SleepMore Inferring Sleep Duration at Scale via Multi-Device WiFi Sensing

2025-05-03 0 0 5.74MB 32 页 10玖币

侵权投诉

193

SleepMore: Inferring Sleep Duration at Scale via Multi-Device WiFi

Sensing

CAMELLIA ZAKARIA,University of Massachusetts Amherst, USA

GIZEM YILMAZ,National University of Singapore, Singapore

PRIYANKA MARY MAMMEN,University of Massachusetts Amherst, USA

MICHAEL CHEE,National University of Singapore, Singapore

PRASHANT SHENOY,University of Massachusetts Amherst, USA

RAJESH BALAN,Singapore Management University, Singapore

The availability of commercial wearable trackers equipped with features to monitor sleep duration and quality has enabled

more useful sleep health monitoring applications and analyses. However, much research has reported the challenge of

long-term user retention in sleep monitoring through these modalities. Since modern Internet users own multiple mobile

devices, our work explores the possibility of employing ubiquitous mobile devices and passive WiFi sensing techniques to

predict sleep duration as the fundamental measure for complementing long-term sleep monitoring initiatives. In this paper,

we propose SleepMore, an accurate and easy-to-deploy sleep-tracking approach based on machine learning over the user’s

WiFi network activity. It rst employs a semi-personalized random forest model with an innitesimal jackknife variance

estimation method to classify a user’s network activity behavior into sleep and awake states per minute granularity. Through

a moving average technique, the system uses these state sequences to estimate the user’s nocturnal sleep period and its

uncertainty rate. Uncertainty quantication enables SleepMore to overcome the impact of noisy WiFi data that can yield

large prediction errors. We validate SleepMore using data from a month-long user study involving 46 college students and

draw comparisons with the Oura Ring wearable. Beyond the college campus, we evaluate SleepMore on non-student users

of dierent housing proles. Our results demonstrate that SleepMore produces statistically indistinguishable sleep statistics

from the Oura ring baseline for predictions made within a 5% uncertainty rate. These errors range between 15-28 minutes for

determining sleep time and 7-29 minutes for determining wake time, proving statistically signicant improvements over prior

work. Our in-depth analysis explains the sources of errors.

CCS Concepts:

•Computing methodologies →Machine learning

;

•Human-centered computing →Ubiquitous and

mobile computing;•Applied computing →Consumer health.

Additional Key Words and Phrases: mobile health, sleep, WiFi

ACM Reference Format:

Camellia Zakaria, Gizem Yilmaz, Priyanka Mary Mammen, Michael Chee, Prashant Shenoy, and Rajesh Balan. 2022. SleepMore:

Inferring Sleep Duration at Scale via Multi-Device WiFi Sensing. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 6, 4,

Article 193 (December 2022), 32 pages. https://doi.org/10.1145/3569489

Authors’ addresses: Camellia Zakaria, University of Massachusetts Amherst, USA, nurcamellia@cs.umass.edu; Gizem Yilmaz, National

University of Singapore, Singapore, gizem.yilmaz@nus.edu.sg; Priyanka Mary Mammen, University of Massachusetts Amherst, USA,

pmammen@cs.umass.edu; Michael Chee, National University of Singapore, Singapore, michael.chee@nus.edu.sg; Prashant Shenoy, University

of Massachusetts Amherst, USA, shenoy@cs.umass.edu; Rajesh Balan, Singapore Management University, Singapore, rajesh@smu.edu.sg.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that

copies are not made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page.

Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy

otherwise, or republish, to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. Request permissions from

permissions@acm.org.

2474-9567/2022/12-ART193 $15.00

https://doi.org/10.1145/3569489

Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 6, No. 4, Article 193. Publication date: December 2022.

arXiv:2210.14152v3 [eess.SP] 16 Nov 2022

193:2 •Zakaria et al.

1 INTRODUCTION

Sleep is essential to one’s physical, emotional, and mental health. The National Sleep Foundation (NSF) recom-

mends between 7 to 9 hours as appropriate for adults to maintain general wellness [

]. However, sleep deciency

is a common public health problem, with approximately 50 to 70 million Americans suering from chronic sleep

disorders [

]. The risk factors associated with insucient sleep are performance and cognitive decits, and its

long-term eects are correlated with severe consequences such as obesity, stress, depression, and stroke [

Thus, accurately determining sleep habits remains a topic of immense interest.

