Context-aware Sleep Analysis with Intraday Steps and Heart Rate

Time Series Data from Consumer Activity Trackers

Zilu Liang

1,2 a

, Huyen Hoang Nhung

, Lauriane Bertrand

and Nathan Cleyet-Marrel

Ubiquitous and Personal Computing Lab, Kyoto University of Advanced Science (KUAS), 621-8555 Kyoto, Japan

Institute of Industrial Science, The University of Tokyo, 113-8654 Tokyo, Japan

National Institute of Electrical Engineering, Electronics, Computer Science, Hydraulics and Telecommunications

(INP-ENSEEIHT), 31000 Toulouse, France

Keywords: Sleep Tracking, Data Mining, Personal Informatics, Ubiquitous Computing, Wearable Computing.

Abstract: Wearable consumer activity trackers have become a popular tool for longitudinal monitoring of sleep quality.

However, sleep data were routinely visualized in isolation from other contextual information. In this paper,

we proposed a sleep analytics method to identify the associations between sleep quality and the contextual

data that are readily measurable with a single Fitbit device. Different from prior studies that only focused on

the daily aggregation of the contextual factors (e.g., total step counts), our method considers the intraday

temporal patterns of these factors. Time-domain, frequency-domain, and nonlinear features were derived

using the minute-by-minute intraday step and heart rate time series. The results showed that some of the

identified contextual features such as the zero-crossing of steps and the absolute energy of heart rate could

lead to actionable insights. While the nonlinear features—such as the average and longest diagonal line length

derived through the recurrent quantitative analysis of the step time series—may not lead to insights that can

be immediately acted on, they generated new hypotheses for further scientific studies. The results also showed

that when dealing with data of consumer wearables, the individual-level analysis could generate more

personally relevant insight than the cohort-level analysis.

1 INTRODUCTION

Getting enough and quality sleep is critical for

people’s physical and mental health (Buysse, 2014).

While traditional sleep monitoring technologies such

as polysomnography (PSG) and actigraphy were only

available in medical settings, recent advances in

consumer wearable technologies have expanded

sleep monitoring to daily life. Consumer activity

trackers such as Fitbit are affordable, easy to use, and

provide an intuitive user interface for data

visualization. These devices have achieved great

popularity not only among individual users but also

recently in the scientific research community (Peach

et al., 2018; Weatherall et al., 2018). As the latest

models can achieve comparable accuracy against

medical devices, these devices are increasingly used

in research studies to generate new insights into sleep

health (Liang, 2021; Liang & Ploderer, 2020;

Yurkiewicz et al., 2018)

https://orcid.org/0000-0002-2328-5016

https://orcid.org/0000-0002-5805-2087

Despite their popularity, consumer sleep tracking

technology is yet recognized as an effective tool that

helps people improve their sleep quality. Most sleep

trackers rely on motion-sensing technology

(accelerometer or gyroscope) to gauge how often a

user moves during sleep. Therefore, they may

overestimate or underestimate sleep and wake. For

example, a user wakes up in the middle of the night

but lying still could get an imprecise sleep summary

the next day. Furthermore, a previous study pointed

out ‘not identifying reasons for sleep problems’ and

‘not knowing how to act’ as two main barriers to

improving sleep with consumer activity trackers

(Liang & Ploderer, 2016). From a data science

perspective, addressing these two barriers requires the

analysis of users’ sleep data within their lifestyle

context (Liang, Ploderer, et al., 2016). Despite of

being able to collect multiple streams of behavioural

and physiological data (e.g., steps, heart rate, calorie

expenditure), Fitbit only allows users to visualize

170

Liang, Z., Nhung, H., Bertrand, L. and Cleyet-Marrel, N.

Context-aware Sleep Analysis with Intraday Steps and Heart Rate Time Series Data from Consumer Activity Trackers.

DOI: 10.5220/0010892900003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 170-179

ISBN: 978-989-758-552-4; ISSN: 2184-4305

these data separately, leaving it difficult to explore the

relationships among different streams of data. Figure

1 illustrates how sleep data are presented in isolation

from other streams of data that can potentially

provide contextual information. It is worth

mentioning that this problem is not specific to Fitbit

but rather universal to all consumer activity trackers.

Figure 1: A screenshot of the Fitbit dashboard.

