KOZEY, SARAH L.1; STAUDENMAYER, JOHN W.2; TROIANO, RICHARD P.3; FREEDSON, PATTY S.1
Accurate assessment of physical activity (PA) in a free-living environment is a critical feature of PA research. Specifically, conducting surveillance of population activity levels, determining the efficacy of programs to increase PA, and quantifying the relationship between PA dose and chronic disease all depend on accurately assessing PA (24). In the past 10 yr, motion sensors, such as pedometers and accelerometers, have emerged as valid tools to objectively measure PA (2,3). Motion sensors are not dependent on the cognitive ability of study participants for accurate PA measurement and are not subject to recall bias associated with self-report tools (4). Accelerometers provide objective measures of the frequency, duration, and intensity of PA, which allow researchers to quantify the dose-response relationship between PA and health outcomes (14). Pedometers provide a measure of total steps per day, which is associated with numerous health outcomes including weight status (22).
The ActiGraph 7164 (AM1; Pensacola, FL) is a commercially available uniaxial accelerometer used extensively in PA research. This monitor measures vertical acceleration in units called counts and has a pedometer function to measure steps. The count output from AM1 is typically calibrated in a laboratory by establishing linear regression associations between accelerometer counts and a measured physiologic variable (e.g., METs) (6,7,10,15,17,18). On the basis of linear regression models, cut points are developed to estimate time spent in various PA intensity levels. The cut points define the ranges of count output that correspond to light, moderate, and vigorous intensity (<3 METs, 3-6 METs, and >6 METs, respectively). In the field, the most widely used regression approach is the method developed by Freedson et al. (7) using AM1. The 2003-2006 National Health and Nutrition Examination Survey used AM1 and cut points similar to those of Freedson et al. to report the first nationally representative data of objectively measured PA in the United States (12,19).
In 2005, ActiGraph introduced the GT1M (AM2), a new model of accelerometer with numerous technological advances. Notably, AM2 contains a solid-state monolithic accelerometer and uses microprocessor digital filtering, replacing the AM1's piezoelectric bimorph beam accelerometer that uses analog circuit filtering. Detailed specifications of AM1 are provided elsewhere (13,20). The advantage of the AM2 solid-state device and digital filtering system is that, upon installing the accelerometer in the circuit, its response to the 1g acceleration of the earth is fixed and does not drift, thereby eliminating the need for unit calibration. In contrast, the analog components in AM1 fluctuate and thus require regular external calibration. There is four times more memory in AM2 compared with AM1, and AM2 can collect at least 2 wk of data between charges. The unit battery recharging, initialization, and data downloading for AM2 are performed through a USB port, whereas AM1 requires the use of a reader interface unit for accelerometer-to-computer communication. In order to enhance the reconstruction of the digital signal, AM2 samples 30 times per second and contains a 12-bit analog-to-digital converter. In comparison, AM1 samples 10 times per second and has an 8-bit analog-to-digital converter (1).
It is important to determine whether the monitor output is comparable between models to allow trend analysis of population-based PA levels and comparison across studies because both models are currently in use. Specifically, it is important to determine whether regression equations developed for AM1 may be applied to output from AM2 to yield valid estimates of time spent in various PA intensity categories. There are numerous technological advantages in AM2 compared with AM1, but information about the comparability of the two ActiGraph models is limited.
In one recent model comparison study, Rothney et al. (16) compared the intraunit and intermodel differences in count output using a mechanical oscillator. They reported significant differences between models in count output at low frequencies, different slopes at varying radii, and different slopes at all frequencies except 120 rpm. This type of testing provides precise control of condition parameters and simultaneous data collection for multiple monitors. However, human subject studies are needed because the mechanical oscillator properties are not completely representative of how these monitors behave when used in human subjects.
In a monitor model comparison study in human subjects, Corder et al. (5) compared both ActiGraph models in 30 adolescents and reported a high correlation (r = 0.95), similar average counts, and similar time spent in moderate and vigorous PA when both monitors were worn for 7 d (P > 0.05). Overall count output was 9% lower for AM2 compared with AM1, which was statistically significant. The estimate of time spent in sedentary behavior was significantly higher for AM2 compared with AM1, whereas time spent in light-intensity activity was significantly higher for AM1 (P < 0.05). In a sample of 16 young endurance-trained males, Fudge et al. (8) reported that count output for both AM1 and AM2 leveled off at fast running speeds. However, the speeds at which the plateau occurred were different between monitors (>14-16 km·h−1 for AM2 and >10-12 km·h−1 for the AM1).
