Automatic stage scoring of single-channel sleep EEG based on multiscale permutation entropy Chih-En Kuo and Sheng-Fu Liang
Abstract—Multiscale entropy is a recently developed method to estimate complexity associated with the long-range temporal correlation of a time series. Since sleep EEG patterns also change regularly from light to deep sleep states, we firstly applied multiscale permutation entropy (MPE) to analysis sleep EEG to investigate the relations between changes of sleep stages and the MPE values. It was observed that correlation coefficient between the averaged MPE values of sleep EEG and the manual scoring of sleep stages can reach over 0.7. Then a MPE-based sleep scoring method for single channel EEG was developed. After training based on the data from 10 subjects, the overall sensitivity of the proposed automatic sleep scoring method combining MPE, autoregressive models, and linear discriminant analysis can reach 89.1% evaluated by the data of the other 10 subjects. Due to high accuracy and requiring only single-channel EEG, the proposed method has good applicability for sleep monitoring and home cares. Index Terms － Multiscale permutation entropy (MPE), automatic sleep scoring, single channel EEG, autoregressive (AR) model, linear discriminant analysis (LDA).
H
I. INTRODUCTION
uman beings spend approximately one third of their lives sleeping. Sleep diseases, such as insomnia and obstructive sleep apnea, seriously affect patients’ quality of life. The prevalence of insomnia symptoms without restrictive criteria is approximately 33% in the general population [1]. For diagnosis of sleep problems, all-night polysomnographic (PSG) recordings including electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG) are usually taken from the patients and scored by a well-trained This work was supported by the National Science Council of Taiwan under grant NSC 98-2221-E-006 -161- MY3, NSC 99-2220-E-009-032, and Aim for the Top University Plan of the National Chiao Tung University and Ministry of Education,Taiwan. S.-F. Liang is with the Department of Computer Science and Information Engineering & the Institute of Medical Informatics, National Cheng Kung University, Tainan 701, Taiwan and Biomimetic Systems Research Center, National Chiao Tung University (phone: 886-6-2757575 Ext: 62549; fax: 886-6-2747076; e-mail:
[email protected]). C.-E Kuo is with the Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan 701, Taiwan (e-mail:
[email protected]). All correspondence should be addressed to Professor S.-F. Liang.
978-1-4577-1470-2/11/$26.00 ©2011 IEEE
expert according to Rechtschaffen & Kales (R&K) rules [2] to classify each epoch into one of the sleep stages, including wakefulness (Wake), non-rapid eye movement (stages 1-4) and rapid eye movement (REM). Recently, stages 3 and 4 were combined to become the new slow wave sleep stage (SWS). Because visual sleep scoring is a time-consuming and subjective process, automatic sleep staging methods based on multi-channel signals, including EEG, EMG, and EOG [3-9] or single EEG channel [6, 10], were developed. These methods contain two processes: feature extraction to analyze the recording epoch and classification to identify the sleep stage of the epoch. According to the R&K standard, some features have been proposed for sleep staging, including the alpha ratio [6], Spindle ratio [8], and SWS ratio [9]. In addition, spectral power, power ratio, spectral, and frequency [6] have been used in previous methods. In classification, many methods have been proposed, such as linear discriminant analysis (LDA) [11], artificial neural network (ANN) [6], fuzzy system [6], and decision tree [12]. Recently, a new signal analysis method called multiscale entropy (MSE) has been proposed [13, 14] to estimate the complexity associated with the long-range temporal correlation of a time series or various biomedical signals such as EEG [15-16], ECG [14], and heart rate [17]. However, sleep EEG has never been analyzed with the MSE and MSE never been as feature for sleep scoring. The relationship between MSE and sleep EEG is an interesting problem. MSE can be computed from the different types of entropy with multiple coarse-grained sequences, such as approximate entropy (ApEn) [18], sample entropy (SampEn) [19] or permutation entropy (PEn) [20]. Many of the changes observed in the EEG under general anesthesia are also seen in natural sleep, thus implying that similar neurophysiological mechanisms are involved [21]. Thus, it is envisaged that PEn will be able to capture some of the differences in the EEG patterns that are observed during different sleep stages. In this paper, we first identified the relationship between sleep stages and the corresponding MPE values. Then a MPE-based automatic sleep scoring method [22] for signal-channel EEG was proposed. The advantage of multi-scale entropy compared with single-scale entropy for sleep EEG analysis was also discussed.
