Deep Sleep Detection Using Only Respiration

Yanjun Li, Xiaoying Tang and Zhi Xu

(1.Key Laboratory of Convergence Medical Engineering System and Healthcare Technology, the Ministry of Industry and Information Technology, School of Life Science, Beijing Institute of Technology, Beijing 100081, China;2.State Key Laboratory of Space Medicine Fundamentals and Application, China Astronaut Research and Training Center, Beijing 100094, China)

Abstract: Although polysomnogram (PSG) is the gold standard method for the evaluation of sleep quality, it becomes very difficult to clean the residual conductive gel in the hair after collecting brain electricity in the space weightlessness environment. This paper explores the feasibility of detecting deep sleep by using respiratory signal alone. Respiratory signals of oronasal airflow and abdomen movements were analyzed on ten healthy subjects from an open-access sleep dataset, namely ISRUC-Sleep. Deep sleep segments were detected by linear support-vector machine (LSVM) with three indices, including the amplitude variability in the time domain, the energy ratio of main respiratory band in the frequency domain, and the information entropy in the time-frequency domain. The Cohen’s Kappa coefficients were 0.43, 0.41 and 0.45 by general LSVM with feature vectors derived from oronasal airflow, abdomen movements and both respiration above, respectively. Moreover, the corresponding Cohen’s Kappa coefficients were 0.48, 0.41 and 0.49 by individual LSVM, respectively. Respiration-based method can achieve a moderate accuracy for the detection of deep sleep, with individual LSVM a little better than the general LSVM. Using this approach, detecting deep sleep automatically is attainable by respiratory signals from unconstrained and contact-free measurement. It can be applied to the sleep monitoring for astronauts on orbit.

Key words: sleep quality; sleep scoring; sleep stage classification; deep sleep detection; time-frequency analysis

Sleep is an active process that relates to the central nervous system (CNS) and the autonomic nervous system (ANS)[1]. The activity of the CNS is mainly analyzed on the bases of the electroencephalogram (EEG), the electrooculagram (EOG) and the electromyogram (EMG)[2]. Nevertheless, the activity of the ANS is revealed by the sympathovagal balance, which is usually quantified by the heart rate variability (HRV) and the blood pressure variability (BPV). Sleep is a global and systematic behavior[3]as different sleep stages not only affects the CNS but also partially influence the ANS.

Sleep quality could be revealed by either macrostructure or microstructure. The macrostructure consists of sleep stages, including wakefulness, rapid eye movement (REM) sleep and non-rapid eye movement (NREM) sleep. REM sleep is also called the active sleep (AS), while NREM sleep is also called the quiet sleep (QS)[4]. NREM sleep is further divided into light sleep (stage N1 and N2) and deep sleep (stage N3). According to the Manual for the Scoring of Sleep and Associated Events[5], macrostructure is scored mainly with EEG, EOG and EMG. Besides, the microstructure comprises the assessment of the cyclic alternating pattern (CAP)[6]. CAP is characterized by sequences of transient electrocortical events during NREM sleep, which reflects the instability of the sleep process[6].

Sleep disorders have negative effects on health, as low quality sleep is associated with fatigue, increased risk of cardiovascular disease, impaired cognitive functioning, reduced alertness, mood disturbances, impaired work productivity, decreased immune functioning, and so on. Poor sleep is usually revealed in the form of increased sleep fragmentation and light sleep, prolonged sleep latency, reduced deep sleep and REM sleep, difficulties in initiating and maintaining sleep, nightmares, movement disorders and frequent awakenings[7].

Although polysomnogram (PSG) is the gold standard method for scoring sleep stages and diagnosing sleep disturbance, it is an expensive, time-consuming, and labor-intensive procedure for data collection and data interpretation, which is usually performed in sleep centers or hospitals[8]. Moreover, multiple electrodes and various sensors are attached directly to the body throughout the night for PSG acquiring, which makes the subjects very uncomfortable and severely affects sleep quality and sleep structure[9]. It is important to monitor health status noninvasively in the living environment for long-term health management[10]. Hence, development of non-intrusive and even non-contact techniques for sleep stages identification is desirable[11], especially for the older, the sick and the disable[2]. Nowadays, bio-signals may be collected by embedding sensors into a watch, a hat, a belt, a shirt or a bed mattress, and so on. For example, the static-charge-sensitive bed and biological radars are two well-known, non-intrusive, contact-free and unconstrained methods to measure body movement, heart rates and respiratory fluctuations, providing a possibility of estimating sleep stages without disturbing sleep.