On the other hand, several research studies have

attempted to address the above limitation of consumer

activity trackers. In these studies, researchers

developed web and mobile applications that allow

users to explorer the correlations among multiple

streams of health data readily collected with

consumer activity trackers (Bentley et al., 2013;

Daskalova et al., 2016; Kay et al., 2012; Liang,

Ploderer, et al., 2016). Both linear correlation analysis

and data mining techniques have been employed to

identify relationships between sleep and lifestyle

context (Daskalova et al., 2016; Liang, Chapa-

Martell, et al., 2016; Liang, Ploderer, et al., 2016).

Here we coin the term ‘context-aware sleep

computing’ as the umbrella of all the research studies

that attempt to analyse sleep within the context of

users’ lifestyle, physiological and psychological

states, and living environment.

Current context-aware sleep computing research

is limited to the daily aggregation of contextual

factors. Prior studies have only considered the

associations of sleep to the total number of steps,

calories expenditure, or minutes spent in various heart

rate zones in a day (Bentley et al., 2013; Daskalova et

al., 2016; Kay et al., 2012; Liang, Ploderer, et al.,

2016). While daily aggregations provide important

information on day-to-day variability, the intraday

temporal patterns of these factors—which may

potentially correlate to sleep quality at night—were

largely overlooked. This study aims to fill in this gap.

We performed a two-week data collection experiment

with 16 participants using Fitbit Charge 3. The

minute-by-minute time series data of steps and heart

rate were retrieved using a special Fitbit web API that

requires permission from the Fitbit company. Time-

domain, frequency-domain, and nonlinear features

were derived from the time-series data to capture the

intraday temporal patterns of these two factors in

different dimensions. We also proposed an ensemble

feature selection method to identify the important

intraday features that significantly correlate to sleep

quality at night. The contribution of this study is two-

fold.

 The proposed context-aware sleep analysis

method bridges a methodological gap in

persona informatics by considering the

intraday temporal patterns of lifestyle factors.

 We demonstrated how the proposed method

could help generate not only actionable insights

for individuals, but also interesting research

hypothesis that may inspire further studies in

sleep science.

2 RELATED WORKS

Sleep plays a critical role in human health and has

strong associations with learning, memory, and

metabolism. Many studies have been conducted to

help people understand more about sleep. However,

sleep experiments performed in sleep labs had some

drawbacks since the environment in which sleep

occurs was very different from a bedroom

environment. The findings of these studies might not

be generalized to real situations and result in poor

ecological validity. Recently, the development of

commercial sleep-tracking devices provides

researchers with a tool to track sleep as well as

daytime activities in naturalistic settings. While the

companion apps of these devices only present

different streams of data independently, several

research studies have developed third-party web and

mobile applications that help users to learn about the

relationship between sleep metrics and contextual

factors (Bauer et al., 2012; Bentley et al., 2013; Kay

et al., 2012; Liang, Ploderer, et al., 2016). These

studies have demonstrated both feasibility and merits

in investigating the effects of multiple categories of

factors along with sleep. To collect data without

disturbing participants’ daily activities, many studies

Context-aware Sleep Analysis with Intraday Steps and Heart Rate Time Series Data from Consumer Activity Trackers

171

used wristband activity trackers (Fitbit, Xiaomi Mi

Band) to collect behavioural and lifestyle information

in addition to sleep. Additional sensors were also used

to explore the home sleep environment (Kay et al.,

2012; Liang, Ploderer, et al., 2016; Park et al., 2019;

Wang et al., 2021). (Kay et al., 2012), built up a

system called Lullaby which can be installed in

participants’ bedrooms to capture light, temperature,

noise, and motion signals. Contextual factors varied

from study to study and some of them cannot be

detected automatically, such as consuming caffeine,

nicotine use, relaxation, and food intake. This

problem can be solved by asking the participants to

write down their observations manually (Bauer et al.,

2012; Bentley et al., 2013; Daskalova et al., 2016;

Liang, Ploderer, et al., 2016; Park et al., 2019).

However, missing data is a big challenge since

manual logging did not occur frequently. Apart from

using wristband devices, some studies used available

sensors in smartphones and developed their own

widget so that they can reduce the need for external

devices (Bauer et al., 2012; Daskalova et al., 2016).

