In recent decades, sleep and physical activity researchers have independently used the same technology (i.e., actigraphy and accelerometry) to objectively measure the specific behavior most relevant to their own research, effectively advancing the technology in parallel. On the physical activity side of this technological advancement, monitoring protocols that originally focused on waking hours only presented analytical issues related to children’s relative compliance to wear time requirements and ultimately how to define a “valid day” of data collection (12). These operational definitions were also crucial for the accurate measurement of sedentary behavior, which is particularly sensitive to reduced wear time (8,17). Recently, some physical activity researchers have adopted a 24-h wear protocol with the intention to increase wear time. This advancement created the opportunity to bridge sleep and physical activity behaviors previously isolated by study design. Despite the apparent advantages of studying these behaviors in tandem, this merger is not without challenges. Case in point, researchers now collecting 24-h accelerometer data are confronted with the measurement conundrum that time spent in the nocturnal sleep period must first be identified and accounted for before physical activity and sedentary behavior accumulated during the awake portion of the day can be determined (16).
Wrist-worn accelerometers have been the standard in children’s sleep research (9), whereas waist-worn accelerometers are more common in physical activity and sedentary behavior research (1,6). Physical activity researchers are also now using wrist-worn triaxial accelerometers to measure physical activity and sedentary behavior in children with relative success (3,10,11). Conversely, sleep researchers are also experimenting with using waist-worn accelerometers to measure sleep. For example, Kinder et al. (7) applied the Sadeh algorithm (13) (a sleep scoring algorithm) to waist-worn accelerometer data collected from free-living children with acceptable results. However, their process still required children’s or their parents’ manual data entry of reported bed/wake times. This requirement is likely a challenging cognitive task for younger children and a laborious effort for researchers faced with replacing missing data by visually inspecting and interpreting individual raw data files without any known standardized rules to guide interpretation. The development of a fully automated algorithm to identify and describe nocturnal sleep-related patterns with waist-worn accelerometers is a necessary step to merging physical activity and sleep research (16).
Embarking on a large multinational epidemiological study, the International Study of Childhood Obesity, Lifestyle, and the Environment (ISCOLE) (5) implemented a 24-h wear protocol with a waist-worn ActiGraph GT3X+ (ActiGraph LLC, Pensacola, FL) accelerometer. We initially developed a fully automated algorithm (16) to detect children’s sleep period time (SPT), defined as the time difference between sleep onset and sleep offset, separate from waking physical activity or sedentary behavior (16). Relative to a criterion of expert visual inspection of raw data, the algorithm produced similar estimates and was within published values of SPT for children. However, limitations of the original study include the focus on a single selected 24-h time block and the application of no external validation of SPT outside of expert visual data inspection. It also became apparent with its subsequent application to larger data sets containing more variable sleep-related behaviors that our initial attempt did not account for the possibility of nocturnal nonwear or extended episodes of wakefulness separating the SPT into multiple sleep episodes. Put in another way, our original algorithm only presented a monophasic SPT without allowing for the possibility of biphasic or multiphasic sleep episodes. In addition, we did not fully consider the potential for misclassifying daytime periods of nonwear or other sedentary behaviors as sleep episodes (i.e., “naps”). Therefore, the purposes of this study were 1) to add layers and features to a previously published fully automated algorithm (16) to exclude extended episodes of night time nonwear/wakefulness and potentially misclassified daytime sleep episodes and 2) to validate the refined sleep algorithm (RSA) against children’s sleep logs.
METHODS
Participants
Forty-five fourth-grade school children (22 boys and 23 girls) from a single school participating in the Baton Rouge, USA, site of ISCOLE (5) were invited to take part in this ancillary study. Written informed consent from a parent/guardian and written assent from each child were obtained prior to any data collection. The study protocol and procedures were approved by the Pennington Biomedical Research Center’s Institutional Review Board.