For decades, various unconstrained methods based on cardiopulmonary signals have been introduced for the improvement of health management at home. Recent researches showed that the cardiorespiratory system is related with sleep stages. A hidden Markov model (HMM) achieved an accuracy of 79% for wake-NREM-REM classification by heartbeat interval and movement activity from the bed sensor[12]. Using combined features derived from RR-intervals, ribcage respiratory effort and electrocardiogram-derived respiration (EDR) signal, moderate accuracies of 79% and 67% were achieved on 37 subjects with obstructive sleep apnea for wake-NREM-REM classification by subject-specific classifier and subject-independent classifier, respectively[11]. An accuracy of 89%, a specificity of 93%, and a sensitivity of 79% were achieved with support vector machine (SVM) in sleep staging by a combination of 27 indices of time domain, time-frequency, and fractal features derived from HRV and respiratory signals[13]. Phase synchronization is also related with sleep stages. Hamann et al.[14]suggested that a decrease in phase-locking is caused by the desynchronized activity of the nervous system during REM sleep between the cardiac and respiratory signals. The percentage of cardiorespiratory coordination and the duration of coordinated epochs are higher during slow-wave sleep (SWS) compared with REM sleep, and 4∶1 is the most frequent phase locked ratio of heart rate to respiratory rate[15]. However, Censi et al.[16]suggested that phase locking phenomena among respiration, HRV and BPV tended to be transient rather than permanent.

According to American Academy of Sleep Medicine staging, deep stage refers to N3 stage. The prevalence of EEG delta wave is an indicator of sleep depth. However, the hygienic cleaning of the residual conductive gel in the hair after EEG collecting is very difficult in the space weightlessness environment. Encouragingly, recent research has shown that EEG delta activity could be partially revealed by the ANS activity. The ANS oscillations were significantly coupled with a “mirror-image” to the overnight oscillations in EEG delta wave[17], while cardiac changes preceded the changes in EEG several minutes[18]. Although some researchers had used cardiorespiratory activity for sleep stage classification, seldom researchers have used respiratory signal alone for sleep stage classification. Although sleep scoring based on respiratory signal alone is a great challenge, specific sleep stage detection is possible. We put forward the hypothesis that deep sleep stage might be characterized by more regular breathing than non-deep stages. In this paper, the possibility of detecting deep sleep was explored by using respiratory signal alone, which was not usually used for this purpose. The regularity of respiration was measured by indices derived from the time, frequency and time-frequency domains.

1 Methods

1.1 Database

Respiratory signals of oronasal airflow and abdomen movements from ten records (10 healthy subjects, 9 males and 1 female, age 40±10 years and one record per subject) were studied in this paper from the open-access sleep dataset ISRUC-Sleep[19]. ISRUC-Sleep contains both PSG and annotations of sleep stages, which could be downloaded freely from the URL “http:∥sleeptight.isr.uc.pt/ISRUC Sleep/”. Both oronasal airflow (labeled “DC3” in the dataset) and abdomen movements (labeled “X8” in the dataset) were sampled at a frequency of 25 Hz per channel. Each sleep stage lasts 30 s. Therefore, the moving step for analysis is 30 s.

1.2 Feature extraction

1.2.1Amplitude variability in the time domain

If the peak and the valley in each respiratory cycle could be identified, the amplitude variability in the time domain seems easy to be obtained. However, the respiratory waveform usually shows up in multiple morphologies, especially in wakefulness and REM sleep. Therefore, a generalized index of amplitude variability is defined:

The respiration rate usually lies in the range of [0.15, 0.4] Hz, i.e., the respiratory cycle lies in the range of [2.5, 7.0] s. Each stage with duration of 30 s is divided into 4 fragments, and each fragment lasts 7.5 s. Therefore, each fragment with duration of 7.5 s contains at least one complete respiratory cycle. The difference between the maximum and the minimum value in theith fragment is defined asDi, wherei=1, 2, 3, 4. The amplitude variability is defined as follows

vA=std(X)/mean(X)

(1)

where “std” is the standard deviation (SD) and “mean” represents the average value of the vector, and X=[D1D2D3D4].