This study compares the two models of ActiGraph accelerometer during self-paced locomotion of varying speeds among adults. The primary purpose of this study was to determine whether the count output and estimates of PA intensity levels were comparable between AM1 and AM2. As a secondary analysis, the step output was compared between models. We also examined how body mass index (BMI) and sex affect the monitor comparison.
Participants were recruited from the University of Utah and surrounding community. All participants read and signed an informed consent document. The study protocol and documents were approved by institutional review boards of the University of Utah, University of Massachusetts, Westat (a contractor for this project), and the National Cancer Institute (the study sponsor). All participants completed a health history questionnaire and had their resting blood pressure measured. Participants whose resting blood pressure was greater than 150 mm Hg systolic and/or greater than 100 mm Hg diastolic were ineligible to participate in the study. Height and body mass were measured for all qualified participants. The age of participants ranged from 17 to 74 yr, and mean ± SD BMI was 26.1 ± 5.44 kg·m−2. Complete participant characteristics are shown in Table 1.
Each participant wore two ActiGraphs: model AM1 and model AM2 (ActiGraph; LLC, Pensacola, FL). Ten units of each model were used for testing. Each AM1 was calibrated once before testing according to the manufacturer's recommendation, whereas AM2 does not require external calibration. Each ActiGraph was initialized to sample during 60-s epochs. The ActiGraph was threaded onto an elastic belt, which was securely positioned at the waist under the participant's clothing. The midpoint of one ActiGraph was positioned in line with the axilla. The second monitor was positioned adjacent and posterior to the first monitor. The order of monitor placement was alternated between subjects. Each participant completed three self-paced locomotion trials at slow, medium, and fast speeds around a 0.47-km indoor circular hallway (one lap per trial). Participants were instructed to maintain a comfortable, constant pace throughout each trial. To compute speed, the time of each trial was recorded using a stopwatch. The order of the trials was balanced among participants (e.g., order for subject 1 was slow, medium, fast; subject 2 medium, fast, slow; subject 3 fast, slow medium), and participants had a minimum of 3 min of rest between trials.
To compare monitor output per minute, the following data cleaning procedures were performed. The first minute and residual seconds per trial were removed, and the middle portion of data was analyzed (e.g., a trial of 7:22 min, minutes 2-6 were used in analyses). If the residual seconds were less than 10 s, then a full 2 min were removed at the end to ensure a steady walking pace was achieved (e.g., a trial of 5:01 min, minutes 2-4 were used). Data were excluded from analyses for 7 of the 108 subjects owing to the monitor not recording (n = 1) or failure to report start and stop times of trials (n = 6). We compared the count output (counts per minute) and step output (steps per minute) measured every 60 s for each model using repeated-measures mixed models. At each walking speed, we used the mixed model to assess the average difference between models (bias). Bias was assessed by using the difference between the outputs (AM2 − AM1) as the response in the mixed model. For each speed, the intercept in the model provides an estimate of bias. If the bias was positive, the AM2 output was higher on average than the AM1. Random effects specific to each participant were included in the model. Ninety-five percent confidence intervals (95% CI) from the mixed model were used to determine significance. If the 95% CI crossed 0, the difference was not statistically significant at α = 0.05. We also performed these analyses for steps per minute and count output on stratifications of the study sample, which were grouped by sex and weight status based on BMI (normal weight: BMI < 25 kg·m−2; overweight or obese: BMI ≥ 25 kg·m−2). If the 95% CI overlapped between groups, the difference was not statistically significant at α = 0.05.
Intensity levels (METs) and count category comparison tables were developed by cross-classifying minute-by-minute results from each model in two ways. First, we compared the number of minutes in arbitrarily chosen count ranges of 1000 (e.g., <1000 counts per minute, 1000-1999 counts per minute). Second, we compared the number of minutes classified in the same MET intensity category (i.e., light, moderate, and vigorous) as determined by the cut points of Freedson et al. (7). Finally, correlation coefficients for count output were computed for the models at each self-selected speed. All analyses were done using R software version 2.8 (http://www.r-project.org/).
