1
SCieNTifiC REPORTS | (2018) 8:12975 | DOI:10.1038/s41598-018-31266-z
www.nature.com/scientificreports
Estimating sleep parameters using
an accelerometer without sleep
diary
Vincent Theodoor van Hees
1
, S. Sabia
2,3
, S. E. Jones
4
, A. R. Wood
4
, K. N. Anderson
5
,
M. Kivimäki
3
, T. M. Frayling
4
, A. I. Pack
6
, M. Bucan
7,8
, M. I. Trenell
9
, Diego R. Mazzotti
6
,
P. R. Gehrman
6,8
, B. A. Singh-Manoux
2,3
& M. N. Weedon
4
Wrist worn raw-data accelerometers are used increasingly in large-scale population research. We
examined whether sleep parameters can be estimated from these data in the absence of sleep diaries.
Our heuristic algorithm uses the variance in estimated z-axis angle and makes basic assumptions
about sleep interruptions. Detected sleep period time window (SPT-window) was compared against
sleep diary in 3752 participants (range = 60–82 years) and polysomnography in sleep clinic patients
(N = 28) and in healthy good sleepers (N = 22). The SPT-window derived from the algorithm was 10.9
and 2.9 minutes longer compared with sleep diary in men and women, respectively. Mean C-statistic to
detect the SPT-window compared to polysomnography was 0.86 and 0.83 in clinic-based and healthy
sleepers, respectively. We demonstrated the accuracy of our algorithm to detect the SPT-window. The
value of this algorithm lies in studies such as UK Biobank where a sleep diary was not used.
Wrist-worn raw-data accelerometers are increasingly used for the assessment of physical activity in large popula-
tion studies such as the Whitehall II study or mega-cohorts such as UK Biobank
1–3
. e decision to use raw-data
accelerometers is motivated by the improved comparability of output across dierent sensor brands
4,5
, and better
control over all steps in data processing
6
. Accelerometers are commonly worn for 24 hours per day, thus providing
information over the day and night; making them potentially valuable for sleep research.
A major challenge in accelerometer-based sleep measurement is to derive sleep parameters without additional
information from sleep diaries
1,3,7
. Standard methods for sleep detection based on conventional accelerometers
(actigraphy) involves asking the participant to record their time in bed, sleep onset, and waking up time
8–10
. In a
previous paper we developed a method to detect sleep guided by sleep diary records
11
. However, the increasing
use of accelerometry in studies worldwide without sleep diaries necessitates the development of novel methods
to derive indicators of sleep behaviour, in the absence of sleep diary records. A crucial step is the detection of
the sleep period time window (SPT-window), which is the time window starting at sleep onset and ending when
waking up aer the last sleep episode of the night. Once the SPT-window can be detected without a diary, our
previously published method can be used to detect sleep episodes within this window
11
. Polysomnography (PSG)
is considered the gold-standard measure of sleep parameters, making it an ideal methodology to validate sleep
detection methods using an accelerometer. Additionally, experiments in daily life can be used to establish con-
current validity with sleep diary.
We aim to develop and evaluate a heuristic algorithm for the detection of the SPT-window from raw data
accelerometers unaided by a sleep diary and to compare sleep parameters (waking up, sleep onset time and
SPT-window duration) with sleep diary records assessed in the daily life of a large cohort of older adults, and with
PSG data collected in a sleep clinic and a group of healthy good sleepers.
1
Netherlands eScience Center, Amsterdam, The Netherlands.
2
INSERM U1018, Centre for Research in Epidemiology
and Population Health, Université Paris-Saclay, Paris, France.
3
Department of Epidemiology & Public Health,
University College London (UCL), London, UK.
4
University of Exeter Medical School, Genetics of Complex Traits,
Exeter, UK.
5
Regional Sleep Service, Freeman Hospital, Newcastle-upon-Tyne, UK.
6
Center for Sleep and Circadian
Neurobiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania, USA.
7
Department of Genetics, Perelman School of Medicine, University of Pennsylvania School of Medicine, Philadelphia,
Pennsylvania, USA.
8
Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania School
of Medicine, Philadelphia, Pennsylvania, USA.
9
Movelab, Newcastle University, Newcastle-upon-Tyne, UK.
