39
SLEEPGUARD: Capturing Rich Sleep Information using
Smartwatch Sensing Data
Sleep is an important part of our daily routine – we spend about one-third of our time doing it. By tracking the sleep-related
events and activities, sleep monitoring provides the decision support to help us understand the sleep quality and the causes of
poor sleep. Wearable devices provide a new way for sleep monitoring, allowing us to monitor sleep from the comfort of our own
home, but existing solutions do not take full advantage of the rich sensor data provided by these portable devices. In this paper,
we develop a novel approach to track a wide range of sleep-related events using smartwatches. We show that it is possible to
track, using a single smartwatch, sleep events like body postures and movements, acoustic events, and illumination conditions.
From these events, a statistical model can be designed to effectively evaluate a user’s sleep quality across various sleep stages.
We evaluate our approach by conducting extensive experiments involved fifteen users across a 2-week period. Our experimental
results show that our approach can track a richer set of sleep events, provide better decision support for evaluating sleep quality,
and help to identify causes for sleep problems compared to prior work. We also show that SLEEPGUARD can help users to
improve their sleep quality by helping them to understand causes of sleep problems.
CCS Concepts: • Human-centered computing → Ubiquitous and mobile computing;
Additional Key Words and Phrases: Smartwatch, Sleep events, Sensing
ACM Reference Format:
. 2018. SLEEPGUARD: Capturing Rich Sleep Information using Smartwatch Sensing Data. Proc. ACM Interact. Mob. Wearable
Ubiquitous Technol. 9, 4, Article 39 (February 2018), 30 pages. https://doi.org/0000001.0000001
1 INTRODUCTION
Sleep plays a vital role in good health and personal well-being throughout one’s life. Lack of sleep or poor quality
of sleep can lead to serious, sometimes life-threatening, health problems [
7
,
20
,
44
], decrease level of cognitive
performance [
4
,
6
], and affect mood and feelings of personal well-being [
55
,
56
]. Besides having an adverse
effect on individuals, insufficient or poor quality sleep has a significant economic burden, among others, through
decreased productivity, and medical and social costs associated with treatment of sleep disorders [
47
]. Indeed, to
highlight the significance of sleep quality, the Centre for Disease Prevention (CDC) has declared insufficient sleep
as a public health problem in the US [
24
], and the concern is widely shared amongst other industrialized countries.
Traditionally, sleep monitoring is performed in a clinical environment using Polysomnography (PSG). In
PSG, medical sensors attached to human body are used to monitor events and information such as respiration,
electroencephalogram (EEG), electrocardiogram (ECG), electro-oculogram and oxygen saturation [
29
,
45
,
53
,
61
].
These information sources can then be used to determine sleep stages, sleep efficiency, abnormal breathing, and
overall sleep quality. While PSG is widely considered as the gold standard for sleep monitoring, and while it is
extensively used to support clinical treatments of sleep disorders, it has some disadvantages that make it unsuitable
for longitudinal and large-scale sleep monitoring. Firstly, the sensing instruments are time-consuming and laborious
to put on, and they are prone to disrupting sleeping routines. Secondly, PSG is rather expensive to use and requires
a clinical environment and highly trained medical professionals to operate. Due to these disadvantages, PSG is only
suitable as a way to support severe disorders that clinical care is required.
Author’s address:
ACM acknowledges that this contribution was authored or co-authored by an employee, contractor, or affiliate of the United States government.
As such, the United States government retains a nonexclusive, royalty-free right to publish or reproduce this article, or to allow others to do so,
for government purposes only.
© 2018 Association for Computing Machinery.
2474-9567/2018/2-ART39 $15.00
https://doi.org/0000001.0000001
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Vol. 9, No. 4, Article 39. Publication date:
February 2018.
39:2 •
Recently, sleep monitoring based on off-the-shelf mobile and wearable devices has emerged as an alternative way
to obtain information about one’s sleeping patterns [
62
,
68
]. By taking advantage of diverse sensors, behaviours and
routines associated with sleeping can be captured and modelled. This in turn can help users understand their sleep
behaviour and provide feedback on how to improve their sleep, for example, by changing routines surrounding
sleep activity or improving the sleeping environment. What makes self monitoring particularly attractive is the
non-invasive nature of the sensing compared to PSG. Examples of consumer-grade sleep monitors range from apps
running on smartphones or tablets to smartwatches and specialized wearable devices [10, 32, 34, 58, 67, 69].