Instrumentation
The ActiGraph GT3X+ is a triaxial accelerometer capable of collecting raw acceleration data at sampling frequencies up to 100 Hz over a dynamic range of ±6g. Device data were downloaded using ActiLife software version 6.5.2 (ActiGraph LLC). Activity counts were recorded for each axis, and vector magnitude was calculated combining data from all three axes (
;)
). Steps and inclinometer position (“off,” “lying,” “sitting,” or “standing”) were also available for each epoch (selected time interval) of output. ActiGraph GT3X+ acceleration data can be processed using the default filter or the low frequency extension (LFE) filter, which increases sensitivity of detecting accelerations, especially low force acceleration (e.g., slow walking and other types of slow and/or low force movement events). As suggested by Hjorth et al. (4), ActiGraph sleep-related data were processed using the LFE filter.
In addition to wearing the accelerometer, participants were asked to write down the time they went to bed (bedtime) each night and the time they got out of bed (wake time) each morning in a simple sleep log. Children were asked to record only these two time points; they were not required to record details about any night wakening episodes or daytime naps.
Procedures
Each participant’s height and body mass were measured using a portable stadiometer (Seca 213; Seca GmbH & Co. KG., Hamburg, Germany) and a digital scale (SC-240; Tanita Corporation, Tokyo, Japan), respectively. Body mass index (BMI) was calculated as kilograms per meter squared. Each participant’s age was queried from their parent/guardian.
Accelerometer monitoring procedures were identical to those described in our previous studies (5,16). The participants in this study represent a subsample of the study described by Katzmarzyk et al. (5). Briefly, participants were asked to wear the accelerometer 24 h·d−1 for seven consecutive days, except when bathing or during water-based activities. Data were collected at a frequency of 80 Hz and subsequently downloaded and integrated into 60-s epochs with the LFE filter enabled using the ActiLife Software.
Participants were asked to complete the sleep log for the entire 7-d monitoring protocol. Verbal instructions were given daily in the classroom by study staff to remind children to properly record their bedtime and wake time (to the nearest minute) for each night of sleep.
Data Treatment
Sleep log.
The sleep log was used to obtain self-reported bedtime and wake time. Log SPT was calculated as the elapsed time between children’s self-reported bedtime and wake time.
Sleep log + accelerometry.
Traditional sleep actigraphy in free-living settings uses an approach that combines objectively monitored data with self-reported bedtime and wake time (1). Logged bedtimes and wake times were used to focus the accelerometer monitoring frame in order to establish an objective marker of sleep onset (accelerometer-defined time to sleep). Specifically, accelerometry data were processed using the ActiLife software that uses the Sadeh algorithm (13) to determine a probability-based sleep score for each minute (i.e., each minute is scored as “sleep” or “wake”) between the logged bedtime and wake time. Sleep onset was then identified as the first minute scored as sleep following the reported bedtime (Larry Hawk, ActiGraph, personal communication). Log + Accel SPT was then computed by determining the elapsed time between sleep onset and the logged wake time.
RSA.
The RSA was built upon our previous algorithm (16) that fully automated the identification of accelerometer-determined markers for sleep onset and sleep offset, but it provided no allowance for the possibilities of multiple sleep onsets and offsets. The algorithm was implemented in SAS (version 9.3; SAS Institute Inc., Cary, NC). This was done by first applying the Sadeh algorithm (13) and the inclinometer function from the accelerometer to identify the probability of sleep for each minute. A number of data-driven standards were created to provide the most accurate results when compared to the other sleep measurement tools. Sleep onset was defined as the first of five consecutive minutes scored as sleep from the Sadeh + inclinometer algorithm (SIA). Sleep offset was defined as the first of 10 consecutive minutes of SIA-scored wake minutes. In addition, SPT from the original algorithm (16) was only identified if ≥160 min elapsed between sleep onset and sleep offset.