The gold standard for measuring sleep is using polysomnography (PSG), a multichannel, multimodal approach

performed in controlled environments like a sleep clinic by trained technicians. These factors make PSG challeng-

ing to use for day-to-day sleep monitoring. More recently, low-cost sleep wearable devices (e.g., FitBit, Jawbone)

[

] have become popular for sleep monitoring. However, wearables demand behavioral change in users

putting on the device to bed, which can cause reluctance for some [

]. To overcome the challenges of continuous

sleep monitoring on a longitudinal basis, researchers have explored contactless methods using radio, and radar

signals [

]. However, the instrumentation of dedicated solutions in the building infrastructure limits their

support for scalability. Recently, using WiFi signals such as channel state (CSI) and backscatter information has

been proposed as viable solutions for sleep tracking [

]. This is made possible by both smartphones becoming

common across most income levels [

] and WiFi solutions becoming prevalent in public, institutional and

residential locations [

]. These solutions detect when the user has fallen asleep by checking for vibrations

and other signals. Unfortunately, they require custom hardware and are thus hard to deploy at scale, despite

yielding reasonable accuracy.

Smartphones have been used for sleep detection and monitoring. Prior approaches use smartphone activity as

a proxy for user activity and long periods of device inactivity to infer sleep periods. These approaches include

both client-side methods, based on monitoring screen activity [

], as well as network-side methods, based on

monitoring WiFi network activity of the device [

]. Since users tend to habitually use their mobile device before

they sleep and upon waking up, device activity, or lack thereof, is a feasible approach for detecting sleep periods

[

]. While prior work has shown the feasibility of such approaches, they are known to suer from signicant

errors, often more than an hour, when determining sleep time. These challenges motivate the need to develop an

accurate but generally accessible sensing approach that monitor users’ daily sleeping behavior without changing their

routines. Such functionality primarily contributes as a critical resource to the longitudinal requirement in almost

all sleep medicine studies, facilitating complete data collection of fundamental sleep measures in real-time.

In this paper, we present SleepMore, a practical approach to sleep monitoring using passive observations of a

user’s device WiFi activity. Our approach leverages the growing number of mobile devices owned by each user in

recent years (e.g., phone, tablet, laptop, e-readers) [

]. While smartphones remain the primary mobile device

for most users, they may use a combination of devices over the day (e.g., use a tablet to stream content prior to

bedtime and use the phone for other activities). We hypothesize that the higher errors of a single device sleep

detection approach [

] can be minimized by observing network activity from all the user’s devices to infer

sleep and wake times accurately.

SleepMore collects network activity information of all devices directly from the WiFi access points (AP) as

features. Then, it employs a two-pronged technique, running both a random forest machine learning classication

and moving average estimation models to predict sleep duration. Specically, it rst classies users in sleep or

awake states and computes condence intervals for these predictions using an innitesimal jackknife variance

estimation method. Outcomes with less than 95% condence level are noted as low condence predictions. From

applying a moving average, the most extended sequence of sleep states is observed as the user’s nocturnal sleep

period, with the start of the sequence denoting bedtime,

𝑇𝑠𝑙𝑒𝑒𝑝

, and the end of sequence denoting wake time,

Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 6, No. 4, Article 193. Publication date: December 2022.

SleepMore: Inferring Sleep Duration at Scale via Multi-Device WiFi Sensing •193:3

𝑇𝑤𝑎𝑘𝑒

. Simultaneously, the uncertainty rate of this estimation is instanced by the number of low condence

prediction states present in this sequence.

Our biomedical sleep research experts conducted an IRB-approved study over four weeks of an academic

semester during the COVID-19 pandemic restriction phase. Accordingly, we evaluate the performance of SleepMore

among 46 undergraduate students who resided on campus. We also conducted a small-scale user study among

three non-student participants to evaluate our system’s performance in home settings. As part of the study

protocols, student participants wore the Oura ring (gen 2) [

] wearable sleep tracker for baseline. They were

required to connect their devices to the campus WiFi while in their respective residences. While our study

specied no criteria on participants’ sleep habits, all users must own multiple personal devices. Participants

were a mix of habitual and irregular sleepers. They were asked to provide the MAC addresses of smartphones,

laptops, and tablets so that we could identify these devices directly from the WiFi infrastructure and extract their

network event logs. By default, all WiFi traces are anonymized. Our home participant chose to either provide

diary logs or use their personal Fitbit. They provided their WiFi network logs directly to us. To the best of our

knowledge, this is the rst work to accurately predict sleep using a scalable WiFi-based technique using inputs

from multiple user-owned devices. In designing, implementing, and evaluating SleepMore, our paper makes the

following contributions:

(1)

We present a random forest ML-based algorithm with innitesimal jackknife variance estimation and

moving average smoothing technique to predict users’ nocturnal sleep from WiFi network activity data

of multiple user devices in residential spaces. Our ML classier can predict the state of a user as sleep or

awake and estimate sleep duration based on the most extended sequence of sleep states within a 24-hour

period. We employ the variance estimation method to measure how condent a prediction is being made,

agging predictions beyond a 95% condence interval as low-condence outcomes and calculating the

uncertainty rate of sleep estimates by the number of low-condence outcomes present in the sequence of

sleep states. These techniques accurately estimate a user’s sleep duration and determine their bedtime and

wake time.

