of epochs than that in [6] [8].

good discriminatory capability among the six classes. The boxwhisker plots of in Fig. 2 serve as a visual hypothesis test and also confirm the efficacy of the selected features. Since the notches do not overlap in Fig. 2, the selected set of features blatantly possesses good discriminatory capability. V. C LASSIFICATION The classifier used in this work is Adaptive Boosting (AdaBoost). AdaBoost was pioneered by Freund and Schapire [27]. AdaBoost generates a set of hypotheses such that subsequent hypotheses focus on increasingly difficult instances [28]. Afterwards, it combines them through weighted majority voting of the classes predicted by the individual hypotheses. Each of the hypothesis is generated by training a weak classifier using instances drawn from an iteratively updated distribution of the training data. This distribution update makes sure that instances misclassified by the previous classifier are more likely to be included in the training data of the next classifier. Consequently, consecutive classifiers’ training data are geared towards increasingly difficult-to-classify instances. This is the essence of AdaBoost learning algorithm. Freund and Schapire also propounded two extensions of AdaBoost [27], namely- AdaBoost.M1 and AdaBoost.M2. AdaBoost.M1 requires that all classifiers have a weighted error no greater than 0.5. AdaBoost.M2 resolves this problem by removing the weighted error restriction. Instead, it defines

B. Discussions The proposed scheme was able to eradicate some problems of prior studies and so showed good accuracy values. The

4

TABLE III P ERFORMANCE C OMPARISON OF VARIOUS S TATE - OF - THE - ART M ETHODS . T HE H IGHEST ACCURACY VALUES ARE I NDICATED IN B OLDFACE Authors

Features

Classifier

6-Class 5-Class 4-Class

Herrera et al. [25]

Symbolic Representation of EEG

Support Vector Machine

-

-

69.80%

Berthomier et al. [6]

Spectral and Temporal

Fuzzy Classifier

-

71.2%

74.50%

Huang et al. [26]

Features from Forehead EEG

Support Vector Machine

-

-

77.12%

Ronzhina et al. [8]

Power Spectrum Density

-

81.42%

Proposed Method

Spectral Features

Artificial Neural Network 76.70% Boosted Decision Trees

EEG activities of the two sleep stages- S1 and REM are almost indistinguishable by visual inspection [29]. This poses an enormous challenge to an automatic sleep staging algorithm and hinders it from obtaining high values of accuracy. Despite demonstrating high levels of accuracy as we can see from Table III, it was reported in [30] that it is much harder to distinguish S1 and REM stages by using DVGs. Hence, this method yields only 15.8% sensitivity for S1 in five-state classification of sleep stages. Liang et al. [4] also reports only 18.75% sensitivity for S1 in five-state classification of sleep stages. The proposed algorithm, on the other hand, has not altogether solved the problem. But it has been able to distinguish between S1 and REM better than most prior studies such as [4] [30]. The five-class confusion matrix in Table V reveals that our method meliorates the S1 sensitivity to 22.85% using identical data settings. The accuracy value in the six-state confusion matrix in Table IV is also similar and hence corroborates with this fact. What is more, the proposed scheme gives higher sensitivity values of AWA and S2 stage as opposed to those of [4] for five sleep stages. We can see from Table V that our method correctly classifies 93.40% of AWA epochs and 80.84% of S2 epochs whereas [4] reported 91.99% and 70.19% accuracies for AWA and S2 respectively. Furthermore, spectral features successfully capture statistical differences of the epochs of various classes as our prior discussions and analyses reveal. Boosting, on the other hand, is very resistant to overfitting. All these factors contribute to the high levels of accuracy of the proposed method. Some of the prior studies such as [10] [31] use multiple physiological signals to classify sleep stages. Using EEG, EOG and EMG for feature extraction can be a potential source of problem mainly for two reasons. Firstly, it complicates the preparation procedure for the subject [4]. Secondly, more electrodes cause more inference and thus degrade the quality of the recordings. Since our study uses single-channel EEG signal, it is free of both of the above problems.

