Physical activity measured by accelerometers requires basic assumptions to translate the output into meaningful measures. We used accelerometer data from the Osteoarthritis Initiative to investigate in the context of knee osteoarthritis (OA) the following data processing assumptions derived from the general US adult population: nonwear (a period the monitor was removed), based on zero activity exceeding 60 minutes; and a valid day of monitoring, based on wear time evidence exceeding 10 hours.
We examined the influence of nonwear thresholds ranging from 20 to 300 minutes of zero activity on mean daily activity minutes (counts >0), mean daily activity counts, and mean daily moderate to vigorous physical activity minutes. The effect of selecting minimums of 8, 10, or 12 wear hours to signify a valid day of monitoring on data retention was examined.
Our sample of 3,536 days of accelerometer data from 519 persons with knee OA showed that mean daily activity minutes increased with the nonwear threshold until stabilizing at 463 minutes per day, corresponding to the 90-minute nonwear threshold. Similar patterns were observed for mean daily activity counts. Varying the nonwear threshold had no effect on mean daily moderate to vigorous physical activity minutes. Choosing the 90-minute nonwear threshold and a minimum of 10 wear hours to constitute a valid day provided 94% data retention.
Data supported applying the 90-minute nonwear threshold to the knee OA population instead of the 60-minute threshold for the general population, while retaining the 10-hour valid day threshold.
The number of US adults with arthritis-related activity limitations is expected to escalate from 17 million in 2006 to nearly 25 million by the year 2030 (1). A leading cause of arthritis-related activity limitation is knee osteoarthritis (OA), which affects an estimated 12% of US adults ages 60 years or older (2). Physical activity has been found to reduce pain and improve function in persons with knee OA, which fuels the growing interest in the measurement of physical activity in this population (3).
Objective measurement is central to the accurate assessment of physical activity. Sole reliance on self-report is problematic because subjects are known to both underestimate their daily walking distance (4) and overestimate the amount of energy expended during moderate-intensity daily activity (5, 6). Accelerometer technology validated from expensive gold standard methods (e.g., doubly-labeled water) can be used successfully in community populations (6–10). Advantages of accelerometer monitoring to objectively measure physical activity are minimal participant burden and capture of salient behavior characteristics, including the frequency, intensity, and duration of physical activity.
One day of accelerometer monitoring can produce 1,440 minute-by-minute recordings of “accelerometer counts.” A major challenge in the use of accelerometers is translating the downloaded output into meaningful physical activity outcomes. Often this analytic step is a “black box,” requiring blind reliance on algorithms provided by the accelerometer manufacturer. Basic data decisions embedded in these analytic algorithms can substantially alter the calculated physical activity outcomes (11). Reliance on a one-size-fits-all “black box” may be inappropriate and could lead to erroneous conclusions.
The accelerometer algorithm released by the National Cancer Institute (NCI), which was developed for the landmark physical activity study from the National Health and Nutrition Examination Survey (NHANES) sample (12, 13), provides an important benchmark to translate accelerometer output into physical activity measures. The NCI algorithm made the translation process transparent (14). A major caveat for OA research is that parameter values in the NCI algorithm were derived from the general US adult population (13). Two foundational NCI algorithm parameters are the thresholds to identify 1) “nonwear” and 2) a “valid day” of accelerometer monitoring. The nonwear threshold (NCI value = 60 minutes of zero activity counts) is an attempt to objectively distinguish periods of inactivity (e.g., sitting) from periods when the monitor was removed (i.e., nonwear). Underestimating the nonwear threshold may potentially lead to inappropriate discarding of periods of inactivity and underestimate sedentary time. The valid day threshold (NCI value = 10 hours of evidence of monitoring) forms the basis to determine if accelerometer monitoring occurred for a sufficient time to represent a full day of physical activity. If the estimated monitoring time in a day is below the designated threshold, accelerometer data from that day are considered invalid. Therefore, the valid day threshold directly affects data loss.