Despite the popularity of consumer-grade sleep monitors, currently the full potential of these devices is not
being realized. Indeed, while current consumer-grade sleep monitors can capture and model a wide range of sleep
related information, such as estimating overall sleep quality, capturing different stages of sleep, and identifying
specific events occurring during sleep [
42
,
67
,
74
], they offer little help in understanding the characteristics
that surround poor sleep. Thus, these solutions are unable to capture the root cause behind poor sleep or to
provide recommendations on how to improve sleep quality. This is because current solutions focus on monitoring
characteristics of the sleep itself, without considering behaviours occurring during sleep and the environmental
context affecting sleep, e.g., ambient light-level and noise. Indeed, sleep quality has been shown to depend on a
wide range of factors. For example, intensity of ambient light [
39
], temperature [
70
], and noisiness [
50
] of the
environment can significantly affect sleep quality. Similarly, the user’s breathing patterns, posture during sleep, and
routines surrounding the bedtime also have a significant impact on sleep quality. Without details of the environment
and activities across sleep stages, the root cause of poor sleep cannot be captured and the user informed of how to
improve their sleep quality. To unlock the full potential of consumer-grade sleep monitoring, innovative ways to
take advantage of the rich sensor data accessible through these devices are required.
The present paper contributes by presenting the design and development of SLEEPGUARD, a holistic sleep
monitoring solution that captures rich information about sleep events, the sleep environment, and the overall quality
of sleep. SLEEPGUARD is the first to solely rely on sensor information available on off-the-shelf smartwatches
for capturing a wide range of sleep-related activities (see Table 1). The key insight in SLEEPGUARD is that sleep
quality is strongly correlated with characteristics of body movements, health related factors that can be identified
from audio information, and characteristics of the sleep environment [
62
]. By using a smartwatch, the sensors
are close to the user during all stages of the night, enabling detailed capture of not only sleep cycles, but body
movements and environmental changes taking place during the sleep period. capturing these sleeping events
from sensor data, however, is non-trivial due to changes in sensor measurements caused by hand motions during
sleep. To overcome this challenge, changes in sensor orientation relative to user’s body need to be tracked and
opportune moments where to capture sensor data need to be detected. To address these issues, SLEEPGUARD
integrates a set of new methods for analysing and capturing sleep-related information from sensor measurements
available on a smartwatch. SLEEPGUARD also incorporates a model that uses the detected events to infer the
user’s sleep stages and sleep quality. While some prior research has examined the use of smartwatches for sleep
monitoring [
16
,
36
,
57
,
62
], these approaches have only been able to gather coarse-grained information about
sleep and often required additional highly-specialized devices, such as pressure mattresses or image acquisition
equipment to supplement the measurements available from the smartwatch. In this paper, we demonstrate that, for
the first time, using only a smartwatch, it is possible to capture an extensive set of sleep-related information – many
of which are not presented in prior work. Having a more comprehensive set of sleep-related events and activities
available enables users to gain a deeper understanding of their sleep patterns and the causes of poor sleep, and to
make recommendations on how to improve one’s sleep quality.
We evaluate SLEEPGUARD through rigorous and extensive benchmark experiments conducted on data collected
from fifteen participants during a two week monitoring period. The results of our experiments demonstrate that
SLEEPGUARD can accurately characterize body motions and movements during sleep, as well as capture different
acoustic events. Specifically, the lowest accuracy for SLEEPGUARD in our experiments is 87%, with the best event
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Vol. 9, No. 4, Article 39. Publication date:
February 2018.
SLEEPGUARD: Capturing Rich Sleep Information using Smartwatch Sensing Data • 39:3
detection accuracy reaching up to 98%. We also demonstrate that SLEEPGUARD can accurately detect various sleep
stages and help users to better understand their sleep quality. During our experiments, six of the
15
participants
suffered from some sleep problems (
4
with bad and
2
with general sleep quality), all of whom were correctly
identified by SLEEPGUARD. Moreover, we also demonstrate that SLEEPGUARD is able to correctly identify the
root cause of sleep problems for the
4
participants with bad sleep quality, whether it is due to suboptimal hand
position, body posture or sleeping environment. Compared to state-of-the-art sleep monitoring systems, such as
Fitbit and Sleep Hunter, the main advantage of SLEEPGUARD is that can report a wider range of sleep events and
provide a better understanding for the causes of sleep problems.
Contributions
This paper makes the following contributions:
•
We present the design and development of SLEEPGUARD (Section 2), the first holistic sleep monitoring
system to rely solely on sensors available in an off-the-shelf smartwatch to capture a wide range of sleep
information that characterizes overall sleep quality, user behaviours during sleep, and the sleep environment.
