Journal of Neuroscience Methods 246 (2015) 142–152

Contents lists available at ScienceDirect

Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth

Computational Neuroscience

Using off-the-shelf lossy compression for wireless home sleep staging Kun-Chan Lan a,b , Da-Wei Chang a,b , Chih-En Kuo c , Ming-Zhi Wei a , Yu-Hung Li a , Fu-Zen Shaw c , Sheng-Fu Liang a,b,∗ a b c

Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan 701, Taiwan Institute of Medical Informatics, National Cheng Kung University, Tainan 701, Taiwan Department of Psychology, National Cheng Kung University, Tainan 701, Taiwan

h i g h l i g h t s • • • • •

We examine the effects of off-the-shelf lossy compression on an all-night PSG dataset, in the context of automated sleep staging. The popular compression method Set Partitioning in Hierarchical Trees (SPIHT) was used. A rule-based automatic sleep staging method was used to classify the sleep stages. The result shows that the system can achieve more than 60% energy saving and a high accuracy (>84%) in classifying sleep stages. The feasibility of using off-the-shelf lossy compression for wireless home sleep staging was demonstrated.

a r t i c l e

i n f o

Article history: Received 9 May 2014 Received in revised form 6 March 2015 Accepted 9 March 2015 Available online 16 March 2015 Keywords: Home sleep staging Sleep EEG PSG SPIHT compression

a b s t r a c t Background: Recently, there has been increasing interest in the development of wireless home sleep staging systems that allow the patient to be monitored remotely while remaining in the comfort of their home. However, transmitting large amount of Polysomnography (PSG) data over the Internet is an important issue needed to be considered. In this work, we aim to reduce the amount of PSG data which has to be transmitted or stored, while having as little impact as possible on the information in the signal relevant to classify sleep stages. New method: We examine the effects of off-the-shelf lossy compression on an all-night PSG dataset from 20 healthy subjects, in the context of automated sleep staging. The popular compression method Set Partitioning in Hierarchical Trees (SPIHT) was used, and a range of compression levels was selected in order to compress the signals with various degrees of loss. In addition, a rule-based automatic sleep staging method was used to automatically classify the sleep stages. Results: Considering the criteria of clinical usefulness, the experimental results show that the system can achieve more than 60% energy saving with a high accuracy (>84%) in classifying sleep stages by using a lossy compression algorithm like SPIHT. Comparison with existing method(s): As far as we know, our study is the first that focuses how much loss can be tolerated in compressing complex multi-channel PSG data for sleep analysis. Conclusions: We demonstrate the feasibility of using lossy SPIHT compression for wireless home sleep staging. © 2015 Elsevier B.V. All rights reserved.

1. Introduction Point-of-care (POC) patient monitoring refers to near-patient testing, usually outside the central hospital or primary care facility. One important POC application is to measure and track sleep

∗ Corresponding author at: Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan 701, Taiwan. E-mail address: sfl[email protected] (S.-F. Liang). http://dx.doi.org/10.1016/j.jneumeth.2015.03.013 0165-0270/© 2015 Elsevier B.V. All rights reserved.

quality. Human beings spend approximately one third of their lives sleeping, and condition such as insomnia and obstructive sleep apnea can seriously affect the quality of life. According to a survey, 50–70 million people suffer from sleep disorders in the United States (Colten and Altevogt, 2006). Sleep staging is probably the most important piece of information in the analysis of sleep. Polysomnography (PSG) is the gold standard used in hospitals for sleep staging, and this method continuously and simultaneously records multiple physiological signals during sleep, such as electroencephalogram (EEG),

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

Fig. 1. Typical polygraphic recordings during the wake, stage 1 (S1), stage 2 (S2) light sleep, slow-wave sleep (SWS) associated with deep sleep, and rapid-eye movement (REM) states. Each column shows EEG, EOG and EMG obtained during the corresponding sleep stage.

electromyogram (EMG), electrooculogram (EOG), electrocardiogram (ECG), and blood oxygen saturation data. EEG records electrical activities of the brain which have distinct patterns in different sleep stages, such as a sleep spindle or delta rhythm. EOG records eye movement during sleep, which is critical to differentiate the rapid eye movement (REM) sleep stage from the other sleep stages. EMG records bioelectrical signals generated by the activities of skeletal muscles, and play an important role in distinguishing Wake from REM sleep. Fig. 1 shows typical PSG recordings corresponding to various sleep stages. Diagnosis of sleep disorders using PSG is typically performed in hospitals and sleep centers, and subjects are thus often kept on a waiting list for a considerable period of time, which in the United States, ranges from a few weeks to more than one year (Flemons et al., 2004). Moreover, sleeping in an unfamiliar environment, such as a hospital, may cause the first-night effect in subjects (Agnew et al., 1966), leading to less REM sleep, shorter total sleep time, and lower sleep efficiency. Sleep recording at home can reduce the first-night effect (Edinger et al., 1997) and decrease the waiting time for sleep evaluation. Recently, there has been increasing interest among both the academics (Chang et al., 2012; Griessenberger et al., 2013; Kelly et al., 2012; Lubecke and Boric-Lubecke, 2009) and practitioners in at-home sleep monitoring through the use of lightweight, wearable wireless PSG systems. However, with PSG signals, even a small amount of recording can generate very large amounts of data, and wireless transmission is a major contributor to power consumption in portable devices (Casson and RodriguezVillegas, 2007). In addition, with the surging popularity of cloud computing, many studies have proposed providing remote sleep monitoring as a cloud service for elderly patients at nursing homes (Biswas et al., 2010; He et al., 2013; Hossain, 2013). However, since most ISPs charge subscribers based on the volume of traffic, minimizing the amount of data to be transmitted is clearly a desirable aim (Peng et al., 2012). Lossy compression generally achieves much higher compression ratios than lossless compression, but at the cost of imperfections in the reconstructed signal. A trade-off thus exists between the amount of loss in signal fidelity that can be tolerated, and the compression ratio that can be achieved. Percentage Root-Mean Squared Difference (PRD) is a common measure of the loss of signal fidelity between two signals. The smaller the PRD, the lower the distortion introduced by the compression process, and although higher compression ratios are wanted, they result in