These accelerometer algorithm parameters require assessment for the knee OA population, whose physical activity patterns differ from the general population due to pain and stiffness. Because the knee OA population may be more sedentary (15, 16), criteria based on the general adult population to distinguish nonwear from inactivity may overstate nonwear and understate inactivity that is a consequence of the symptoms and/or joint structural changes of OA. The purpose of this study was to investigate empirically if the NCI accelerometer algorithm threshold values for nonwear and valid day derived from the general population are appropriate to apply to accelerometer data from persons with knee OA for evaluating physical activity outcomes, and to explore the impact of applying alternate threshold values to that accelerometer data. The availability of an algorithm examined in the context of knee OA to translate accelerometer recordings into physical activity measures will improve the quality of objective physical activity measures and reduce the likelihood that important and useful data will be lost due to inappropriate threshold application.
MATERIALS AND METHODS
Study population and sample.
This study analyzed data from the Osteoarthritis Initiative (OAI), a prospective natural history study investigating the development and progression of knee OA in persons ages 45–79 years at enrollment with or at higher risk to develop knee OA. Annual OAI interviews began in 2004 at 4 clinical sites: Baltimore, Maryland, Columbus, Ohio, Pittsburgh, Pennsylvania, and Pawtucket, Rhode Island, and are currently ongoing (online at http://www.oai.ucsf.edu/datarelease/About.asp). The OAI excluded subjects with rheumatoid or inflammatory arthritis; with severe joint space narrowing in both knees on the baseline knee radiograph, or unilateral total knee replacement and severe joint space narrowing in the other knee; with bilateral total knee replacement or plans to have bilateral knee replacement in the next 3 years; unable to undergo a 3.0T magnetic resonance imaging examination of the knee because of contraindications; with a positive pregnancy test; unable to provide a blood sample; with the use of ambulatory aides other than a single straight cane for more than 50% of the time in ambulation; with comorbid conditions that might interfere with the ability to participate in a 4-year study; and with current participation in a double-blind randomized trial. All of the OAI participants underwent knee radiography; the baseline visit identified 2,679 participants with radiographic knee OA (i.e., radiographic evidence based on a Kellgren/Lawrence scale grade of ≥2 calculated from separate scores for osteophytes and joint space narrowing in a knee) from the total OAI enrollment of 4,796 persons. The protocol for baseline radiographic measures may be found online at http://www.oai.ucsf.edu/datarelease/OperationsManuals.asp. A subsequent physical activity ancillary study to the OAI collected accelerometer data on participants returning for the ongoing 48-month followup visit.
A total of 981 persons consented to participate in accelerometer monitoring at the OAI 48-month followup visit during an evaluation period from September 2008 through April 2009, representing 79% of eligible participants (n = 1,240) during this period. These evaluation analyses were restricted to 519 participants with baseline radiographic knee OA. Accelerometer data were merged with OAI public data (from baseline to the most recent 36-month visit) to determine evaluation of sample characteristics. For analysis purposes, body mass index (BMI) missing at 36 months (n = 21 participants [4.1%]) was imputed by using data from previous visits. The evaluation sample in comparison to the entire OAI radiographic knee OA cohort was almost identical in baseline age and BMI, and was slightly more likely to be women (62% versus 58%).
Accelerometer measures and procedures.
Physical activity was monitored at the OAI 48-month followup visit using a GT1M Actigraph accelerometer (Actigraph). The accuracy (walking speed &lsqbr;17&rsqbr;) and test–retest reliability (18) of Actigraph accelerometers under field conditions have been established in many populations, including persons with OA (19). The GT1M Actigraph accelerometer is a small uniaxial accelerometer that measures vertical acceleration and deceleration (20). Accelerometer output is an activity count, which is the weighted sum of the number of accelerations measured over a time period (in this case, 1 minute), where the weights are proportional to the magnitude of measured acceleration.
Accelerometer recordings were translated on a minute-by-minute basis into physical activity intensity categories (light, moderate, and vigorous) from activity counts. The influence of different thresholds to assess physical activity intensity has been previously examined (21–23). For the purposes of standardization, we applied intensity thresholds used by the NCI (13) for light (1–2,019 counts), moderate (2,020–5,998 counts), and vigorous (5,999 counts or greater) intensity.