A number of changes were made to eliminate the shortfalls (catalogued in the introduction) of the previously published algorithm in order to create the RSA. First, to only focus on nocturnal sleep, sleep onset could only occur between 7:00 p.m. and 5:59 a.m. Second, the definition of sleep offset was refined and identified as the first of 10 or 20 consecutive SIA-scored wake minutes, depending on the time of day (10 min—5:00 a.m. to 11:58 a.m.; 20 min—9:40 p.m. to 4:59 a.m.). Although SPT from the original publication was described as a single block of time bordered by a single sleep onset and offset, the RSA refinements allowed identification of extended episodes of wakefulness separating the SPT into distinct sleep episodes with multiple sleep onsets and offsets. If two sleep episodes were separated by less than 20 min of SIA-scored wake minutes, then they were combined into a single sleep episode starting with the first minute of the first sleep episode and ending with the final minute of the last sleep episode. Sleep episodes that were separated by at least 20 min of SIA-scored wake minutes were distinct within the SPT and were not combined. Total sleep episode time (TSET) represented the total minutes from all sleep episodes; in cases where there was a single sleep episode, or all sleep episodes were separated by less than 20 min of SIA-scored wake, the TSET was equal to RSA SPT.
Noncompliance to accelerometer wearing regimens remains an important threat to validity in free-living populations; therefore, it is important to correctly classify data patterns indicative of nonwear. To eliminate misclassification of a nonwear period as sleep, a second algorithm was used to identify periods of nonwear within a previously defined sleep episode. Nonwear was identified when 90 consecutive minutes of 0 activity counts were encountered while allowing for up to 2 min of nonzero activity counts. The nonwear period ended when a third minute of nonzero activity counts was detected. If at least 90% of a sleep episode was categorized as nonwear, then all minutes within that sleep episode were redefined as nonwear and not included in the calculation of TSET.
Analytic sample.
For the analyses conducted herein, we excluded data from those participants who failed to return their accelerometer (n = 2) or whose accelerometer malfunctioned (n = 1). We also excluded those who failed to return a completed sleep log (n = 5) or whose logs were illegible (n = 3). In total, the analytic sample was composed of 34 participants.
Statistical analysis.
Participants’ data were averaged across days, providing just one value per participant for each of the variables described above. In addition, average daily data were also presented by the day of the week (representing the day that the participants went to bed). Similar to our previous publication (16), a number of statistical procedures were used to compare the results obtained by the different methods. Pearson product–moment correlations were calculated to assess the magnitude of associations, and previously published standards (14) were used classify the associations. In addition, mean differences, mean absolute differences, and mean absolute percent differences were calculated to determine differences between methods. Paired t-tests were used to compare mean Log SPT and mean Log + Accel SPT to mean RSA TSET. Cohen’s d effect sizes (2) were also computed to provide a context for the magnitude of differences between the variables obtained with the different methods. In addition, Bland–Altman plots were prepared for both pairings. With the exception of mean absolute difference and mean absolute percent difference, the same calculations were conducted to compare log wake time and bedtime to RSA sleep onset and offset, as well as Log + Accel sleep onset to RSA sleep onset. All statistical analyses were conducted using IBM SPSS Statistics (version 20.0; SPSS, Chicago, IL), and the level of significance was defined as P < 0.05.
RESULTS
Descriptive characteristics of the analytical sample are presented in Table 1. Only two children had one night each with more than one sleep episode, which made the overall RSA TSET approximately the same as RSA SPT. Daily means and overall means of SPT for Log and Log + Accel as well as the RSA TSET are presented in Table 2. The correlation between Log SPT and the RSA TSET was 0.32 (P = 0.07) and between the Log + Accel SPT and RSA TSET was 0.34 (P = 0.05).
TABLE 1: Descriptive characteristics of participants.
TABLE 2: Mean SPT and TSET calculated by different methods.