Taking advantage of mobile phones, factors like

location, weather, free/busy hours, communication

records can be extracted automatically (Bentley et al.,

2013; Kay et al., 2012; Liang, Ploderer, et al., 2016;

Park et al., 2019; Wang et al., 2021). In these studies,

some participants were amazed by how little they

knew about sleep despite having sleep every day (Kay

et al., 2012). Participants were able to see the links

between sleep hours with emotion and physical

activity for the next day (Bentley et al., 2013).

Studying contextual factors not only benefits healthy

subjects but also contributes to sleep disorder

research. (Park et al., 2019) found that contextual

factors such as calories consumed, walk, distance,

stairs, and active ratio could be useful for predicting

sleep efficiency and ranking the risk level of insomnia

for the next night’s sleep. Some contextual factors

such as age, gender, subjective perception of sleep

quality and heart rate were shown to affect the

accuracy of sleep trackers and were used to develop

more accurate sleep staging algorithms (Liang &

Chapa-Martell, 2019, 2021).

Some limitations exist and demand further work

to improve. First, existing studies have only

considered the daily aggregation of lifestyle

contextual factors, such as the total number of steps,

the total calories expenditure, and the total minutes

spent in each activity intensity zone. The intraday

variability and the temporal patterns were largely

neglected. Second, the data from different

participants were usually merged into one large

dataset for analysis, assuming the homogeneity of the

cohort. While such cohort-level analysis is widely

adopted, it is found that the results are usually not

generalized well to individuals, especially when the

intra-personal variability is larger than the inter-

personal variability (Molenaar, 2004). In this study,

the intraday temporal patterns of the time series data

of steps and heart rate were captured using a diversity

of time-domain, frequency-domain, and nonlinear

features. In addition, we performed contextual-aware

sleep analysis for participants individually to identify

correlations between sleep and lifestyle for each

person. In what follows, we demonstrate the

usefulness of intraday features of the time series data

of lifestyle factors, as well as the importance of a

research paradigm shift from cohort informatics

towards personal informatics.

3 METHODS

An overview of the proposed context-aware sleep

analysis is illustrated in Figure 2. All data were

collected using Fitbit Charge 3. We constrained the

contextual factors to steps and heart rate because they

are readily measurable together with sleep data using

a single Fitbit device, which best represents the usage

scenario of consumer activity trackers in real life. In

what follows, we detail the data collection experiment,

data preprocessing, feature construction and the

original feature selection algorithm.

3.1 Data Collection and Retrieval

Due to the lack of high-quality open-access datasets

that serve our purpose, we conducted a 14-day data

collection experiment on our own with 16 participants

using Fitbit Charge 3. The participants were recruited

through personal connections and word of mouth.

Applicants with diagnosed sleep problems were

excluded. The cohort consist of 9 women and 7 men,

with an average age of 30 years. Ethics approval was

obtained from the Ethics Committee of the Kyoto

University of Advanced Science.

We mailed a Fitbit Charge 3 device to each

participant and instructed them to set up the device

and the companion Fitbit app on their smartphones.

The participants were required to log in to the Fitbit

app using a provided email account that our research

assistants created exclusively for the data collection

experiment. The subjects were encouraged to wear

the Fitbit Charge 3 as often as possible and to

synchronize the device daily. Participants who

successfully completed the data collection

experiment were allowed to keep the Fitbit device as

HEALTHINF 2022 - 15th International Conference on Health Informatics

172

a reward, and they were instructed to re-login on their

Fitbit apps using their personal email account, so that

their data would not be synchronized to the

experiment account afterward.

The daily aggregation of sleep data was retrieved

using Fitbit public web API. We selected three sleep

metrics—total sleep time (TST), wake after sleep

onset (WASO), and deep sleep ratio—as indicators of

sleep quality, as prior studies showed that many users

to consumer sleep trackers rely on these metrics to

assess their sleep quality (Bauer et al., 2012; Bentley

et al., 2013; Kay et al., 2012; Liang & Ploderer, 2020;

Liang, Ploderer, et al., 2016). Prior validation studies

found that Fitbit are reasonably accurate in measuring

the daily aggregation of sleep metrics (De Zambotti

et al., 2019; Liang & Chapa-Martell, 2018).

The intraday time series of steps and heart rate

were retrieved using a special Fitbit web API that

requires getting permission from the Fitbit company.

While a third-party service has no limitations in

accessing the aggregated data, permission is needed

to access the intraday time series. Both the steps and

heart rate time series were retrieved at one-minute

resolution.