80.34% 82.03% 82.83%

TABLE IV C ONFUSION M ATRIX FOR 6- STAGE S LEEP C LASSIFICATION

Expert

S1 S2 S3 S4 AWA REM

S1 65 22 1 0 47 66

S2 48 1477 131 36 62 131

Computer S3 0 62 119 46 21 5

S4 0 31 44 193 21 2

AWA 84 118 37 38 3794 146

Sen 21.52% 81.56% 35.42% 61.46% 94.19% 56.52%

REM 105 101 4 1 83 455

TABLE V C ONFUSION M ATRIX FOR 5- STAGE S LEEP C LASSIFICATION

Expert

S1 S2 S3-S4 AWA REM

S1 69 18 1 46 43

Computer S2 36 1464 145 67 126

SWS 2 117 438 53 5

AWA 79 92 58 3762 133

REM 116 120 8 100 498

Sen 22.85% 80.84% 67.38% 93.40% 61.86%

signal classification and various pathological state detection problems, seismic signal analysis etc. How the algorithm behaves for novel variants of AdaBoost and other boosting algorithms- can be an intriguing topic of further research. We thus come into a conclusion, as the experimental results blatantly suggest that the proposed algorithm is effective, efficient and has the potential of making a low-power, wearable sleep monitoring device feasible. R EFERENCES [1] A. Rechtschaffen and A. Kales, “Manual of standardized terminology, techniques and scoring systems for sleep stages of human subjects,” U. G. P. Office, Washington DC Public Health Service, 1968. [2] A. L. C. C. Iber, S. Ancoli-Israel and S. F. Quan, “Westchester, usa: American academy of sleep medicine,” The AASM Manual for the Scoring of Sleep and Associated Events: Rules, Terminology and Technical Specification, 2007. [3] C.-T. Lin, L.-W. Ko, J.-C. Chiou, J.-R. Duann, R.-S. Huang, S.-F. Liang, T.-W. Chiu, and T.-P. Jung, “Noninvasive neural prostheses using mobile and wireless eeg,” Proceedings of the IEEE, vol. 96, no. 7, pp. 1167– 1183, July 2008. [4] S.-F. Liang, C.-E. Kuo, Y.-H. Hu, Y.-H. Pan, and Y.-H. Wang, “Automatic stage scoring of single-channel sleep eeg by using multiscale entropy and autoregressive models,” Instrumentation and Measurement, IEEE Transactions on, vol. 61, no. 6, pp. 1649–1657, June 2012. [5] T. Penzel and R. Conradt, “Computer based sleep recording and analysis,” Sleep Medicine Reviews, vol. 4, no. 2, pp. 131–148, 2000.

VII. C ONCLUSIONS AND F UTURE W ORK In the present paper, a spectral feature based feature generation scheme is proposed for automatic sleep scoring. Boosting is introduced as a classification model for automatic sleepstaging. The proposed approach can be implemented to solve other signal classification problems such as ECG, EMG, MEG

