Accelerometer-based Sleep/Wake Detection in an Ambulatory

Environment

Jan Cornelis

, Elena Smets

1,2

and Chris Van Hoof

1,2,3

Imec, Leuven, Belgium

Electrical Engineering-ESAT, KU Leuven, Belgium

Imec, Holst Centre, The Netherlands

Keywords: Sleep Detection, Wearable Devices, Actigraphy, Large Scale.

Abstract: It has been shown that poor sleep quality and stress are major causes for mental and physical health problems

in developed countries. Thanks to advancements in wearable technology, remote patient monitoring has

become possible, without the need of cumbersome and expensive equipment. A method for sleep/wake

detection is proposed, using chest-worn accelerometer sensors. A total of 1727 nights from 580 individuals

were analysed, resulting on the identification of an average sleep time of 463 min (SD=±80 min) per day. Our

algorithm was able to automatically detect 483 min (SD=±97 min) of sleep. Results show that actigraphy with

an accelerometer located at the chest has potential for sleep monitoring, though further research is required

for further validation, preferably using polysomnography as a benchmark.

1 INTRODUCTION

Stress is regarded as one of the elementary factors for

primary insomnia (Morin, et al., 2003). It has been

shown that insomnia can have a significant negative

impact on the life quality of an individual, including

reduced work and cognitive performance (Léger, et

al., 2002), (Durmer and Dinges, 2005) and an

increased risk of developing obesity (Phillips, 2006),

cardiovascular diseases (Li, et al., 2014) and

depression (Morawetz, 2003). To date, the golden

standard for investigating human sleep patterns is

polysomnography. However, this procedure can be

experienced as cumbersome, is expensive, and

usually deprives the subjects from their familiar

environment, which can lead to changes in their

sleeping patterns. (Le Bon, et al., 2001). Over the past

30 years, the use of wearable technology has

significantly improved, allowing ambulatory sleep

investigation. Therefore, researchers are able to

conduct experiments on a larger scale, outside a

controlled laboratory environment, possibly resulting

in more viable data as the first night effect could be

reduced (Le Bon, et al., 2001). Actigraphy is

considered to be a reliable method for sleep/wake

detection (Littner, et al., 2003). Most actigraphy units

are constructed in a watch like band shape that is

either worn at the wrist or at the ankle. Sleep and

wake patterns are estimated from periods of activity

and inactivity based on registered movement in the

device. (Littner, et al., 2003) Typically, actigraphy

shows an accuracy for detecting sleep epochs

between 87 and 90 percent compared to a

polysomnography. (Meltzer, et al., 2012). This paper

investigates the possibility for sleep classification

using a chest-worn health device rather than a wrist

or ankle-worn device, based on accelerometer data

(ACC). The advantage of using such device, is that it

is also capable of registering electrocardiogram

(ECG) signals besides registering ACC data, which at

a later stage, could provide more insight into

physiology correlated issues with insomnia and sleep

stages. A new method for sleep detection is required,

since the chest oscillation during breathing is

registered in the ACC data, causing false wake

labelled positives, an issue that does not occur in

traditional wrist worn devices. The aim of the study

is to achieve a sleep detection accuracy equal to

traditional actigraphy for the chest worn device. Data

from 1002 volunteers over a 5 day consecutive period

was used. The outcome of the algorithm was

compared to diary input from the volunteers.

Cornelis, J., Smets, E. and Van Hoof, C.

Accelerometer-based Sleep/Wake Detection in an Ambulatory Environment.

DOI: 10.5220/0007398603750379

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 375-379

ISBN: 978-989-758-353-7

375

2 METHODS

2.1 Subject Recruitment

Volunteers (n=1,002) were recruited from the active

working population from 11 technology, banking,

and public sector companies located in Belgium and

the Netherlands. People were encouraged to

participate through means of internal company

communication and the distribution of flyers.

Participants had a chance to win a dinner or travel

voucher (11 vouchers for every 200 participants). The

collected sample contained 481 males (48%) and 446

females (45%).75 participants (7.5%) did not report

their gender. The participants were between the age

of 21 and 65 (x

=39.5 ± 9.8). An informed consent was

obtained from the participants prior to their

participation in the experiment.