Figure 2: An overview of the proposed context-aware sleep

analysis method with Fitbit.

3.2 Data Preprocessing

The data preprocessing protocol described below was

performed individually on the dataset of each subject

to handle missing data and to ensure the correct

timestamp match between the contextual data and the

sleep data.

Missing data was an occasional issue when no

sleep stage data was recorded throughout the night, or

no resting heart rate was recorded upon waking up on

a day. The Fitbit API supports the retrieval of two

kinds of sleep data. The ‘stage’ data consist of sleep

stage levels include ‘light’, ‘deep’, ‘rem’, and ‘wake’,

while the ‘classic’ data consist of sleep pattern levels

include ‘asleep’, ‘restless’, and ‘wake’. In other

words, when the sensor did not record sufficient

signals to infer sleep stages of a night, it only roughly

classified sleep and awake. The target sleep metrics

that were related to sleep stages were all filled in with

NAs on nights with no sleep stage information.

Missing heart rate data were set to NA as well.

Afterward, the NAs were imputed with the mean of

the intraday time series.

The contextual data and the sleep data needed to

be matched by date. According to the data scheme of

Fitbit, the sleep data of day N corresponds to the sleep

that ends in the morning of day N (not the sleep that

starts on the night of day N). Hence, the contextual

data of sleep on day N refers to the steps and heart

rate data between the end time of sleep on day N-1

and the start time of sleep on day N. Depending on

whether a user goes to bed before midnight (case 1)

and after midnight (case 2), the sleep start time of day

N could be either on day N-1 (case 1) or on day N

(case 2). The corresponding contextual data that

matched to the sleep on day N hence differed between

case 1 and case 2. In addition, the raw sleep data

retrieved only consist of sleep stages in minutes. We

calculated the ratio of deep sleep (DR) by dividing the

minutes of deep sleep by TST.

3.3 Feature Construction

We derived features from the intraday time series of

steps and heart rate. A full list of derived features is

summarized in Table 1.

The time-domain features were directly derived

from the preprocessed time series data. These features

capture the statistical and morphological

characteristics of the intraday time series data.

Frequency-domain features were derived from the

Fourier transform of the original time series data.

These features capture the spectral characteristics of

the intraday time-series data. Nonlinear features were

Context-aware Sleep Analysis with Intraday Steps and Heart Rate Time Series Data from Consumer Activity Trackers

173

derived after phase space construction by applying

Taken’s time-delay embedding to the time-series data

(Dingwell & Cusumano, 2000). Several nonlinear

dynamic system analytic techniques were applied for

deriving nonlinear features. These techniques

included recurrence quantitative analysis (RQA),

Poincaré plots (PP) (Hoshi et al., 2013), detrended

fluctuation analysis (DFA) (Hardstone et al., 2012),

as well as several measures of entropy (López-Ruiz

et al., 1995). These features capture the chaotic

characteristics and the complexity of the intraday

steps and heart rate time-series. The infinite and

missing values were first unified as ‘NA’ and then

imputed by the mean of the corresponding features.

3.4 Feature Selection

Feature selection is a critical step in identifying the

contextual features that are relevant and have the

strongest predictive power of the target sleep metric

(Guyon & Elisseeff, 2003). Existing feature selection

algorithms fall into three main categories: wrappers,

filters, and embedded methods. Each category has its

merits and demerits. Wrapper methods build a

predictive model to score feature subsets, which

usually provide the best-performing feature set but

are computationally intensive. Filter methods achieve

a trade-off between computational speed and the

usefulness of the feature set. Embedded methods

perform feature selection as part of the model

construction process and the computational

complexity is between the previous two categories. In

this study, we proposed an ensemble feature ranking

and selection method illustrated in Figure 3. The

proposed algorithm leverages six feature selection

algorithms to generate an average importance score

for each feature and performs feature pruning based

on a set of criteria.