448
II. MATERIALS AND METHODS A. Subjects and recordings All-night PSG sleep recordings were obtained from 20 healthy subjects (12 males and 8 females, aged 21.2 ± 1.1 years). The subjects were interviewed about their sleep quality and medical history. None of them reported any history of neurological or psychological disorders. The PSG recordings of each subject, including six EEG channels (F3-A2, F4-A1, C3-A2, C4-A1, P3-A2, and P4-A1, according to the international 10-20 standard system), two EOG channels (the above right and below left outer canthus), and a chin EMG channel, were acquired through the Siesta 802 PSG (Compumedics, Inc.). The sampling rate was 1 KHz with 16-bit resolution. The filter settings of the cut-off frequencies were 0.5–30 Hz for EEG/EOG and 5–100 Hz for EMG. These nine-channel signals were used for manual scoring and only the data of the C3-A2 EEG channel were used for the developed single channel sleep staging system. B. Manual scoring The 20 PSGs were scored by two sleep experts who worked independently from each other and according to R&K rules [2]. Each 30-s epoch was classified into one of the five sleep stages, including Wake, REM, S1, S2, SWS, and movement artifact. In our experiments, only epochs belonging to the five sleep stages were used and movement artifact epochs were rejected [3, 6]. Agreement was 92.3% between two sleep experts. In this paper, we compared the automatic classification and the manual scoring of expert 1 because the results of comparison between expert 2 and automatic analysis did not differ statistically from those computed with expert 1. C. Single-channel sleep staging system Figure 1 shows the flow chart of the proposed single-channel sleep staging method that includes three parts: (1) pre-processing, (2) feature extraction, and (3) classification. The following figure presents each part in greater detail. Preprocessing Input: EEG (C3-A2)
Feature Extraction
Downsampling (256Hz) Band pass filtering (0.5-30Hz)
Part 2.1. Multiscale permutation entropy Multiscale permutation entropy (MPE) measures the complexity of a time series by taking into account the permutation entropy with respect to multiple temporal scales [24]. Given an EEG time sequence with N samples, x {x1 , x2 , ..., x N } , the original times series is divided into non-overlapping windows of length τ, defined as the scale factor. The data points inside each window are then averaged. Each element of the coarse-gained time series yτ(j) is calculated by the following equation: y ( j )
Linear discriminant analysis
Autoregressive
Contextual rule smoothing
j
1
N
x , 1 j
i i ( j 1) 1
.
(1)
After obtaining each element of the coarse-gained time series, we calculate the permutation entropy that defined by Bandt and Pompe [20] for each of the coarse-grained time series. For scale τ, the corresponding coarse-gained time series N has N/τ points and is denoted as Y { y (1),..., y (i ),..., y ( )} , an embedding procedure forms vectors Y (t ) { y (t ), y (t n),..., y (t mn)} with the embedding dimension m and lag n. Then, Y (t ) can be arranged in an increasing order. For m different numbers, there will be m! possible order patterns π. For a permutation with number π, let f(π) denote its probability in the time series. Then, the relative N probability is pi ( ) f ( ) /( (m 1)n) . The permutation
entropy is defined as follows: K
H p (m)
Classification
Multiscale permutation entropy (scale:1-20)
Model (order 8)
analyses: 1) multiscale permutation entropy (MPE) and 2) autoregressive model (AR). The MPE is the principal analysis for the method and the AR model is the complementary feature to raise the classification accuracy of stage 1.
p ( ) ln p ( ) i
i
(2)
i 1
where K is the distinct number of symbols should be less than or equal to m!. The PEn value will be 1 when all permutations have equal probability. Conversely, PEn will be small if the time series is regular. Thus, the more regular the time series, the smaller the PEn value. The parameters m 3 and n 1 were used to calculate MPE values in this study.
Scoring Result
Theta band EEG (4-8Hz)
Fig 1. The flow chart of single-channel sleep staging method.
Part 1: Pre-processing After down-sampling the signals to 256 Hz for simplicity, an eighth-order Butterworth band-pass filter with 0.5-30 Hz pass-band is used to filter the down-sampled signals for MSE analysis. In addition, an eighth-order Butterworth band-pass filter with 4-8 Hz pass-band was also utilized to extract the theta band components for the autoregressive (AR) model [23] to complement the MPE in accurately recognizing stage 1 [2].