2.2 Data Collection Protocol

The data was collected over a period of two years,

from 2015 till 2017. Prior to the start of the

experiment, a survey had to be filled out containing

personal information such as gender, age, health

information, work related conditions and lifestyle.

The experiment lasted over a period of five days,

starting on Thursday and ending on Monday. During

the experiment, participants were requested to fill in

a diary using Ecological Momentary Assessments

(EMAs) on a smartphone application. EMAs allow

researchers to do frequent sampling of the behaviors

of the participant in real-time. (Shiffman, et al., 2008)

The application asked the participants 12 times per

day, at random times, to rate their perceived stress

level from the past hour on a 5-point Likert scale.

Additionally, each morning, the participants were

asked to fill in a sleep diary in which they had to

annotate the time they went to bed, how long it took

to fall asleep, the number of times they woke up and

the time at which they woke up in the morning. The

participants

were able to fill in their sleep times

Figure 1: A health device used in the experiment.

freely, by using the smartphone keyboard. The time it

took to fall asleep was a multiple-choice: 0-10

minutes, 11-30 minutes, 31-60 minutes or >60

minutes. If they reported it took more than 60 minutes

to fall asleep or if they woke up at least once during

the night, additionally the reason for not being able to

fall asleep or waking up was asked.

2.3 Sensor Information

Each participant was asked to wear a health device at

the chest, for the duration of the experiment, i.e. five

days continuously (fig1). This is a regulatory

approved device, for recording ECG (256Hz) and

triaxial accelerometer (ACC) (32Hz) signals. The

data was stored on an SD card, and read out after the

experiment was concluded. Before the start of the

experiment, the internal clock of the device was

synchronised to UTC. Participants were asked to

remove the sensor in case they participated in a

vigorous physical activity, in order to prevent

potential damage from sweating.

2.4 Sleep Wake Classification

The most commonly referred sleep/wake detection

methods for wrist actigraphy are those of Cole et al

(Cole, et al., 1992) and Sadeh et al (Sadeh, et al.,

1994) The findings of Cole et al are based on previous

findings of Webster et al, who used equation eq. 1.

(Webster, et al., 1982)

A = 0.025(0.15 X

t-4

+ 0.15 X

t-3

+ 0.15 X

t-2

+ 0.08

t-1

+ 0.21 X

+ 0.12X

t+1

+ 0.13 X

t+2

(1)

In this equation, X(t) represents the sum of the digital

activity values of the Medilog1 recorder for all 30 2-

s data epochs in 1 min at time t. (Webster, et al.,

1982).

Activity indicator A is considered sleep if

A<1. (Webster, et al., 1982). However, above stated

activity recognition methods are all based on wrist-

based activity. When the activity is measured from

the chest, there is a natural oscillation due to the

breathing pattern. Therefore, there was a need for a

modified sleep/wake detection, with a lower

sensitivity. The analysis was performed in MATLAB

and the classification is determined by the ACC

recordings of the health device. Each 60 seconds the

ACC signal was scored for activity (A). For each axis,

the difference between the minimal and maximal g

value was calculated, and the activity was determined

by eq. 2.

HEALTHINF 2019 - 12th International Conference on Health Informatics

376

A= 0.025(0.2X

t-3

+0.2X

t-2

+0.2X

t-1

+ 0.2X

0.2X

t+1

)

(2)

Activity indicator A is considered sleep if A<1. X is

the average maximal difference in g for each axis. t

represents the time epoch in minutes of the signal,

with t0 as the current minute. The sleep/wake state

was evaluated based on activity indicator A (eq1), and

stored as a Boolean true/false. From the moment the

first 30 minutes of the Boolean stored sleep/wake

indicator where label as sleep, the participant was

considered asleep until the data indicated that the

participant was up for at least 30 minutes, with a

minimum of 120 minutes of registered sleep, in order

to exclude potential daytime naps from the dataset. If

participants did not fill in the sleep diary correctly in

the morning, the data of the previous night was

removed. In total 1727 nights were included for

analysis.