As illustrated in Figure 3, the six feature selection

algorithms include one wrapper (i.e., recursive

feature elimination (RFE)), two filters (i.e., F-test and

mutual information (MI)), and three embedded

methods (i.e., multivariate linear regression, Lasso

regression, and Ridge regression). All features were

scaled between [0, 1] before being passed to the

feature selection algorithm. Each algorithm k

generated an importance score 

,



for a feature x in

relation to a target sleep metric y. The 

,



of all six

algorithms were scaled to the range [0, 1] and then

averaged to generate an average importance score for

feature x in relation to sleep metric y. In the meantime,

the support 

,

of feature x in relation to sleep

metric y—defined as the number of algorithms that

generated a scaled 

,



above 0.5—was also

computed. Pearson’s correlation coefficient and the

correspondent p-value were calculated to quantify the

linear relationship between feature x and sleep metric

Table 1: Features constructed using Fitbit intraday time

series data.

Category Feature Denotation

Time-

domain

mean mean

median median

standard deviation std

variance variance

peak to peak p2p

maximum max

minimum min

absolute energy absEnergy

mean absolute

difference

meanAbsDiff

zero cross zc

skewness skew

kurtosis kurt

order moment mmt5th

Frequenc

y-domain

total spectrum totalSpec

maximal spectrum maxSpec

peak ratio peakRatio

Nonlinear recurrence rate recurRate

percent

determinism

det

average diagonal

line length

avgDiagLine

longest diagonal

line length

longestDiagLine

entropy of diagonal

lines lengths

entropyDiagLine

laminarity lam

trapping time trappingTime

longest vertical line

length

longestVertLine

entropy of vertical

lines lengths

entropyVertLine

ratio between

determinism and

recurrence rate

ratioDetRecurRate

ratio between

laminarity and

determinism

ratioLamDet

correlation

dimension

corDim

scaling exponent alpha

scaling exponent

with 50% overlap

alphaOverlap

Hurst exponent hurstExpK

Shannon entropy shannonEn

sample entropy sampEn

permutation entropy permuEn

system complexity sysComplexity

HEALTHINF 2022 - 15th International Conference on Health Informatics

174

Figure 3: The proposed ensemble method for selecting the most important intraday contextual features in relation to the target

sleep metrics.

Three conditions were defined to select the most

important features: (1) ̅

,

> 0.5, (2) 

,

≥ 3, and

(3) p < 0.05. The outputs of the ensemble feature

selection method were the selected contextual

features and the corresponding Pearson’s correlation

coefficients in relation to each sleep metric.

In this study, feature selection was performed at

both the individual level and the cohort level. At the

individual level, the cleaned dataset of each subject

was fed directly into the ensemble feature selection

method. At the cohort level, the datasets of all

subjects were merged before being fed into the

ensemble feature selection method. It is worth noting

that at the cohort level analysis, the repeated measures

correlation was used in place of Pearson’s correlation

to handle the dependence among observations. The

parameter α was set to 0.5 for Lasso and Ridge, and

the RFE was set to stop the search when 5 features

were left. Missing values were removed in a pair-wise

manner in correlation analysis.

4 RESULTS

The contextual features that were significantly

associated with each sleep metrics are shown in

Figure 4~6. The features were selected using the

proposed ensemble method. Red, blue, and grey cells

indicate significantly and positively correlated

important features, significantly and negatively

correlated important features, and unimportant

features, respectively. The shades of red and blue

indicate the strength of correlation. The first column

shows the result at the cohort level, and the

subsequent columns show the result for each subject.

As can be seen from figures 4-6, the identified

correlations exhibit great inter-participant differences,

while no correlation was found between the

contextual and sleep metric for P2 and P9.

Figure 4 shows the identified important

contextual features of TST. At the cohort level,

ratioDetRecurRate was the only contextual feature

Context-aware Sleep Analysis with Intraday Steps and Heart Rate Time Series Data from Consumer Activity Trackers

175

Figure 4: Contextual features that significantly correlate to TST. The value and colour shade of a cell indicate the correlation

coefficient and the correlation strength, respectively.

Figure 5: Contextual features that significantly correlate to WASO. The value and colour shade of a cell indicate the

correlation coefficient and the correlation strength, respectively.