143

larger PRD. This paper considers a remote sleep monitoring system, as shown in Fig. 2. Our architecture consists of a wearable PSG system that records, compresses and wirelessly transmits the PSG data to a remote “server” that is responsible for analyzing the PSG signal. Through the use of an automatic rule-based sleep staging method (Liang et al., 2012), our goal is to examine how the sleep staging performance is affected by different compression levels. We used the popular off-the-shelf Set Partitioning in Hierarchical Trees (SPIHT) (Said and Pearlman, 1996) compression algorithm to compress PSG signal. SPIHT consists of two steps: first, SPIHT employs the Discrete Wavelet Transform (DWT) to decompose the signal into the wavelet coefficients. It then encodes the resulting coefficients into a binary stream based on the Embedded Zerotree Wavelet (EZW) coder (Shapiro, 1993). All-night PSG sleep recording data from 20 healthy subjects was used for testing. The data were compressed at a range of levels and passed through the automatic sleep staging system. The objective of this work is to understand the effectiveness of an off-the-shelf compression algorithm with regard to the performance of sleep stage detection. Numerous studies have attempted the compression of different physiological signals, such as those obtained from EEG (Daou and Labeau, 2012; Higgins et al., 2013; Srinivasan et al., 2013, 2011), ECG (Adel et al., 2012; Isa et al., 2012, 2014; Rubio et al., 2013; Shridhar and Mohankrishnan, 1984), EMG (Norris and Lovely, 1995), EOG (Bhandari et al., 2007) and echocardiogram (Cavero et al., 2013), and these generally focus on how to maximize the compression ratio using different coding techniques (Wegener, 2010). For example, Cárdenas-Barrera et al. (2004) proposed an algorithm to minimize power consumption when compressing EEG data. Our study is close to the work of Higgins et al. (2010), which investigated how an off-the-shelf lossy compression algorithm, such as JPEG2000 (Skodras et al., 2001), will affect the performance of seizure detection using EEG data. However, while all the prior works only consider the use of single-channel data, such as EEG or ECG, more phenomena can be discovered using multi-channel data (Kuo et al., 2013). For example, the characteristics of in S1 and REM stages are very similar in EEG data (see Fig. 1) but different and can be easily distinguished in EOG and EMG data. Our study focuses on the amount of loss in signal fidelity that can be tolerated when compressing complex multi-channel PSG data for sleep staging analysis.

2. Data source and methodologies 2.1. Subjects and recordings All-night PSG sleep recordings were obtained from 20 healthy subjects (12 males and 8 females) ranging from 19 to 23 years in age (mean = 21.2 ± 1.1). The data from five subjects were used to generate the system, and data from the other fifteen subjects were used for testing. The recordings included 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 (positioned 1 cm lateral to the left and right outer canthi), and a chin EMG channel (Siesta 802 PSG, Compumedics, Inc.). The sampling rate was 256 Hz with 16-bit resolution. The 20 PSG sleep recordings were visually scored by a sleep specialist using the Rechtschaffen & Kales (R&K) rules (Rechtschaffen and Kales, 1968) with a 30-s interval (termed the epoch). According to AASM and R&K rules’ suggestion which is the gold standard of manual sleep scoring, C3-A2 is used as the main EEG channel for manual sleep scoring. Therefore, we follow their suggestion and adopt the use of C3-A2 as the EEG channel in our work.

144

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

Fig. 2. Block diagram of an automatic sleep scoring system architecture.

We analyze the data from three channels in our automatic sleep staging system: the central EEG (C3-A2), the difference of the two EOGs, and the chin EMG. The EEG and EOG data were filtered with an eighth-order Butterworth band-pass filter with a cutoff frequency of 0.5–30 Hz, and the EMG data were filtered with a 5–100 Hz eighth-order Butterworth band-pass filter. The continuous time signals were segmented with every 30-s epoch. 2.2. Compression Compression of PSG signals is useful with regard to the storage and transmission of the data. The Discrete Wavelet Transform (DWT) (Shensa, 1992) is a popular technique used by many compression algorithms to pre-process the signal. DWT decomposes a signal into a set of basic functions known as wavelets. The initial function, also known as the mother wavelet, is used to construct the other wavelets by means of dilation and shifting. The DWT coefficients are defined as the inner product of the original signal and the selected basis functions. These coefficients provide an alternative representation of the original signal and give good localization of the signal’s energy components from both the time and the frequency domain. This work uses SPIHT (Said and Pearlman, 1996) to compress the PSG data. SPIHT is a lossy compression method based on Discrete Wavelet Transform (DWT). As compared to other compression methods, DWT is generally more scalable as the transform process can be applied to a signal as many times as needed and hence very low compression ratios can be achieved (Talukder and Harada, 2010). In addition, as shown in the prior work (Garry et al., 2013; Khanam and Ahmad, 2012; Ranjeet et al., 2011), wavelet transform usually produces better results than the other methods for medical signals like EEG and ECG (Garry et al., 2013; Khanam and Ahmad, 2012; Ranjeet et al., 2011). Among all the popular waveletbased compression algorithms (Chagas et al., 2000; Higgins et al., 2013; Kim et al., 2006), it has been shown that SPIHT has the lowest compression ratio (Higgins et al., 2013), which makes SPHIT particularly useful to be implemented on a low-powered device for sleep monitoring. Finally, SPIHT arranges the bits in order of significance, with the most significant bits being encoded first. Therefore, if the encoding or transmission is interrupted at any point, the signal can be reconstructed to a level of fidelity appropriate to the number of bits received, which makes SPIHT very suitable for transmitting large PSG data over the Internet since users with slower connection

speeds can download only a fraction of the data, obtaining much more usable results compared to other codecs such as Progressive JPEG (Bilgin et al., 2003). SPIHT exploits the inherent similarities across the subbands in a wavelet decomposition of an image. SPIHT performs binary partitioning decisions in order to determine the “significance” of each of the coefficients produced by the DWT. The goal of the partitioning decisions is to keep insignificant coefficients in large subsets; the larger the subset, the more efficiently they can be represented in the coded bit-stream. At each level, a threshold value is used by SPIHT to check the “significance”, which allows the wavelet coefficients to be encoded as binary numbers, through progressive bit-plane analysis. SPIHT orders the bits by significance, with the most significant ones being encoded first, so that an increasingly refined copy of the original image can be obtained progressively. In our experiments, we employed the Cohen–Daubechies–Feauveau wavelet filter to encode the PSG data and used four bytes to store the DWT coefficients. We chose a window size of 1024 in compressing the PSG data. Generally speaking, the choice of a larger window size requires more computational power, but leads to a better compression ratio (which suggests less transmission power). A window size of 1024 gives the minimum power consumption without losing signal fidelity. Table 1 shows a simple pseudo code to illustrate the procedure of the SPIHT algorithm.