Part 2.2. Autoregressive model The autoregressive (AR) model is a parametric model used to describe a stationary time series, and it is a popular tool for EEG analysis [6, 23]. The model parameters can be used to determine the EEG states. In this study, the order of the AR model is 8, and the inputs of the AR model are the theta band signals (4-8 Hz) extracted by an eighth-order Butterworth band-pass filter in the pre-processing.
Part 2: Feature extraction
Part 3: Classification A linear classifier, linear discriminant analysis (LDA), was utilized to classify the extracted MSE values and AR coefficients into five sleep stages. LDA uses a hyperplane to determine the linear combination of features that best separates
The continuous filtered signals were segmented into non-overlapping 30-s windows (called epochs) for feature extraction. The feature extraction part contained two major
449
III. EXPERIMENTAL RESULTS The experiments consist of three parts: (1) the relation between the MPE values of different scales and sleep stages (2) only the MPE values corresponding to scales 1-20 are used as the features for the LDA classifier to evaluate the effect of MPE on sleep individually and (3) MPE (scale factor 1-20) and AR model coefficients (order 8) are combined as the features for the LDA to evaluate the performance of the proposed sleep staging method.
when the numbers of scale factors used as features increased from 1 to 20. The average sensitivity is higher than 73% when the number of used scale factors is more than 4. When the number of selected scale factors approaches 12, the average sensitivity approaches complete flatness. Although single scale PEn shows better performance on the sensitivities of SWS, other stages are very low. These results show that multi-scale entropy is more distinguishable and stable compared with only single scale entropy. It is noted that the agreement of S1 is still less than 15% even 20 scale factors were used. Therefore, the complementary features are required to improve the performance. Table 1. The confusion matrix between the automatic scoring method using MPEs as features and the manual sleep scoring. Computer Wake Wake 174 48 S1 Human
A. Relationship between the MPE values and sleep stages Figure 2 shows the relation between the MPE values of different scales and sleep stages of one subject (No. 9). Fig. 2(a) shows the results of averaging the 20 MPE values in each epoch after smoothing by moving average algorithm [26]. Fig. 2(b) shows the manual sleep scoring by the expert. The correlation coefficient between the averaged MPE values and the manual scoring of sleep stages reaches a maximal level of 0.7282. This shows that the MPE values and sleep stages are highly correlated. This encouraging result motivates us to apply the MPE to the development of automatic sleep staging methods.
22
35
243
112
3714
90
12
15
259
1185
30 REM Average SE
27
63
1
Wake
Sensitivity
80 70 60 50 2
4
6
300
400
500
600
700
800
900
30 20 10
0
2
4
6
Scale
S2
SWS
70
94 92
60 50 40 30 20
90 88 86 84 82 80 78 76
10 0
2
4
6
0
8 10 12 14 16 18 20 22
2
4
6
100
200
300
400
500
600
700
800
900
REM
100
Overall
90 80
Sensitivity
Fig 2 .The relation between the MPE and sleep stages. (a) Results of averaging the 20 MPE values in each epoch; (b) manual sleep staging reviewed by the expert.
B. Sleep scoring models based on MPE The MPE values corresponding to 20 scale factors of the single channel EEG were used as features of the LDA classifier to construct a sleep scoring model. Table 1 shows the confusion matrices of five-stage epoch classification by automatic staging versus manual scoring. The sensitivity of computer scoring corresponding to each stage, and the average sensitivities are given. The average sensitivity can reach up to 80.61%. Apart from S1, the sensitivities of the other stages are more than 76%. Figure 3 shows the sensitivity curves of each stage and the average result by using different numbers of MPE scales as features for sleep scoring. It was found that the sensitivity curve of the average results rose from approximately 40% to 83%
8 10 12 14 16 18 20 22
Scale
Scale 0
8 10 12 14 16 18 20
Scale
Sensitivity
200
40
8 10 12 14 16 18 20
80
100
S1
0 0
90
0
4629 80.23% 1510 78.47% 1741 1862 93.5% 39
50
40
Mov Wake REM S1 S2 SWS
470
8480 80.61%
90
2.2 2.0 1.8 1.6 1.4 1.2 1.0
SWS REM Total SE 0 20 226 76.99% 0 148 253 8.69%
S2
Sensitivity
(b)
S2 29
SWS
100
Sensitivity
mean Average
(a)
S1 3
Sensitivity
two or more classes of objects or events. The PSG data of ten subjects were used to train the LDA classifier, and the other ten subjects were used to verify the performance of our proposed method. Finally, sleep staging has periodicity and continuity from light to deep [2]. After classifying the sleep stage by LDA, some misclassified epochs can be corrected according to the temporal contextual information and R&K rules. According to the rules presented in [6, 25], a total of 11 rules were utilized to smooth and fine-tune the classification results of LDA and increase the accuracy of our method.