WhenvAis close to 0, the amplitude of respiration peaks tend to be consistent and the epoch is likely to be deep sleep. Conversely, whenvAis huge, the amplitude of respiration peaks tends to be variable and the epoch is likely to be wakefulness or REM sleep.

1.2.2The energy ratio of main respiratory band in the frequency domain

The power spectrum distribution (PSD) of the respiratory signal was calculated for each segment with 30 s by the fast Fourier transform (FFT). Then the respiratory frequencyfcthat corresponds to the maximum value in PSD curve was obtained by finding the frequency value located in the range of [0.15, 0.4] Hz. Afterward, the energy ratio of respiration peak in the PSD with a bandwidth 0.06 Hz to the energy within the band [0.15, 0.4] Hz was defined as follows

(2)

WhenrRis close to 1, the respiratory rate tends to be consistent and the epoch is likely to be deep sleep. Conversely, whenrRis close to 0, the respiratory rate tends to be variable and the epoch is likely to be wakefulness or REM sleep.

1.2.3The information entropy in the time-frequency domain

To obtain the information entropy (Shannon entropy), the following steps were involved:

Firstly, the PSD in the range of [0, 0.4] Hz of each segment with 30 s was regarded as a column vector, as follows

xi=[s1,is2,is3,i…sk,i…sM-1,isM,i]T

(3)

whereirepresents theith 30 s segment and the superscript T represents the transpose matrix.

Secondly, each elementsi,jwas transformed topi,jwhich is in the range of [0, 1], as follows:

(4)

Matrix P is constructed withpi,j.

(5)

The information entropy (Shannon entropy) in each column is calculated by

(6)

1.3 Deep sleep detection

1.3.1General LSVM

The leave-one-record-out strategy was used for general LSVM training. Deep sleep stage is labeled as “-1” and other stages are labeled as “1”. The LSVM model was trained with an input vector [vArRIj] from 9 records, and it was tested with the remaining record, ensuring that results derived from LSVM were validated on a fresh record to avoid over optimized thresholds. This process is repeated 10 times so that each record is trained on the general LSVM derived from the rest 9 records.

1.3.2Individual LSVM

For each record, the odd number data segments consist dataset A (the first, the third, …, segments), and even number data segments consist dataset B (the second, the fourth, … segments). Firstly, dataset A was used for training and dataset B for testing. Consequently, dataset B was used for training and dataset A for testing. Finally, the above two time detections consists the complete detection results for one record. As a result, the test data set was withheld from the training dataset, making test sample set fresh from the training sample set.

In addition, LSVM classifier was implemented through LIBSVM Toolbox[20], which is available freely online at “http:∥www.csie.ntu.edu.tw/～cjlin/libsvm/”.

1.4 Validation and performance assessment

Deep sleep detection results were compared with the annotation on the open source database ISRUC-Sleep to get four indices of true positives (TP), true negatives (TN), false positives (FP), and false negative (FN), respectively. Then accuracy (ACC), sensitivity (SEN), specificity (SPE) and positive predictive value (PPV) were calculated according to the following equations, respectively.

ACC=(TP+TN)/(TP+TN+FP+FN)

(7)

SEN=TP/(TP+FN)

(8)

SPE=TN/(TN+FP)

(9)

PPV=TP/(TP+FP)

(10)

TP is the number of deep sleep epochs annotated in the database that are actually classified as deep sleep epochs by the classifier. TN is the number of non-deep sleep epochs annotated in the database that are actually classified as non-deep sleep epochs by the classifier. FP is the number of non-deep sleep epochs annotated in the database that are actually classified as deep sleep epochs by the classifier. FN is the number of deep sleep epochs annotated in the database that are actually classified as non-deep sleep epochs by the classifier.