Daily and overall mean comparisons between the different methods are presented in Table 3. When compared to the Log SPT, the RSA TSET had a mean absolute percent difference of 6.0% ± 4.2% and a significant mean difference of −20 ± 36 min (t(33) = −3.2, P = 0.003; Cohen’s d = 0.65) . All daily mean absolute percent differences were less than 10%. The Bland–Altman plot for this pairing is presented in Figure 1A. Only three data points were outside the acceptable range of ±1.96 SD from the mean. When compared to the Log + Accel SPT, the RSA TSET had a mean absolute percent difference of 5.1% ± 4.0% and a nonsignificant mean difference of −9 ± 36 min (t(33) = −1.5, P = 0.15; Cohen’s d = 0.29). All daily mean absolute percent differences were less than 10%. The Bland–Altman plot for this pairing is presented in Figure 1B.
TABLE 3: Difference between estimated SPT and TSET calculated by different methods.
FIGURE 1: A, Bland–Altman plot of Log and RSA sleep estimates. B, Bland–Altman plot of Log + Accel and RSA sleep estimates.
Detailed information on wake times and bedtimes as well as sleep onset and sleep offset times is presented in Table 4. Log bedtime and RSA sleep onset correlations were moderately high (r = 0.71), and Log + Accel sleep onset was also moderately high with RSA sleep onset (r = 0.74). The correlation between log wake time and RSA sleep offset was 0.61. The differences between the Log + Accel and the RSA sleep onsets (t(33) = 1.5, P = 0.15) and between the log wake time and the RSA sleep offset (t(33) = −1.5, P = 0.16) were not significant.
TABLE 4: Comparisons of bedtime versus sleep onset and wake time versus sleep offset from the Log, Log + Accel, and RSA methods.
DISCUSSION
We were able to refine (i.e., add layers and features) a previously published fully automated algorithm for detecting children’s sleep (16) to now allow for determination of disrupted nocturnal sleep episodes and avoid misclassification of daytime nonwear or sedentary behavior as sleep. We also achieved acceptable levels of accuracy (<10% mean absolute percent difference) when the RSA values were compared to congruent Log and Log + Accel parameters. Similar to information obtained from a sleep log, the original algorithm (16) treated SPT as a single large block of time. This approach ignored the potential for interrupting the sleep episode with extended episodes of wakefulness during the night, resulting in them being misclassified as continuous sleep rather than more correctly classifying them as wakeful physical activity behaviors of various intensities. Although only two children had detectable patterns of nighttime wakefulness in this small select sample, the present refinement provides a distinct advantage when compared to results from sleep logs and also provides a unique opportunity to study disrupted sleep and other sleep-related behaviors in more detail in future studies of larger samples. This refinement also correctly reclassifies movement detected during nocturnal waking episodes to physical activity/sedentary behavior-focused analyses and permits an unobstructed focus on nocturnal sleep patterns, eliminating the potential for misclassified daytime sleep episodes (i.e., nonwear or other sedentary behaviors misclassified as naps) from the final TSET. Finally, the last refinement was implemented to prevent misclassification of nocturnal nonwear as sleep.
Although a moderately high correlation (r = 0.74) between Log bedtime and RSA sleep onset time was found, the magnitude of the mean difference (approximately 20 min) was statistically significant. However, this difference was anticipated. Children were asked to record an approximate time that they went to bed, and it is possible that they remained awake and moved around for some time thereafter. In contrast, the RSA is programmed to interpret sleep onset from a definite cessation of movement pattern observed within a specified time period. As expected, in most cases (i.e., 80%), the RSA detected a later time for sleep onset than what the children recorded as their Log bedtime. This delay between self-reported bedtime and RSA-determined sleep onset was confirmed with the calculation of sleep onset using the Log + Accel method. Because this latter method calculated sleep onset after defining sleep latency, there ultimately were no significant differences observed between Log + Accel and RSA estimates of sleep onset time. The significant difference observed between Log bedtime and RSA sleep onset could also be attributed to possible recording errors in the sleep log information provided by the children (see an example in Fig. 2). Using the RSA, sleep onset was estimated to be at 9:56 p.m., while the Log bedtime was at 8:25 p.m. and the sleep onset using Log + Accel was 8:58 p.m. The figure clearly demonstrates that the child was not sleeping (inferred from an obvious movement pattern) at the time indicated in the sleep log, and therefore, RSA should be considered a more reliable estimate of sleep onset. Not only were they not sleeping but the figure also shows that the child accumulated a few minutes at >20 steps per minute after their recorded Log bedtime, which is a clear indication that the child was not yet in bed.