Table 1 The pseudo code of SPIHT algorithm. Compression procedure Input: Compression level: Lv Input: Raw Data: [d1 . . . d1024 ] [c1 . . . c1024 ] = DWT([d  1 . . . d1024 ]) ; n = max( log2 ci );

O = ; for k = 1 to Lv Threshold = 2n ; if ∃ci > Threshold then Sk = encode(ci ); ci = 0; else Exit for-loop; end if O = O ∪ Sk ; n = n − 1; end for Ouput: O

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

E: EEG O: EOG M: EMG

145

Features

1 Wake, S1, S2, REM

SWS, S1, S2, REM

Alpha E 8-13 E

2 Wake, S1, S2

3 SWS, S2

REM, S1, S2

Alpha E 5-30 M

4

5

S2, S1

Wake, S1

REM, S1

0.5-4 E

6 0.5-30 E Spindle E SWS E

S2, S1

SWS

8

9 0.5-4 E Spindle E

10

Mean(fre.) E Mean(fre.) M

(9)

11

0.5-4 E 22-30 E

7 0.5-30 E Spindle E SWS E 0.5-4 O

REM, S1

S2, S1

0.5-30 E Spindle E SWS E

S2

(10) 12

0.5-4 E Spindle E

Amp M

REM, S1, S2

13 0.5-4 E Spindle E

Amp M

S1

S2

Wake

S1

REM

S1

S1

S2

REM

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(11)

(12) (13)

S1

S1

S2

(14)

No.

Type

Feature

Source

Label

No.

Type

Feature

Source

Label

1

PS

Total power of 0.5-30 Hz

EEG

0.5-30 E

7

SF

Mean frequency of 0.5-30 Hz

EEG

Mean(fre.) E

2

PS

Total power of 5-30 Hz

EMG

0.5-30 M

8

SF

Mean frequency of 5-30 Hz

EMG

Mean(fre.) M

3

PR

0.5-4 Hz/0.5-30 Hz

EEG

0.5-4 E

9

DR

Alpha ratio

EEG

Alpha E

4

PR

8-13 Hz/0.5-30 Hz

EEG

8-13 E

10

DR

Spindle ratio

EEG

Spindle E

5

PR

22-30 Hz/0.5-30 Hz

EEG

22-30 E

11

DR

SWS ratio

EEG

SWS E

6

PR

0.5-4 Hz/0.5-30 Hz

EOG

0.5-4 O

12

EMG energy

Mean amplitude

EMG

Amp M

Fig. 3. Diagram of the decision tree (top) and the 12 features for automatic sleep scoring (bottom) (Liang et al., 2012). * PS (power spectrum), PR (power ratio), SF (spectral frequency), DR (duration ratio).

2.3. Sleep staging method For the diagnosis of sleep issues, all night PSG recordings, including EEG, EOG and EMG data, are usually taken from subjects and the recordings are scored by a well-trained expert according to the R&K rules (Rechtschaffen and Kales, 1968). According to these, each epoch (i.e. 30 s of data) is classified into one of the sleep stages, including wakefulness (Wake), non-rapid eye movement (stages 1–4, from light to deep sleep) and rapid eye movement (REM). Stages 3 and 4 are frequently combined as the slow wave sleep (SWS) stage. In recent years, the R&K rules have started to be replaced by those developed by the American Academy of Sleep (AASM) which are fast-becoming the new clinical standard (Moser et al., 2009). Because visual sleep scoring is a time consuming and subjective process (Danker-hopfe et al., 2009), a number of automatic sleep staging methods have been developed, including rule-base (Liang et al., 2012), numerical classification (Agarwal and Gotman, 2001; Berthomier et al., 1998; Schaltenbrand et al., 1996) and hybrid approaches (Park et al., 2000). In rule-based methods,

signal information and human knowledge are combined to deduce a reasonable sleep state. Rule-based methods require the detection of specific patterns, such as K-complexes, sleep spindles in EEG and rapid eye movements in EOG. It can be time-consuming to construct a rule-based system, especially one based on human knowledge. In contrast, numerical classification methods do not require a set of rules or any human knowledge as spectral analysis is commonly used for feature extraction. However, without human knowledge and microstructural pattern recognition, some situations might not be taken into consideration, such as the transition between S1 and S2. Although hybrid systems take advantage of both methods, they are more complicated to implement and the results may show only incremental improvements over rule-based methods. In addition, the overall agreement of these methods is generally in the range of 80–85%, which is within the inter-scorer agreement range (Norman et al., 2000). In this work, we employ a rule-based method (Liang et al., 2012) by using a decision tree to identify different sleep stages. The main purpose of this study is to investigate if the quality of PSG

146

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

Spectral frequency (SF): After FFT, the mean frequency of spectral power (SF) was calculated for the EEG and EMG, using the following equation

30 SF =

f =0

f · PS(f )

30

f =0

(3)

PS(f )

Duration ratio (DR):

Fig. 4. Flow chart of the automatic staging system.

signal after a lossy compression can still be maintained at the level of the expert scoring. Therefore, we chose the rule-based method, which in principle is most similar to the classification model of a human expert, as our automated sleep staging method. Moreover, the rule-based method allows us to easily trace where misclassification occurs in the decision tree, which is very helpful to us in analyzing how the PSG signal and features are affected by the chosen compression method. The stages that were easier to identify were processed at the nodes in the upper layers, while the stages that were more difficult to distinguish were processed at the nodes in the lower ones. A distance measure is used to select the effective features of the nodes in the decision tree. There are a total of 13 decision nodes in the decision tree and 12 features for automatic sleep scoring, as shown in Fig. 3 (Liang et al., 2012). The automated staging procedure is shown in Fig. 4. The classification process is carried out in two steps: feature extraction and classification. 2.3.1. Feature extraction To identify the sleep stage, we need to extract the features in the signal. Before extracting spectral features, the signal was segmented into non-overlapping intervals of 2 s for a 512-point fast Fourier transformation (FFT) calculation. The spectra corresponding to the fifteen 2-s segments which were averaged to represent the spectrum for a 30-s epoch. The bottom of Fig. 3 lists the 12 features which can be generally divided into five different types including: (1) power spectrum, (2) power ratio, (3) spectral frequency, (4) duration ratio, and (5) EMG energy. Power spectrum (PS): After the FFT, the power spectrum (dB) was summed in the band 0.5–30 Hz for the EEG and EOG, and 5–30 Hz for the EMG; this was considered the total power. The total powers in the EEG and EMG were used as features and for the calculation of the power ratio. The total was obtained using the following equation: PStotal =

30 

PS(f )

(1)

f =0

where PS(f) is the power of the frequency f. Power ratio (PR): After the FFT, for one frequency band in a 30-s epoch, which has 15 powers, the mean power of each frequency band was collected. The ratio of each band to the total power 0.5–30 Hz was then calculated and considered as a feature. The power ratio is given by the following equation shown as below:

j PR =

f =i

30

PS(f )

f =0

(2)

PS(f )

in which, i and j indicate the range of the respective spectral power band for the PR features. The total bands of the power ratio in our features are 0.5–4 Hz, 8–13 Hz, and 22–30 Hz for EEG, and 0.5–4 Hz for EOG.

