DEVELOPMENT OF A SLEEP MONITORING SYSTEM WITH

WEARABLE VITAL SENSOR FOR HOME USE

Takuji Suzuki

, Kazushige Ouchi

, Ken-Ichi Kameyama

and Masaya Takahashi

1,2

Humancentric Laboratory, Corporate Research & Development Center, Toshiba Corporation, Japan

National Institute of Occupational Safety and Health, Japan

Keywords: Daily sleep care, Sleep monitoring, Wearable sensor and Data viewer.

Abstract: This paper describes a new sleep monitoring system for home use. The basic system consists of a wearable

physiological sensor and PC software for analyzing sleep quality from user’s wrist motion and heart rate

variability. Different from a conventional sleep monitoring device used in a hospital, the sensor is so small

and easy-to-use that a normal person can use it at home. This means that the system is useful for a sleep

specialist who wants to check a patient's daily sleep pattern. The system can also be used for self-care. We

have developed a wrist-watch-shaped physiological sensor that monitors user’s wrist motion and pulse wave

interval. We have also developed the algorithm for computing the quality of sleep from these physiological

data on PC. Although sleep is a kind of brain activity and our sensor can not directly measure it, the output

of our algorithm is close to medically evaluated sleep quality. We performed dozens of comparison

experiments and found that its accuracy was about 73.5% on average. The value of the accuracy is enough

for assessing a normal person’s sleep quality.

1 INTRODUCTION

In recent years, many people have been suffering

from sleep disorder caused by mental stress,

irregular lifestyle or shift work. However, it is not

easy to determine the quality of sleep because deep

sleep is not always good sleep and shallow sleep is

not always bad sleep. For example, it is natural that

a person cannot sleep well because of jet lag.

However, a person who is always sleepy in the

daytime for a period exceeding one month might

have a health problem. Therefore, it is important for

a doctor to check a patient’s sleep habits for several

days in order to diagnose and cure his/her sleep

disorder properly. Moreover, it is necessary for a

person to check his/her own sleep habits and to

change his/her lifestyle (self-care).

However, there is no good system to record and

analyze daily sleep. For example, most medical

sleep sensors, such as those employed for

polysomnography (PSG), are for recording many

kinds of physiological data (EEG, EMG, EOG and

so on) for only one or two nights, not for recording

sleep habits with natural state in daily life. It is also

too difficult for a normal person to handle PSG at

home because it involves the use of many electrodes

for measuring the physiological data. A doctor can

attempt to learn a patient's sleep habits by

interviewing him or her, but this is an inherently

unreliable approach. A simple and easy-to-use sleep

monitoring system that can be used in the home is

strongly desired in order to get objective data on

sleep habits.

In order to develop such a system, we have

created a wrist-watch-shaped wearable physiological

sensor that monitors user’s wrist motion and pulse

wave intervals (Pulse-to-Pulse Intervals: PPIs). The

sensor can be made small and simple because wrist

motion and pulse wave can be easily measured

compared to the case of using PSG. We have also

developed the algorithm for computing the quality of

sleep from these physiological data. The algorithm

can distinguish sleep stage (wake /REM /NREM

[shallow /deep]) using the relationship between

autonomic nervous activity and sleep stages.

Although sleep is a kind of brain activity and our

sensor cannot directly measure it, the output of our

algorithm is close to medically evaluated sleep

quality.

In the following sections, the way of expressing

sleep data, related works, our system’s hardware and

software, and the validation result of the sleep

estimation are discussed.

326

Suzuki T., Ouchi K., Kameyama K. and Takahashi M. (2009).

DEVELOPMENT OF A SLEEP MONITORING SYSTEM WITH WEARABLE VITAL SENSOR FOR HOME USE.

In Proceedings of the International Conference on Biomedical Electronics and Devices, pages 326-331

DOI: 10.5220/0001784203260331

 SciTePress

2 SLEEP DATA

Generally speaking, sleep is a kind of brain activity

and its purpose is recovery from brain fatigue.

Therefore, sleep state is measured mainly by EEG,

and is classified into several stages. Sleep state is

roughly divided into REM (rapid-eye movement)

sleep and NREM (Non-REM) sleep. NREM sleep is

divided into 4 stages. Stages 3 and 4 of NREM sleep

are so called deep sleep, and stages 1 and 2 are

shallow sleep. These stages are decided by a sleep

specialist using PSG data (Rechtschaffen, 1968), and

their change is shown in a graph called a hypnogram

(shown in Figure 1).