Uniform scripted instructions were given on the wear and positioning of the accelerometer. The participants were instructed to wear the accelerometer upon arising in the morning and continuously until going to bed at night for 7 consecutive days. The unit was worn on a belt at the natural waistline on the right hip in line with the right axilla. The participants maintained a daily log to record the dates of accelerometer monitoring. Skipped days reported on the log were excluded from the analysis.
Accelerometer algorithm parameter definitions.
Nonwear refers to a sustained period of little or zero activity that may represent an interruption in accelerometer monitoring. The benchmark NCI nonwear threshold applied by Troiano and colleagues (13) to US adult accelerometer data consisted of an interval of at least 60 minutes of zero activity counts that contained no more than 2 minutes of low (<100) activity counts. A rolling window algorithm (i.e., scan each minute and begin a potential nonwear period when a zero activity count is found) identified nonwear periods. Consistent with the NCI approach, a nonwear period ended with either a third minute of low activity counts or a 1-minute activity count greater than or equal to 100. The term “zero activity” was used to designate a period of continuous zero activity counts or near-zero activity that allows up to 2 minutes of <100 activity counts. Nonwear parameter thresholds of 20, 30, 45, 60, 90, 120, 150, 180, 210, 240, 270, and 300 minutes were investigated. A period of zero activity longer than the threshold time designated nonwear. Wear hours were calculated on a daily basis as 24 hours minus the nonwear hours.
A valid day refers to daily wear hours of estimated accelerometer monitoring that meet/exceed a threshold. Accelerometer data from nonvalid days are considered unreliable to describe a full day of physical activity. Therefore, the definition of a valid day directly impacts data loss. The NCI benchmark definition for a valid day from Troiano and colleagues for US adults was a single day containing 10 or more hours of monitoring (also called wear time) (13). Importantly, wear time did not need to be continuous to be considered valid. We tested valid day thresholds of 8, 10, and 12 hours.
Mean daily wear hours and 3 physical activity outcomes were assessed: 1) mean daily activity counts, 2) mean daily minutes of activity (i.e., minutes with accelerometer activity counts >0), and 3) mean daily moderate to vigorous physical activity minutes. A mean daily outcome represents the outcome summed over all wear hours divided by the number of monitored days for each person.
In order to assess whether the general population–based nonwear and valid day thresholds could be applied in the OA population, we investigated the minimum nonwear threshold that resulted in stable physical activity outcomes and the influence of different threshold values on data loss.
Descriptive analyses graphically depict the relationship between nonwear thresholds and the outcomes described above. This process was used to identify candidate nonwear thresholds that were associated with stable results in regard to physical activity outcomes when applied to accelerometer data from participants with knee OA. Similar descriptive analyses were performed to examine stability among stratified BMI, age, and sex subgroups. Data retention as a function of candidate nonwear and valid day thresholds was also plotted.
Statistical analysis of nonwear threshold.
We compared the physical activity outcomes derived from candidate nonwear thresholds (e.g., 90 minutes, 120 minutes) with outcomes derived from the NCI 60-minute nonwear US adult population benchmark. Statistical testing using Studentized t-statistics to maintain overall testing at an alpha level of testing of 0.05 (24) was applied when evaluating within-person paired differences in physical activity outcomes. Due to skewed distributions, nonparametric quantile regression (25) was used to estimate the difference in medians of physical activity outcomes and the SEMs of those differences. Results of statistical testing are reported as confidence intervals to reflect the precision of the data and a plausible range of findings; a 95% confidence interval that excludes zero indicates a statistically significant result.
Statistical analysis of valid day threshold.
A linear weighted kappa coefficient was used to describe the agreement in the estimated number of valid days resulting from the application of different threshold values (26).
We used Stata/SE, version 10.0 (27), for quantile regression, and SAS, version 9.2 (28), for other analyses.
A total of 519 OAI participants with definite knee OA collectively contributed 3,536 days of accelerometer data for analysis. This evaluation sample was predominantly women (62%), college educated (60%), and white (86%), with a mean ± SD age of 62 ± 9 years. Accelerometer assessment employed up to 7 days of continuous monitoring. More than 96% of the participants had 6–7 days of monitoring. These 519 participants compared with the remaining 2,160 OAI participants with knee OA at baseline were almost identical in the distribution of age and self-reported physical activity, but included more women (62% versus 57%) and whites (86% versus 76%) and had a slightly lower average BMI (29 kg/m2 versus 30 kg/m2).