(1) Alpha ratio: The PSG signal in a 30-s epoch was also segmented into 15 2-s segments. The alpha ratio is the ratio between the number of alpha segments and the total segments in an epoch. Two eighth-order bandpass Butterworth filters with passbands of 8–13 Hz and 22–30 Hz were designed. In addition to the alpha band of 8–13 Hz that is normally used, a beta band of 22–30 Hz was added as a feature because we found that Wake also had high power in the 22–30 Hz band. The two filtered signals were then combined and a threshold (0.5) was used to detect whether or not a segment was an alpha one. More specifically if the value of the absolute amplitude for the combined signal relative to the original signal was higher than the threshold, the segment was set as alpha. (2) Spindle ratio: The spindle ratio is the ratio between the number of spindle segments and the total segments in an epoch. FFT and Butterworth bandpass filtering among the sigma band of 12–15 Hz were used to calculate the spindle ratio. FFT was used to find whether the power of the sigma band (12–15 Hz) was high, and the filtering signal was used to detect any large, sudden amplitude changes. If both were high, the segment was set as a spindle. (3) SWS ratio: The SWS ratio is the ratio between the number of SWS segments and the total segments in an epoch. We designed a third-order bandpass Butterworth filter with a passband of 0.5–2 Hz. If the absolute amplitude of the filtered signal was higher than the threshold (0.2), the segment was set as SWS. This was mainly used to separate SWS from the other stages. EMG energy: The mean value of the absolute amplitude of total data points in an epoch was calculated from the EMG signal and considered as a feature. During sleep, particularly during the REM stage, EMG activity decreases as compared to when one is awake. This feature can also be used for artifact detection as the EMG energy increases during body movement. After feature extraction, normalization of features was employed to reduce the effects of individual variability. For each feature, the mean of the maximal 10% data was calculated as the maximum value (max) of the feature for the subject, while the mean of the minimal 10% data was calculated as the minimum one (min). Once the min and max of a feature are calculated, they are used to normalize other values. Any value higher than max will be set to 1, while any value lower than min will be set to 0. This could prevent extremely high or low values from affecting our results. 2.3.2. Classification The extracted features are then used in a decision tree, as shown in Fig. 3, to separate the 30-s epochs into two different clusters for each decision node from top to bottom. Each decision node aims to separate two different stages. The details of the decision tree we employ in this work are described in Liang et al. (2012). After classifying the sleep stage using the decision tree, a smoothing process that, considers the temporal contextual information, was applied for continuity. These rules refer to the relationship between epochs before and after the current one. For example, three consecutive epochs of S2, REM, and S2 will be replaced with the sequence S2, S2, and S2. Similarly, consecutive

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

147

Fig. 5. The 10-s original and reconstructed signal of stage 2 for EEG, EOG and EMG data with different compression levels.

epochs of REM, S1, and REM will be replaced with the sequence REM, REM, and REM. 3. Results

SE =

3.1. Metrics used in our analysis 3.1.1. Compression ratio (CR) The compression ratio (CR) measures the efficiency of the compression process and is defined as CR =

Compressed data size Original data size

(4)

Generally speaking, a small CR is desirable. Most lossy compression algorithms, including SPIHT, are based on sub-band coding ´ 1995). In the traditional sub-band (SBC) (Vetterli and Kovaˇcevic, image coding algorithm, which uses a scalar quantizer, high frequency sub-band signals are usually distorted, which cause noticeable high-frequency noise in the reconstructed signal (Saito et al., 1991). SPIHT offers a range of threshold levels and while running it at a lower level results in a lower CR, this comes at the cost of a higher PRD (i.e. more high-frequency noise). On the other hand, using a higher threshold level can reduce the amount of loss in signal fidelity during the compression process, but will then increase the computation overhead (and power consumption) and result in a higher CR (which implies more energy consumption in wireless transmission). 3.1.2. Percentage Root-Mean Squared Distortion (PRD) Percentage Root-Mean Squared Distortion (PRD) measures the similarity between the original and reconstructed signal and is defined as follows: PRD =

||∂ − ˛|| ||∂||

staging method is evaluated by the confusion matrix that computes the sensitivity (SE) of overall agreement for each sleep stage. SE is defined as follows:

(5)

where ∂ and ˛ are the original and reconstructed signals, respectively, and  represents the Euclidean norm. By comparing the CR and PRD of the reconstructed signal, it can be seen how the decreasing CR is related to a loss in signal quality. 3.1.3. Sleep Staging Agreement The confusion matrix is the typical evaluation method for multi-classification problems. The performance of the sleep

TP , TP + FN

(6)

where TP, total number of correctly detected positive events; FN: total number of erroneously negative detections. 3.2. Compression performance Fig. 5 illustrates the effect of compression on the original and reconstructed PSG signal for different compression levels. We extracted a 10-s EEG, EOG, and EMG signal from stage 2 for compression levels from 7 to 4. As shown in Fig. 5, the features of the EEG signal in stage 2 such as spindle and K-complex start disappearing when the compression level is less than 7. As described previously, a lower compression level results in more loss in signal fidelity of the reconstructed data. We also can see that the noise increases as we reduce the compression level. Fig. 6 shows the plots of the CR vs. the PRD for different channels and difference sleep stages. Each point in Fig. 6 is the average of the CR and PRD of a total of 20 subjects in each sleep stage for one particular compression level (levels 9 to 5, from left to right). The duration of sleep data for each subject is from 7 to 8 h. We first manually identify each stage (i.e. Wake, Stage1, Stage2, SWS and REM) in different channels (i.e. EEG, EGO and EMG) for each subject and then compress it with different compression levels. As shown in Fig. 6, the interval between the PRDs of two adjacent compression levels increases with the reduction of the compression level, while the interval between the CRs of two adjacent compression levels decreases with the reduction of the compression level. In addition, for the same compression level, EEG and EOG data have similar CRs, and the compressed EMG signal has a higher CR than that of EEG across all sleep stages. Finally, for the same compression level, the EEG and EOG signals have similar PRDs (e.g. for S2 and SWS) while EMG data exhibits a better PRD than that of EEG. 3.3. Performance of sleep staging with different compression levels Table 2 shows a total number of 30-s epochs in each sleep stage of the 15 test subjects. The total number of epochs in our

148

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

Fig. 6. Compression ratio vs. PRD for each channel. Each point is the average of the CR and PRD in each sleep stage for compression level 9 to 5 (from left to right).