A doctor mainly uses a hypnogram for

evaluating a person’s sleep quality. For example, the

doctor checks the quantity of deep sleep if a patient

complains about oppressive drowsiness in the

daytime. If the patient frequently wakes up in the

night and experiences difficulty in breathing, he/she

might be suffering from sleep apnea syndrome. If

REM sleep always occurs soon after falling asleep,

there might be a problem concerning the patient’s

nervous system. From the viewpoint of healthcare, it

is important to check the balance of deep sleep,

REM sleep or sleep cycle. Therefore, a sleep

monitoring system for home use can also show the

result of one night’s data in a graph similar to a

hypnogram.

Figure 1: Sleep hypnogram.

There are many studies on the relationship between

physiological parameters and sleep stages. For

example, Baharav et al. stated that autonomic

nervous activity level derived from heart rate

variability (HRV) during sleep changes in response

to the sleep stages (Baharav, 1995). A value of

LF/HF shows the activity of the sympathetic nerve.

During a REM sleep, a value of LF/HF and the

variability of that are large, and the value of LF/HF

decreases during a NREM sleep, particularly in the

case of deep sleep (Slow Wave Sleep). Since the

brain stem controls both the cerebrum and the

autonomic nervous system, it may be possible to

estimate the sleep stage using HRV.

3 RELATED WORKS

A number of trials have been conducted with a view

to developing sleep monitors for home use. For

example, body/wrist motion has been used for

wake/sleep identification.

The amount of activity

(number of subtle wrist motions per minute)

measured

from acceleration sensors is often used for

monitoring wake/sleep rhythms (Sadeh, 1989)

although the sleep stages (ex. REM sleep / NREM

sleep) cannot be determined from the data.

More recently, researchers have focused on

measuring heart/pulse rate and analyzing its

variability: HRV (Watanabe, 2004, Michimori, 2003

and Wakuda, 2007). The sleep stages can be

calculated from HRV if the indices of HRV are

properly mapped for the sleep stages.

However, there are two problems in this

approach. One is that body/wrist motion often

disturbs heart/pulse sensing and the HRV value can

not be calculate correctly. The other is that the level

of autonomic nervous activity differs according to

age, sex and body/mental condition. For example,

the autonomic nervous system of the young is

generally more active than that of the old. Sleep

stages cannot be classified using static thresholds.

Our sensor measures both pulse wave interval

and wrist motion. The wrist motion data are used not

only for counting the amount of activity, but also for

detecting errors in HRV data. This solves the first

problem mentioned above.

For the second problem, we employ a statistical

method for deciding sleep stages (Suzuki, 2007). We

assume that there are several stages in a certain

period of sleeping time since the sleep stage

cyclically repeats about every 90 minutes. It means

that the data of autonomic nervous activity can be

classified into several groups if we have any 90-120

minute dataset. In this way, the thresholds for

dividing sleep stages are changed flexibly along with

the dataset.

4 THE OVERVIEW OF THE

SYSTEM

4.1 Wearable Physiological Sensor

Figure 2 shows our wearable physiological sensor.

The size of the sensor is 50mm*60mm*13mm and

the weight is only 35g. A rechargeable battery is

used as an electrical power source. It is possible to

measure physiological data for over 40 hours after

Wake

REM

Stage1

Stage2

Stage3

Stage4

DEVELOPMENT OF A SLEEP MONITORING SYSTEM WITH WEARABLE VITAL SENSOR FOR HOME USE

327

full charge. The sensor incorporates a photoelectric

pulse wave sensor and a 3-axis accelerometer.

Besides, it has an external pulse wave sensor.

Therefore, pulse waves can be measured on the

user’s wrist or on his/her finger, depending on

his/her preference. The front panel serves as a wrist-

watch displaying date and time, and as a sensor

displaying time and the amount of activity. The

sensor has only two buttons; namely, one is a light

switch, and the other is a switch to start/end sensing.

Figure 2: Wearable physiological sensor.

The sensor measures pulse waves and accelerations

on a user’s wrist and stores the computed pulse-to-

pulse intervals (PPIs) and the amount of activity in a

flash memory (4MiB). Both analog and digital filters

are used to remove the fluctuations of the amplitude

and the basal line of pulse waveform, which makes

PPIs more precise. As the size of the data measured

in one night (7 hours) is 256 KiB, the sensor can

store almost 2 weeks’ data in the flash memory.

The sampling rate of the pulse wave and 3-axis

accelerations is 64Hz. However, the resolution of the

PPI is 0.1 ms by using linear interpolation to detect

pulse peak.

The amount of activity is calculated as the

number that the scalar of the 3-axis acceleration is

larger than 0.01 G, which is the same as Actigram

(Cole, 1992).