Correspondence and requests for materials should be addressed to V.T.v.H. (email: v.vanhees@esciencecenter.nl)
Received: 8 March 2018
Accepted: 8 August 2018
Published: xx xx xxxx
OPEN
www.nature.com/scientificreports/
2
SCieNTifiC REPORTS | (2018) 8:12975 | DOI:10.1038/s41598-018-31266-z
Methods
Study population. In order to assess the validity of our algorithm in dierent settings and against both data
from sleep diary and polysomnography, data are drawn from three dierent study populations described below.
e Whitehall II cohort study
12
: full details on data collection were previously described
11
. Briey, acceler-
ometer measurement was added to the study at the 2012/2013 wave of data collection for participants seen at
the central London clinic and for those living in the South-Eastern regions of England who underwent a clinical
evaluation at home
2
. Of the 4879 participants to whom the accelerometer was proposed in the Whitehall II Study,
388 did not consent and 210 had contraindications (allergies to plastic or metal, travelling abroad the following
week). Of the remaining 4281 participants who wore the accelerometer, 4204 (98.2%) had valid accelerometer
data (a readable data le). Among them, sleep diary data were missing for 80 participants and 29 additional par-
ticipants did not meet criteria for accelerometer wear time (at least one night dened as noon-noon with >16 h of
wear time). Of the remaining 4095 participants (a total of 27,966 nights) 342 did not have complete demographic
data (age, BMI and sex). erefore, the main assessment of discrepancies between the accelerometer and the sleep
diary was undertaken in 3752 participants (76.9% of those invited) with a total of 25,645 nights
11
. e resulting
participants (75.2% men) were on average 69.1 (standard deviation (SD) = 5.6) years old and had a mean body
mass index (BMI) of 26.4 (SD = 4.2) kg/m
2
.
Sleep clinic patients: these data come from 28 adult patients who were scheduled for a one-night polysom-
nography (PSG) assessment at the Freeman Hospital, Newcastle upon Tyne, UK, as part of their routine clinical
assessment and were subsequently invited to participate in the study
11
. All 28 patients recruited for the pol-
ysomnography study (11 female) had complete accelerometer data for the le wrist and 27 had complete data
for the right wrist and were aged between 21 and 72 years (mean ± sd: 45 ± 15 years). Diagnosed sleep disorders
included: hypersomnia (N = 2), insomnia (N = 2), REM behaviour disorder (N = 3), sleep apnoea (N = 5), nar-
colepsy (N = 1), sleep apnoea (N = 4), parasomnia (N = 1), restless leg syndrome (N = 5), and sleep paralysis
(N = 1), and nocturnia (N = 1). ree patients had more than one sleep disorder.
Healthy good sleepers: these data come from 22 adults who underwent a one-night PSG assessment at the
University of Pennsylvania Center for Sleep. Twenty-two participants recruited for the polysomnography study
(68% female) had complete accelerometer data for the non-dominant wrist and were aged between 18 and 35
years (mean ± sd: 22.8 ± 4.5 years).
Ethics Statement. In all three studies participants were provided with instructions and an information sheet
about the study and were given time to ask questions prior to providing written informed consent. e studies
were approved by the University College London ethics committee (85/0938) and the NRES Committee North
East Sunderland ethics committee (12/NE/0406), and University of Pennsylvania ethics committee (819591)
respectively. All experiments were performed in accordance with relevant guidelines and regulations.
Instrumentation. Participants in the Whitehall II Study were asked to wear a tri-axial accelerometer
(GENEActiv, Activinsights Ltd, Kimbolton, UK) on their non-dominant wrist for nine (24-h) consecutive days.
ey were asked to complete a simple sleep diary every morning which consisted of two questions: ‘what time did
you rst fall asleep last night?’ and ‘what time did you wake up today (eyes open, ready to get up)?’ e acceler-
ometer was congured to collect data at 85.70 Hz with a ±8 g dynamic range. A more complete description of the
accelerometer protocol can be found in our earlier publication
2
.
In the second and third study, polysomnography (Embletta
®
, Denver) was performed using a standard pro-
cedure, including video recording, a sleep electroencephalogram (leads C4-A1 and C3-A2), bilateral eye move-
ments, submental EMG, and bilateral anterior tibialis EMG to record leg movements during sleep. Respiratory
movements were detected with chest and abdominal bands measuring inductance, airow was detected with
nasal cannulae measuring pressure, and oxygen saturation of arterial blood was measured. Airow limitation and
changes in respiratory movement were used to detect increased upper-airway resistance. All respiratory events
and sleep stages were scored according to standard criteria so that EEG determined total sleep time could be
measured
9
. Participants in the second study (PSG in sleep clinic) were asked to wear the same brand of accelerom-
eter as in the rst study (GENEActiv, Activinsights Ltd, Kimbolton, UK) on both wrists throughout the one-night
polysomnography assessment. Here, the accelerometer was also congured to record at 85.70 Hz. Accelerometer
data were collected on both wrist to assess the role of sensor location on classication performance, unfortunately
no information on handedness was recorded. Participants in the third study (PSG in healthy good sleepers) were
asked to wear an accelerometer of the brand Axivity (Axivity Ltd, Hoults Yard, UK) on the non-dominant wrist
throughout the one-night polysomnography assessment. Here, the accelerometer was congured to record at
