Abstract

Introduction

Low physical activity is associated with an increased risk of morbidity and mortality [1]. However previous studies are predominantly based on self-reported participation in leisure time activity [2] from which it is difficult to quantify total physical activity across different domains [3]. This uncertainty makes it difficult to convert epidemiological association results into public health recommendations about the minimum level of physical activity required for health and the benefits of engaging in different durations of activity of different intensity. The development of objective methods for assessing physical activity has provided an opportunity to quantify the dose-response relationship of activity with health as a complement to the subjective assessment of self-reported participation in specific activities.

Accelerometry is the most widely used method for objective assessment of physical activity in population studies [4,5], and large studies from the UK [6–8], US [9], and Canada [10] indicate age gradients and differences between men and women; time-of-day and day-of-week differences in physical activity. Most earlier studies used accelerometers which were worn around the waist and during awake-time only, a protocol which can result in relatively large amounts of missing data [11]. Therefore, wrist-worn accelerometers are becoming more widely used as an objective measure of physical activity in cohorts in the UK [12], US [11], and Brazil [13]. These devices are water-proof and worn continuously day and night, resulting in higher levels of participant compliance [11,12]. Wrist-worn accelerometers have also been validated against established measures of physical activity energy expenditure [14,15].

Cohort studies which include hundreds of thousands of participants followed up over time are required in order to describe the relationship between physical activity and health outcomes that have a number of potential lifestyle, environmental, and genomic causes [16]. Objective assessment of physical activity in such large population-based cohorts has previously not been undertaken because of the challenges of cost and the feasibility of collecting, processing and analysing data on this large scale. In this paper we describe the methods used to collect and analyse physical activity by wrist-worn accelerometry in the UK Biobank cohort study and report the variation in activity in more than 100,000 participants by age, sex, and time.

Methods

Accelerometer & Data Collection

For objective assessment of physical activity, we used the Axivity AX3 wrist-worn triaxial accelerometer (see Fig 1), a commercial version of the Open Movement AX3 open source sensor (https://github.com/digitalinteraction/openmovement) designed by Open Lab, Newcastle University. This device demonstrated equivalent signal vector magnitude output on multi-axis shaking tests [17] to the GENEActiv accelerometer used in the Whitehall II [12], Fenland [15] and Pelotas cohorts [13]. The Axivity device facilitates transparent data processing analysis due to its open-source firmware platform and unforced sampling of raw measurement data. We set up the Axivity accelerometers to start at 10am two working days after postal dispatch, and capture triaxial acceleration data over a seven day period at 100Hz with a dynamic range of +-8g.

Open in a separate window

Fig 1

UK Biobank triaxial accelerometer and processing steps to extract physical activity information.

Axivity AX3 triaxial accelerometer worn on dominant hand as used in UK Biobank (top left). Time series trace of processed accelerometer values after one week of wear (top right). Overview of process to extract proxy physical activity information from raw accelerometer data (bottom).

Participants were informed in the invitation email and device mail-out letter that the accelerometer should be worn continuously and that they should carry on with their normal activities. Participants were asked to start wearing the accelerometer immediately after receiving it in the post and to wear the monitor on their dominant wrist. They were also informed that the device was configured to automatically turn itself on soon after its arrival and off seven days later. Finally, participants were asked to mail the device back to the co-ordinating centre, in a pre-paid envelope, after the seven day monitoring period.

Results

A total of 236,519 UK Biobank participants were approached, of whom 106,053 agreed to wear a physical activity monitor (44.8%). The median time between each participant being invited to take part and being sent a device was 113 days (IQR: 73–137 days). Fig 2 shows that 103,712 datasets were received for data analysis. 123 participants were excluded as they were aged less than 45 years. Eleven participants were excluded from further analysis; eight because the calibration by the preceding or subsequent measurement was not possible due to insufficient data; and three participants due to unreliable device data. A total of 4043 devices were used on a median number of 27 occasions (IQR: 8–39). The median time between each device being posted was 17.0 days (IQR: 15.8–19.8) with a median of 832 devices (IQR: 629–994) posted each week.

Open in a separate window

Fig 2

Participant flow chart; the UK Biobank study 2013–2015 (n = 103,712).

Calibration of the data to local gravity greatly reduced the error in the assessment of acceleration with the root mean square error of stationary points falling from an average of 81.8 mg (95% CI: 81.6–82.1) to an average of 2.6 mg (95% CI: 2.6–2.6). However, 2.9% (n = 3049) of participants had insufficient stationary data to inform the calibration. These individual records were calibrated using stationary episodes from the previous (n = 2887) or next (n = 154) use of the same device by different participants. The influence of clips (readings beyond the sensor’s dynamic range of +-8g) before (median: 160, IQR: 62–393) and after (median: 169, IQR: 67–410) calibration, interrupts (median: 0, IQR: 0–0), and errors such as clips or missing readings (median: 200, IQR: 66–355) was negligible, with respect to the median of 58.6 million data readings (IQR: 56.0–60.1 million).