80
60
40 0
2
4
6
8 10 12 14 16 18 20 22
Scale
70 60 50 40 30 0
2
4
6
8 10 12 14 16 18 20 22
Scale
Fig 3. Sensitivity curves of each stage and the average result by using different numbers of MPE scales as features for sleep scoring.
C. Automatic stage scoring method for single-channel EEG The complete automatic sleep scoring method combined MPE and autoregressive models as features, and linear discriminant analysis was utilized as classifier. Table 2 shows the confusion matrices of five-stage epoch classification by the automatic staging combining MPE 1-20, 8 AR coefficients and smoothing process versus manual scoring. The sensitivity of overall is 89.1%. Compared to Table 1, the sensitivities of each
450
[5]
stage and the overall sensitivity were all increased by AR coefficients and smoothing process.
[6]
Table 2. The confusion matrix between the automatic sleep scoring after smoothing and the manual sleep scoring (MPE 1-20 and 8 AR coefficients).
[7]
Computer
Human
Wake
S1
S2
Wake
200
15
5
SWS REM Total 0
6
226
88.49%
S1
50
48
56
0
99
253
18.97%
S2
10
86
4279
40
214
4629
92.43%
SWS
0
19
259
1211
21
1510
80.19%
REM
1
8
34
1
1818 1862
97.63%
8480
89.1%
Average SE
SE
[8]
[9] [10]
[11]
[12]
IV. DISCUSSION AND CONCLUSION In the present study, a novel approach based on multiscale permutation entropy analysis of a single EEG channel for automated sleep scoring was developed. Until now, this is the first time that MPE was applied to automatic sleep scoring. It was observed that correlation coefficient between the averaged MPE values of sleep EEG and the manual scoring of sleep stages can reach over 0.7. The sensitivity of the average results rose from approximately 40% to 83% when the numbers of scale factors used as features increased from 1 to 20 for sleep staging. The results show that the performance of multiscale entropy is much better than single scale entropy for sleep analysis. Compared to conventional sleep scoring methods that require multiple physiological signals, including EEG, EOG and EMG for feature extraction [4-6], the single channel sleep scoring approach has the advantage of reducing interference in sleep quality due to the use of fewer electrodes and the simplification of the preparation procedure. Sensitivities of single-channel automated sleep staging methods [6, 10] were reported in the range of 74-82.9%. The proposed method combining MPE, autoregressive models, and linear discriminant analysis can reach 89.1% evaluated by the data from 10 subjects. The proposed algorithm can combine with a portable EEG recording device to provide more portability and wearability for sleep quality evaluation at home in the future. REFERENCES [1] [2]
[3]
[4]
[13]
[14] [15]
[16]
[17]
[18] [19]
[20] [21]
[22]
[23]
M.M. Ohayon, “Epidemiology of insomnia: what we know and what we still need to learn,” Sleep Medicine Reviews, vol. 6, pp. 97-111, 2002. A Manual of Standardized Terminology, Techniques and Scoring Systemfor Sleep Stages of Human Subjects, A. Rechtschaffen and A. Kales,Eds., Brain Inform. Service/Brain Res. Inst., Univ. California, Los Angeles,1968. N. Schaltenbrand et al., “Sleep stage scoring using the neural network model: Comparison between visual and automatic analysis in normal subjects and patients,” Sleep, vol. 6, pp 26-35, 1996. H.J. Park, K.S. Park, and D.U. Jeong, “Hybrid neural-network and rule-based expert system for automatic sleep stage scoring,” Proceedings of the 22nd Annual EMBS International Conference, pp. 1316-9, 2000.