FIGURE 2: Example of a single child’s self-recorded and accelerometer determined sleep-related behaviors.
A moderate correlation (r = 0.61) was apparent between log wake time and RSA sleep offset time. Additionally, there were no significant mean differences in these clock times. The mean log wake time of 6:44 a.m. (±29 min) was only 6 min later than the RSA sleep offset time, which again was somewhat expected as the children logically may have initiated detectable movements and then delayed a few minutes after waking up before actually recording their wake time as the time they left their bed. Based on the RSA TSET, children slept an average of 9 h, which is close to the same amount of time we reported in our previous study (16) (i.e., 9.2 h) and is within expected values for children of this age (15).
The RSA can be implemented by researchers interested in capturing physical activity, sedentary behavior, and also nocturnal sleep information using a 24-h waist-worn accelerometer protocol. The methodology provided here (SAS syntax available at http://www.pbrc.edu/pdf/PBRCSleepEpisodeTimeMacroCode.pdf) in combination with previously published information (16) can be used to determine nocturnal sleep onset and offset times, TSET, wakefulness episodes during the night, and nocturnal nonwear periods. Although this study serves as a basis for the unification of physical activity/sedentary behavior and sleep research, additional research is needed to continue to refine best practices for analysis of these sleep-related data generated from 24-h waist-worn accelerometer protocols.
Although this study was carefully conducted, it is not free from limitations. The RSA-derived variables were compared to information obtained from Log and Log + Accel methods and it was found that data were comparable. However, these reference methods are not considered gold standards for sleep assessment. Instead, these methods can be considered to be epidemiological standards, appropriate for field-based data collection. Regardless, future studies should compare the results obtained with the RSA (which makes inferences about sleep/wake based on accelerometry-detectable movement/nonmovement patterns) to clinical measures like polysomnography (which incorporates electroencephalography, electrooculography, electromyography, and electrocardiography). We did not make any attempts to distinguish naps from sedentary behavior during daytime. We intentionally constructed the algorithm to ignore daytime behaviors that might otherwise be classified as sleep to avoid misclassification of nonwear or time spent in other types of sedentary behavior. It is problematic to distinguish naps from other sedentary behaviors because there is no criterion measure in free-living setting. This means that the RSA is not appropriate for researchers interested in studying daytime napping behaviors. We focused on determining clock times for nocturnal sleep onset and offset time and, except for detecting extended episodes of wakefulness, largely “ignored” all the detailed information available from accelerometry between these two time points. Additional refinements to this methodology could feasibly shift the analysis of these overlooked movement/nonmovement patterns beyond simply the definition of TSET based on detectable sleep onset and offset times. Data collection was conducted during a week when children were attending school and had a regular schedule. It is not possible to determine how the RSA will function with an irregular schedule, for example, when children are on summer breaks. It is also not possible to generalize these findings to other age groups.
In summary, we were able to refine a previously published fully automated algorithm for detecting children’s sleep (16), add to its utility, and ultimately achieve acceptable levels of accuracy for RSA-determined sleep parameters when compared to Log and Log + Accel parameters. The RSA presented in this study can be used by researchers who wish to use a 24-h wear protocol with waist-worn accelerometers to determine the clock-based anchors of nocturnal sleep onset and sleep offset, identify nocturnal patterns of sleep episode time and wakefulness, and effectively separate data for distinct analyses of physical activity/sedentary behavior within the same data set. The algorithm also sets the foundation for further exploration of the movement/nonmovement patterns contained within the nocturnal SPT, moving toward generating estimates of disrupted sleep and other sleep-related behaviors that will be useful in epidemiological analyses.
We acknowledge the contributions of the ISCOLE Coordinating Center at the Pennington Biomedical Research Center, Gina Pennington, Coordinator of the Baton Rouge site, and our local data collection team.