The stored PPIs and amount of activity data are

sent to PC via USB.

We evaluated the performance. Firstly, the

correlation coefficient between the amount of

activity counted by the sensor worn on the left

forearm and that measured by an actigraph

(Micromini-Motionlogger Actigraph, Ambulatory

Monitoring Inc.) worn on the right forearm during

sleep was 0.95 (average of 3 healthy subjects).

Figure 3 shows an example of the result.

Figure 3: Actigram during sleep. Upper graph shows the

result measured by Actigraph and lower graph shows that

measured by our sensor.

Besides, the correlation coefficient between the PPIs

computed by the pulse wave measured by the sensor

and the R-R intervals computed by a simultaneously

measured electrocardiogram during sleep was

evaluated. Single-channel ECG was measured by

CM5 lead using PSG (Polymate AP1124, TEAC

Corporation, sampling rate: 1 kHz) simultaneously

with the PPI measured by our sensor. R-R intervals

were computed using commercially available R-R

interval analysis software for the PSG (NoruPro

Light Systems, Japan). The correlation coefficient is

0.96 (average of 3 healthy subjects). Figure 4(a)

shows the correlation plot, and (b) shows the Bland

& Altman plot between R-R intervals of ECG and

PPIs.

These values are accurate enough to use the sensor

as a medical device.

4.2 PC Software

Figure 5 shows the flow of the algorithm for

computing sleep stages from the data of PPIs and the

amount of activity.

We employ Cole’s algorithm for wake/sleep

identification from the amount of activity data (Cole,

1992). This algorithm cannot determine wake/sleep

in real time, but its accuracy is about 90%. At the

same time, the indices of autonomic nervous activity

are derived from frequency analysis of the

variability of PPIs. Firstly, sampled PPIs’ dataset in

a minute is interpolated at even intervals by cubic

spline interpolation by the minute. Next, Fast

Fourier Transformation (FFT) is executed for the

even-interval PPIs to get the frequency spectrum. In

the frequency domain, the integral value of the

power from 0.04Hz to 0.15Hz is called LF (low

frequency), which shows both sympathetic and

parasympathetic nervous activities. The integral

value of the power from 0.15Hz to 0.4Hz is called

HF (high frequency), which shows parasympathetic

nervous activity. Therefore, we can get the balance

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

328

R - R in te rv a ls o f E C G (s )

Pulse-to-P ulse Intervals (s)

1.501.251.000.750.50

1.50

1.25

1.00

0.75

0.50

(a)

Average of P-P interval and R-R interval of EC G (s)

Difference in P-P interval and R-R interval of EC G (s)

1.51.31.10.90.70.5

0.5

0.4

0.3

0.2

0.1

0.0

-0.1

-0.2

-0.3

(b)

Figure 4: Correlation plot (a) and Bland & Altman plot (b)

between R-R intervals of ECG and PPIs.

of sympathetic and parasympathetic nervous activity,

which is related to sleep stages as we mentioned

above. In order to classify the sleep stages from the

dataset of LF and HF values, the k-means clustering

method is adopted. Firstly, REM/NREM sleep is

divided from 2-hour data set, and then, shallow/deep

sleep is divided from its NREM dataset.

We developed sleep analysis software on

Windows XP/Vista using this algorithm. The

program was coded and compiled by Visual Basic

Ver.6 (Microsoft Corp.).

Figure 5: Algorithm for computing sleep stages.

Figure 6 and 7 are the picture image of the

software. Figure 6 is one night’s data, which shows

pulse rate, variability of pulse rate, LF and HF

trends, amount of activity and the simplified

hypnogram.

Figure 6: Result of one night’s data.

Figure 7 is a summary of a 2-week hypnogram

showing sleep habits. This is the most useful

function for sleep care, which cannot be

implemented in the conventional sleep monitoring

systems.

DEVELOPMENT OF A SLEEP MONITORING SYSTEM WITH WEARABLE VITAL SENSOR FOR HOME USE

329

Figure 7: Result of summary of a 2-week hypnogram.

5 VALIDATION OF THE SLEEP

ESTIMATION

Correlation between the sleep stage estimated by the

proposed method using our wearable physiological

sensor and the sleep stage estimated using PSG by

sleep specialists was evaluated. EEG, EOG, chin

EMG, ECG, respiration and SpO2 by PSG

(Polymate AP1000, TEAC Corporation, Sampling

rate: 250Hz), the pulse wave and acceleration by our

sensor was recorded simultaneously in a night (8

hours). The test was held in two cites (Showa

University East Hospital, Tokyo, Japan and The

Institute for Science of Labour, Kawasaki, Japan).