5

[6] C. Berthomier, X. Drouot, M. Herman-Stoca, P. Berthomier, J. Prado, D. Bokar-Thire, O. Benoit, J. Mattout, and M.-P. dOrtho, “Automatic analysis of single-channel sleep eeg: Validation in healthy individuals,” Sleep, vol. 30, no. 11, p. 15871595., Nov 2007. [7] A. Krakovsk and K. Mezeiov, “Automatic sleep scoring: A search for an optimal combination of measures,” Artificial Intelligence in Medicine, vol. 53, no. 1, pp. 25 – 33, 2011. [8] M. Ronzhina, O. Janouek, J. Kolrov, M. Novkov, P. Honzk, and I. Provaznk, “Sleep scoring using artificial neural networks,” Sleep Medicine Reviews, vol. 16, no. 3, pp. 251 – 263, 2012. [9] S. Imtiaz and E. Rodriguez-Villegas, “A low computational cost algorithm for rem sleep detection using single channel eeg,” Annals of Biomedical Engineering, vol. 42, no. 11, pp. 2344–2359, 2014. [10] H. Koch, J. A. Christensen, R. Frandsen, M. Zoetmulder, L. Arvastson, S. R. Christensen, P. Jennum, and H. B. Sorensen, “Automatic sleep classification using a data-driven topic model reveals latent sleep states,” Journal of Neuroscience Methods, vol. 235, no. 0, pp. 130 – 137, 2014. [11] A. R. Hassan and M. I. H. Bhuiyan, “Computer-aided sleep staging using complete ensemble empirical mode decomposition with adaptive noise and bootstrap aggregating,” Biomedical Signal Processing and Control, vol. 24, pp. 1 – 10, 2016. [12] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. e215–e220, 2000 (June 13). [13] B. Kemp, “The sleep-edf database online.” [Online]. Available: http://www.physionet.org/physiobank/database/sleep-edf/ [14] B. Kemp, A. Zwinderman, B. Tuk, H. Kamphuisen, and J. Oberye, “Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg,” Biomedical Engineering, IEEE Transactions on, vol. 47, no. 9, pp. 1185–1194, Sept 2000. [15] G. Peeters, “A large set of audio features for sound description (similarity and classification) in the cuidado project,” in Tech. Rep., IRCAM, 2004. [16] A. Lerch, An Introduction to Audio Content Analysis: Applications in Signal Processing and Music Informatics. Wiley-IEEE Press, 2012. [17] A.-C. Conneau and S. Essid, “Assessment of new spectral features for eeg-based emotion recognition,” in Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on, May 2014, pp. 4698–4702. [18] A. R. Hassan, S. K. Bashar, and M. I. H. 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. 1–6. [19] A. R. Hassan and M. A. Haque, “Computer-aided sleep apnea diagnosis from single-lead electrocardiogram using dual tree complex wavelet transform and spectral features,” in Electrical and Electronic Engineering (ICEEE), 2015 International Conference on, Nov. 2015, pp. 1–4. [20] ——, “Computer-aided gastrointestinal hemorrhage detection in wireless capsule endoscopy videos,” Computer Methods and Programs in Biomedicine, pp. –, 2015. [21] S. K. Bashar, A. R. Hassan, and M. I. H. Bhuiyan, “Identification of motor imagery movements from eeg signals using dual tree complex wavelet transform,” in Advances in Computing, Communications and Informatics (ICACCI), 2015 International Conference on, Aug 2015, pp. 1–6. [22] W. H. Kruskal and W. A. Wallis, “Use of ranks in one-criterion variance analysis,” Journal of the American Statistical Association, vol. 47, no. 260, pp. 583–621, 1952. [23] A. R. Hassan, “Automatic screening of obstructive sleep apnea from single-lead electrocardiogram,” in Electrical Engineering and Information & Communication Technlogy (ICEEICT), 2015 International Conference on, May 2015, pp. 1–6. [24] A. R. Hassan and M. A. Haque, “Epilepsy and seizure detection using statistical features in the complete ensemble empirical mode decomposition domain,” in TENCON 2015, 2015 IEEE Region 10 Conference, Nov. 2015, pp. 1–6. [25] L. Herrera, A. Mora, C. Fernandes, D. Migotina, A. Guillen, and A. Rosa, “Symbolic representation of the eeg for sleep stage classification,” in Intelligent Systems Design and Applications (ISDA), 2011 11th International Conference on, Nov 2011, pp. 253–258. [26] 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

[27] [28]

[29]

[30]

[31]

6

eeg signals,” in Computational Intelligence, Cognitive Algorithms, Mind, and Brain (CCMB), 2013 IEEE Symposium on, April 2013, pp. 1–5. 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. A. R. Hassan, “A comparative study of various classifiers for automated sleep apnea screening based on single-lead electrocardiogram,” in Electrical and Electronic Engineering (ICEEE), 2015 International Conference on, Nov. 2015, pp. 1–4. M. Corsi-Cabrera, Z. Muoz-Torres, Y. del Ro-Portilla, and M. Guevara, “Power and coherent oscillations distinguish rem sleep, stage 1 and wakefulness,” International Journal of Psychophysiology, vol. 60, no. 1, pp. 59 – 66, 2006. G. Zhu, Y. Li, and P. Wen, “Analysis and classification of sleep stages based on difference visibility graphs from a single channel eeg signal,” Biomedical and Health Informatics, IEEE Journal of, vol. PP, no. 99, pp. 1–1, 2014. J. Virkkala, J. Hasan, A. Vrri, S.-L. Himanen, and K. Mller, “Automatic sleep stage classification using two-channel electro-oculography,” Journal of Neuroscience Methods, vol. 166, no. 1, pp. 109 – 115, 2007.