Table 2 Total number of 30-s epochs in each sleep stage from 15 testing subjects. Percentage of the epochs corresponding to each sleep stage was also given.

Total number of 30-s epochs Percentage

Wake

S1

S2

SWS

REM

Total

312 2.44%

450 3.52%

6904 53.93%

2292 17.90%

2844 22.21%

12,802

experiments is 12,802. A percentage of the epochs corresponding to each sleep stage were also given. The performance of sleep staging with different compression levels is shown in Fig. 7. Here the scoring from an expert on the uncompressed data is used as the ground truth. The overall sensitivity (SE) is 86.83% when the uncompressed data is used. The overall sensitivity increases as the compression

100 90 80

70 60 50 40 30 20 10 SE (%)

0

Lv 5 Lv 6 Lv 7 Lv 8 Lv 9 Uncompression

Wake 67.63 78.53 93.59 92.31 92.31 91.67

S1 43.47 30.63 29.28 34.46 34.46 35.56

S2 76.15 82.02 84.06 86.38 86.31 86.72

SWS 83.06 88.69 92.24 91.49 91.71 90.53

REM 49.86 69.26 86.33 91.21 91.25 91.7

Overall 70.48 78.61 84.3 86.61 86.62 86.83

Fig. 7. The sensitivity (SE) of each stage and overall with different SPIHT compression levels.

level increases from 5 to 9. When the compression level is equal to or higher than 7, the overall sensitivity and the sensitivity for all the stages are higher than 84% except for S1 (the overall sensitivities between the uncompressed data and the compressed data with the compression level 9 are almost the same); the overall sensitivity is acceptable in the criteria of the clinical usefulness since it is within the inter-scorer agreement range (80–85%) (Bruyneel et al., 2011). Relatively, the overall sensitivity cannot be accepted in the clinical application when the compression level is lower than 7. Generally speaking, it is difficult to classify S1 since it is a short-transforming state from Wake to S2. While many prior studies showed high accuracies with regard to overall agreement, their sensitivity for S1 is normally only around 20% (Schaltenbrand et al., 1996). In addition, as shown in Section 2, the extracted features from the signal are used in a decision tree to separate two different stages at the leaf node based on a threshold. Fig. 8 shows the distributions of the feature “Amp M” extracted from the original signal (Fig. 8(a)) and the signal reconstructed signal from different compression levels (Fig. 8(b)–(d)) in the real S1 and the REM stages (through human manual inspection). The x-axis shows the feature value and the yaxis is the total number of one particular feature value. The vertical red line indicates the threshold that was used, in this case 0.15. When the “Amp M” value of an epoch is greater than 0.15, then it is classified as the S1 stage, and if not as the REM stage. As shown in Fig. 8, we can see the total number of feature values in the REM stage which are greater than 0.15 increases as we decrease the compression level, which suggests more epochs in the

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

1

The distributions of feature "Amp M" in s1 and REM stage (Th=0.15)

2000 1500 1000 500 0

Wireless Transmission Energy

0.9

S1 REM 0

0.1

0.2

0.3

0.4 0.5 0.6 (a) Original signal

0.7

0.8

0.9

1

S1 REM 0

0.1

0.2

0.3

0.4 0.5 0.6 (b) compressiom Lv 7

0.7

0.8

0.9

1

Normalized Energy Consumption

2000 1500 1000 500 0

Compression Energy 0.8 0.7 0.6 0.50 0.5

0.42 0.37

0.4

0.33 0.28

0.3 0.2 0.1 0 Uncompression

2000 1500 1000 500 0

2000 1500 1000 500 0

149

5

6

7

8

9

Compression Levels S1 REM 0

0.1

0.2

0.3

0.4 0.5 0.6 (c) compressiom Lv 6

0.7

0.8

0.9

Fig. 9. Normalized power consumption when without SPIHT compression and using different compression levels (9 to 5 from right to left) in SPIHT.

1

S1 REM 0

0.1

0.2

0.3

0.4 0.5 0.6 (d) compressiom Lv 5

0.7

0.8

0.9

1

Fig. 8. The distribution of “Amp M” feature extracted from the original signal and the reconstructed signal from different compression level (7 to 5) in S1 and REM stage.

REM stage will be misclassified to the S1 one at a lower compression level. Table 3 shows the total number of misclassifications in the S1 and REM stages, using the feature “Amp M” extracted from the original signal and the reconstructed signal from different compression levels in the S1 and REM stages. Table 3 clearly shows that the total number of misclassifications increases when we employ a lower compression level, which results in a lower sensitivity (SE). Generally speaking, the effect of loss compression leads to more high frequency noise which in turns results in the distortion of feature values in EEG, EOG, and EMG signals. Therefore, a tradeoff needs to be made between the compression ratio and the accuracy of the automatic staging method. 3.4. Power consumption SPIHT is a lossy compression algorithm and encodes the data in an iterative way. The meaning of ‘level’ in SPIHT is the number of iterations used in compressing the data. A higher number of iterations suggests that more information is encoded in the raw data so that the final decoded data will have a smaller degree of distortion (but at the cost of a poorer compression ratio). The amount of Table 3 Total number of misclassifications using the feature “Amp M” extracted from the original signal and the reconstructed signal from different compression level in S1 and REM stage. Amp M

Original signal Compression Lv 7 Compression Lv 6 Compression Lv 5

Total number of misclassification S1

REM

Total

196 205 214 143

151 237 663 1209

347 442 877 1352

computation in SPIHT will increase as the number of iterations is increased. Therefore, compressing data at a higher level will introduce more energy consumption as compared to running SPHIT at a lower level. To evaluate the amount of energy saving we can achieve from SPHIT, we compare the energy used to transmit 1024-sample uncompressed data with the energy used to first compress the data and then transmit it. The results are then normalized to the energy consumption of transmitting the uncompressed data. The energy consumption of the wireless transmission is about 33 nJ for transmitting one bit using the Nordic nRF8001 transceiver a Bluetooth Low Energy (LE) transceiver. The energy consumption of compression was obtained by measuring the power consumption and execution time of the SPIHT program. As shown in Fig. 9, both the energy for compression and that for wireless transmission increase with increasing compression levels. Based on our measurement, compressing 1024-sample data at level 5 requires 124 ␮J and can reduce the data size by 92.66%. Compressing the data at level 7 requires 207 ␮J, and can reduce the data size by 88.65%. These results indicate a reduction in energy consumption of 72% and 63% for compression levels 5 and 7, respectively, as compared to uncompressed data transmission. This could offer significant energy savings with regard to providing a day-to-day remote sleep monitoring service in nursing homes (Biswas et al., 2010; He et al., 2013; Hossain, 2013). Taking into account the criteria of the clinical usefulness and the results from Figs. 7 and 9, we can concluded that the best tradeoff between the increased ratio of energy saving and the reduced overall sensitivity of the automatic sleep staging method is when the compression level is 7. 4. Discussion As discussed previously, the amount of high-frequency noise increases as we reduce the compression level. When compressing EEG, EOG and EMG with the same compression level, our results show that the overall performance is acceptable when level 7 is used, as shown in Fig. 7. In addition, when the compression level is decreased from 9 to 5, there is a significant drop in sensitivity (SE) for the Wake and REM stages. This is because when a low compression level is used, the Wake and REM stages could be misclassified as S1. Comparatively speaking, the EMG signal has more highfrequency components than the EEG one. In addition, even in the same signal (such as EEG), the amount of high-frequency components is different for different stages. Given that the nature of the