100 Hz.
Accelerometer data preparation. A previously published method was used to minimize sensor calibra-
tion error
13
and to detect and impute accelerometer non-wear periods
2,14
. Arm angle was estimated as follows:
π=
+
⋅
−
angletan
a
aa
180/,
z
z
xy
1
22
where a
x
, a
y
, and a
z
are the median values of the three orthogonally positioned raw acceleration sensors in gravi-
tational (g) units (1 g = 1000 mg) derived based on a rolling ve second time window. Here, the z-axis corresponds
to the axis positioned perpendicular to the skin surface (dorsal-ventral direction when the wrist is in the anatom-
ical position). Next, estimated arm angles were averaged per 5 second epoch and used as input for our algorithms
for detecting sleep period time (SPT-window) and sleep episodes.
www.nature.com/scientificreports/
3
SCieNTifiC REPORTS | (2018) 8:12975 | DOI:10.1038/s41598-018-31266-z
Heuristic algorithm to detect the SPT-window. ere are several challenges in the development of an algorithm
to detect the SPT-window: absence of hard data labels to train a classier under daily life conditions (not in a
clinic), consideration of daily life behaviour, e.g. how to handle sleep scattered across the full 24-hour day and
ensure that the algorithm is not over tted to a specic population or accelerometer brand. us an algorithm was
developed by visually inspecting twenty random accelerometer multi-day recordings from dierent studies and
accelerometer brands (ten from the Whitehall II Study as reported in this paper and ten from UK Biobank study
1
)
while iteratively enhancing the algorithm to best detect the visible data segment of no movement without using
or looking at sleep diary data.
e resulting heuristic algorithm, which we will refer to as Heuristic algorithm looking at Distribution of
Change in Z-Angle (HDCZA), applied per participant is illustrated in Fig.1 and works as follows. Step 1–2:
Calculate the z-angle per 5 seconds. Steps 3–5: Calculate a 5-minute rolling median of the absolute dierences
between successive 5 second averages of the z-angle. ese rst ve steps make the algorithm invariant to the
potentially unstandardized orientation of the accelerometer relative to the wrist and aggregate it as the roll-
ing variance over time. Step 6–7: Calculate the 10
th
percentile from the output of step 5 over an individual day
(noon-noon), and multiply by 15. is is used as a critical individual night derived threshold to distinguish
periods of time involving many and few posture changes. Detect the observation blocks for which the output
from step 5 was below the critical threshold, and keep the ones lasting longer than 30 minutes. Step 8: Evaluate the
length of the time gaps between the observation blocks identied by step 7, if the duration is less than 60 minutes
then count these gaps towards the identied blocks. Step 9: e longest block in the day (noon-noon) will be the
main SPT-window, dened as the time elapsed between sleep onset (start of the block) and waking time (end of
the block). ese last four steps reect assumptions from us as researcher about the nature of sleep.
Our motivation for the design of the algorithm is as follows. By visually inspecting the angle-z values over
a day some individuals seemed inactive or sleeping throughout the day with minimal variation in angle, while
other individuals had more distinct inactive (night time) and active (daytime) periods. ese dierences pre-
sumably reect the degree of sedentary lifestyle and amount of sleep in a day. Using a percentile as part of the
threshold calculation allows the threshold to account for between-individual dierences in z-angle distribution.
e factor 15 in step 6 of the algorithm was derived iteratively using visual inspection of the classication. e
30-minute time period is motivated by the assumption that people are typically not in bed for less than 30 minutes
for their nocturnal time in bed, as opposed to daytime napping, and the 60-minute time period is motivated by
the assumption that sleep separated by awake periods greater than 60 minutes ought to be treated as two distinct
sleep episodes to avoid adding early evening naps or aernoon naps to the SPT-window. A sensitivity analysis on
HDCZA parameter settings and their inuence on algorithm performance across the datasets can be found in
Supplementary information (page 81).
