This paper aims to estimate sleep stage with high accuracy. We improves the existing estimation method which analyses heart rate variability features obtained from the pressure sensor by the trigonometric function regression model for achieve it. Specifficaly, we focused on huge heart rate variability in order to take into account the tendency of the sleep cycle repeated in sleep, and to estimate sleep stage with each sleep cycle. The purpose of this improvement is to calculate sleep rhythms obtained by the intermediate frequency components of heart rate more accurately. To prove the effectiveness, we conducted the subject experiment for the two subjects and evaluated sleep by the 4 stages classification, REM, N1, N2, N3. The results showed the accuracy was 61.1 % which was higher 6.5 points than the existing method. This result suggests that estimating the sleep stage in each sleep cycle, in this case two divided part(the beginning and the latter sleep,) was effective to improve the estimate accuracy of sleep stage.
Sleep quality is one of the most significant factors on
wellbeing. Because it relates to human body and activities. For
example, sleep quality contributes to the physical and mental
systems and influences our health directly. Dementia is one
example of that sleep disturbance increases the risk
For measuring sleep condition, the professional people determine sleep condition as the evaluation of its depth. They have patients to get the examination called polysomnography (PSG) and analyze the obtained data such as brain activities, eye movement and muscles activity, then sleep depth is indicated. The analysis for measuring sleep depth must be conducted by several professional people, which takes deals of time. In addition, PSG gives huge burden to subjects, because subjects have to be attached some electrodes, which is unusual condition for them. Therefore, it is difficult in that way to measure sleep stage in the long term and to adapt it to several people such as children.
To overcome these shortages, sleep stage estimations have
been researched using the non-contact ways which are able
to obtain some biological data in place of brain wave and
so on. For example, one of the sleep stage estimations
employs the visual processing method and apply the obtained
biological data from image data
To solve these problems in some regression models of statistical analysis, we employ the existing sleep stage estimation applying another biological knowledge in addition to the relationship between the intermediate frequency component of heart rate based on the sleep cycle. Then, we attempt to improve this existing method and achieve the higher accuracy with the regression model. In the proposed method, we focus on the difference in a cycle of sleep stage. Sleep stage has the cycle repeated from the light sleep to deep sleep several time in sleep. Each cycle occurred during sleeping has different features in the value of the amplitude and the acrophase. To determine these features of each sleep cycle, we build the hypothesis that the heart rate variability changes with huge range when the sleep cycle switches. Based on it, we divide sleep into two parts and estimate in each part. By dividing as described above, the regression model fits the heart rate in comparison with the existing sleep stage method we employed, and the achieve to improve the accuracy of sleep stage estimation.
The rests of this paper are organized as follows. In the section of Sleep Mechanism, we explain sleep stage and the biological rhythms involved it. Next, some statistic approaches are introduced in section Related work, then we move on the our proposed method inspired by the related works. After that, we indicate experiment and their results, and analyze these in the section of Discussion. Finally, we conclude this paper.