•
We develop novel and lightweight algorithms for capturing sleep-related information on smartwatches taking
into consideration changes in orientation and location of the device during different parts of the night. We
show how to overcome specific challenges to effectively track events like sleep postures (Section 2.1.1), hand
positions (Section 2.1.2), body rollovers(Section 2.2.1), micro body movement(Section 2.2.2), and acoustical
(Section 2.3) and lighting conditions (Section 2.4).
•
We extensively evaluate the performance of SLEEPGUARD using measurements collected from two-week
monitoring of
15
participants (Section 3). Our results demonstrate that SLEEPGUARD can accurately capture
a wide range of sleep events, estimate different sleep stages, and produce meaningful information about
overall sleep quality (Section 4). We show that SLEEPGUARD successfully reveals the causes of poor sleeps
for some of our testing users and subsequently helps them improve their sleep by changing their sleep
behaviours and sleeping environment (Section 4.2.4).
2 SLEEPGUARD SLEEP MONITORING PLATFORM
SLEEPGUARD is a novel smartwatch-based sleep monitoring system that aims at estimating sleep quality and
capturing rich information about behaviours and events occurring during sleep. By capturing this information,
SLEEPGUARD can analyze potential reasons for sleep problems and provide the user with suggestions on how
to improve their sleep routine or sleep environment. To achieve its design goal, SLEEPGUARD exploits a wide
range of sensors that are common on commercial off-the-shelf smarwatches: (i) accelerometer, gyroscope, and
orientation sensor are used to collect body and hand movements; (ii) microphone is used to measure level of
ambient noise and to capture acoustic events; and (iii) ambient light sensor is used to monitor illumination within
the sleep environment. The different sensors and information extracted from them are summarized in Table 1. In
the following we discuss the different subcomponents of SLEEPGUARD in detail.
2.1 Detecting Sleep Postures and Movements
One’s sleeping position, also referred to as sleep posture, and the extent of body movements are important factors
in determining overall sleep quality. Suboptimal posture has been shown to affect the severity of sleep disorders
and is widely used in medical diagnoses to analyse effects of sleep disorders [
30
,
52
] while having a good sleep
posture has been shown to correlate with subjective assessments of sleep quality [
25
]. Similarly, high degree of
body movements during sleep likely reflects restlessness, which results in poor sleep quality. SLEEPGUARD uses
motion sensors (accelerometer, gyroscope, and orientation sensor) to capture user’s sleep posture and habits. In the
following we detail the techniques we use for capturing the body posture and movements. SLEEPGUARD, currently
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Vol. 9, No. 4, Article 39. Publication date:
February 2018.
39:4 •
Table 1. Sleep events targeted in this work
Event Type
Sleep postures Supine, Left lateral, Left lateral, Prone
Hand positions Head, Chest, Abdomen
Body rollover Count
Micro body movements Hand moving, Arm raising, Body trembling
Acoustic events Snore, Cough, Somniloquy
Illumination condition Strong, Weak
Smartwatch
Left Lateral ProneSupine Right Lateral
Fig. 1. Four sleep body postures.
Smartwatch
ChestAbdomen Head
Fig. 2. Three hand positions.
Right
Lateral
Left
Lateral
Fig. 3. Body rollover from the left side
to the right side.
supports the 4 basic sleep postures (see Fig. 1); 3 hand positions (see Fig. 2); 6 types of body rollovers (see Fig. 3
for an example); and 3 types of body micro movements.
2.1.1 Sleep Posture Detection . Dreaming and sleep quality are associated with underlying brain functions,
which in turn are affected by body posture [
3
]. Sleep posture also varies across individuals and should fit personal
and physical needs of the individual [
43
,
63
]. For example, sleeping in a prone position is unsuitable for people with
ailments, such as heart disease or high blood pressure. On the other hand, people can consciously avoid postures
that would be beneficial for health and sleep quality [
27
]. Having an effective way to detect the current posture
and track changes in it would thus be essential for estimating overall sleep quality, and avoiding potential harm.
SLEEPGUARD captures the four basic sleep postures, which are supine, left lateral, right lateral, and prone. This is
illustrated in Fig. 1. Detecting these postures using a single wrist sensor, however, is non-trivial because the sensor
cannot accurately track movement of the entire body. To accomplish posture detection, we observe that the arm
position strongly correlates to the sleep postures, i.e., the arm is typically located in a specific, stable location for a
given sleep posture. This observation suggests that we can first identify the user’s arm position and the duration
of a specific arm location, and then map the information to a sleep posture. Later in this paper, we show that our
approach can achieve a high accuracy in identifying sleep postures.