Second algorithm for reference. When comparing our algorithm to the sleep diary we also considered a second,
but more naïve heuristic algorithm, which we will refer to as L5 ± 6. e algorithm is based on the raw signal
metric Euclidian Norm (vector magnitude) Minus One with negative values rounded to zero (ENMO), which in
formula corresponds to
++ −
{}
()
maxacc accacc 1,0,
xyz
222
with acc
x
, acc
y
, and acc
z
referring to the three orthogonal acceleration axes pointing in the lateral, distal, and ven-
tral directions, respectively
14
. Metric ENMO has previously been demonstrated to be correlated with magnitude
of acceleration as well as human energy expenditure in the present generation of wearable acceleration sensors
14
.
L5 ± 6 takes the 12 hour window centred around L5 (least active ve hours in the day based on metric ENMO)
and then searches within this window for sustained inactivity periods which were previously described
11
. In short,
sustained inactivity periods are calculated as the absence of change in arm elevation angle (same angle-z as used
above) larger than 5 degrees for more than 5 minutes
11
. Next, the SPT-window is dened from the start of the rst
to the end of the last occurrence of a sustained period of inactivity in the 12-hour window.
Sleep episodes within the SPT-window. Sleep episodes were dened as the sustained periods of inactivity within
the SPT-window, as dened in the previous section
11
. From this, the number of sleep episodes within each
Figure 1. Steps of the heuristic algorithm HDCZA for SPT-window detection.
www.nature.com/scientificreports/
4
SCieNTifiC REPORTS | (2018) 8:12975 | DOI:10.1038/s41598-018-31266-z
SPT-window detected (HDCZA, L5 ± 6) was calculated as well as sleep eciency within the SPT-window calcu-
lated as the percentage of time asleep within the SPT-window
11
.
Statistical analysis. Comparison with sleep diary. e SPT-window derived from both the HDCZA and
L5 ± 6 were compared separately with sleep diary records with a multi-level regression to account for the vari-
ation in availability of night time data and to include both night and person level predictors. For SPT-window
duration (dierence between sleep onset and waking time), sleep onset and waking time, the dierence between
diary and accelerometer-based detection was used as the dependent variable, while population demographics
(sex, age, BMI), season (winter or summer) and weekend versus weekday were used as predictors. Here, we used
function lme from R package nlme. Further, correlation coecients with diary valuesand mean absolute error
(MAE) for sleep onset, waking time, and SPT-window duration were calculated. Additionally, the c-statistic,
also known as the Area Under the Curve (ROC), was calculated from the epoch-level binary classications of
theSPT-window <1> or not <0> by diary and the HDCZA and L5 ± 6. e c-statistic, was rst calculated per
day and then aggregated as average per participant. Additionally, to investigate whether more wakefulness time
within the SPT-window corresponds to a larger HDCZA-sleep diary dierence in SPT-window duration we cal-
culated the amount of wakefulness categorised as [0-1), [1-2), [2-3), [3-4), and at least 4 hours, and compared this
with the dierence in SPT-window duration between sleep diary and the HDCZA. e notation [a-b) is used to
denote an interval that is inclusive of ‘a’ but exclusive of ‘b’.
Evaluation with polysomnography. e recording time of PSG is typically constrained to the time in bed win-
dow, which means that our heuristic algorithm (HDCZA) may not detect sucient data corresponding to time
out of bed to derive its critical threshold and accurately detect the SPT-window. We addressed this concern by
adding simulated wakefulness data to the beginning and ending of the accelerometer and PSG recording. e
PSG and accelerometer data were expanded with 90 minutes of simulated data at the beginning and ending that
would not trigger the SPT-window detection: simply the class wakefulness for PSG, and a sine wave with ampli-
tude 40 degrees and period 15 minutes complemented with random numbers (mean = 0, standard deviation = 10)
for accelerometer-based angle-z. Note that the specic shape of the simulated values is not critical as long as it
does not trigger the detection of sleep and the 10
th
percentile of all the data (step 6 of HDCZA) reects real and
not simulated data. e addition of simulated data is needed because the heuristic detection algorithm eectively
searches for the beginning and end of a large time period without body movement, if the full PSG represents sleep
then the algorithm would not be able to detect such a transition in movement level. Additionally, the algorithm’s
threshold that scales with the variance in the data was constrained to a range corresponding to the 2.5
th
and 97.5
th
percentile of the distribution of the threshold value observed in a sample of daily life accelerometer recordings,
0.13 and 0.50, respectively. is was done because the in-clinic PSG does not provide a full 24-hour cycle of body
movement to derive this threshold. In the PSG evaluation we did not evaluate L5 ± 6, because it requires more
than 12 hours of (non-simulated) data, which most PSG recordings do not oer. Aer sleep classication with
HDCZA and before running the comparison between HDCZA and PSG, 60 minutes of simulated data were
removed at the beginning and end.