When comparing the count output between models, a small but statistically significant bias of 2.7% (95% CI = 0.8%-4.7%) occurred across all speeds (range = 0.22-3.8 m·s−1; 0-14,593 counts per minute). This bias is equivalent to the AM2 recording 58 counts per minute higher than the average count value of 2152 counts per minute recorded by AM1 (Table 2). Overall, the AM1 and AM2 count outputs were highly correlated (r = 0.99; Fig. 1; P < 0.05).
The average count output for each model was compared at self-paced slow, medium, and fast speeds (Table 2). Additional analysis was done by absolute speed tertile, and the relationship between AM2 and AM1 output did not change. Thus, all values reported are based on self-selected speeds. The coefficient of variation for each model at each speed was less than 3.2%. The count output for AM2 relative to AM1 at each speed is presented in Table 3. The count output from the models was not significantly different at self-paced slow speeds or fast speeds. However, a significant positive bias occurred at medium speeds, where AM2 count output was 5.3% higher than that of AM1 (95% CI = 3.4-7.2). At all speeds, the monitors were significantly correlated (r = 0.96, 0.95, and 0.97 at slow, medium, and fast speeds, respectively; P < 0.05). Figure 2 shows monitor output comparison at slow speeds (<1000 counts per minute, <0.52 m·s−1). At very slow speeds, AM1 recorded counts when the AM2 recorded zeros.
Cross-classification of count output for the categorical ranges (e.g., <1000, 1000-1999 counts per minute) agreed 85.5% of the time. For 11.2% of the minutes, the AM2 output was higher than that of AM1. AM1 recorded higher output than AM2 for 3.2% of the minutes (Table 4). As shown in Table 5, when the monitors were cross-classified based on the Freedson cut points (7) to estimate activity intensity (i.e., light, moderate, and vigorous), the models agreed 96.1% of the time. Estimated activity intensity from AM2 was higher than AM1 for 2.3% of the minutes, whereas the AM1 estimated a higher-intensity category for less than 1% of the minutes recorded.
The step output was significantly higher for AM1 than AM2 across all speeds (bias = −19.8%, 95% CI = −23.2 to −16.4), which was due to a large bias at slow speeds (bias = −59.5%, 95% CI = −50 to −72.2). There was no difference in steps per minute between models for medium or fast speeds, as shown in Table 6. Figure 3 shows that the differences in steps per minute were large for speeds less than 0.89 m·s−1, whereas at faster speeds, there was no difference between models.
For the normal-weight group, AM2 recorded significantly higher count output than AM1 for both medium (bias = 5.7%, 95% CI = 3.4-0.8) and fast speeds (bias = 3.1%, 95% CI = 1.3-4.8). Within the overweight or obese group, the average count output for AM2 was higher at medium speeds (bias = 4.8%, 95% CI = 1.6-8.1) but not at other speeds. However, because the CI overlapped at all speeds, there were no significant differences between BMI groups. Count output was significantly higher with the AM2 for females at medium (bias = 7.4%, 95% CI = 4.3-10.5) and fast speeds (bias = 3.8%, 95% CI = 1.1-6.5). For males, AM2 counts were significantly higher only at medium speeds (bias = 3.0%, 95% CI = 1.0-5.0). Similar to BMI groups, the differences in count output between sexes were not statistically significant at any speed. Step output was significantly lower for AM2 at slow speeds for all sex and BMI groups. There were no differences in step output for medium or fast speeds by sex or BMI. There were no statistically significant differences between males and females or between BMI categories for step output.
Direct comparison of AM1 to AM2 during self-paced locomotion is important for researchers who wish to compare data using both models across studies. The main finding of this study was that a small difference in count output between models did not result in meaningful between model differences in time spent in different activity intensity categories. The primary application of monitor data is in surveillance and clinical studies to quantify time spent in moderate to vigorous PA. Thus, it appears that data reporting time in activity intensities collected using AM2 to objectively quantify PA can be compared to data collected using AM1.