To separate sleep postures, SLEEPGUARD considers a set of feasible hand positions for each posture. In the
supine position, we assume the user’s hand to be on the left side of the body, on the abdomen, on the chest or on
the head; in the left and right lateral positions, we assume the user’s hand is close to the pillow, on the chest or on
the waist. Finally, in the prone position we assume the user’s hand is on the side of the head or above his/her head.
These positions were selected based on a pilot carried out in our test environment (see Sec. 3). Fig. 1 shows one
possible hand position for each of the postures.
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Vol. 9, No. 4, Article 39. Publication date:
February 2018.
SLEEPGUARD: Capturing Rich Sleep Information using Smartwatch Sensing Data • 39:5
0 1 2 3 4 5 6 7
0
20
40
60
80
100
120
140
160
180
Time (s)
Angle (deg)
X−axis Y−axis Z−axis
(a) Supine
0 1 2 3 4 5 6 7
0
20
40
60
80
100
120
140
160
180
Time (s)
Angle (deg)
X−axis Y−axis Z−axis
(b) Left Lateral
0 1 2 3 4 5 6 7
0
20
40
60
80
100
120
140
160
180
Time (s)
Angle (deg)
X−axis Y−axis Z−axis
(c) Right Lateral
0 1 2 3 4 5 6 7
0
20
40
60
80
100
120
140
160
180
Time (s)
Angle (deg)
X−axis Y−axis Z−axis
(d) Prone
Fig. 4. The tilt angle characteristics of four body postures.
Similarly to SleepMonitor [
67
], we use tilt angles readings from three dimensions to detect postures. To identify
which of the postures the data collected within a time window corresponds to, we average all the calculated tilt
angle values of that window in each dimension. We then calculate the Euclidean distance of the input values to
a set of posture profiles, which are based on measurements collected in a pilot study that involved 10 users (see
Sec. 3.2). We then use the body posture associated with the nearest neighbor as the detection outcome.
Fig. 4 shows the angle values of the four sleep postures targeted in this work. This diagram suggests that the tilt
angles of three axes have obvious differences. The sleep posture thus can be inferred based on the position of the
smartwatch and the created angle mapping. However, a limitation of this approach is that the hand positions during
supine and prone postures are similar when the hand is located on the side of the head (Fig. 4(a) and 4(d)), thus the
classification accuracy will be affected.
To improve detection accuracy between supine posture and prone postures, SLEEPGUARD integrates orientation
data as auxiliary feature. This is based on the observation that hand directions in the supine and prone positions are
different. When the result of the previous step is prone or supine, and the hand is detected to be located next to
the body, we combine the tilt angle with three axes data obtained from the direction sensor as a new feature, and
classify these postures using a template-based distance matching approach. Specifically, we return the position
corresponding to the template with minimum Euclidean distance with current sensor measurements as the user’s
posture.
2.1.2 Hand Position Recognition. The hand position during sleep can disclose potential health problems,
and an improper hand position can even result in health issues [
51
]. For instance, placing the hand on the abdomen
may indicate discomfort whereas placing the hand on the chest can increase the likelihood of nightmares due to
long-term pressure on the heart. Similarly, placing the hand on the head can put excess pressure on shoulder nerves
and cause arm pain as blood flow is restricted. This can lead to eventual nerve damage, with symptoms including a
tingling sensation and numbness [51].
SLEEPGUARD is designed to recognize three common hand positions – if the hand is placed on the abdomen,
chest or head when the user is in the supine posture, as shown in Fig. 2. We have chosen these three hand positions
because there are found to be the most common and representative positions in our pilot study (Section 3.3). Our
hand position recognition algorithm is based on sensor data of rotation angles, tilt angles, and respiratory events. It
works by first using the rotation and tile angles to detect if the hand was placed on the head. If the hand was detected
to be not put on the head, it then uses the respiratory events to detect if the hand was placed on the abdomen or the
chest, but not elsewhere before it utilizes the rotation angles to distinguish the abdomen position from the chest
position. We now describe how to detect each of the three positions in more details.
Detect the head position. Fig. 5 show the change of the rotation angle gathered from the x, y, and z directions
using the gyroscope for one of our pilot study users when his hand was initially placed next to the body and
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Vol. 9, No. 4, Article 39. Publication date:
February 2018.