•

SleepMore is implemented as a cloud-based web service, building a semi-personalized model that requires

40% of users’ data for training, equivalent to 9 days of training data for 23 days of prediction.

•

Predictions made within a 5% uncertainty rate range between 15-28 minutes of sleep error and 7-29

minutes of wake error, proving statistically signicant improvements over prior work [

]. Note that

predictions within 5% uncertainty rate make up 80% of our predicted outcomes.

•

Our system evaluation is supplemented with results comparing SleepMore to prior AI techniques. These

comparative analyses conclude conditions under which each technique would thrive in predicting sleep

using WiFi network data.

(2)

We conduct an extensive experimental evaluation of SleepMore on the student population residing on campus.

Our results show that SleepMore can accurately determine the state of users sleeping with approximately

90% recall. These state predictions help to accurately estimate users’ sleep duration, bed and wake times.

•

We extend our solution deployment to two residential setting. The models for home users are tested in a

cross-environment, utilizing students’ data as the training set. Our results demonstrate the feasibility of

SleepMore in real-world settings and with actual residents. Further, we provide insights into using smart

home devices that are increasingly present in homes and utilize WiFi connections.

•

We characterize key factors that impact the performance of our approach, including the use of one to

many devices and the lack of training data to learn night-owl sleep schedules that did not often occur

among our participants.

Conclusively, SleepMore yields statistically indistinguishable sleep duration predictions compared to the

commercialized wearable trackers. However, it is essential to emphasize that our prediction mechanism of

Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 6, No. 4, Article 193. Publication date: December 2022.

193:4 •Zakaria et al.

estimating sleep duration is a partial function of the entire set of ne-grained features that Oura ring and similar

wearable sleep trackers can oer. In situations where sleep studies require regular data logging over an extended

period, SleepMore is competent as a lightweight supplementary tool without requiring users to wear a sleep

tracker or install a dedicated mobile app on their smartphone.

2 MOTIVATION AND BACKGROUND

This section provides background on sleep monitoring applications and sensing techniques, and motivates our

multi-device passive-sensing approach.

2.1 Sleep Health, a Call-to-Action

Much work has reported sleep deprivation as a public health burden [

]. Researchers sought to understand the

reasons and consequences of insucient sleep in dierent populations by measuring standard sleep parameters

such as time in bed, sleep duration, wake frequencies, and sleep latency [

]. These studies have noted

insucient sleep among adults as a result of lifestyles and work schedules [

], while adolescents’ sleep loss is

positively associated with more device use and online activities [

]. At the very least, sleep duration is documented

as the most fundamental and critical predictor of dierent health outcomes with longitudinal associations to weight

gains [27], quality of life [35], cardiovascular illnesses [7], cognitive impairments [16], to name a few.

The paradigm shift of recognizing sleep as a critical predictor of signicant health consequences [

] has led to a

fast-growing trend of consumer products oering digital sleep health options for monitoring and improving sleep.

From a research perspective, it has spurred a clear call to action for clinicians to comprehensively assess sleep

health among various age groups. Many of these works raised concerns over support for the basic understanding of

sleep, specically, improving the eectiveness of sleep screening [

] over longitudinal periods and developing

new technologies to accomplish this task [32].

2.2 Sleep Sensing Technologies

Sleep studies in clinical practice generally utilize retrospective scales, daily sleep logs, and/or polysomnography.

However, it is challenging for everyday consumers to personally monitor their sleep behavior with these

methods. Further, it remains highly burdensome for sleep studies to conduct long-term evaluation. Hence, sleep

technologies oer a low-burden approach to automatically detect a user’s sleep in the comfort of their own homes.

In a comprehensive study over two months, Massar et al. [

] compared and contrasted sleep measurements from

a consumer sleep-tracker, smartphone-based ecological momentary assessment, and user-phone interactions of

198 users. Their investigation identied stable interindividual dierences in sleep behavior, underscoring the

utility of these modalities in characterizing population sleep and peri-sleep behavior. We discuss the pros and

cons of these options as follows:

(1) Contact-based wearables:

Commercial alternatives have shown feasibility in monitoring sleep. Zam-

botti’s dened sleep wearable trackers as “over-the-counter, relatively low-cost devices available without

prescription or clinical recommendations,” [

] varying from wristbands to smartwatches, earbuds to

rings. The study protocol employing sleep wearables typically involves loaning these devices to enrolled

participants over a xed duration for reusability in future studies. The implications of such practice could

result in behavior modication during the study. Specically, users not used to wearing a device to bed and

raising the challenge of a high user attrition rate [12,29,61].