According to the AASM, Sleep is divided into five stages,
Wake, Rapid Eye movement (REM) stage, N1, N2, N3 from
light to deep sleep
The human body has many biological rhythms which
involved in the physical activities. Sleep activities are affected
by two kinds of biological cycles called an ultradian rhythm
and a circadian rhythm. An ultradian rhythm is approximate
90 minutes cycles which composes sleep activities. Human
repeat the cycle of deep and light sleep in sleep based on this
rhythm. It can be calculated with heart rate variability in
specific frequencies range A circadian rhythm is approximate
24 hours cycle, which involved with the physiological
activities in a day, including sleeping and waking. It is acquired
by a core body temperature. Moreover, several researchers
showed the body core temperature is similar movement to
that of heart rate, Therefore, it is possible to observe body
core temperature by heart rate data as Figure1 shows
Watanabe proposed non-contact method for estimating sleep
stage based on the researches, that idea indicates the heart
rate relates sleep stage strongly
Real-time Sleep Stage Estimation (RSSE) can estimate
sleep stage with a non-contact device in real time
N X an cos( n=1 2 t L=n ) + bn sin(
(1) 2 t L=n ) h denotes the heart rate estimated at time t with the model parameter, which composes three kinds of parameters: (1) L denotes the maximum period in the intermediate frequency component of heart rate, (2) N denotes the number of the trigonometric functions composed of the prospective intermediate frequency, and (3) a set of parameters phi = a1; b1; :::; an; bn denotes the most approximated coefficients to adapt the raw heart rate, which are provided as the following likelihood function is minimized;
T J ( ) = 1 X(HR(t)
T t=1 h(t; ))2 +
N
N X(a2n + b2n) (2) n=1 where T denotes the time elapsed after a subject starts to sleep, HR denotes the obtained heart rate at time t. Next, the sleep stage is estimated by the discretization of the predictive heart rate h based on the equation: s(t) = >85 <
0
>:d (h(t; ) ave:) + 2e
stdev:
d (h(ts;td)eva:ve:) + 2e > 5;
d (h(ts;td)eva:ve:) + 2e < 0;
otherwise:
stdev: =
ave: =
s
max(T; L)
1
1
where s(t) denotes the sleep stage at time t. This
discretization formula is derived according to the previous
research
Finally, RSSE independently classifies Wake and REM stage. Wake detection focuses on the huge body movement (BM) during sleeping, which is calculated by the standard deviation of body movement in every minute and the average of body movement until current sleep according to the following function.
BMstd > 1:0 BMave (6)
For REM classification, RSSE focuses on the increasing ratio of heart rate during recent minutes, and detects the start sleep and the end of sleep to estimate the term for REM stage. To detect start of REM stage, RSSE compares the median heart rate within recent x minutes with the median one from recent 2x minutes to x minutes in every minute. In detail, when the increasing ratio exceeds a certain threshold, RSSE decides that the REM stage starts. With continuing to observe the increasing ratio, then RSSE decides the end of REM stage when the ratio starts to increase after the ratio gets zero. Additionally, to reduce misclassification between REM stage and Wake stage, RSSE cancels the REM detection in the start point if huge body movement occurs during 2x minutes before, because of the features that Wake stage occurs when heart rate increases rapidly.
(HRmreecdent
> 0:04 (7)
Real-time Sleep Stage Estimation (RSSE) obtains biological
data (heart rate and body movement) from biological
pressure sensor, and estimates sleep stage utilizing them
Therefore, in this paper, RSSE based on Circadian
Rhythm (RSSECR) is an improving method of RSSE to
estimate sleep stage, which focuses on circadian rhythm to take
into account time elapse. Specifically, the following
estimation methods are added to RSSE; (1) extraction of low
frequency components of heart rate, (2) change of distribution
in standardization for sleep stage and (3) revision by
comparing of intermediate frequency component and low
frequency component of heart rate.
(1) Extraction of low frequency components of heart rate
The low frequency component of heart rate as circadian
rhythm, while circadian rhythm is generally obtained by
measuring core body temperature
f (t) = fh(t; MF ) h(t; LF )g (8) h(t) denotes the heart rate at time t, M F denotes the intermediate frequency component of the heart rate, and LF denotes the low frequency component of the heart rate. is set between 0:1 and 1:0, and is indicated the degree of influence of the circadian rhythm.
RSSECR achieves to improve the accuracy of sleep stage estimation in comparison with RSSE, however, the sleep stage waveforms are not enough to observe sleep situation. Because the effects of correction in (3) are too strong and counteract the details movements in sleep stage waveform. Although, the correction in (3) works as an adjustment for a gap of the intermediate frequency components of heart rate compared with heart rate, the accuracy of the adjustment frequency tends to be depended on the accuracy of the intermediate frequency. Therefore, the high accuracy intermediate frequency components of heart rate is essential to maintain sleep stage waveforms with high accuracy of sleep stage estimation.