45 normal healthy subjects (30 males and 15

females, 19-72 years old) are measured. All subjects

had informed consent.

The sleep stages of PSG were distinguished

manually by sleep specialists (doctor, clinical

laboratory technologist, or sleep researcher) who

belong to those cite based on Rechtschaffen & Kales

method (Rechtschaffen, 1968) by the minute. Our

sensor also estimated the sleep stages also by the

minute.

We defined coincidence ratio as an evaluation

function to compare the estimation result by our

sensor with the result by PSG.

The coincidence ratio is defined as a correlation

coefficient of moving average of sleep stages (20-

minutes time window) between the stages estimated

by this method and those from PSG. Table 1 shows

the result of the comparison.

Table 1: Result of the comparison.

Figure 8 shows an example of the estimation results.

Figure 8: An example of the estimation results (upper:

sleep stage distinguished by medical professionals using

PSG, lower: sleep stage distinguished automatically by the

sensor).

6 DISCUSSIONS

“Beat-to-beat” pulse interval detection is necessary

to obtain autonomic nervous activity. Our algorithm

has enough ability to get “beat-to-beat” pulse

intervals. However, this algorithm is applicable for

healthy subjects, except for cardiac disease,

peripheral blood circulation disorder.

The coincidence ratio of our sleep stage

estimating algorithm is 0.735. Although it is a rather

low value, PSG results also varied depending on the

examiner (variance is about 20%), and therefore it

seems to be acceptable for home healthcare use.

However, it is also applicable for healthy subjects,

except for autonomic nerve disorder, cardiac disease.

7 CONCLUSIONS

Measurement of sleep habits is a promising new

medical field. However, there are no systems

suitable for it. This is because current sleep

monitoring systems cannot satisfy the needs for

accurate analysis of sleep and convenience in use. In

order to provide a solution, we developed a small

and easy-to-use sensor device. We also developed an

algorithm for analyzing sleep stages. We confirmed

sufficient accuracy in the detection of PPIs by

comparison with R-R Intervals by ECG, and that in

the estimation of the sleep stage by comparison with

the result of PSG. The software can display the sleep

data of one night and the summary of a 2-week

hypnogram. This function is useful not only for a

doctor analyzing a patient's sleep habits, but also for

a user analyzing his/her sleep.

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

330

REFERENCES

Rechtschaffen A., and Kales A., A Manual of standardized

terminology, techniques and scoring system for sleep

stages of human subjects. Public Health Service, U.S.

Goverment Priting Office, 1968.

Sadeh A., Alster J., Urbach D. and Lavie P.,

Actigraphically based automatic bedtime sleep-wake

monitor scoring: validity and clinical applications. J

Ambul Monit 2, pp. 209–216, 1989.

Watanabe T. and Watanabe K., Noncontact Method for

Sleep Stage Estimation, IEEE Transactions on

Biomedical Engineering, Vol.51, No.10, pp.1735-

1748, 2004.

Michimori A., Fukushima K. and Hagiwara H., Sleep

Monitoring System by Using Heart Rate Variability

Analysis, Matsusita Technical Report, No.82, 29-33,

2003 (in Japanese).

Wakuda Y., Noda A., Hasegawa Y., Arai F., Fukuda T.,

and Kawaguchi M., Biological Rhythm Based

Wearable Sleep State Observer, Journal of Advanced

Computational Intelligence and Intelligent

Informatics, Vol.11, No.2, pp.232-241, 2007

Baharav A., Kotagal S., Gibbons V., Rubin B.K., Pratt G.,

J. Karin and S. Askelrod, Fluctuations in. autonomic

nervous activity during sleep displayed by. power

spectrum analysis of heart rate variability, Neurology

45:66, 1183-1187, 1995.

Suzuki T, Ouchi K, Moriya A, Kameyama K and

Takahashi M. Development of a Sleep-Stage

Estimation Method using Heart RateVariability and

Actigraphy measured by Wearable Sensor, Sleep and

Biological Rhythms; 5 Suppl.1:A38, 2007.

Cole RJ, et al., Automatic sleep/wake identification from

wrist actigraphy, Sleep, 15(5), 461-469, 1992.

Heart rate variability: standards of measurement,

Physiological interpretation and clinical use. Task

Force of the European Society of Cardiology and the

North American Society of Pacing and

Electrophysiology. Circulation. 93, pp.1043–1065,

1996.

DEVELOPMENT OF A SLEEP MONITORING SYSTEM WITH WEARABLE VITAL SENSOR FOR HOME USE

331