150

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

Table 4 The distribution distances (DDs) of the features for each node in the decision tree based on the uncompressed data and the compression data from level 9 to 5, respectively. Node no.

Separate stages

Feature

Uncompressed

Lv 9

Lv 8

Lv 7

Lv 6

Lv 5

1

Wake/SWS

2

Wake/REM

3

SWS/REM, S1

4 5

S2/Wake REM/S2

6

SWS/S2

Alpha E 8–13 E Alpha E 5–30 M 0.5–4 E 22–30 E 0.5–4 E 0.5–30 E Spindle E SWS E 0.5–30 E Spindle E SWS E 0.5–4 O

0.90 0.86 0.75 0.65 0.79 0.69 0.78 0.54 0.58 0.34 0.64 0.67 0.71 0.68

0.90 0.83 0.61 0.59 0.77 0.68 0.69 0.52 0.56 0.30 0.63 0.66 0.68 0.66

0.85 0.81 0.40 0.51 0.76 0.58 0.68 0.52 0.54 0.3 0.63 0.62 0.68 0.63

0.77 0.73 0 0.30 0.72 0.52 0.63 0.52 0.41 0.27 0.63 0.52 0.67 0.55

7 8

REM/S2 S1/S2

9

Wake/S1

0.5–4 E Spindle E Mean(fre.) E Mean(fre.) M Amp M

0.23 0.34 0.53 0.66 0.44

0.90 0.90 0.85 0.85 0.75 0.75 0.64 0.64 0.78 0.77 0.69 0.68 0.70 0.70 0.54 0.53 0.57 0.57 0.33 0.31 0.64 0.63 0.66 0.66 0.70 0.68 0.67 0.67 Same as node no. 5 0.19 0.15 0.31 0.31 0.51 0.51 0.65 0.65 0.32 0.32 Same as node no. 8 Same as node no. 10 Same as node no. 8

0.13 0.30 0.38 0.60 0.19

0.13 0.30 0.23 0.51 0

0.10 0.19 0.04 0.40 0

REM/S1 S1/S2 REM/S1 S1/S2

10 11 12 13

SPIHT algorithm is to compress the data by filtering out the highfrequency components in the signal, the compressed EEG data can achieve a better CR than that of the EMG data for the same compression level. However, a better CR comes at the cost of more information loss in the reconstructed signal and results in a worse PRD. In order to improve the performance of rule-based method, we systematically explore the effects of compression on the features used by each node of the decision tree (see Fig. 3). A distribution distance (DD) measure (Liang et al., 2012) is calculated for selecting the proper features at each nodes. More specifically, assuming that two stages A and B (e.g. Wake/SWS) are separated at the node 1 of the decision tree, the distribution distance (DD) of the feature with respect to A and B is calculated as follows:

 DD(A, B) =

1− 0

A + B ¯ 2|A¯ − B|

¯ if A + B ≤ 2|A¯ − B| else

¯ B) ¯ and ( A ,  B ) are the means and the standard deviaHere (A, tions (SD) of the feature used for separating stages A and B. A feature with a large DD value suggests that the values for this feature in stages A and B are significantly different. The analysis of DD can be used for selecting proper features at each node. We calculate DD of the features for each node in the decision tree from compression level 9 to 5. The results are shown as in Table 4. We can get the following observations from Table 4: (1) At a low compression level such as level 5 and 6, for most the features, the DD computed based on the compressed data is significantly different from that of original uncompressed data. In addition, at these levels, the computed DDs from the compressed data are mostly small (e.g. the DDs of node 10 are zero). This observation suggests, when the compression levels are set as low as level 5 or level 6, the data will be too noisy and

distorted so that most of the features we employ in this study become useless. (2) On the other hand, at a high compression level such as level 8 and 9, for most the features, the DD computed based on the compressed data is very similar to that of original uncompressed data (e.g. most of the features have a difference of less than 0.05 between the original and the compressed data). This finding suggests, when the compression levels are high (such as level 8 and level 9), most of the features exhibit similar performance between compressed data and original uncompressed data. Therefore, there is no need to change these features for the data compressed at level 8 or 9. (3) Finally, at the compression level 7, there are four features with significantly different DDs between compressed data and the original uncompressed data, which suggests these features might not be ‘good’ features for the compressed data. They are “Alpha E” at node 2, “0.5–4 E” at node 8, “Mean(fre.) E” at node 9, and “Amp M” at node 10. We tried to remove the first three features but not the last one from the decision tree (because “Amp M” is the only feature at node 10). The results based on the original decision tree and the modified decision tree for the compression level 7 are shown in Table 5, which indicates that, generally speaking, the modified decision tree cannot improve the sensitivity of each stage and overall agreement. Based on the above observations, we would like to conclude that, when the compression level is set higher than (or equal to) 7, our rules work well for both compressed and uncompressed data. On the other hand, with a compression level lower than 7, the compressed data simply becomes too noisy so that most of our rules perform poorly for separating the sleep stages. In order to understand how the compressed signal affects the scoring of a sleep specialist, we asked a sleep technologist to score the compressed data, as a comparison to the results from our automatic stager. Note that, some prior studies in sleep staging [19]

Table 5 The sensitivity of each stage and overall agreement with the use of the original and the modified decision trees at the compression level 7.