For the problem in RSSECR, we propose the improving method to estimate the intermediate frequency with more fitted waveform to the raw heart rate. The improvement in the proposed method is to divide whole sleep data into two parts depended on each sleep cycle. This improvement is effective when the heart rate variability is huge during sleep. Because, the synthetic wave for the ultradian rhythm is calculated to minimize the gap between the raw heart rate and the intermediate frequency components of heart rate in the regression model, which is difficult to capture each sleep cycle. To solve it, we estimate the intermediate frequency components of heart rate on the suitable durations by dividing sleep data. The proposed method employs the estimation process of RSSECR excluding the correction in (3), and the step is added before estimating the intermediate frequency components of heart rate. In order to estimate sleep cycle as the added step, we focus on the detection for REM stage in RSSE to estimate sleep cycle. Each sleep cycle gradually gets deep stage and become light stage, and whose aggregation composes of sleep. Therefore, the REM detection is utilized to distinguish when the cycle changes from the deep stage to light stage. Based on it, sleep is divided as a one cycle when the REM detection works first after 30 minutes from falling asleep. According to the above statement, the following sequence shows how to estimate sleep stage in the proposed method; (step 1) to obtain the heart rate from the biological sensor, (step 2) to divide sleep into two parts by the REM detection, (step 3) to estimate the intermediate frequency and the low frequency components of the prospective heart rate, (step 4) to calculated the difference between the intermediate frequency and the low frequency, (step 5) to standardize the difference and (step 6) to estimate sleep stage by the discretization of z values. In this paper, we propose two ways to estimate the adjusted frequency components of heart rate in the step 5 as following; Case1: to calculate the average and standard deviation with whole sleep data. Case2: to calculate the average and the standard deviation in each sleep cycle. Both cases attempt to improve the estimation of the intermediate frequency components of heart rate, however the aim is different. Case1 considers the relationship with the whole sleep data, on the other hands, the aim in Case2 is to focus on each sleep cycle only excluding the consideration of the variation in whole sleep.
To verify the effectiveness of the proposed method subject experiments are conducted. In this experiment, one day’s worth of sleep was measured for each of two healthy subjects, and that data was used. At that time, in order to get used to the measuring equipment like PSG, the subjects put on the equipment for two days and go to bed in advance. Those data were not used in the experimental results. Each subject wore the PSG of measuring sensor and slept with laying the biological sensor under the bed so as to measure their heart rate at the same time. This experiment used EMFit sensor as to measure the biological sensor. EMFit sensor developed by the VTT Technical Research Center of Finland as a non-contact biological sensor, which can measure heart rate as pressure signal every one second. On the day before the sleep measurement day, the subjects were not let activities that disrupted the living rhythm such as all night, and excessive exercise and drinking on the day.
The proposed method is compared the accuracy and the sleep stage between RSSECR and the following two cases;(1) The first part was estimated with the sleep data from falling asleep to have a REM flag, (2) the first part was estimated with the overall sleep data. The accuracy is evaluated sleep stage in four stages excluding Wake(REM, NREM1-3) by the comparison in the sleep stage estimation with PSG as the correct answer. Because Wake is a sudden occurrence, it can not be estimated by the biological rhythm that the proposed method focuses on. And NREM4 tends to appear more often in younger people, the sleep of these subjects did not appear 4.
Table1 shows the result. In each method, the proposed methods, Case1 and Case2, showed higher accuracy than RSSECR. In particular, the result of subject 1 was improved a high accuracy.
Figure2 shows the sleep stage with PSG in the Subject2 and the result of the REM detection. The blue line represents the sleep stage estimated using PSG, and the orange line represents the Rem detection. The horizontal axis is the sleep time, and the vertical axis for the blue line represents the sleep stage and for orange the presence or absence of detection. The REM stage by estimating using PSG was separated overall of the sleep stage into four parts. And three of them were estimated correctly by the REM detection excluding in the last part. The reason why the first REM detection was incorrect is that heart rate and body movement are unstable after bedtime, so different estimation methods are needed. However, in order to know the disturbance of the rhythm, there is no problem since REM sleep can be accurately estimated by this estimation method.