e following performance metrics for SPT-window detection were used: dierence in onset, waking time,
and duration, accuracy, c-statistic, t-test, and mean absolute error (MAE). Performance estimates accuracy and
c-statistic were derived from both the data, as well as from the data expanded with wakefulness time to simu-
late performance estimates in a 24 hour recording. Sleep classication within the SPT-window was evaluated as
dierence in duration (t-test) and as the percentage of time spent in sleep stages REM, and non-REM stages 1,
2, and 3 (N1, N2, and N3) correctly classied by the algorithm as part of SPT-window. Sleep eciency within
the SPT-window by PSG and algorithm was compared via t-test and MAE. A P-value of <0.005 was considered
signicant
15
. Further, method agreement was evaluated with modied Bland-Altman plots
16
with PSG criterion
values on the horizontal axis.
Considering the relatively small sample size in our PSG analysis, we also report the minimal detectable dier-
ence between estimated and PSG reference values given the sample size, observed standard deviation, observed
correlation, a required signicance level of 0.005, and a required power of 0.80 using R package pwr and the
algorithm for power calculation for paired t-tests as described by Cohen
17
.
Code availability. Both SPT-window detection algorithms are implemented and available in open source
R package GGIR version 1.5-23 (https://cran.r-project.org/web/packages/GGIR/)
18
, see the software’s doc-
umentation on input arguments ‘loglocation’ and ‘def.noc.sleep’ for further details on the use of L5 ± 6 and
HDCZA. e R code used for our comparisons with sleep diary can be found at: https://github.com/wadpac/
whitehall-acc-spt-detection-eval. e R code used for our comparisons with polysomnography can be found at:
https://github.com/wadpac/psg-ncl-acc-spt-detection-eval, with the code used for the Newcastle data in the mas-
ter branch of the repository and its adaptation for the dierently formatted Pennsylvanian data in the psg-penn
branch.
Results
Comparison between accelerometer results and that from sleep diary. Demographic character-
istics of the three study cohorts are described in Table1. e probability density distribution for the dierence
between sleep parameter estimates from algorithm and sleep diary is more symmetrical around zero compared
with the L5 ± 6 approach, see Fig.2. e heuristic algorithm HDCZA estimates sleep onset on average 12.5 and
7.5 minutes earlier than that reported in the sleep diaries by men and women, respectively, 3.9 minutes per ten
years of age relative to mean age, and 3.0 minutes for a weekend day, see Table2. Dierence between sleep diary
estimates and HDCZA estimates in waking time and SPT window duration were associated with sex, age, and
www.nature.com/scientificreports/
5
SCieNTifiC REPORTS | (2018) 8:12975 | DOI:10.1038/s41598-018-31266-z
BMI, see Table2. e L5 ± 6 method estimates sleep onset on average 86.4 and 78.5 minutes earlier than that
reported in the sleep diary for men and women, respectively. Dierence between sleep diary and L5 ± 6 estimates
of SPT-window, sleep onset, and waking time were associated with sex and BMI, but inconsistently with weekday,
see Table2. e Pearson’s correlation coecients and c-statistics between accelerometer derived sleep parameters,
and sleep diary, are higher for HDCZA compared with L5 ± 6, see Table3. e combined MAE from onset and
waking time was 34.8 and 75.6 minutes for HDCZA and L5 ± 6, respectively.
For nights with [0-1), [1-2), [2-3), [3-4), and at least 4 hours of accumulated wakefulness an average dierence
in SPT-window duration between sleep diary records and our heuristic algorithm (HDCZA) was observed as
27, 3, −58, −154, and −236 minutes corresponding to 57.9, 32.1, 7.5, 1.6, and 0.7% of 25,645 recorded nights,
respectively. Here, the last two categories, corresponding to at least 3 hours of accumulated wakefulness, reect
8.5% of the participants.
