Harmonization of three different accelerometers to classify the 24 h activity cycle

Sleep, sedentary behavior (SED), light physical activity (LPA), and moderate to vigorous physical activity (MVPA) are four behaviors that can occur during a 24 h day. Integrating these behaviors into one model has been termed the 24 h activity cycle (24-HAC) (Rosenberger et al 2016, 2019). The 24-HAC behaviors are dynamic, especially among young adults, and can be influenced by individual, social and environmental factors (Rosenberger et al 2016, 2019).

The behaviors that make up the 24-HAC are independently associated with several important health outcomes (Chastin et al 2021). A large and growing body of evidence indicates that engaging in high SED time (>9.5 h d⁻¹) or having a short sleep duration (<7 h d⁻¹) is associated with negative health outcomes (e.g. obesity, diabetes, hypertension, and early mortality) independently of MVPA (Yin et al 2017, Jakicic et al 2019, Katzmarzyk et al 2019, Pescatello et al 2019, Ekelund et al 2020). In contrast, greater time spent in LPA and MVPA is moderately-to-strongly associated with many positive health outcomes including improved sleep quality and reduced risks of anxiety, depression, type 2 diabetes, and cardiovascular disease (Yin et al 2017, Jakicic et al 2019, Katzmarzyk et al 2019, Pescatello et al 2019).

Traditional approaches to measure the 24-HAC include self-report of the component behaviors (e.g. international physical activity questionnaire, Pittsburgh sleep quality index) and/or the use of a single wrist worn device (e.g. Actiwatch, GENEActiv, ActiGraph); however, two important limitations of these approaches should be considered. First, self-report measures of the 24-HAC components are subject to recall error or other biases that may lead to substantial measurement error (Haskell 2012). Second, a single wearable device is unlikely to capture the different 24-HAC components with equal precision. For example, while wrist-worn actigraphy is commonly used to assess free-living sleep behaviors, it may generate biased estimates of physical activity intensity and misclassify body postures compared to waist and/or thigh worn accelerometers (Lyden et al 2012, Ancoli-Israel et al 2015, Migueles et al 2017, Marcotte et al 2020). To overcome these challenges, integrating data from multiple wearable devices may generate more accurate estimates of the 24-HAC behaviors (Dooley et al 2023). Few studies have explored the advantages and disadvantages of using more than two accelerometers to estimate 24-HAC behaviors in free-living settings (De Craemer and Verbestel 2021, Micklesfield et al 2022, Dooley et al 2023, Pulsford et al 2023).

To address this gap, the present study aimed to: (a) evaluate the feasibility of using three research grade wearable monitors to measure key components of the 24-HAC, (b) develop a hierarchical algorithm to integrate data from the multiple measurements into minute-by-minute estimates of 24-HAC behavior and (c) use the resulting data to examine daily and weekly patterns of 24-HAC behaviors in a university student sample.

Participant recruitment and enrollment

Research design and procedures

24 h activity cycle measures

Data processing

For this analysis, all three data sources (ActiGraph GT3X+, activPAL3, and ActiGraph GT9X Link) were processed using the previously noted software packages and then combined in 1 min epochs. Single accelerometer studies have long used 1 min epochs for assessing physical activity, sedentary behavior, and sleep in healthy adults as reflected in many published algorithms (i.e. Freedson 1998, Troiano 2008, VANE, CREA Sadeh, Cole-Krepke). Each minute of monitored time was then classified as either sleep, SED, LPA, or MVPA using the approach detailed in figure 1. Briefly, non-sleep minutes were classified as (a) SED if the activPAL identified the participant as sitting for the entire minute or, absent activPAL data, the ActiGraph GT3X+ had a vector magnitude ≤150 counts per minute (cpm); (b) LPA if the activPAL registered upright posture for any portion of the minute or if the ActiGraph GT3X+ registered 151–2690 vector magnitude cpm; and (c) MVPA if the ActiGraph GT3X+ registered ≥2691 vector magnitude cpm. Supplemental figure 2 provides the SAS code used to operationalize the algorithm. Only days in which a behavior could be assigned for ≥1380 min using the combined data sources were retained for analysis as 'valid days'. Valid days with <1440 min of data, due to missing data or spring daylight savings, were standardized to a 1440 min day by multiplying each 24-HAC estimate by 1440/#valid minutes. Valid days with more than 1440 min of data, due to fall daylight savings, were standardized to a 1440 min day by multiplying each 24-HAC estimate by #valid minutes/1440.