HEALTHINF 2022 - 15th International Conference on Health Informatics

176

Figure 6: Contextual features that significantly correlate to the deep sleep ratio. The value and colour shade of a cell indicate

the Pearson’s correlation coefficient and the correlation strength, respectively.

that exhibits a significant correlation. This factor was

also an important factor for P3, though the correlation

strength at the cohort level was much lower than that

at the individual level. At the individual level, the

same contextual feature may demonstrate the

opposite correlation direction for different

participants. For example, meanAbsChange_HR was

negatively correlated to TST for P1 (r = -0.72, p =

0.006) but positively correlated to TST for P4 (r =

0.57, p = 0.035). The strongest correlation was found

between absEnergy_HR and TST for P14 (r = -0.79,

p < 0.001). No correlation was found between the

contextual features and TST for P2, P6, P7, P9, and

P11. Figure 5 shows the identified important

contextual features of WASO. At the cohort level, no

factor was significantly correlated to WASO. Similar

to TST, the same contextual feature may demonstrate

opposite correlation direction for different

participants. It is shown that entropyDiagLine_Steps

was a negatively correlated factor for P3 (r = -0.70, p

= 0.008) but a positively correlated factor for P6 (r =

0.60, p = 0.010), and sampEn_Steps was a negatively

correlated factor or P1 (r = -0.63, p = 0.020) but a

positively correlated factor for P15 (r = 0.50, p =

0.020). Significant inter-subject differences were

observed. P4 and P5 had the highest number of

correlated features, while no correlation was found

for P2, P9, and P13.

Figure 6 shows the identified important

contextual features of the deep sleep ratio. No

contextual feature was selected at the cohort level. At

the individual level, no contextual feature was

selected for 9 out of 16 participants. P3 and P11 had

the highest number of selected features for the deep

sleep ratio.

5 DISCUSSIONS

With the burgeon of consumer sleep tracking

technologies, there has been an increasing analytical

need to interpret personal sleep data within a user’s

behavioural and physiological context. In response to

this need, several prior studies have considered the

relationships between sleep and the daily

aggregations of contextual factors (Bauer et al., 2012;

Bentley et al., 2013; Kay et al., 2012; Liang, Ploderer,

et al., 2016), but the intraday temporal patterns of the

contextual factors were largely neglected. In this

study, we directed the focus to the intraday temporal

patterns and characteristics of the heart rate and step

time-series data, which can be readily measured

together with sleep data using consumer activity

trackers such as Fitbit. We derived time-domain,

frequency-domain, and nonlinear features from the

minute-by-minute intraday time series and proposed

an ensemble feature selection method to identify the

most important intraday features that were

significantly associated to target sleep metrics.

This study yielded two principal findings. First,

the intraday temporal patterns of the behavioural and

physiological data collected with consumer activity

trackers encoded valuable contextual information for

sleep analysis. Second, the correlation analysis results

generated at the cohort level are likely to deviate from

the correlations at the individual level.

Some of the identified contextual features could

lead to intuitive interpretations that generated

actionable insights. The zero-crossing of the intraday

step time-series was an important contextual factor at

the individual level for TST and WASO. At the

individual level, it shows that a decrease in zero-

crossing was associated with increased sleep hours

Context-aware Sleep Analysis with Intraday Steps and Heart Rate Time Series Data from Consumer Activity Trackers

177

for P10 and P14, but increased wake time for P12.

Since zero-crossing is an indicator of the noisiness of

a signal (Liang, 2021), it is indicated that P10 and P14

were likely to achieve longer sleep hours by

improving the regularity of their daily physical

activity, while P12 may pursue the opposite to reduce

wake time during sleep. Zero-crossing has been an

important feature in EEG-based automatic sleep

staging (Şen et al., 2014). Our finding suggests that

the zero-crossing of intraday step time series

collected using consumer activity trackers may serve

as a predictor of night sleep, though it requires further

analysis to confirm this hypothesis.

The mean absolute difference of the intraday heart

rate time series was an important contextual factor of

all the target sleep metrics at the individual level. An

increase in the mean absolute difference of the

intraday heart rate was associated with increased TST

for P4, decreased WASO for P1 and P14, and

increased deep sleep ratio for P11. Since being

engaged in more intense physical activity during the

day is linked to the increased mean absolute

difference in heart rate, these participants may

attempt to integrate exercise into daily routines for

better sleep at night.

Contextual factors such as the absolute energy of

the intraday heart rate time series also yielded

actionable insights. An increase in the absolute

energy of the heart rate time series was positively

associated with WASO for P8 but was negatively

associated with WASO for P12. Correspondently, P8

may benefit from spending more time in the low heart

rate zone while P12 may benefit from the opposite.

On the other hand, some of the nonlinear features

may not provide insights that can be immediately

acted on, but they may generate interesting

hypotheses that inspire further scientific studies.