Lv7 (original) Lv7 (modified)

Wake (%)

S1 (%)

S2 (%)

SWS (%)

REM (%)

Overall (%)

93.59 92.76

29.28 37.93

84.06 81.64

92.24 91.51

86.33 86.16

84.3 83.08

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152 Table 6 The sensitivity of the overall result with decreasing compression levels for the EOG signal. Compression level

Overall SE

EEG = 7, EOG = x, EMG = 7

x = 7, SE = 83.57% x = 6, SE = 83.2% x = 5, SE = 82.99% x = 4, SE = 78.86%

showed that the characteristics of the EEG signal disappear when the PRD of the compressed signal is more than 7%. As shown in Fig. 6, the PRDs of the compressed EEG, EOG, and EMG signal are larger than 7% when the compression level is less than 7. Therefore, we only asked the technologist to score the sleep data compressed at level 7. Again, the uncompressed PSG signal is used as the ground truth, and the overall agreement between the manual scorings from uncompressed and compressed signal is 93.66%. Moreover, the sensitivities of Wake, S1, S2 SWS, and REM are 93.75%, 71.43%, 94.72%, 91.75%, and 94.96%, respectively. The agreement from this manual scoring is generally higher than that from our automatic sleep stager, which is not surprising, since a rule-based sleep stager is typically very sensitive to information loss due to the compression. Therefore, a rule-based method like ours tends to exhibit poorer agreement than other methods like manual scoring or numerical-classification-based sleep staging methods (Agarwal and Gotman, 2001; Berthomier et al., 1998; Schaltenbrand et al., 1996). Finally, we discuss the system performance when using different compression levels for different channels. As shown in Fig. 3, the feature extracted from EOG signal is only used in node 6. Based on this observation, we try to reduce the compression level of EOG, and the results are shown in Table 6. Initially, the compression level is set at 7 for all channels. The overall sensitivity was 83.57%, 83.2%, 82.99%, and 78.86% when the compression level of EOG was decreased from 7 to 4, respectively. The overall sensitivity is still acceptable when the compression level of EOG is reduced to five (higher than 80%). However, we were unable to reduce the compression levels of the EEG and EMG signals because the sensitivity of the overall results will not be acceptable when these are lower than 7, due to the above-mentioned reasons. In addition, features extracted from EEG and EMG signal are used in many nodes at the top of the decision tree, as shown in Fig. 3. If a misclassified epoch happens at the top of the decision tree, such errors cannot be corrected at the bottom of the decision tree. The performance of a decision-tree-based method thus strongly relies on the correct judgment of the node at the top of the tree. 5. Conclusion and future work In this paper, we discuss the use of a lossy compression algorithm for automatic sleep stages classification. The SPIHT algorithm was used to compress and decompress signals from a multi-channel PSG database, and we employed a rule-based method using 12 features extracted from multi-channel PSG data to classify the sleep stages. To the best of our knowledge, our study is the first work that focuses on how much loss can be tolerated in compressing complex multi-channel PSG data for sleep analysis. Considering the criteria of the clinical usefulness, the best tradeoff between the increased ratio of energy saving and the reduced overall sensitivity of the automatic sleep staging method is when the compression level is 7 (with which we can achieve more than 60% energy saving and a high (i.e. >84%) overall sensitivity). In addition, we propose a channel-independent compression scheme for multi-channel data, such as PSG to further reduce the amount of data that needs to be transmitted or stored. In our future

151

work, we would like to examine the effects of various parameters used in the compression algorithm (e.g. window size, bit-stream length, and size of the coefficient) on the performance of a sleep staging system. In addition, we plan to study the feasibility of using different compression levels for signals in different sleep stages. Acknowledgments This work was supported by the National Science Council of Taiwan and Ministry of Science and Technology of Taiwan under Grants NSC 101-2221-E-006-232, MOST 104-2811-H-006-004, NSC 101-2221-E-006-198-MY2, MOST 103-2221-E-006-178, NSC 1012221-E-006-098-MY3 and NSC 101-2221-E-006-150-MY3. References Adel H, Zahran O, Al-Nauimy W, El-Halafawy S, El-Sayed T, El-Samie FEA. SPIHT compression of ECG signals using an image processing concept. In: 8th international conference on informatics and systems (INFOS); 2012. Agarwal R, Gotman J. Computer-assisted sleep staging. IEEE Trans Biomed Eng 2001;48:1412–23. Agnew HW, Webb WB, Williams RL. The first night effect: an EEG study of sleep. Psychophysiology 1966;2:263–6. Berthomier C, Prado J, Benoit O. Automatic sleep EEG analysis using filter banks. Biomed Sci Instrum 1998;35:241–6. Bhandari A, Khare V, Santhosh J, Anand S. Wavelet based compression technique of electro-oculogram signals. In: 3rd Kuala Lumpur international conference on biomedical engineering 2006. Springer; 2007. p. 440–3. Bilgin A, Marcellin MW, Altbach MI. Compression of electrocardiogram signals using JPEG2000. IEEE Trans Consum Electron 2003;49:833–40. Biswas J, Maniyeri J, Gopalakrishnan K, Shue L, Eugene PJ, Palit HN, et al. Processing of wearable sensor data on the cloud – a step towards scaling of continuous monitoring of health and well-being. In: 2010 annual international conference of the IEEE Engineering in Medicine and Biology Society (EMBC); 2010. p. 3860–3. Bruyneel M, Sanida C, Art G, Libert W, Cuvelier L, Paesmans M, et al. Sleep efficiency during sleep studies: results of a prospective study comparing home-based and in-hospital polysomnography. J Sleep Res 2011;20:201–6. Cárdenas-Barrera JL, Lorenzo-Ginori JV, Rodríguez-Valdivia E. A wavelet-packets based algorithm for EEG signal compression. Inform Health Soc Care 2004;29:15–27. Casson AJ, Rodriguez-Villegas E. Data reduction techniques to facilitate wireless and long term AEEG epilepsy monitoring. In: 3rd international IEEE/EMBS conference on neural engineering, 2007 (CNE’07); 2007. p. 298–301. Cavero E, Alesanco A, Castro L, Montoya J, Lacambra I, Garcia J. SPIHT-based echocardiogram compression: clinical evaluation and recommendations of use. IEEE J Biomed Health Inform 2013;17:103–12. Chagas A, Da Silva E, Nadal J. ECG data compression using wavelets. In: IEEE computers in cardiology 2000; 2000. p. 423–6. Chang D-W, Liu Y-D, Young C-P, Chen J-J, Chen Y-H, Chen C-Y, et al. Design and Implementation of a modularized polysomnography system. IEEE Trans Instrum Measure 2012;61:1933–44. Colten HR, Altevogt BM. Sleep disorders and sleep deprivation: an unmet public health problem. National Academies Press; 2006. Danker-hopfe H, Anderer P, Zeitlhofer J, Boeck M, Dorn H, Gruber G, et al. Interrater reliability for sleep scoring according to the Rechtschaffen & Kales and the new AASM standard. J Sleep Res 2009;18:74–84. Daou H, Labeau F. Pre-processing of multi-channel EEG for improved compression performance using SPIHT. In: 2012 annual international conference of the IEEE Engineering in Medicine and Biology Society (EMBC); 2012. p. 2232–5. Edinger JD, Fins AI, Sullivan RJ Jr, Marsh GR, Dailey DS, Hope TV, et al. Sleep in the laboratory and sleep at home: comparisons of older insomniacs and normal sleepers. Sleep J Sleep Res Sleep Med 1997. Flemons WW, Douglas NJ, Kuna ST, Rodenstein DO, Wheatley J. Access to diagnosis and treatment of patients with suspected sleep apnea. Am J Respir Crit Care Med 2004;169:668–72. Garry H, McGinley B, Jones E, Glavin M. An evaluation of the effects of wavelet coefficient quantisation in transform based EEG compression. Comput Biol Med 2013;43:661–9. Griessenberger H, Heib D, Kunz A, Hoedlmoser K, Schabus M. Assessment of a wireless headband for automatic sleep scoring. Sleep Breath 2013;17:747–52. He C, Fan X, Li Y. Toward ubiquitous healthcare services with a novel efficient cloud platform. IEEE Trans Biomed Eng 2013;60:230–4. Higgins G, Faul S, McEvoy RP, McGinley B, Glavin M, Marnane WP, et al. EEG compression using JPEG2000: how much loss is too much? In: 2010 annual international conference of the IEEE Engineering in Medicine and Biology Society (EMBC); 2010. p. 614–7. Higgins G, McGinley B, Faul S, McEvoy RP, Glavin M, Marnane WP, et al. The effects of lossy compression on diagnostically relevant seizure information in EEG signals. IEEE J Biomed Health Inform 2013;17:121–7. Hossain MA. A survey on sensor-cloud: architecture, applications, and approaches. Int J Distrib Sens Netw 2013.