(2) Contactless sensing:

Low compliance with wearables has motivated the design of contactless sensing

approaches. Examples include the use of cameras [

], RF or radar sensing to characterize sleep stages [

]

and posture [

]. Despite these advances, RF sensing systems have yet to accelerate commercial adoption,

while camera-based solutions are more likely to intrude on privacy interests.

Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 6, No. 4, Article 193. Publication date: December 2022.

SleepMore: Inferring Sleep Duration at Scale via Multi-Device WiFi Sensing •193:5

(3) Smartphone-based sensing:

By contrast, smartphone-based sensing arose as an option from the de-

veloping smartphone dependency [

] among everyday users, leading to researchers using the “phone-

as-a-sensor.” Eorts specic to sleep monitoring include distinguishing respiratory patterns through a

microphone [

] or inferring user sleep behavior from monitoring screen interactions and application

usage [

]. In all these cases, the solution presents a dedicated mobile application that must be

installed in users’ smartphones and its data acquisition running in the device’s background.

These works collectively underscore the advancement in sleep monitoring technologies. While promising,

much tension in utilizing the modalities above in sleep studies arises from the challenge of user retention. Massar

et al. describe designing an incentive structure to continuously encourage regular data logging across such

modalities over the study period. With many people operating e-devices for minutes to hours before actually

intending to sleep, we seek to establish an accurate and generally accessible sensing approach that monitors

users’ daily sleeping behavior without changing their routines. It is important to emphasize that our work aims

not to replace the utilization of existing modalities that can obtain key physiological measures. However, it is

positioned as a complementary mechanism to support sleep monitoring scenarios over longitudinal periods.

2.3 Passive WiFi-based Sensing

Attempts to bypass the high attrition rate in smartphone-based sensing led to the utilization of WiFi-based

passive sensing techniques. Prior work by Mammen et al. has shown that it is possible to use WiFi connection logs

through users’ single smartphone collected from the WiFi infrastructure to predict sleep [

], however yielding

only 88.50% accuracy. Separately, broader surveys on behavioral monitoring via smart devices have argued that

utilizing single-device for monitoring is not fully comprehensive, in part because of inadequate device coverage

from users owning multiple devices [40,41,55].

(1)

Prior work only uses data collected from a single smartphone based on the assumption that users spend

most of their time online and on their smartphones [

]. We hypothesize that

including all devices

owned by the user will signicantly improve the accuracy of sleep detection.

(2)

There is a signicant gap between the accuracy of smartphone-based methods and those achieved by

sleep wearable trackers, which can achieve 96% recall at detecting sleep [

]. With this result in mind, it

is essential that our proposed solution, while using coarse-grained data source, achieves no statistically

signicant dierence from a wearable sleep tracker

in order to serve its purpose. In this study, we

use the Oura Ring (gen 2) as a representative wearable device for sleep monitoring.

2.4 Design Rationale

Figure 1shows an example of the network connection frequency for a typical user with multiple devices every

15-minutes through 24 hours between Day

, 6 pm to Day

, 6 pm. The observation that the personal smartphone

is the last device used before bedtime [

] makes it viable as a sleep monitoring sensor. However, users tend to

own multiple devices, and in this case, a secondary tablet device denotes the rst active usage upon the user

waking up. This behavior presents practical reasoning to infer users’ sleep behavior more accurately through

multiple device usage. Hence, this work utilizes the WiFi network device activity collected from multiple devices

to estimate an individual’s nocturnal sleep duration over a longitudinal basis as the most fundamental feature

supporting sleep studies.

Specically, it monitors the devices that are associated with the network to infer the user’s bedtime and

wake-up time (referred to as

𝑇𝑠𝑙𝑒𝑒𝑝

and

𝑇𝑤𝑎𝑘𝑒

). Adding new user-owned devices is easy as they would only need

to be connected to the same WiFi network. In addition, our solution signicantly preserves privacy as it does not

require decoding the actual content of any packets sent by a device. For each user-owned device (denoted by

Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 6, No. 4, Article 193. Publication date: December 2022.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

193SleepMore:InferringSleepDurationatScaleviaMulti-DeviceWiFiSensingCAMELLIAZAKARIA,UniversityofMassachusettsAmherst,USAGIZEMYILMAZ,NationalUniversityofSingapore,SingaporePRIYANKAMARYMAMMEN,UniversityofMassachusettsAmherst,USAMICHAELCHEE,NationalUniversityofSingapore,SingaporePRASHANTSHENOY,Universi...

展开>> 收起<<

SleepMore Inferring Sleep Duration at Scale via Multi-Device WiFi Sensing.pdf

共32页,预览5页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

SleepMore Inferring Sleep Duration at Scale via Multi-Device WiFi Sensing

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: