International children's accelerometry database (ICAD): Design and methods

After a principle investigator or research group had consented for their data to be pooled, contact was made with their data manager to transfer data. Data was transferred using a secured FTP drop-site. Raw accelerometer. DAT files and additional phenotypic information (usually in SPSS, excel or SAS format) were deposited. Contextual information on the methods of data collection and variable coding was also garnered through telephone administered questionnaires.

World-wide, 24 studies were approached and invited to contribute data. The eligibility criteria for inclusion were: 1) physical activity data in the form of raw accelerometer (.dat) files from a version of a waist worn Actigraph accelerometer (e.g. 7164, 71256, GT1M¹) on children 3-18 years, (2) accompanying data, at a minimum, of gender, age and measured height and weight. Three of the invited groups chose not to contribute data. One further study could not be included because the data did not meet the minimum inclusion criteria. Hence, 20 of the 24 invited studies contributed data. Formal data sharing agreements and policies were established between ICAD and all partners. Each partner consulted with their research ethics board to confirm sufficient ethical approval had been attained for contributing the data. All individual data within the pooled data set were allocated a unique and non-identifiable participant ID to ensure anonymity of data.

The Actigraph family of accelerometers have become the de facto standard for measuring physical activity, having been used in the studies large and small the world over. These Actigraph models are uniaxial² accelerometers that detect vertical accelerations in the magnitude of 0.05-2.13 g with a frequency response of 0.25-2.50 Hz [23]. These Actigraph accelerometers are small (e.g. GT1M model is 4.5 × 3.5 × 1.0 cm and 43g) and in each of the contributing studies were worn on a belt fastened around the waist. When the Actigraph is accelerated, a voltage signal is generated proportional to the intensity of the acceleration. The acceleration signal is sampled at 10 Hz (7164 and 71256 model) or 30 Hz (GT1M model) and summarized in user-defined intervals (epochs) according to the users specific needs. Once downloaded, these temporally stamped data can be analyzed to provide measures of activity level (volume), and intensity, as well as daily, weekly and seasonal activity patterns [24]. These Actigraph models have demonstrated acceptable levels of technical reliability [25‐28]. In addition the Actigraph has been shown to be valid in both children and adolescents [29]. Ekelund et al. [30] assessed the validity of the Actigraph in free-living children using energy expenditure measured by doubly labeled water (DLW) as the criterion measure. Accelerometer output (counts/min) was related to physical activity level (r = 0.58, P < 0.01). This can be compared with correlations of self report versus gold standard measurements that are usually in the range of r = 0.0-0.2 [30].

The characteristics of the contributing studies are outlined in Table 1. The majority of the studies are located in Europe (N = 13). The other studies are located in the United States (N = 4), Brazil (N = 1) and Australia (N = 2). Study designs include cross-sectional, longitudinal, closed cohort, and intervention studies. The information provided in Table 1 refers to the data actually deposited in ICAD. For example, in the 1993 Pelotas Birth Cohort Study [31‐33] accelerometry was only conducted at one wave (at 13 years) of the longitudinal study and thus, within ICAD, this study is described as cross-sectional. In total, 44,454 viable baseline and repeated measured files were contributed from a total of 31,976 participants (including 12,022 boys and 19,954 girls) aged 2.5-18.4 yrs. With the exception of the female only sample, Project TAAG, the contributing studies collected data on both males and females.

Data were analyzed using KineSoft version 3.3.20 (KineSoft, Saskatchewan, Canada; http://​www.​kinesoft.​org).

Partners were requested to contribute all raw Actigraph data files (i.e. files with a. dat file extension). All studies were able to meet this request with the exception of the NHANES studies (see below for discussion of the special treatment for this study). In total 46,131 files were deposited (Figure 1). Of these, 298 (0.6%) duplicate files for the same participant were removed. This occurred when two monitors were placed on an individual, either as a part of a validation study or to increase the chances of obtaining a reliable file (i.e. if one monitor fails, there will be data from another monitor). In these instances the files were viewed and if one was corrupt, the non-corrupt file was used. If both files were 'not corrupt' then the first labeled file was used. A total of 419 (0.9%) files were not included because they did not have the minimum accompanying variables. Finally, 219 (0.5%) files deemed to have no wear time data (discussed in detail below) and 5 (0.01%) corrupt files were unable to be processed and were not included in the database. This resulted in the processing of 45,190 accelerometer files.

The files in the database were 'flagged' if they had one of more of the following characteristics (see Figure 2 for an example of an 'OK' file and Figures 3, 4 and 5 for examples of files flagged as spurious):

1. Overnight wear of ≥ 10 minutes at hrs 2, 3 and/or 4 am. Visual graphing of these files was used to determine the reason for the overnight wear. The four reasons for overnight wear were: i) Legitimate wearing of the monitor overnight, ii) 'Temporally shifted' file (Figure 3), iii) Potentially spurious device that does not return to baseline (zero) (Figure 4), iv) Malfunctioned unit (Figure 5). A 'temporally shifted' file is when the time stamping of the data is shifted by a number of hours which results in greater than expected consecutive zeros during the day and activity counts during the night on each day of monitoring.

2. Plateau (3 consecutive counts at the same number) at a count ≥ 10. This was a good indicator of technical faults with the device (with plateaus occurring most often at 32767 counts) (Figure 5).

A variable was included in the database which indicates files considered spurious (556 (1.2%)) and 'temporally shifted' (180 (0.4%)). It is recommended that these files are excluded from further analysis. Thus the final database includes data obtained from 44,454 'viable' accelerometer files (Figure 1).

To accompany the accelerometer files, all projects contributed additional participant information. Studies were included in ICAD if they had a minimum of age, gender, height and weight. However, within a study, individuals were included if they had age and gender at a minimum (i.e. could be missing height and weight). Where available, additional information including body composition (e.g. circumference measures, skinfolds), health markers (e.g. blood pressure, cholesterol), economic indicators (e.g. household income), parental information (e.g. height, weight) and behavioral variables (e.g. school travel mode, TV viewing) were also requested.

For a variable to be included in the database, it had to be present in at least 3 contributing studies. The coding procedure for all variables was standardized across projects to ensure uniformity within the database. Where necessary, the appropriate formulae were used to re-calculate the unit of measurement for body composition variables. To create standardized categorical variables the response categories used by different projects were compared and, where appropriate, combined category options were created for ICAD. Original participant responses were then re-coded to the new ICAD categories. See Additional file 1 for a full description of all non-accelerometer variables and further re-coding details.

Figure 6 shows the sample size at each chronological age for girls and boys. The numbers include all the viable data (i.e. excluding spurious and temporally shifted files) in the final database (total = 44,454). Figure 6a shows the age distribution of all files, including repeated measures on the same individual. Figure 6b shows the distribution of the baseline measures (N = 31,976) only (thus participants are only featured once). At most ages there is a fairly equal representation of girls and boys with the exception of age 13 and 14, where more females are represented. This is primarily due to the contribution of data at these ages from Project TAAG, a female-only study.

The 20 contributing studies collected accelerometer data using three versions of the Actigraph family of monitors (see Table 1). Of the 20 studies, the Actigraph (7164, also known as the CSA and MTI) was used in 8, the newer Actigraph GT1M, model was used in 8, and 4 studies used a mixture of both models including one where the 71256 model was used (which only differs from the 7164 with respect to the size of the memory).

The fact that epoch length influences the interpretation of accelerometer data has been well documented [34‐40]. An epoch, refers to the amount of time over which activity counts are summed and stored. Epochs varied from 5 seconds to 60 seconds (see Table 1). The older studies tended to collect at 60 second epochs, because the older generation accelerometers were only capable of storing data collected using epoch lengths of < 60 seconds for a limited number of days. Newer generations of the Actigraph have increased storage capacity allowing for higher resolution data to be collected while still being able to measure physical activity for at least 7 days. For the purpose of the pooled dataset it was necessary that all outcome variables be comparable; therefore, all files which used an epoch < 60 seconds were reintegrated up to 60 seconds for analysis using an automated tool in the KineSoft toolbox.

As previously mentioned, an obvious advantage of pooling accelerometer data is a very practical one: to increase sample size. However, this comes at a price: in comparison to a single large-scale study, ICAD will have inflated variability in its accelerometer measurements. This is because, although much variability is reduced by the utilization of standardized analytical techniques, variation in the deployment strategies used between contributing studies can generate significant between-study variability in accelerometer outputs. One potential cause of between-study variability is the use of different models (i.e. generations) of the Actigraph. In fact Rothney et al. 2008 [48] found differences in the outputs from the 7164 and the GT1M monitors when undergoing mechanical testing. Unfortunately, understanding these differences is hampered by the fact that the unlike the Actigraph 7164 (see [23]) there is no published paper detailing the specifications of the Actigraph GT1M. Fortunately, recent studies [28, 49‐51] provide insight into technical specifications of the GT1M. Despite the between model differences in the Actigraph, studies in free-living individuals showed that during self-paced locomotion at a range of speeds, the different Actigraph models did not result in meaningful differences in the classification of physical activity intensity [50] or activity count output [49]. Increased variability could also be caused by variation in the season (e.g. spring, summer, winter, fall) in which the accelerometer was worn [52, 53].

The above limitations notwithstanding, the increased variability in ICAD can be considered a real advantage. For example, the main strength of the pooled ICAD database is its enhanced social and cultural diversity, compared to any of the individual studies. The investigation of the influence of these socio-cultural influences may lead to new insights into how children's physical activity patterns are shaped. Furthermore, including participants from a variety of geographical locations may lead to further insights into the possible influence of the physical environment. Lastly, analysis of pooled data can be useful to quantitatively examine the generalizability of findings. For example, findings that are replicated across individual studies carry increased validity, as they are more likely to be linked to the research question being addressed than to differences in study design and data collection procedures (such as different climatic zones, different seasons, different days of the week, different number of days of deployment, etc.).

Unfortunately, with the large variation in the tools used to assess nutrition and psychological constructs, these data are not suitable for pooling and were not collected. This is a limitation of the database. If repositories of accelerometer data become more mainstream investigators may see added benefit of considering measures that are consistent between large studies (as seen in recent projects such as the Public Population Project in Genomics (P3G) [54]). Furthermore, this may accelerate the work aimed at equalizing scales and measures, thus enabling analysis of nutrition and psychological data, as it relates to physical activity, in the future.

The ICAD partners (i.e. contributors of the data) have sole access to the pooled database for 12 months (March 2011-Feb 2012), after which time the database is open to any approved user (see http://​www.​mrc-epid.​cam.​ac.​uk/​Research/​Studies/​ for more information on the procedures for using ICAD).