Similar to what has been reported previously (5,16), AM2 required a higher acceleration to record a nonzero count than AM1. However, the precise cutoff between sedentary time and light activity between monitors cannot be determined because this investigation evaluated only the monitors during walking activities. Future research should focus on developing cut points for sedentary and light activity for both models to determine whether the extra acceleration required from AM2 to elicit a nonzero count corresponds to actual movement. For example, the higher filter in AM2 may not be sensitive enough to distinguish light intensity from sedentary behavior. This is an important finding considering recent evidence highlighting the relationship between time spent in light-intensity activity and positive health outcomes (9,12). Consistent with Rothney et al. (16), the current study revealed a crossover effect where AM2 counts were lower on average than AM1 at counts <1000 and then higher on average as count values increased. The 9% lower AM2 counts reported by Corder et al. (5) was potentially driven by this underestimation of AM2 counts at very low speeds. Thus, increasing AM2 count values by 9% as the authors suggest could be problematic, particularly in a highly active population, because the AM2 produced higher counts at medium and fast speeds.
We did not have a sufficient number of participants running at speeds greater than 14 km·h−1 to examine the leveling off of counts at high speeds reported by Fudge et al. (8). No bias occurred at the fast speeds in our sample, suggesting that the impact of differences in filtering procedures between generations of the ActiGraph is minimal during self-selected speeds for the general population. However, future investigations should compare the count output between models to determine where the linear relationship between speed and count output is compromised. Although it is important to determine whether the monitors have different responses at very high speeds, there may be little practical significance for population-level comparisons because the general population rarely participates in vigorous activity (19).
A surprising finding was the very large between model differences in step output at very low speeds. According to the manufacturers, the lower filtering range was changed from 0.21 Hz in AM1 to 0.25 Hz in AM2, which may explain this difference (1). Studies comparing AM1 to measured steps and to research grade pedometers have concluded that AM1 records extraneous steps and results in higher step counts than the criterion measures (11,21). This was recently supported using the National Health and Nutrition Examination Survey data collected with AM1 that reported extremely high average step counts for the population that are not supported in other samples (23). It is possible that the higher filter in AM2 reduces the error in step counts, explaining the differences we found in this investigation. However, we did not have a criterion measure of steps in our sample, so we can only conclude that the two models produce different step output and studies that compare steps per day using different models should be interpreted with caution. Future research should validate the step count feature in AM2 compared with measured steps and research quality pedometers. It should be noted that recent firmware updates may affect results, particularly at the low speeds where large step differences were observed.
Several limitations of this study should be noted. We evaluated the accelerometers with specific locomotion trials at three relatively constant self-selected speeds, which did not allow for comparisons between monitors during intermittent and more variable motion observed during many activities of daily living. However, because the majority of calibration studies have included primarily locomotion activities, this study design allowed us to examine the validity of previously published regression equations for two different models of the ActiGraph. The study was not designed to compare monitor models for sedentary and light-intensity movement. The higher threshold for nonzero counts in the AM2 compared with the AM1 may lead to underestimation of time spent in light-intensity activity or it may reduce the likelihood that noise is reported as a signal. Another limitation is that we did not obtain a criterion measure of steps, so we could not determine whether one model more accurately measured steps. However, the purpose of the study was to simultaneously compare the models to determine whether they can be compared across studies, and we did determine they have different responses at slow speeds.
This study also had important strengths. The sample was fairly large (n = 116) and included a wide range of ages (17-74 yr) and BMI (17-47 kg·m−2). The participants completed a wide range of speeds, consistent with what would be observed in free-living situations. In addition, both ActiGraph models were worn simultaneously, which allowed for within-individual comparisons.
The step output between models was not comparable at slow speeds, and comparisons of step data collected with both models should be interpreted with caution. The count output from AM1 was slightly but significantly higher than AM2 during self-paced locomotion. However, this difference in count output did not result in meaningful differences in time spent in various activity intensity classifications. Thus, data collected with AM1 appear to be comparable to AM2 for the count output when translating the data into estimates of PA intensity categories. Data collected on population levels of PA and across using both models are comparable across studies.
The authors thank Susie McNutt, Joan Benson, and the University of Utah graduate students for their assistance in data collection.
This project was funded by contract no. HHSN 261200700343P from the National Cancer Institute.