Figure 1 shows an example of a complete day (1,440 minutes) of accelerometer output that contains prolonged periods of inactivity. Because the participants were instructed to take off the accelerometer during the evening upon retiring, the graph indicates that the monitor was first worn at approximately 6:00 AM and was last taken off at approximately 10:20 PM. However, the strings of zero activity counts during waking hours could either be due to nonwear or very sedentary activities. In this study, participants were permitted to remove the units if they took a nap, bath, or shower during the day. On average, the units were removed 1.4 times per day, which included taking the unit off at bedtime (calculated from on/off times reported by the participants via logs).
Nonwear threshold results.
Candidate nonwear thresholds ranging from 20 to 300 consecutive minutes of zero activity were investigated to identify potential interruptions in wear time as part of the data reduction process. The relationship between candidate nonwear threshold values and wear hours and physical activity outcomes are displayed graphically in Figures 2A and B. As expected, mean daily wear hours (Figure 2A) increased with the length of the nonwear threshold. The average daily activity counts increased with the length of the nonwear threshold until it stabilized near the 90-minute to 120-minute threshold. Similarly, the average daily minutes of (nonzero) activity increased with the nonwear threshold, but stabilized near the 90-minute to 120-minute threshold (Figure 2B). It was evident that the average of daily moderate to vigorous physical activity minutes was invariant to the nonwear thresholds. This invariance was expected because the nonwear period was terminated by definition when a minute of moderate to vigorous physical activity occurred. As a result, the same amount of moderate to vigorous physical activity minutes over a single day was captured by all of the nonwear thresholds.
Stratified analyses that investigated the nonwear threshold among the BMI categories are shown in Figure 3. Participants with normal weight (BMI <25 kg/m2) had the highest level of physical activity in terms of mean daily activity minutes. Overweight (BMI 25–30 kg/m2) and obesity (BMI ≥30 kg/m2) were associated with reduced physical activity, i.e., a mean daily loss of approximately 10 and 40 activity minutes, respectively. Across all of the BMI groups, average daily activity minutes increased as the nonwear threshold increased, and stabilized near the 90-minute to 120-minute threshold. This is consistent with the overall results shown in Figure 2B. Additional stratified analyses in age (49–54, 55–64, 65–74, and ≥75 years) and sex also revealed stabilization near the 90-minute to 120-minute nonwear thresholds (data not shown).
Statistical comparisons of physical activity outcomes based on 60-minute (the general population benchmark), 90-minute, and 120-minute nonwear thresholds using the Studentized t-test are shown in Table 1. The magnitude of the outcome differences associated with the 90-minute versus 120-minute threshold was extremely small, as indicated by the narrow confidence intervals. For example, there was a less than 0.1% difference between the averaged physical activity outcomes based on the higher 90-minute or 120-minute thresholds. Statistical testing was precluded for moderate to vigorous physical activity minutes, which were identical for all of the thresholds.
Table 1. Physical activity outcomes by nonwear parameter values from 519 persons with osteoarthritis contributing 3,536 days of accelerometer monitoring
Because the distributions of the differences are skewed, medians and SEMs of medians from quantile regression are used in the Studentized t-tests that maintain an alpha level of testing of 0.05 with multiple statistical tests. A 95% confidence interval (95% CI) that excludes zero indicates that the median difference is statistically different from zero at an overall alpha level of testing of 0.05.
Statistical test precluded by zero variance on differences due to identical results in groups A, B, and C.
Mean daily activity counts
Mean daily (nonzero) activity minutes
Mean daily moderate to vigorous physical activity minutes‡
Valid day threshold results.