Additionally, the Cohen’s Kappa was used to assess the epoch-by-epoch agreement between the reference deep sleep staging and the computed deep sleep staging. Usually the result is excellent when Kappa is over 0.75, and it is fair or good when Kappa is from 0.40 to 0.75. However, the result is poor when Kappa is below 0.40.

2 Results

Fig.1 is one example of deep sleep detection from record “3”. In Fig.1a, non-deep sleep epochs are characterized by scattered frequency bands distributed in a large area of the PSD, which reflects the irregularity of the changes in respiratory rate. In contrast, deep sleep epochs are characterized by one main frequency band distributed in a small area, implying the regularity of respiration.

Fig.1 Deep sleep detection by general LSVM with oronasal airflow on record “3”

Deep sleep epochs in the database annotation are marked as blue in Fig.1b. W represents wakefulness, R represents REM sleep, L represents light sleep, and D represents deep sleep. In Fig.1c, mostvAduring deep sleep is lower than that during other stages. In Fig.1d, mostrRduring deep sleep is higher than that during other stages. In Fig.1e, most entropy during deep sleep is lower than that during other stages, representing that respiration during deep sleep is associated with higher certainty. It is very obvious in Fig.1 that deep sleep epochs are consistent with lowvAand entropy as well as highrR. Encouragingly, the final deep sleep detection results by general LSVM with oronasal airflow in Fig.1f are close to the true deep sleep in the database annotation, with the accuracy of 84.3%, the sensitivity of 77.0%, the specificity of 89.0%, the PPV of 81.3% and the Cohen’s Kappa coefficient of 0.67, as shown in Tab.1. In Fig.1f, D represents deep sleep and ND other stages. The top two-lines are derived from the database annotation, and the bottom two-lines are the deep sleep detection results.

Fig.2 shows the deep sleep detection for record “10” by general LSVM with oronasal airflow, abdomen movements and both of them respectively. The normalized time-frequency matrix is similar in appearance in Fig.2a and Fig.2b. As shown in Tab.1, the Cohen’s Kappa coefficients are 0.62, 0.64 and 0.64, respectively, meaning that abdomen movements achieve better performance than oronasal airflow. However, for most records, performance of oronasal airflow is a little better than that of abdomen movements. In Fig.2d, the top two-lines are derived from the database annotation, the three pairs of two-lines below that are deep sleep detection results from the oronasal airflow, the abdomen movements and the combined of them, respectively.

Fig.3 shows the deep sleep detection for record “2”. As shown in Tab.1 and Tab.2, the Cohen’s Kappa coefficients for record “2” with oronasal airflow by general LSVM and individual LSVM are 0.27 and 0.50, respectively. For most records, individual LSVM is a little better than general LSVM. In Fig.3c, the top two-lines are database annotation, the two pairs of two-lines below that are deep sleep detection results by general LSVM and individual LSVM, respectively.

Tab.1 Comparison of the algorithm’s performance among oronasal airflow, abdomen movements and combined signals with general LSVM model

Fig.2 Deep sleep detection by general LSVM on record “10”

Fig.3 Deep sleep detection by general LSVM and individual LSVM on record “2” with oronasal airflow

According to Tab.1, accuracies were 77.7%, 77.2% and 78.3%, and specificities were 80.6%, 81.3% and 81.1% for oronasal airflow, abdomen movements and the two combined, respectively. However, corresponding sensitivities were only 67.9%, 63.2% and 68.5%, respectively, and positive predictive values were only 50.9%, 50.1% and 51.9%, respectively. The Cohen’s Kappa coefficients were 0.43, 0.41 and 0.45 for oronasal airflow, abdomen movements and the two combined, respectively, indicating a fair or moderate epoch-by-epoch agreement between the reference and the computed results. It also held true for Tab.2.