2.5 Validation

The outcome of the sleep-wake classification

algorithm is compared with the diary entries from the

EMAs. The maximum falling asleep times were

added to the reported time to bed, e.g. if the

participant indicated it took 0-10 minutes to fall

asleep, 10 minutes were added to the time to bed to

find the time the participant actually fell asleep. We

investigated for which percentage of the nights the

reported sleep and wake times matched with the

detected sleep and wake times based on the

accelerometer data. Since self-reporting is not always

accurate, a tolerance of 0, 10, 30 and 60 minutes was

introduced, allowing the classification to differ 0, 10,

30 or 60 min respectively from the self-reported wake

and sleep times.

Figure 2: Flowchart for the sleep identification.

3 RESULTS

The tolerance scores of the algorithm are shown in

table 1. A total of 1727 nights of 580 unique

participants were analysed, and 81% of the nights fell

within a 60 minute range in wake up and sleep time

Table 1: Population within tolerance.

Accelerometer-based Sleep/Wake Detection in an Ambulatory Environment

377

Figure 3: Scatterplot comparing the total sleep time (TST) of the diary and the algorithm for the entire population (A)-

(n=1723) and the population within the 60 minute tolerance (B) (n=1427). Regression equation for the entire population is

y=0.82x+102 and for the population within the 60 min tolerance y=0.93x+45. The root mean square values (RMSEs) are

75.72 and 38.08 for A and B respectively.

with regards to the diary entry. The average total

sleep time (TST) of the population was 462 min for

the data reported in the diary, with a SD of ±80 min,

and 483 min for the data predicted with the algorithm,

with an SD of ±97 min. The population that fell

within the one hour tolerance had a TST of 460 min

(SD=±76 min) for the diary and 472 min (SD=±78

min) for the algorithm. Reported average falling

asleep times for the populations was 22 minutes

(SD=±15 min). The RMSEs are 75.72 min and 38.08

min for A and B respectively. The correlation

coefficient for the total estimated sleep time based on

diary and algorithm was 0.69. For the population that

fell within the diary boundary of 60 minutes, the

absolute mean difference is 29 minutes (SD=±26

min), and the correlation coefficient between the

sleep times is 0.90. A scatterplot comparing the

algorithm and the diary TST is presented in figure 3.

4 DISCUSSION

On average, the algorithm overestimated the sleep

period by 20 min. The overestimation of the TST is

in line with other research. This is a known issue with

accelerometer data, as it is difficult to distinguish the

sleep onset (SO), wake after sleep onset (WASO) and

stage 1 sleep, as activity is generally limited when one

is falling asleep (Lockley, et al., 1999). Nevertheless,

the agreement rate is within acceptable range,

especially considering that the maximum range for

falling asleep was subtracted from the TST, which is

likely to be an overestimation of the reported SO. The

current study has limitations regarding validation of

the results, i.e. the lack of a comparison to a golden

standard (polysomnography). Further research should

investigate how the polysomnography, actimetry and

self-reported sleep times are associated. Since the

device used in this study also recorded the ECG, this

could be used to further enhance the sleep/wake

detection of the algorithm. Research has shown that

the inclusion of ECG-based analysis can further

improve the sleep wake detection, and could enable

the differentiation between light sleep (stage 1 and 2),

slow wave sleep (stage 3 and 4) and rapid eye

movement sleep (Tal, et al., 2017). The data could

potentially also be used for the detection of health

hazards, opening the path for further usage of

wearable sensors in ambulatory healthcare

monitoring (Mezick, et al., 2013).

5 CONCLUSION

We have collected ambulatory physiological data of

1,002 subjects during 5 consecutive days and 4

nights, in combination with background information,

and smartphone-based self-reports. 580 subjects from

this dataset were eligible for this analysis. This paper

provides a method to distinguish sleep and non-sleep

HEALTHINF 2019 - 12th International Conference on Health Informatics

378

periods on basis of accelerometer data, which can be

used independent from the diary input. The usage of

a chest located health device rather than a

conventional wristband has the advantage that

additional signals such as ECG can be recorded,

without the need for additional sensors, which

decreases the subject’s discomfort during

measurements. Our paper presents an important first

step for further research in linking continues

monitored physiological night-time data with

psychological self-reports. This could be used to

create a model for individual based feedback,

granting personalised health information to the user

of the device. The ECG data from this dataset could

be used to further enhance the detection of potential

health hazards, contributing for increased usage of

wearable sensors for healthcare monitoring purposes

in the future.