All authors are members of a larger study funded by the Coca-Cola Company.
The results of the present study do not constitute endorsement by the American College of Sports Medicine.
REFERENCES
1. Cain KL, Sallis JF, Conway TL, Van Dyck D, Calhoon L. Using accelerometers in youth physical activity studies: A review of methods.
J Phys Act Health. 2013; 10 (3): 437–50.
2. Cohen J.
Statistical power analysis for the behavioral sciences. 2nd ed. Hillsdale, N.J.: L. Erlbaum Associates; 1988.
3. Ekblom O, Nyberg G, Bak EE, Ekelund U, Marcus C. Validity and comparability of a wrist-worn accelerometer in children.
J Phys Act Health. 2012; 9 (3): 389–93.
4. Hjorth MF, Chaput J-P, Damsgaard CT, et al. Measure of sleep and physical activity by a single accelerometer: Can a waist-worn ActiGraph adequately measure sleep in children?
Sleep Biol Rhythms. 2012; 10 (4): 328–35.
5. Katzmarzyk PT, Barreira TV, Broyles ST, et al. The International Study of Childhood Obesity, Lifestyle and the Environment (ISCOLE): design and methods.
BMC Public Health. 2013; 13 (1): 900.
6. Kim Y, Beets MW, Welk GJ. Everything you wanted to know about selecting the “right” ActiGraph accelerometer cut-points for youth, but …: a systematic review.
J Sci Med Sport. 2012; 15 (4): 311–21.
7. Kinder JR, Lee KA, Thompson H, Hicks K, Topp K, Madsen KA. Validation of a hip-worn accelerometer in measuring sleep time in children.
J Pediatr Nurs. 2012; 27 (2): 127–33.
8. Masse LC, Fuemmeler BF, Anderson CB, et al. Accelerometer data reduction: a comparison of four reduction algorithms on select outcome variables.
Med Sci Sports Exerc. 2005; 37 (Suppl 11): S544–54.
9. Meltzer LJ, Montgomery-Downs HE, Insana SP, Walsh CM. Use of actigraphy for assessment in pediatric sleep research.
Sleep Med Rev. 2012; 16 (5): 463–75.
10. Phillips LR, Parfitt G, Rowlands AV. Calibration of the GENEA accelerometer for assessment of physical activity intensity in children.
J Sci Med Sport. 2013; 16 (2): 124–8.
11. Rosenberger ME, Haskell WL, Albinali F, Mota S, Nawyn J, Intille S. Estimating activity and sedentary behavior from an accelerometer on the hip or wrist.
Med Sci Sports Exerc. 2013; 45 (5): 964–75.
12. Rowlands AV. Accelerometer assessment of physical activity in children: an update.
Pediatr Exerc Sci. 2007; 19 (3): 252–66.
13. Sadeh A, Sharkey KM, Carskadon MA. Activity-based sleep–wake identification: an empirical test of methodological issues.
Sleep. 1994; 17 (3): 201–7.
14. Safrit MJ, Wood TM.
Introduction to Measurement in Physical Education and Exercise Science. 3rd ed. St. Louis: Mosby; 1995. p. 71.
15. Scholle S, Beyer U, Bernhard M, et al. Normative values of polysomnographic parameters in childhood and adolescence: quantitative sleep parameters.
Sleep Med. 2011; 12 (6): 542–9.
16. Tudor-Locke C, Barreira TV, Schuna JM Jr.,, Mire EF, Katzmarzyk PT. Fully automated waist-worn accelerometer algorithm for detecting children’s sleep-period time separate from 24-h physical activity or sedentary behaviors.
Appl Physiol Nutr Metab. 2014; 39 (1): 53–7.
17. Winkler EA, Gardiner PA, Clark BK, Matthews CE, Owen N, Healy GN. Identifying sedentary time using automated estimates of accelerometer wear time.
Br J Sports Med. 2012; 46: 436–42.