The influence of the valid day threshold on data retention was graphically examined. Figure 4 shows the proportion of data retained for the valid day threshold set at 8, 10, and 12 hours of evidence of monitoring as a function of nonwear thresholds. These valid day thresholds approximately represent wearing a monitor one-half, two-thirds, and three-quarters of waking hours, respectively, under the common situation of 8 hours of sleep and 16 hours awake. For example, at the 60-minute nonwear benchmark threshold, the 10-hour valid day threshold resulted in 10% data loss (i.e., 90% data retention). Increasing the threshold to 12 hours more than doubled the data loss to 26%. The curves from this knee OA sample related to the 8-hour and 10-hour valid day threshold stabilized and captured more than 94% of monitored days near the 90-minute to 120-minute nonwear thresholds, whereas the 12-hour threshold resulted in 85% data retention. These graphs suggest that the benchmark threshold of 10 or more wear hours used in the general US adult population to identify a valid day is applicable in the context of knee OA, even when nonwear thresholds of 90 or 120 minutes are used.
Agreement in the distribution of valid days for this knee OA sample was evaluated using weighted kappa coefficients for nonwear thresholds of 60, 90, and 120 minutes in Table 2. Consistent with Figure 4, the frequency of valid days was greatest for the 120-minute nonwear threshold, followed by the 90-minute and 60-minute thresholds. There was good agreement between the 60-minute nonwear threshold and the higher thresholds (weighted κ = 0.67–0.76) and very good agreement in the highest 90-minute and 120-minute nonwear thresholds (weighted κ = 0.91).
Table 2. Distribution of valid days (with >10 of hours wear time) by nonwear parameter values from 519 persons with osteoarthritis with accelerometer monitoring over 3,536 days
Weighted kappa on the agreement for specified parameter values of the distribution of valid days.
This study examined basic assumptions used to translate physical activity measured by accelerometers into meaningful physical activity outcomes for a knee OA population. The presented methodology facilitated decisions on nonwear and valid day thresholds that minimized data loss while capturing stable and meaningful physical activity outcomes. Our findings supported 1) applying a higher threshold of at least 90 minutes of zero/little activity to signify accelerometer nonwear rather than the general population benchmark of 60 minutes, and 2) retaining the general population valid day threshold of 10 wear hours in this knee OA population.
Nonwear represents interruptions in accelerometer data collection. It is important to note that both nonwear and very sedentary activities register as “0” counts in accelerometer readouts. It is difficult to accurately attribute long continuous periods of “0” activity counts to nonwear or to a prolonged inactive or “nonmovement” time period such as reading, watching TV, or sedentary office work. The released NCI accelerometer algorithm provides an efficient approach that is independent of participant report to objectively capture a participant's activity pattern by distinguishing wear time from nonwear periods.
Compared with the general adult population, people with OA may be more likely to have reduced levels of daily physical activity due to symptoms of joint pain and stiffness and/or structural changes. In fact, the average daily activity of our arthritis cohort is less than 80% of that of the published activity counts among the NHANES general population of adults ages ≥40 years (13). Prolonged periods of accelerometer zero activity counts in this knee OA cohort may be more likely to reflect extended periods of sitting rather than nonwear than in the general population. Our findings suggested that an alternate nonwear threshold of at least 90 minutes of zero activity may be more appropriate for persons with knee OA than the 60-minute nonwear threshold based on the general adult population. By increasing the minimum minutes of zero activity counts required to signify nonwear, more inactivity minutes were included within wear time and less data were discarded.
This adjustment will not only better capture the sedentary lifestyle of this population, but will improve accelerometer data retention. In fact, our data showed that the average daily inactive minutes during wear time (i.e., wear time minus activity minutes) increased from 339 minutes to 392 minutes when the nonwear threshold increased from 60 to 90 minutes (Figures 2A and B). It is worth noting that although wear time increases as the nonwear threshold increases, physical activity outcomes investigated here achieved stability near a 90-minute nonwear threshold value. Although our findings support a higher threshold in this OA population than the general population threshold, the stability of results suggests that it is not necessary to increase the nonwear threshold beyond 90 minutes.