Fig 3 illustrates that 80.6% of participants wore the device for at least 150 hours out of a scheduled 168 hours. Men wore the device for a median of 166.3 hours (IQR: 157.7–168.0) and were slightly more compliant than women who wore the device for a median of 165.6 hours (IQR: 156.7–167.0). Table 1 shows that older age groups had marginally higher levels of compliance than younger age groups. Analysis of wear time compliance by age on a linear scale shows that on average there was a difference of 2 hours 18 minutes (1.6%) for each decade. In addition, Table 1 indicates minimal differences in the wear time compliance by time-of-day and week-weekend day. No wear-time differences were found by season. We removed 6978 (6.7%) participants who had insufficient wear data for our remaining analyses on accelerometer measured physical activity.

Open in a separate window

Fig 3

Cumulative distribution function of accelerometer wear time compliance; the UK Biobank study 2013–2015 (n = 103,578).

Table 1 describes the variation in mean vector magnitude, the summary measure of accelerometer measured physical activity, by age and sex in the sub-group of 96,600 participants who had good wear time compliance. Vector magnitude was higher in women than men, apart from those aged 45–54 years (p = 0.98). The mean effect size for these sex differences was small (0.09), ranging from 0.01 for 45–54 years to 0.15 for 75–79 years. There was strong evidence of accelerometer measured physical activity differing by age group in both men and women. The mean physical activity in the age group 45–54 years was 31.17 mg (SD 9.10) and was, on average 7.5% or 2.35 mg lower per decade. The mean effect size for these age differences was large, at 0.89 for women and 0.9 for men. Fig 4 shows the distribution of the data within age and sex strata, highlighting that although there appears to be an overall decline in average physical activity with increasing age, there is considerable overlap in the distributions with many older participants being more active than those in the youngest age category.

Open in a separate window

Fig 4

Acceleration vector magnitude by sex and age; the UK Biobank study 2013–2015 (n = 96,600).

Fig 5 shows the mean physical activity level by hour of day averaged across the whole measurement period by age and sex. It shows that the effect size for physical activity differences between age groups are most apparent in the afternoon (0.74 for women and 0.69 for men) and evening (1.06 for women, 1.12 for men) with smaller differences by age group in the morning (0.56 for women, 0.46 for men). Weekdays and weekend days differed, with vector magnitude higher at weekdays except for those aged 45–54 years. However, the mean effect size for these day differences was small (0.10), ranging from 0.04 to 0.15 across female age groups and 0.11 to 0.18 for male age groups (see Fig 6). Seasonality also differed, with vector magnitude lower during winter months except for women aged 75–79. However, the mean effect size for these season differences was small (0.21), ranging from 0.09 to 0.18 across age groups in women and 0.17 to 0.41 across age groups in men (see Fig 6).

Open in a separate window

Fig 5

Variation in mean acceleration across the day by age and sex: the UK Biobank study 2013–2015 (n = 96,600).

Shading bounds represent two standard errors.

Open in a separate window

Fig 6

Acceleration vector magnitude by day of the week (top), season (bottom), age, and sex: the UK Biobank study 2013–2015 (n = 96,600).

To illustrate time spent at different physical activity intensities, Fig 7 plots the empirical cumulative distribution function of the five second sample values for each subgroup. The bottom part of this figure shows sex differences in the distribution of physical activity intensity, for each age group. For example, men spend more time at or below 25 mg than women (122.6 versus 119.3 hours), but also slightly more time above 225 mg than women (2.18 versus 2.09 hours).

Open in a separate window

Fig 7

Cumulative time spent in various acceleration categories by sex and age (top), and sex differences by age and intensity level (bottom); the UK Biobank study 2013–2015 (n = 96,600).

Discussion

Developments in the technology supporting objective assessment of physical activity have now made it possible to consider assessing this behaviour objectively in large scale population-based cohort studies as an adjunct to more traditional assessment of self-reported participation in activities within different domains of life. However, even with those technological developments, it has previously been unknown whether it would be possible for this approach to be acceptable to participants and whether it would prove to be feasible to collect, analyse and interpret data from over one hundred thousand participants. This report from the UK Biobank study shows that 45% of participants who were invited to wear a monitor accepted the invitation. It also shows that measuring activity with a wrist worn device is highly acceptable to participants as manifest by the very high proportion of people in whom the data were of high quality and completeness. By necessity in the UK Biobank Study participants were invited to wear the monitor some time after recruitment to the baseline visit. As with all add-on measurements that are conducted on a different occasion, there will be participants who do not accept the invitation to participate. Other studies in which wrist worn accelerometers are part of the protocol for a baseline visit, rather than a separate add-on, will be likely to achieve higher participation levels.