[24]
[25]
[26]
451
R. Agarwal, and J. Gotman, “Computer-assisted sleep staging,” IEEE Trans. Biomed. Eng., vol. 48, pp. 1412-23, 2001. C. Berthomier et al., “Automatic analysis of single-channel sleep EEG: validation in healthy individuals,” Sleep, vol. 30, pp. 1587-95, 2007. J. Virkkala, J. Hasan, A. Värri, S. Himanen and K. Müller, “Automatic sleep stage classification using two-channel electro-oculography,” J. Neurosci. Methods, vol. 166, pp. 109-15, 2007. F. Duman, A. Erdamar, O. Erogul, Z. Telatar and S. Yetkin, “Efficient sleep spindle detection algorithm with decision tree,” Expert Systems with Applications, vol. 36, pp. 9980-5, 2009. C. Berthomier, J. Prado and O. Benoit, “Automatic sleep EEG analysis using filter banks,” Biomed. Sci. Instrum., vol. 35, pp. 241-6, 1999. A. Flexer, G. Gruber and G. Dorffner, “A reliable probabilistic sleep stager based on a single EEG signal,” Artificial Intelligence in Medicine, vol. 33, pp. 199-207, 2005. Šušmáková Kristína and Krakovská Anna, “Discrimination ability of individual measures used in sleep stages classification,” Artificial Intelligence in Medicine, vol. 44, pp. 261-77, 2008. P. Anderer et al., “An E-health solution for automatic sleep classification according to Rechtschaffen and Kales: validation study of the Somnolyzer 24 x 7 utilizing the Siesta database,” Neuropsychobiology, vol. 51, pp. 115-33, 2005. M. Costa, A.L. Goldberger and C.K. Peng, “Multiscale entropy analysis of complex physiologic time series,” Phys Rev Lett, vol. 89:068102, 2002. M. Costa, A.L. Goldberger and C.K. Peng, “Multiscale entropy analysis of biological signals,” Phys Rev E, vol. 71:021906, 2005. T. Takahashi et al., “Age-related variation in EEG complexity to photic stimulation: a multiscale entropy analysis,” Clin. Neurophysiol., vol. 120, pp. 476-83, 2009. J.H. Park, S. Kim, C.H. Kim, A. Cichocki and K. Kim, “Multiscale entropy analysis of EEG from patients under different pathological conditions,” Fractals-Complex Geom. Patterns Scaling Nat. Soc., vol. 15, pp. 399-404, 2007. P.R. Norris et al., “Heart rate multiscale entropy at three hours predicts hospital mortality in 3,154 trauma patients,” Shock, vol. 30, pp. 17-22, 2008. S.M. Pincus, “Approximate Entropy (ApEn) as a Complexity Measure,” Chaos, vol. 5, pp. 110-7, 1995. J.S. Richman and J.R. Moorman, “Physiologic time series analysis using approximate entropy and sample entropy,” Am. J. Physiol. Heart Circ., vol. 278, pp. 2039-49, 2000. C. Bandt and B. Pompe, “Permutation entropy: a natural complexity measure for time series,” Phys Rev Lett, vol 88: 174102, 2002. L.J. Voss and J.W. Sleigh, “Monitoring consciousness: the current status of EEG based depth of anaesthesia monitors,” Best Pract. Res. Clin. Anaesthesiol., vol. 21, pp. 313-25, 2007. N. Nicolaou and J. Georgiou, “The use of permutation entropy to characterize sleep electroencephalograms,” Clin. EEG Neurosci., vol. 42, pp. 24-8, 2011. J. Pardey , S. Roberts and L. Tarassenko, “A review of parametric modelling techniques for EEG analysis,” Med. Eng. Phys, vol. 18, pp. 2-11, 1996. Duan Li, Xiaoli Li, Zhenhu Liang, J. Logan and S. Jamie, Multiscale permutation entropy analysis of EEG recordings during sevoflurane anesthesia, J. Neural Eng., vol. 7, pp. 046010, 2010. The AASM manual for the scoring of sleep and associated events, C. Iber, S. Ancoli-Israel, A. Chesson and S.F. Quan, Westchester IL: American Academy of Sleep Medicine, 2007. Ya-lun Chou, Statistical Analysis, Holt International, 1975, ISBN 0030894220, section 17.9.