Several selected nonlinear features were derived from

the intraday step time series using recurrence

quantitative analysis (RQA). For example, the

average diagonal line length (negatively associated to

TST for P12 and positively associated to WASO for

P16), the longest diagonal line length (positively

associated to WASO for P5 and to TST for P16), the

entropy of the vertical line length (positively

correlated to WASO P6 and to deep sleep ratio for

P12) were important contextual features of sleep for

certain participants. Chaos-based analysis of human

physiological data has become widely adopted for

diagnosing motor-control and cardiovascular diseases

(Dingwell & Cusumano, 2000; Wu et al., 2009).

Similarly, the nonlinear chaotic features derived from

the intraday personal health data may represent a

promising method for predicting sleep quality or

diagnosing sleep problems in daily life settings.

6 CONCLUSIONS

In this study, we demonstrated the importance of

considering the intraday temporal patterns of steps

and heart rate for context-aware sleep analysis with

personal health data. The statistical, spectral,

morphological, and nonlinear features of the intraday

time series could all provide valuable predictive

information of sleep at night and should be routinely

included in personal informatics analysis. While

some intraday features provided actionable insights

that could guide behaviour change for better sleep,

others may generate interesting hypotheses that

inspire further scientific studies. In the meantime, the

individual-level analysis may be preferred over

cohort-level analysis for generating personalized

insights on sleep health.

ACKNOWLEDGEMENTS

This work was supported by JSPS KAKENHI Grant

Number 16H07469, 19K20141, and 21K17670.

REFERENCES

Bauer, J., Consolvo, S., Greenstein, B., & al., e. (2012).

ShutEye: encouraging awareness of healthy sleep

recommendations with a mobile, peripheral display. In

Proceedings of the SIGCHI Conference on Human

Factors in Computing Systems, Austin, Texas, USA.

Bentley, F., Tollmar, K., Stephenson, P., Levy, L., Jones,

B., Robertson, S., Wilson, J. (2013). Health Mashups:

presenting statistical patterns between wellbeing data

and context in natural language to promote behavior

change. ACM Transactions on Computer-Human

Interaction (TOCHI), 20(5), Article No. 30.

Buysse, D. J. (2014). Buysse DJ. Sleep Health: Can We

Define It? Does It Matter? Sleep . 2014;37(1):9-17.

doi:10.5665/sleep.3298. Sleep, 37(1), 9-17.

Daskalova, N., Metaxa-Kakavouli, D., Tran, A., et al.

(2016). SleepCoacher: A Personalized Automated Self-

Experimentation System for Sleep Recommendations.

In Proceedings of the 29th Annual Symposium on User

Interface Software and Technology, Tokyo, Japan.

De Zambotti, M., Cellini, N., Goldstone, A., Colrain, I., &

Baker, F. (2019). Wearable sleep technology in clinical

and research settings. Med Sci Sports Exerc, 51(7),

1538-1557.

HEALTHINF 2022 - 15th International Conference on Health Informatics

178

Dingwell, J. B., & Cusumano, J. (2000). Nonlinear time

series analysis of normal and pathological human

walking. Chaos, 10(4), 848-863.

Guyon, I., & Elisseeff, A. (2003). An introduction to

variable and feature selection. J. Mach. Learn. Res.,

3(JMLR.org), 1157-1182.

Hardstone, R., Poil, S.-S., Schiavone, G., Jansen, R.,

Nikulin, V. V., Mansvelder, H. D., & Linkenkaer-

Hansen, K. (2012). Detrended fluctuation analysis: a

scale-free view on neuronal oscillations. Frontiers in

Physiology, 3, 450.

Hoshi, R. A., Pastre, C. M., Vanderlei, L. C. M., & Godoy,

M. F. (2013). Poincaré plot indexes of heart rate

variability: relationships with other nonlinear variables.

Auton Neurosci., 177(2), 271-274.

Kay, M., Choe, E. K., Shepherd, J., & al., e. (2012).

Lullaby: a capture & access system for understanding

the sleep environment. In Proceedings of the 2012 ACM

Conference on Ubiquitous Computing, Pittsburgh,

Pennsylvania.

Liang, Z. (2021). What does sleeping brain tell about stress?

A pilot fNIRS study into stress-related cortical

hemodynamic features during sleep. Frontiers in

Computer Science.