Based on these results of REM detection, sleep cycles were divided into two parts, which contributed to estimate heart rate variability more correctly with the trigonometric regression function model. Figure3 shows the results of the estimated heart rate. The top graph shows the relation between estimated heart rate with whole sleep data and raw heart rate which was obtained by the biological sensor. The bottom graph shows the result which distinguished sleep cycles to two parts compared with raw heart rate. The wave of the proposed method was more fitted than the wave of RSSECR. The top graph in Figure4 shows the heart rate and its low and middle frequency components, and the bottom graph shows the z value used in the proposed method. Figure 5 shows the results of estimating the sleep stage using these. Figure5 shows the results of sleep stage in each case compared with PSG. The point in these results was the sleep stage estimations in the first part and the last part. The first part as getting sleep was not estimated in RSSECR, however Case1 and Case2 distinguished more accurately. Moreover, the last part as getting awake was also estimated with high accuracies in comparison of RSSECR.
As Figure2 shows, the REM stage was estimated by PSG in the four parts, and the proposed REM detection estimated three of them correctly excluding in the first and last part. The estimation was not exactly correct for example the time of starting REM. Because, the REM detection was employed in order to recognize the sleep cycles, and the aim of it was not estimate REM stage in this experiment. However, the current REM detection excluded some results occurred within 60 minutes from getting sleep to avoid distinguishing in the wrong part. We need to verify the more appropriate threshold, while this threshold was set as 60 minutes based on the general length in an ultradian rhythm.
Focused on the raw heart rate variability, the overall average was 56.6. When considering the heart rate variability in the both 90 minutes from the beginning and end of sleep, the average was 59.1 in the beginning sleep, and it was 53.7 in the end of sleep. The average in the beginning was lower than the average in whole sleep, on the other hands, the average in the end was higher, which was one of the reasons that the estimation in the beginning and the end of sleep was difficult as the top figure shows in the Figure3. As the bottom graph shows, the heart rate variability estimated by the proposed method was more adapted the raw heart rate wave in comparison with the RSSECR, which verified the effectiveness of the sleep cycles.
This paper aims to increase the accuracy in sleep stage estimation based on the periodic regression analysis with heart rate variability which obtained by a pressure sensor. The proposed estimation attributes to find the appropriate parameter to approximate the low frequency so as to reflect the tendency over sleep, and to take into account the sleep cycle by dividing the sleep data. According to the subject experiment to indicate the effectiveness, the average of accuracy was 58.5 % and the following imprecations was revealed; (1) The low frequency should be approximated excluding short waves because it could perform as a noise; and (2) sleep stage must assess in the view of overall trend and the current trend with time while it contributes to concern the tendency with time.
This paper proposed two improvements of the current version of the RSSE method, which is proposed in our previous research. In particular, we proposed the use of the personal sleep feature to construct the model in the RSSE method in order to compensate for a biological information shortage, and the real-time Wake and REM sleep stage classifications depends on obtained heart rate and body movement. To investigate the effectiveness of the proposed method, we conducted the subject experiment. Two subjects participate this experiment, and two data of sleeping is obtained for each subject. The experimental result reveals that the proposed improvements increase the estimation accuracy of the RSSECR, and more than 50% of the estimation accuracy is achieved for all subjects and all experiment days by the integrated method of the RSSE method with the proposed personal sleep feature and the Wake and REM classifications. What should be noted here is that since this result is achieved only from two subjects, a verification with more human subjects is necessary. And some parameters in the Wake and REM classifications, e.g., the threshold of the body movement, are device dependent values, which should be automatically calibrated depending on the trend of the obtained biological data. These have to be pursued in the near future in addition to the following tasks: (1) Adequate and automatic selection or generation of the personal sleep feature for robustness of sleep estimation; and (2) an implementation of the WAKE and REM classifications to reduce misclassification.