152

K.-C. Lan et al. / Journal of Neuroscience Methods 246 (2015) 142–152

Isa SM, Jatmiko W, Arymurthy AM. 3D SPIHT for multi-lead ECG compression. In: 2014 IEEE international conference on robotics and automation (ICRA); 2014. p. 488–93. Isa SM, Jatmiko W, Arymurthy AM. Beat reordering for optimal electrocardiogram signal compression using SPIHT. In: 2012 IEEE international conference on systems, man, and cybernetics (SMC); 2012. p. 226–31. Kelly JM, Strecker RE, Bianchi MT. Recent developments in home sleep-monitoring devices. ISRN Neurol 2012. Khanam R, Ahmad SN. ECG signal compression for diverse transforms. Inform Knowl Manage 2012:1–9. Kim BS, Yoo SK, Lee MH. Wavelet-based low-delay ECG compression algorithm for continuous ECG transmission. IEEE Trans Inf Technol Biomed 2006;10:77–83. Kuo C-E, Liang S-F, Lu S-S, Kuan T-C, Lin C-S. Estimation and prediction of drug therapy on the termination of atrial fibrillation by autoregressive model with exogenous inputs. IEEE J Biomed Health Inform 2013;17:153–61. Liang S-F, Kuo C-E, Hu Y-H, Cheng Y-S. A rule-based automatic sleep staging method. J Neurosci Methods 2012;205:169–76. Lubecke VM, Boric-Lubecke O. Wireless technologies in sleep monitoring. In: IEEE radio and wireless symposium (RWS’09); 2009. p. 135–8. Moser D, Anderer P, Gruber G, Parapatics S, Loretz E, Boeck M, et al. Sleep classification according to AASM and Rechtschaffen & Kales: effects on sleep scoring parameters. Sleep 2009;32:139. Norman RG, Pal I, Stewart C, Walsleben JA, Rapoport DM. Interobserver agreement among sleep scorers from different centers in a large dataset. Sleep 2000;23:901–8. Norris J, Lovely D. Real-time compression of myoelectric data utilising adaptive differential pulse code modulation. Med Biol Eng Comput 1995;33:629–35. Park H-J, Oh J-S, Jeong D-U, Park K-S. Automated sleep stage scoring using hybrid rule-and case-based reasoning. Comput Biomed Res 2000;33:330–49. Peng C, Tu G-H, Li C-y, Lu S. Can we pay for what we get in 3G data access? In: Proceedings of the 18th annual international conference on Mobile computing and networking. ACM; 2012. p. 113–24. Ranjeet K, Kumar A, Pandey RK. ECG signal compression using different techniques. In: Advances in computing, communication and control. Springer; 2011. p. 231–41.

Rechtschaffen A, Kales A. A manual of standardized terminology, techniques and scoring system for sleep stages of human subjects; 1968. Rubio ÓJ, Alesanco Á, García J. Secure information embedding into 1D biomedical signals based on SPIHT. J Biomed Inform 2013;46:653–64. Said A, Pearlman WA. A new, fast, and efficient image codec based on Set Partitioning in Hierarchical Trees. IEEE Trans Circ Syst Video Technol 1996;6: 243–50. Saito T, Higuchi H, Komatsu T. Sub-band coding with improved permutation codes for high fidelity image compression. In: IEEE international conference on acoustics, speech, and signal processing (ICASSP-91); 1991. p. 2641–4. Schaltenbrand N, Lengelle R, Toussaint M, Luthringer R, Carelli G, Jacqmin A, et al. Sleep stage scoring using the neural network model: comparison between visual and automatic analysis in normal subjects and patients. Sleep 1996;19: 26–35. Shapiro JM. Embedded image coding using zerotrees of wavelet coefficients. IEEE Trans Signal Process 1993;41:3445–62. Shensa M. The Discrete Wavelet Transform: wedding the a trous and Mallat algorithms. IEEE Trans Signal Process 1992;40:2464–82. Shridhar M, Mohankrishnan N. Data compression techniques for electrocardiograms. Can Electr Eng J 1984;9:126–31. Skodras A, Christopoulos C, Ebrahimi T. The JPEG 2000 still image compression standard. IEEE Signal Process Mag 2001;18:36–58. Srinivasan K, Dauwels J, Reddy MR. Multichannel EEG compression: waveletbased image and volumetric coding approach. IEEE J Biomed Health Inform 2013;17:113–20. Srinivasan K, Dauwels J, Reddy MR. A two-dimensional approach for lossless EEG compression. Biomed Signal Process Control 2011;6:387–94. Talukder KH, Harada K. Haar wavelet based approach for image compression and quality assessment of compressed image; 2010 arXiv:1010.4084. Vetterli M, Kovaˇcevic´ J. Wavelets and subband coding. Englewood Cliffs, NJ: Prentice Hall PTR; 1995. Wegener A. Compression of medical sensor data [exploratory DSP]. IEEE Signal Process Mag 2010;27:125–30.