Tab.2 Comparison of the algorithm’s performance among oronasal airflow, abdomen movements and combined signals with individual LSVM model

According to Tab.2 and Tab.1, the Cohen’s Kappa coefficient improved from 0.43 to 0.48 by individual LSVM for oronasal airflow, and it improved from 0.45 to 0.49 for the combination of oronasal airflow and abdomen movements. However, the Cohen’s Kappa coefficients were both 0.41 by general LSVM and individual LSVM for abdomen movements.

3 Discussion

Changes in respiration are respective to sleep states. Breathing is more regular during deep sleep than during wakefulness and REM sleep[3]. The respiratory frequency and variability increases from NREM to REM sleep[9]. The SD of the respiratory cycles was 0.27 s during deep sleep while it was 0.42 s during wakefulness on sixteen healthy subjects[21]. Chung et al.[9]assessed REM sleep by measuring the respiration rate variability. The sensitivity, the specificity and the accuracy in normal subjects with nasal airflow were 74%, 92% and 89%, respectively. The corresponding values for belt-type sensors were 71%, 93% and 89%, respectively[9].

As respiration is regular in most segments of deep sleep, three indices were defined in this paper for determining deep sleep stage by the regularity of respiration, involving the time domain, the frequency domain, and the time-frequency domain. The respiratory indices in the time domain and the frequency domain had been proposed in our previous study for NREM sleep detection with oronasal airflow from the expanded sleep database in the European data format (EDF)[22]. The error rate was 28.3% and 30.8% by linear discriminant functions of individual-dependent model and independent model, respectively. However, the Shannon entropy was first applied to the distribution of respiration in the time-frequency domain in our study.

For regular respiration, the amplitude variability of wave peak is small in the time domain, and the energy distribution focuses on a narrow band with the center as the respiratory rate in the frequency domain. Moreover, the number of wave peaks during deep sleep should be single rather than multiple peaks in the PSD. As reported by Kirjavainen et al.[23], there was only 9.4% of irregular breathing in deep sleep, while the percent of irregular breathing in light sleep and REM sleep were 36% and 70.7%, respectively. The basis of the theory in this paper is that respiration tends to be more regular during deep sleep than during other stages, and this may be used for deep sleep detection. The proposed method identified deep sleep with the combined features extracted from the time domain, the frequency domain, and the time-frequency domain of respiratory signal. It detected deep sleep on the basis of breathing regularity. The performance of the proposed algorithm is validated by making comparison with the annotation of the ISRUC-Sleep database.

In this paper, the PPV is only about 50%, meaning that only about half detected deep sleep epochs are consist with the annotation, the other half are all false positive. As shown in Fig.4, about 70% of false positives are light sleep, meaning that some light sleep epochs are also associated with regular respiration.

Fig.4 Distribution of false positives for the combination of oronasal airflow and abdomen movements

The proposed algorithm uses only respiration to detect deep sleep by the characteristic that respiration has high regularity during deep sleep. Although the respiration data in this study was obtained by attaching sensors, both the biological mattress and biological radar are non-intrusive, contact-free and unconstrained methods that not only record respiration data but also record the heartbeat and body activity. Respiration signals are easily obtained without attaching electrodes, suggesting the method’s applicability to non-invasive measurement systems.

The greatest limitation is that the number of false positives is near 50% because a few light sleep segments also have regular respiration, while the number of false negatives is near 30% by general LSVM because a few deep sleep segments accompany with sleep disorders have irregular respiration. In future work, our goal is to discriminate sleep stages of wakefulness, light sleep, deep sleep and REM sleep without attaching any sensors to obtain respiration, heartbeat and body activity.

In the space weightlessness environment, it becomes very difficult to clean the residual conductive gel in the hair after EEG collecting. This paper explores the feasibility of detecting deep sleep by using respiratory signal alone. Results show that deep sleep can be recognized by one single respiratory signal.

Single channel respiratory signal greatly lightens the burden of physical, psychological and mental load of multiple signals collected from traditional PSG. It constitutes a complement for the traditional clinical sleep evaluation, providing a possibility for deep sleep evaluation through signals that are recorded at home with non-intrusive, contact-free and unconstrained methods through systems integrated into the environment (such as biological mattress or biological radar). It also can be applied to the sleep monitoring for astronauts on orbit.