Liang, Z., & Chapa-Martell, M. A. (2018). Validity of

consumer activity wristbands and wearable EEG for

measuring overall sleep parameters and sleep structure

in free-living conditions. Journal of Healthcare

Informatics Research, 1-27.

Liang, Z., & Chapa-Martell, M. A. (2019). Accuracy of

Fitbit wristbands in measuring sleep stage transitions

and the effect of user-specific factors. JMIR Mhealth

Uhealth, 7(6), e13384.

Liang, Z., & Chapa-Martell, M. A. (2021). A Multi-level

classification approach for sleep stage prediction with

processed data derived from consumer wearable

activity trackers. Frontiers in Digital Health, 3,

665946.

Liang, Z., Chapa-Martell, M. A., & Nishimura, T. (2016).

Mining hidden correlations between sleep and lifestyle

factors from quantified-self data. In Proceedings of the

2016 ACM International Joint Conference on

Pervasive and Ubiquitous Computing: Adjunct,

Heidelberg, Germany.

Liang, Z., & Ploderer, B. (2016). Sleep tracking in the real

world: a qualitative study into barriers for improving

sleep. In Proceedings of the 28th Australian

Conference on Computer-Human Interaction,

Launceston, Tasmania, Australia.

Liang, Z., & Ploderer, B. (2020). “How Does Fitbit

Measure Brainwaves”: A Qualitative Study into the

Credibility of Sleep-tracking Technologies. Proc. ACM

Interact. Mob. Wearable Ubiquitous Technol., 4(1),

Article 17.

Liang, Z., Ploderer, B., Liu, W., Nagata, Y., Bailey, J.,

Kulik, L., & Li, Y. (2016). SleepExplorer: a

visualization tool to make sense of correlations between

personal sleep data and contextual factors. Personal

Ubiquitous Comput., 20(6): 985-1000.

López-Ruiz , R., Mancini , H. L., & Calbet , X. (1995).

A statistical measure of complexity. Physics Letters A,

209(5-6), 321-326.

Molenaar, P. C. M. (2004). A manifesto on psychologyas

idiographic science: bringing the person back into

scientific psychology, this time forever. Measurement

Interdisciplinary Research and Perspectives, 2(4): 201-

218.

Park, S., Li, C.-T., Han, S., Hsu, C., Lee, S. W., & Cha, M.

(2019). Learning sleep quality from daily logs. In

Proceedings of the 25th ACM SIGKDD Conference on

Knowledge Discovery and Data Mining, USA.

Peach, H. D., Gaultney, J. F., & Ruggiero, A. R. (2018).

Direct and indirect associations of sleep knowledge and

attitudes with objective and subjective sleep duration

and quality via sleep hygiene. The Journal of Primary

Prevention, 39(6), 555-570.

Wang, C., Lizardo, O., & Hachen, D. S. (2021). Using Fitbit

data to examine factors that afftect daily activity levels

of college students. PLOS ONE 16(1): e0244747.

Weatherall, J., Paprocki, Y., Meyer, T. M., Kudel, I., &

Witt, E. A. (2018). Sleep tracking and exercise in

patients with type 2 diabetes mellitus (step-D): pilot

study to determine correlations between Fitbit data and

patient-reported outcomes. JMIR Mhealth Uhealth,

6(6), e131.

Wu, G.-Q., Arzeno, N. M., Shen, L.-L., Tang, D.-K.,

Zheng, D.-A., Zhao, N.-Q., Poon, C.-S. (2009). Chaotic

Signatures of Heart Rate Variability and Its Power

Spectrum in Health, Aging and Heart Failure. PLOS

ONE 4(2): e4323.

Yurkiewicz, I. R., Simon, P., Liedtke, M., Dahl, G., &

Dunn, T. (2018). Effect of Fitbit and iPad wearable

technology in health-related quality of life in adolescent

and young adult cancer patients. Journal of Adolescent

and Young Adult Oncology, 7(5), 579-583.

Şen, B., Peker, M., Çavuşoğlu, A., & Çelebi, F. V. (2014).

A Comparative Study on Classification of Sleep Stage

Based on EEG Signals Using Feature Selection and

Classification Algorithms. Journal of Medical Systems,

38(3), 18.

Context-aware Sleep Analysis with Intraday Steps and Heart Rate Time Series Data from Consumer Activity Trackers

179