Zoom In Zoom Out Reset image size

Statistical analysis

A total of 695 participants completed the study's inital elgibility survey, 123 participants consented to participate and attended the first study visit, after which three participants withdrew from the study due to skin irritation from device wear. The study's sample was 20.2 ± 1.6 yrs, primarily female (63.8%), primarily identified as white non-hispanic (64.8%), and primarily classifed as normal weight (23.5 ± 4.0 kg.m², 60.1%).

Table 1 details data losses and wear time compliance for the 120 participants who completed the device wear protocol, restricted to the first 6 complete days of wear (i.e. the first partial day of wear was excluded along with data beyond the 7th wear day). Approximately 2%–4% of days were lost due to device/battery failure (5 GT9X, 3 activPAL, 2 GT3X+). Among days with available data, average wear times indicated good compliance with the study protocol, as average wear approximated 23 h d⁻¹ for the GT9X and activPAL (which were to be worn continuously) and 16 h d⁻¹ for the GT3X+ (which was to be worn during waking hours only). When examined according to the proportion of valid measurement days (defined as 23+ h d⁻¹ for the GT9X and activPAL and 14+ h d⁻¹ for the GT3X+), wear compliance was highest for the activPAL (93.2% of total days with 80.3% of participants compliant for all 6 days) and lowest for the GT3X+ (70.7% of total days with only 30.5% of participants compliant for all 6 days).

To inform those considering the use of multiple measurement devices in future studies, table 1 also details the data losses and wear compliance that resulted when the analysis was restricted to different combinations of devices. As expected, data losses increased when non-missing data from multiple devices was required, with approximately 6%–8% of days lost for different two-device combinations and 10% of days lost when non-missing data from all three devices was required. In contrast, no days of data were lost when non-missing data from only 1 of 3 devices was required. Among the different two-device combinations, the GT9X/activPAL combination had the highest percentage of valid days (80%) with overlapping device wear, with 46.4% of participants compliant for all 6 d. As expected, requiring simultaneous wear of all three devices resulted in the lowest percentage of valid days (63.6%), with only 18.9% of participants compliant for all 6 d. When non-missing data from only 1 of 3 devices was required, less than 3% of days were invalid and 92.5% of participants had 6 valid days.

In order to make use of all available data and maximize the number of valid days, each minute during device wear was classified as sleep, SED, LPA, or MVPA using the previously described algorithm (figure 1). This resulted in a sample of 108 participants with at least 4 valid weekdays, and 2 valid weekend days, of 24 h movement behavior data.

On average, participants spent 75% of their day in either SED (41.6%, 599.2 min d⁻¹) or sleep behavior (33.7%, 485.7 min d⁻¹), less than a quarter of the day (21%, 299.2 min d⁻¹) in LPA, and the least amount of time in MVPA (4%, 55.8 min d⁻¹). As shown in table 2, weekdays were characterized by higher SED and MVPA, but lower LPA, compared to weekend days. Most estimates of weekday and weekend 24-HAC behaviors were similar across categories of sex, race/ethnicity, and BMI. Statistically significant exceptions were higher MVPA in males on both weekdays and weekends, lower LPA in males on weekends, and lower weekday MVPA in a sub-sample of 17 Asian students.

As depicted in figure 2, the amount of time spent in the different 24-HAC behaviors varied from day to day. Participants had the shortest total sleep duration on fridays (461.7 min) and the longest on sundays (511.6 min). SED time ranged from 544.7 min on saturdays to 635.8 min on tuesdays while LPA ranged from 278.1 min on thursdays to 361.2 min on saturdays. Finally, participants had the least amount of MVPA time on sundays (41.3 min) and the highest on wednesdays (63.7 min).

Zoom In Zoom Out Reset image size

Figure 3 shows the percentage of students engaged in each of the 24-HAC behaviors within an average weekday and weekend day. Sleep showed a clear nocturnal pattern with an approximate one hour delay in sleep initiation and a slightly longer delay in sleep cessation on weekends. A majority of students demonstrated a 'night owl' sleep pattern, with only half of students initiating sleep before 1 a.m. on weekdays and peak sleep engagement (over 90% of students) occurring from approximately 3–5 a.m. On weekdays, non-sleep behaviors between 10 a.m. and 5 p.m. exhibited a cyclic pattern, likely due to typical course schedules in an academic environment. During other times, non-sleep behaviors varied in an apparently random fashion across the day.

