https://orcid.org/0009-0008-0514-7745
Figures
Abstract
Background
Accelerometers are widely adopted for physical activity (PA) measurement. Accelerometry data require pre-processing before entering formal statistical analyses. Many pre-processing criteria may influence PA outcomes and the processed sample, impacting results in subsequent statistical analyses.
Aim
To study the implications of pre-processing criteria for accelerometer data on outputs of interest in physical activity studies.
Methods
We used the ActiGraph hip-worn accelerometry data from 538 adult Latino participants. We studied four most important domains of pre-processing criteria (wear-time, minimum wear-time, intensity level, and modified bouts). We examined the true sample size in pre-processed data, the moderate-to-vigorous physical activity (MVPA) outcome, and regression coefficients of age and gender predicting MVPA.
Results
Many pre-processing criteria have minimum impact to the output of interest. However, requirements for minimum wear-time can have high influence on subsequent analyses for MVPA. High requirements for wear-time (e.g., minimum of 5 days with more than 12 hours of wear-time per day) lead to weakened statistical efficiency in estimating the relationship between potential predictors and the MVPA outcome. Intensity levels using vector magnitude triaxial counts yielded drastically different results than those using conventional vertical axis counts.
Conclusion
Moderate changes in minimum wear-time can yield notably different output data and subsequently influence analyses assessing the impacts of interventions on MVPA behaviors. Processed data using vector magnitude and conventional vertical axis counts are not directly comparable. Sensitivity analyses using alternative pre-processing scenarios are highly recommended to verify the robustness of analyses for accelerometry data.
Citation: Han B, Perez L, Cohen DA, Seelam R, Derose KP (2025) Influence of pre-processing criteria on analysis of accelerometry-based physical activity. PLoS ONE 20(1): e0316357. https://doi.org/10.1371/journal.pone.0316357
Editor: Zulkarnain Jaafar, Universiti Malaya, MALAYSIA
Received: June 28, 2024; Accepted: December 10, 2024; Published: January 2, 2025
Copyright: © 2025 Han et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting Information files.
Funding: This work is supported by a grant from the National Cancer Institute of the National Institutes of Health (grant no. R01CA218188). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Physical inactivity increases the risk and severity of multiple chronic diseases and is the underlying cause of over 10% of deaths in the United States [1, 2]. Accelerometers are widely used in research to measure human physical activity (PA) and are typically worn on the hip or the wrist [3]. Compared with other objective measurement instruments including lab-based techniques and direct observations, accelerometers are self-administered, more economical, scalable, and suitable for studying a large population longitudinally [4–9]. These unique features make accelerometers the preferred measurement instrument in many studies aimed to quantify individual-level PA behavior and to evaluate the effects of PA interventions. Accelerometers have been used as the gold standard to calibrate the less accurate self-reported PA measures [4, 5, 10].
Count-based accelerometry data are in the format of counts at the epoch level, where each epoch is usually 30 seconds or one minute, for the entire duration the accelerometer has been operating. The count data can contain many zeros, very low values, and erratic values because either the device has not been worn or the wearer has been sedentary or did not wear correctly. The count-based accelerometer data must use a pre-processing procedure to eliminate invalid or meaningless (i.e., non-wear) epochs and to convert counts to more interpretable outcomes, such as time spent in a certain PA intensity level.
The pre-processing procedure for count-based accelerometer data includes multiple steps: differentiating wear-time from non-wear-time; imposing minimum requirements for wear-time; defining PA intensity levels, in particular, the cutoff for moderate-to-vigorous physical activity (MVPA); and clustering PA in modified bouts. Each step consists of further technical details with alternative specifications [11]. Past health and behavioral studies using accelerometer data have not always provided details of their pre-processing steps. Many studies simply conformed to a convention adopted in prior studies [12].