ACKNOWLEDGEMENTS

Special thanks goes out to Elena Smets, who

contributed significantly towards this project.

REFERENCES

Cole, R., Kripke, D., Gruen, W., Mullaney, D. and Gillin,

J., 1992. Automatic Sleep/Wake Identification From

Wrist Activity. SLEEP, 15(5), pp.461-469.

Durmer, J. and Dinges, D., 2005. Neurocognitive

Consequences of Sleep Deprivation. Seminars in

Neurology, 25(01), pp.117-129.

Le Bon, O., Staner, L., Hoffmann, G., Dramaix, M., San

Sebastian, I., Murphy, J., Kentos, M., Pelc, I. and

Linkowski, P., 2001. The first-night effect may last

more than one night. Journal of Psychiatric Research,

35(3), pp.165-172.

Léger, D., Guilleminault, C., Bader, G., Lévy, E. and

Paillard, M., 2002. Medical and Socio-Professional

Impact of Insomnia. SLEEP, 25(6), pp.621-625.

Li, M., Zhang, X., Hou, W. and Tang, Z., 2014. Insomnia

and risk of cardiovascular disease: A meta-analysis of

cohort studies. International Journal of Cardiology,

176(3), pp.1044-1047.

Littner, M., Kushida, C., Anderson, W., Bailey, D., Berry,

R., Davila, D., Hirshkowitz, M., Kapen, S., Kramer, M.,

Loube, D., Wise, M. and Johnson, S., 2003. Practice

Parameters for the Role of Actigraphy in the Study of

Sleep and Circadian Rhythms: An Update for 2002.

SLEEP, 26(3), pp.337-341.

Lockley, S., Skene, D. and Arendt, J., 1999. Comparison

between subjective and actigraphic measurement of

sleep and sleep rhythms. Journal of Sleep Research,

8(3), pp.175-183.

Meltzer, L., Walsh, C., Traylor, J. and Westin, A., 2012.

Direct comparison of Two New Actigraphs and

Polysomnography in Children and Adolescents..

SLEEP, 35(1), pp. 59-166.

Mezick, E., Matthews, K., Hall, M., Richard Jennings, J.

and Kamarck, T., 2013. Sleep duration and

cardiovascular responses to stress in undergraduate

men. Psychophysiology, 51(1), pp.88-96.

Morawetz, D., 2003. Insomnia and Depression: Which

Comes First ?. Sleep Research Online, 5(2), pp. 77-81.

Morin, C., Rodrigue, S. and Ivers, H., 2003. Role of Stress,

Arousal, and Coping Skills in Primary Insomnia.

Psychosomatic Medicine, 65(2), pp.259-267.

Phillips, B., 2006. The Association Between Short Sleep

Duration and Obesity in Young Adults: A 13-Year

Prospective Study. Yearbook of Pulmonary Disease,

2006, pp.256-258.

Sadeh, A., Sharkey, M. and Carskadon, M., 1994. Activity-

Based Sleep-Wake Identification: An Empirical Test of

Methodological Issues. SLEEP, 17(3), pp.201-207.

Shiffman, S., Stone, A. and Hufford, M., 2008. Ecological

Momentary Assessment. Annual Review of Clinical

Psychology, 4(1), pp.1-32.

Tal, A., Shinar, Z., Shaki, D., Codish, S. and Goldbart, A.,

2017. Validation of Contact-Free Sleep Monitoring

Device with Comparison to Polysomnography. Journal

of Clinical Sleep Medicine, 13(03), pp.517-522.

Webster, J., Kripke, D., Messin, S., Mullaney, D. and

Wyborney, G., 1982. An Activity-Based Sleep Monitor

System for Ambulatory Use.

SLEEP, 5(4), pp.389-399.

Accelerometer-based Sleep/Wake Detection in an Ambulatory Environment

379