There is no conflict of interest. The results of the present study do not constitute endorsement by the American College of Sports Medicine.
1. ActiGraph. ActiSoft Analysis Software 3.2 User's Manual
. Fort Walton Beach (FL): MTI Health Services; 2005, p. 17.
2. Bassett DR Jr. Validity and reliability issues in objective monitoring of physical activity. Res Q Exerc Sport
3. Chen KY, Bassett DR Jr. The technology of accelerometry-based activity monitors: current and future. Med Sci Sports Exerc
. 2005;37(11 suppl):S490-500.
4. Corder K, Brage S, Ekelund U. Accelerometers and pedometers: methodology and clinical application. Curr Opin Clin Nutr Metab Care
5. Corder K, Brage S, Ramachandran A, Snehalatha C, Wareham N, Ekelund U. Comparison of two Actigraph models for assessing free-living physical activity in Indian adolescents. J Sports Sci
6. Crouter SE, Clowers KG, Bassett DR Jr. A novel method for using accelerometer data to predict energy expenditure. J Appl Physiol
7. Freedson PS, Melanson E, Sirard J. Calibration of the Computer Science and Applications, Inc. accelerometer. Med Sci Sports Exerc
8. Fudge BW, Wilson J, Easton C, et al. Estimation of oxygen uptake during fast running using accelerometry and heart rate. Med Sci Sports Exerc
9. Hamilton MT, Hamilton DG, Zderic TW. Role of low energy expenditure and sitting in obesity, metabolic syndrome, type 2 diabetes, and cardiovascular disease. Diabetes
10. Hendelman D, Miller K, Baggett C, Debold E, Freedson P. Validity of accelerometry for the assessment of moderate intensity physical activity in the field. Med Sci Sports Exerc
. 2000;32(9 suppl):S442-9.
11. Le Masurier GC, Tudor-Locke C. Comparison of pedometer and accelerometer accuracy under controlled conditions. Med Sci Sports Exerc
12. Matthews CE, Chen KY, Freedson PS, et al. Amount of time spent in sedentary behaviors in the United States, 2003-2004. Am J Epidemiol
13. Melanson EL Jr, Freedson PS. Validity of the Computer Science and Applications, Inc. (CSA) activity monitor. Med Sci Sports Exerc
14. Physical Activity Guidelines Advisory Committee Report to the Secretary of Health and Human Services. Part A: executive summary. Nutr Rev
15. Pober DM, Staudenmayer J, Raphael C, Freedson PS. Development of novel techniques to classify physical activity mode using accelerometers. Med Sci Sports Exerc
16. Rothney MP, Apker GA, Song Y, Chen KY. Comparing the performance of three generations of ActiGraph accelerometers. J Appl Physiol
17. Rothney MP, Neumann M, Beziat A, Chen KY. An artificial neural network model of energy expenditure using nonintegrated acceleration signals. J Appl Physiol
18. Swartz AM, Strath SJ, Bassett DR Jr, O'Brien WL, King GA, Ainsworth BE. Estimation of energy expenditure using CSA accelerometers at hip and wrist sites. Med Sci Sports Exerc
. 2000;32(9 suppl):S450-6.
19. Troiano RP, Berrigan D, Dodd KW, Masse LC, Tilert T, McDowell M. Physical activity in the United States measured by accelerometer. Med Sci Sports Exerc
20. Tryon WW, Williams R. Fully proportional actigraphy: a new instrument. Behav Res Methods Instrum Comput
21. Tudor-Locke C, Ainsworth BE, Thompson RW, Matthews CE. Comparison of pedometer and accelerometer measures of free-living physical activity. Med Sci Sports Exerc
22. Tudor-Locke C, Hatano Y, Pangrazi RP, Kang M. Revisiting "how many steps are enough?" Med Sci Sports Exerc
. 2008;40(7 suppl):S537-43.
23. Tudor-Locke C, Johnson WD, Katzmarzyk PT. Accelerometer-determined steps per day in US adults. Med Sci Sports Exerc
24. Ward DS, Evenson KR, Vaughn A, Rodgers AB, Troiano RP. Accelerometer use in physical activity: best practices and research recommendations. Med Sci Sports Exerc
. 2005;37(11 suppl):S582-8.