Zoom In Zoom Out Reset image size

To our knowledge, this is the first study to evaluate the feasibility and utility of using three distinct research-grade accelerometers, worn at different body locations, to measure 24-HAC behaviors under free-living conditions. We observed a high rate of participant compliance with wear instructions for all three accelerometers, although compliance was lower for the waist worn GT3X+ that was to be removed during sleep bouts. Despite high compliance for individual devices, requiring concurrent wear of multiple devices was found to result in a substantially higher proportion of invalid wear days. To avoid these potential data losses, we developed a hiarchical algorithm to categorize each minute of a 24 h day into one of the four components of the 24-HAC using available data from the research-grade accelerometers and supplementary resources (i.e. sleep logs). This process allowed for the prioritization of the most appropriate data source for a given parameter (e.g. activPAL for SED) while accounting for variations in data availability during the measurement period. This approach resulted in plausible estimates of 24-HAC behaviors both within and between days over a one-week period with minimal data losses.

An abundance of evidence indicates that the behaviors comprising the 24-HAC are interelated, underlining the need for accurate simultaneous measurement of these behaviors to provide a comprehensive understanding of how the 24-HAC influences human health (Ekelund et al 2020, Chastin et al 2021). While best practice guidelines exist for the objective measurement of physical activity, sedentary behavior, and sleep (Lyden et al 2012, Rosenberger et al 2013, Ancoli-Israel et al 2015, Edwardson et al 2017, Marcotte et al 2020), there is no consensus on the best approaches for measuring all components of the 24-HAC simultaneously in free living environments. The most established method to objectively measure free-living physical activity intensity is via a hip-mounted triaxial accelerometer, which boasts superior validity and reliability compared to wrist worn accelerometers (Rosenberger et al 2013, Marcotte et al 2020). However, accelerometers worn on the hip cannot accurately differentiate between SED and LPA due to poor discernment between standing and sitting postures. Thigh worn accelerometers, such as the activPAL, address this limitation by measuring both movement and postural changes—offering insights into transitions from sitting to standing and a more nuanced understanding of prolonged sedentary bouts (Lyden et al 2012). While polysomnography is the gold standard for assessing sleep, wrist-worn actigraphy in conjunction with self-report sleep logs is an established method for assessing sleep duration and quality in non-clinical settings (Ancoli-Israel et al 2015). In the current study, an innovative strategy was developed to harmonize the data from these devices which aimed to reduce misclassification of the component behaviors while maximizing data completeness. In the current study, participant compliance for wearing one device was good but, as expected, compliance for wearing all all three devices was much lower. Based upon this observation, it is necessary to develop approaches that prioritize input from multiple devices when available (and reconcile differences) but to also incorporate estimates from single devices when required in order to maximize data completeness.

Two recent studies have examined alternative approaches for harmonizing data from two different device combinations to classify free-living 24-HAC (Micklesfield et al 2022, Dooley et al 2023). In a study by Dooley et al (2023), a cohort of 1285 community-dwelling older women concurrently wore the ActiGraph wGT3X+ on the waist during waking hours and the Actiwatch 2 continuously on the wrist for up to seven days as part of the Study Of Women's Health Across the Nation (SWAN). This study first developed a two-tier system to differentiate waking and non-waking intervals. The first tier utilized sleep data from the Actiwatch, while the second tier deployed a one-dimensional convolutional neural network when there was missing sleep data. Once wake intervals were determined, Evenson cut points (Evenson et al 2015) were used to classify SED, LPA, and MVPA from the ActiGraph wGT3X+ vector magnitude counts. In contrast, Micklesfield et al (2022) used a different approach to derive 24-HAC estimates from hip-worn Actigraph GT3X+ and thigh-worn activPAL data in 794 middle-age South-African men and women. They first identified sleep timing using a decision tree based on raw acceleration data signals from the activPAL combined with sleep log information. The remaining components of the 24-HAC were then identified based on different raw acceleration (Euclidean norm minus one) signal thresholds. While both studies are commendable for their unique strengths, their harmonization techniques are complex and might not be feasible for some investigators. The current study developed a novel harmonization process for three devices, including a decision tree to identify total sleep duration, and a five-step process for assigning each minute outside of sleep to a 24-HAC category. In addition to its relative simplicity, the proposed approach could be readily adapted to studies using different device combinations and has the advantage of miminizing data losses by leveraging information from all available devices during the assesment period.