We have shown that mean vector magnitude in this population was greater in women than men, apart from those aged 45–54 years. Our findings also suggest that men spend more time than women in what might be considered low or sedentary levels of physical activity, while women spend more time in moderate levels of activity. Whether these results indicate true differences in physical activity between sexes or are a function of a between-sex difference in the relationship between wrist acceleration and true activity remains to be investigated. For example, a recent study of 1695 UK men and women reported physical activity energy expenditure to be 7% lower in women compared to men for the same level of non-dominant wrist acceleration[15]. With respect to age, we observed a marked overall difference in the summary measure of physical activity by age, with older participants having levels of activity that are, on average, 7.5% lower for each 10 year age difference. These differences by age group are similar to other population-based studies [6,9] that have used hip worn accelerometers. With respect to time, older participants are much less active than younger participants during afternoons/evenings than in the morning, which mirrors previous findings in older UK adults using hip worn accelerometers [7]. There were small differences between weekday and weekend day physical activity, and also small seasonal differences in activity. We have not generalised the overall descriptive findings to the UK population since the UK Biobank was established as an aetiological study rather than one aimed at population surveillance [6,9].

We have extracted objective physical activity information from 103,578 participants aged 45–79, who were asked to wear accelerometers for seven days on their dominant wrist. The strengths of this study include its use of objective measures of physical activity, excellent participant compliance, unprecedented scale, and use of reproducible methods. For example, >93% of participants provided more than 72 hours of wear time with no missing data bias by time of day. The overall levels of participant compliance in the UK Biobank mirror findings in other studies that have used wrist-worn accelerometers in thousands of participants [11–13]. Wrist-worn accelerometers are not only highly acceptable to participants, but are also valid measures of physical activity energy expenditure. A recent free-living study reported that wrist acceleration explained 44% of the variance in physical activity energy expenditure estimated from individually calibrated combined heart rate and movement sensing[15]. Laboratory-based studies have demonstrated that the signal from wrist-worn devices correlates with physical activity energy expenditure as well as traditional waist-worn devices (left wrist R = 0.86; right wrist R = 0.83; waist R = 0.87) [22]. Thus the relative validity between wrist-worn devices is similar and the association of accelerometer measured physical activity with health outcomes within a study is not dependent upon which wrist was chosen. However, the comparison of absolute values between studies would need to be mindful of which wrist was selected for individual studies. More robust validation studies of physical activity information from raw wrist-worn accelerometer data are needed to enhance the interpretation of this signal.

Even though we used relatively simple summary measures in these analyses, their derivation still involved several critical data processing decisions, the alteration of which would have large effects on the derived physical activity variables [18]. For example, there is uncertainty on how to address negative values during the gravity removal process. Furthermore, it is not possible to perfectly separate static and dynamic acceleration (for example gravity and physical activity) from the measurement of triaxial acceleration alone. Therefore, we produced summary statistics of vector magnitude which do not attempt this separation. In addition, absolute and truncated Euclidian norm minus one and high-pass filtered vector magnitude (all of which attempt to separate activity from gravity) were generated too. We found that while the magnitude of these variables changes, their correlation was very strong (>0.95), which provides confidence in our chosen metric for association studies. Uncertainty also exists on the best method to identify non-wear episodes, and the size of epoch on which to base distributions of physical activity intensity. While the derived factors are only the most basic variables that can be extracted from the raw 100Hz triaxial acceleration data, future projects will be able to build on this foundation to derive additional parameters describing other aspects of physical activity, sedentary behaviour and sleep.

In conclusion, the collection and processing of this large accelerometer dataset in a prospective cohort study lays the foundation for studies of physical activity and its health consequences. The summary variables that we have constructed are now part of the UK Biobank dataset and can be used by researchers as exposures, confounding factors or outcome variables in future analyses.

Supporting Information

Acknowledgments

Funding Statement

The UK Biobank Activity Project and the collection of activity data from participants was funded by the Wellcome Trust (https://wellcome.ac.uk/) and the Medical Research Council (http://www.mrc.ac.uk/). The analysis was supported by the British Heart Foundation Centre of Research Excellence at Oxford (http://www.cardioscience.ox.ac.uk/bhf-centre-of-research-excellence) [grant number RE/13/1/30181 to AD], the Li Ka Shing Foundation (http://www.lksf.org/) [to AD], the UK Medical Research Council (http://www.mrc.ac.uk/) [grant numbers MC_UU_12015/1 and MC_UU_12015/3 to NW and SB], the RCUK Digital Economy Research Hub on Social Inclusion through the Digital Economy (SiDE) (http://www.rcuk.ac.uk/) [EP/G066019/1 to NH], the EPSRC Centre for Doctoral Training in Digital Civics (https://www.epsrc.ac.uk/)[EP/L016176/1 to DJ], and the National Institute for Health Research (http://www.nihr.ac.uk/) [SRF-2011-04-017 to MIT]. The MRC and Wellcome Trust played a key role in the decision to establish UK Biobank, and the accelerometer data collection. No funding bodies had any role in the analysis, decision to publish, or preparation of the manuscript.

Data Availability

The summary variables that we have constructed are now part of the UK Biobank dataset at http://biobank.ctsu.ox.ac.uk/crystal/label.cgi?id=1008. Furthermore, our analysis is freely available and hosted as an open source software project at https://github.com/activityMonitoring/biobankAccelerometerAnalysis

References