Comparison between accelerometer results and that from polysomnography. In the PSG study
in sleep clinic patients, on average 9.4 (standard deviation 1.6) hours of matching data from PSG and accelerome-
ter were retrieved per participant, with no dierence in recording duration between le and right wrist (P = 0.75).
Sleep onset time, waking time, SPT-window duration, and sleep duration within the SPT-window derived from
the HDCZA algorithm diered all non-signicantly from polysomnography and MAE ranged from 31 minutes
for sleep onset to 71 minutes for SPT-window duration, see Table4. e combined MAE from onset and waking
time was 38.9 and 36.7 minutes for the le and right wrist, respectively. SPT-window duration was estimated for
the le wrist within 2 hours for the majority of individuals (75%) but deviated by more than 2 hours in seven
individuals, six of which had a sleep disorder, as shown in Fig.3 (right wrist: 81%, ve, and four, respectively). On
average, the accuracy and C-statistic for SPT-window classication were 87% and 0.86 in the PSG recording win-
dow, and 94% and 0.94 when expanded with simulated wakefulness as an estimate of 24 hour performance, see
Table4. Further, the average sensitivity to detect sleep as part of the SPT-window was above 91% in both wrists,
see Table4. Results for the PSG study carried out in healthy good sleepers indicated better overall performance as
shown in Table5 and Fig.4. e classications of the HDCZA algorithm in comparison with the PSG sleep stage
classication for all participants are provided in the Supplementary information chapter 1 (page 2) and chapter 2
(page 58) to this manuscript.
e minimal detectable dierence in sleep parameters (le wrist) for sleep clinic patients was 18, 32, 47, and
16 minutes for, respectively, sleep onset time, waking time, SPT window duration, and sleep duration within SPT
(right wrist; 24, 32, 35, and 26 minutes, respectively). e minimal detectable dierence in sleep eciency was
4.4 and 7.3 percent point for the le and right wrist data, respectively. e minimal detectable dierence in the
evaluation with healthy good sleepers was 15, 6, 17, and 17 minutes for, respectively, sleep onset time, waking
time, SPT window duration, and sleep duration within SPT. e minimal detectable dierence in sleep eciency
was 2.9 percent point.
Discussion
In this paper we present a heuristic algorithm, referred to as HDCZA, for detecting Sleep Period Time-window
(SPT-window) from accelerometer data in the absence of a sleep diary. Raw data accelerometers are increasingly
used in population research, and the value of this algorithm lies in studies such as the UK Biobank where a sleep
diary was not used
1
. Although the focus of our analysis is sleep, the present ndings are equally valuable for
physical activity research as it will help to split the observation period between night sleep and daytime inactivity.
In our comparison with sleep diary records in a large cohort of older adults (60–82 years) a small systematic
dierence was found in sleep duration and sleep onset time, dierence that varies slightly as a function of sex,
age, and BMI. Here, the average dierence and the Akaike Information Coecients indicated that the algorithm
is better than our naïve reference method L5 ± 6. Furthermore, the C-statistic was on average 95% for HDCZA.
We acknowledge that the sleep diary cannot be considered a gold standard criterion method, but it is reassuring
to see that dierences between algorithm and sleep diary in a large cohort of elderly individuals are on average
within a quarter of an hour.
An important limitation of the sleep diary study data is that no information is available on daytime sleep or
daytime inactivity behaviour to help better understand the misclassications in SPT-window by our algorithm.
To facilitate such research, future methodological studies are warranted to consider implementing daytime sleep
diaries, and possibly additional sensor technologies such as wearable cameras
19
, RFID proximity sensors
20
or
additional wearable movement sensors to better capture a lying posture
21,22
. In addition, impact of handedness on
the estimates could not be assessed.
Study Daily life (diary) PSG sleep clinic PSG healthy good sleepers
N 3752 28 22
Age (mean ± standard deviation in years) 69.1 ± 5.6 44.9 ± 14.9 22.8 ± 4.5
Sex
2822 males, 930
females
17 males and 11
females
7 males and 15 females
SPT-window duration (mean ± standard deviation) 7.7 ± 1.2 hours 8.4 ± 1.6 hours 6.7 ± 0.9 hours
Sleep onset time (mean in hh:mm ± standard deviation) 23:48 ± 71 minutes 22:32 ± 69 minutes 23:24 ± 54 minutes
Waking time (mean in hh:mm ± standard deviation) 7:28 ± 72 minutes 06:58 ± 76 minutes 06:09 ± 32 minutes
Table 1. Participant characteristics used for the analyses.