Our study also examined the influence of the minimum number of hours of accelerometer wear required each day to represent a complete day of activity, represented by the valid day threshold. The physical activity literature has favored a 10-hour level of evidence, which approximately represents monitoring more than two-thirds of a person's waking hours (11) and was the choice for the NHANES study of the general adult population (13). A valid day of monitoring standard based on a 10-hour rule appeared to be applicable to this OA population. Our study found that increasing the valid day threshold from 10 to 12 hours more than doubled data loss (e.g., from 6% to 15% when applying a 90-minute nonwear threshold), whereas reducing the threshold from 10 to 8 hours did not substantially improve data retention. Because the average wear time for the accelerometers in this knee OA sample was 14 hours, it is reasonable that the 10-hour rule would result in good data retention.
It is important to note that the stratified analyses by BMI, age, or sex groups suggested that the same 90-minute nonwear threshold applies across different strata. Findings from a separate study that examined accelerometer data from rheumatoid arthritis participants also supported a 90-minute nonwear threshold while retaining the 10-hour valid day threshold (29). The similarity of the nonwear and valid day thresholds for both inflammatory arthritis (rheumatoid arthritis) and noninflammatory arthritis (OA) suggests that these thresholds may be appropriate for arthritis populations more broadly.
This study has several implications for future studies. First, our study showed that decisions on how to process accelerometer data influence physical activity outcomes in persons with knee OA, which is consistent with similar work related to other adult populations (11, 29). Accelerometer measurement is becoming a common approach to measuring physical activity in OA populations (19, 30, 31). However, methodology from those studies largely depends on accelerometer validation studies performed in the general population. It is important that publications disclose the parameters for deriving physical activity outcomes when reporting findings based on accelerometer monitoring. This disclosure is particularly important if the main interest is reporting patterns of sedentary activity, which are most strongly influenced by data processing decisions. Greater nonwear threshold values will capture more inactive minutes. Second, comparisons of findings across different studies must consider the data processing assumptions used to translate accelerometer readouts into physical activity outcomes. Differences or lack of differences in physical activity outcomes across studies could be partially masked by the accelerometer translation process. Third, it is noteworthy that moderate or vigorous activity assessment was robust to nonwear thresholds. Finally, the practical effect of varying the valid day threshold was on data retention. The general population standard of 10 hours of wear time evidence yielded 94% data retention in this knee OA sample when combined with a 90-minute nonwear threshold.
There are several limitations to acknowledge in the present study. One limitation of the waist-mounted uniaxial accelerometer used in this study is that water activities cannot be captured. Also, it may underestimate vertical acceleration/deceleration activities such as cycling or weight lifting. Therefore, physical activity was underestimated in persons who regularly participated in such activities. However, walking is the most popular leisure time sports activity in the US. Accelerometer monitoring also captures walking for working and transportation purposes (19, 32). Another limitation is that the methods used by NHANES to obtain the thresholds were unknown and if different, may produce different results. It is recognized the OAI is not a probability sample. However, the participants were recruited from multiple geographic sites using recruitment targets balanced for age and sex groups, and represented a broad spectrum of radiographic knee OA.
In summary, accelerometer data processing decisions influence both the stability of the physical outcomes and the data retention rates. We obtained stable physical activity values from a higher nonwear threshold of 90 minutes than the 60-minute general population threshold with the added benefit of retaining more accelerometer recordings on a daily basis among persons with knee OA. These OA-based thresholds retain more valid days of physical activity data.
All authors were involved in drafting the article or revising it critically for important intellectual content, and all authors approved the final version to be submitted for publication. Ms Song had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Study conception and design. Song, Chang, Mysiw, Jackson, Nevitt, Dunlop.
Analysis and interpretation of data. Song, Semanik, Sharma, Chang, Mysiw, Bathon, Eaton, Jackson, Dunlop.
ROLE OF THE STUDY SPONSOR
The Osteoarthritis Initiative study sponsors (National Institutes of Health, Merck Research Laboratories, Novartis Pharmaceuticals Corporation, GlaxoSmithKline, and Pfizer Inc.) had no role in the study design, data collection, data analysis, or writing of this manuscript. Publication of this article was not contingent on the approval of these sponsors.
The authors gratefully acknowledge the support of the OAI for providing data for this work and the insightful suggestions of Dr. David Berrigan that motivated this investigation.