Previous investigatons using two or more device combinations to examine the 24-HAC in university students are currently limited (Kahlhöfer et al 2016, Hargens et al 2021). To date, Kahlohfer et al (2016) is the sole study that used two research grade accelerometers in a sample of 132 German university students. Participants wore an Actiwatch on the wrist to assess total sleep duration without a sleep diary and an activPAL on the thigh to measure sedentary time, standing time, walking time, and daily steps. The study found that their university student sample slept approximately 400 min d⁻¹, engaged in sedentary behaviors for 17.9 h d⁻¹, spent about 5.8 h d⁻¹ in different upright postures (combined standing and walking time), and took around 9610 steps d⁻¹. Estimates of LPA and MVPA were not reported. A study by Hargens et al (2021) used a single device worn in two body locations to assess physical activity, SED and sleep in a sample of 81 U.S. university students. During waking hours participants were instructed to wear an ActiGraph GT3X on their hip to capture habitual physical activity and sedentary behaviors while, during periods of sleep, participants were instructed to move the device from the hip to the wrist with a provided wrist strap. The study reported that, on average, students slept 388.7 ± 59.9 min d⁻¹ and engaged in 660.4 ± 11.4 min d⁻¹ of SED and in 59.6 ± 25.7 min d⁻¹ of MVPA. In general agreement with the current study, Hargens et al (2021) reported that students had a significantly longer sleep duration (404.1 ± 66.7 min d⁻¹), higher total sedentary time (637.4 ± 130.9 min d⁻¹), and lower MVPA (44.1 ± 30.3 min d⁻¹) on weekend days compared to weekdays (sleep: 378.9 ± 60.4 min d⁻¹; SED: 682.6 ± 90.3 min d⁻¹; MVPA: 65.7 ± 28.0 min d⁻¹). Estimates of LPA were not reported.

A primary strength of the present investigation was the simultaneous application of three research grade monitors to classify the 24-HAC behaviors. As highlighted by Rosenberger et al (2016) and subsequent investigations (Chastin et al 2015, Dumuid et al 2018, Janssen et al 2020), no single device can accurately classify the 24 h day in its entirety; thus, device selection should be informed by the research question and the primary outcome(s) of interest. Merging data from the ActiGraph GT9X worn on the wrist to measure sleep, an activPAL worn on the thigh to assess SED, and an ActiGraph GT3X+ worn on the hip to measure LPA and MVPA was used to leverage information from each device to classify 24-HAC behaviors for each minute of the day if one or two accelerometers were used which may misclassify time spent in an indiviaual component. For example, traditional waist-worn accelerometers are worn during waking hours and removed during water based activities or periods of sleep. Thus, non-wear periods may be misclassifed if a wrist-worn accelerometer is left being worn due to higher counts or if no other device is being worn. Because participants in our study were instructed to wear the activPAL continuously for 24 h, the algorithm developed was able to discern postural changes as a more confident approach to classify other components of the 24-HAC. A second strength of the study is the algorithm's ease of use. ActiGraph accelerometers were chosen in the current study because they are used in physical activity research and national surveillance work in the United States. Unlike previous algorithms using multiple complimentary research grade accelerometers used advanced machine learning skills or data processing skills (i.e. GGIR), the current study's algorithm may have the flexibility to collaborate with other accelerometers device brands, but additional processing steps could be required depending on the device's data output or algorithm used to harmonize and classify the 24-HAC. However, future investigations will be needed to support this concern.

Several caveats and limitations should also be considered when interpreting the results of this study. First, the accuracy and precision 24-HAC estimates generated using the proposed algorithm is unknown due to the lack of a criterion measure of these behaviors in the present study. Previous studies (Micklesfield et al 2022, Dooley et al 2023) using multiple research grade monitors to estimate the 24-HAC have also noted the same limitation. Absent a single validated criterion measure of free-living 24-HAC behaviors, using multiple complementary monitors logically be expected to improve estimates of these behaviors compared to single-measure estimates. However, this area of research could benefit from studies conducted in clinical settings which compare single versus multiple monitor estimates of 24-HAC behaviors using direct observation as the criterion. Finally, it is unclear whether the approach outlined here would generate similar results if different device combinations were used or in appreciably different samples.

In conclusion, this study presents a feasible approach to estimate 24-HAC behaviors using three complimentary research-grade accelerometers. Moreover, we present a hierarchical algorithm that harmonizes these assessments by prioritizing the most appropriate data source for a given 24-HAC behavior while accounting for variations in data availability during the measurement period. Given that the 24-HAC is a relatively new paradigm, future studies should aim to (a) assess the validity of various methodological approaches across different groups, (b) evaluate the trade-off between the increased accuracy of multi-device approaches and the additional investigator (e.g. increased data processing time and device costs) and participant burden (e.g. remembering what device to wear and when to wear it), and (c) identify the optimal combination of devices and wear locations to strike a balance between accuracy and feasibility. The findings and methodologies of this study can inform future efforts to measure and harmonize the composition of the 24-HAC, enhancing our understanding of human health.

All data and supplemental data are included within the article itself. The data that support the findings of this study are available upon reasonable request from the authors.

There are no conflicts of interest to disclose.

No external funding was provided for this study.