The impact of the pre-processing procedure for count-based accelerometer data has been noted and investigated in the literature with mixed findings. Several studies acknowledged that choices of pre-processing criteria may influence statistical analyses further downstream [12–16]. For example, a recent study reported that by altering the MVPA intensity cutoff and using bouts vs. all minutes, the sample proportion meeting the national PA guidelines varied dramatically from 3% to 96% [17]. Variations in pre-processing criteria, such as wear-time validation algorithms, minimum requirement for wear-time, and cutoffs for PA levels, can lead to significant changes in both the mean time and the percent of time in every PA levels [15, 18, 19]. A prior study developed empirical approximation formulas for the MVPA outcome by epoch length and cutoff points with and without bouts [20]. Another study showed that different pre-processing criteria altered the significance level of the relationship between a socio-economic predictor and PA outcome, specifically from statistically significant (p < .001) to non-significant (p>.05) [14]. This last study is particularly concerning since many health studies aimed to estimate the relationship between predictors (e.g., a person-level characteristics, a group-level environmental factor, exposure to an intervention, etc.) and the MVPA outcome, rather than the absolute level of the MVPA outcome. This prior evidence suggests that pre-processing criteria can fundamentally influence the findings in such type of health studies. Contrary to these studies above, a study reported that various pre-processing criteria did not alter the findings of PA patterns among obese diabetic patients but only changed the sample size by about 10%, i.e., participants that did not meet the selected criteria were dropped from analyses [21]. Prior studies also have mixed views on using the vector magnitude using triaxial counts. While one study praised the vector triaxial counts for improving identification of valid wear-time compared with the standard vertical axis counts [22], another study did not find that the vector magnitude with triaxial counts can further improve predictive performance of counts for energy expenditure beyond using the vertical axis counts only [23]. Several studies warned to not directly compare data in vector magnitude across different types of accelerometers [24, 25].
In this paper, we aim to investigate multiple important pre-processing criteria by a comprehensive study design. We study the true sample size in pre-processed data, the absolute level of the MVPA outcome, and regression coefficients of age and gender predicting MVPA. The present study describes a systematic method that can be adopted by other researchers as a sensitivity analysis to guide decisions for pre-processing and analyzing accelerometer data. Our findings can also provide useful reference for future health studies using accelerometer to set up their pre-processing procedures.
Method
Ethics statement
We used primary data collected from an ongoing cluster randomized controlled trial, i.e., the parent study. The parent study was approved by the Human Subjects Protection Committee (Assurance No. FWA00003425, IRB No. IRB00000051 [2018]), which is the institutional review board to review research involving human subjects at the RAND Corporation. The parent study was registered in ClinicalTrials.gov (NCT03858868). The goal of the parent study is to promote PA among adult churchgoing Latinos from six churches in predominantly-Latino neighborhoods in and near East Los Angeles, California [26, 27]. All study participants were adult, from whom written consents were collected before data collection.
Sample and data
Potentially eligible participants from the six participating churches were screened for a history and symptoms of health conditions that could preclude PA [28]. Upon consent, participants eligible for the trial were asked to wear an ActiGraph wGT3X-BT activity monitor around their hip for a week, removing it only during water activities (e.g., showering) and when sleeping. To reduce the likelihood of draining batteries in the field, the accelerometers were set with the sampling rate of 30 Hz, the sleep mode enabled, and the inertial measurement unit sensor disabled. The low frequence extension option was not used since the parent study is focused on MVPA. The primary data consisted of 538 individuals and data were collected in 2019. Table 1 describes the characteristics of the sample. We processed the data using ActiLife software v6.13.4 (ActiGraph, Pensacola, FL) and analyzed data using SAS 9.4 and R 4.1.
Detailed steps in a pre-processing process
A pre-processing procedure can be implemented in a general-purpose statistical software package such as SAS and R. Researchers of ActiGraph accelerometers often use the software ActiLife specialized for pre-processing and analyzing ActiGraph accelerometry data. In this paper we worked with the ActiLife software tool. A pre-processing process can involve many steps and details. In this paper we selected the most important or required steps to study. We used the following key terms to organize these steps in line with the interface in ActiLife: domain, criterion, and option. From the top, a processing procedure is split into four sequential domains, where each domain has a distinct task: validation domain 1 (to find valid wear-time), validation domain 2 (to apply minimum valid wear-time requirement), scoring domain 1 (to implement MVPA intensity thresholds), and scoring domain 2 (to identify modified bouts by MVPA intensity levels). Each domain consists of multiple criteria, where a criterion is a specific decision users need to make. A criterion usually has two or more options, from which the user must choose one. Most criteria have one option set as default in ActiLife. Table 2 lists all domains, criteria, and options under consideration in this paper, which are used throughout the remainder of this paper. Lastly, we further define the term scenario: a scenario is a unique set of options selected across all criteria of the four domains. Two scenarios are different if at least one criterion has different options selected between the two scenarios.
Validation domains 1: To find wear-time.
Non-wear-time includes periods when the device was not worn by a participant during the indicated timespan. In this paper, we considered two criteria in validation domain 1: an algorithm to identify valid wear-time and the vector magnitude for counts. First, there are two built-in options in ActiLife to identify non-wear-time, referred to as Troiano_2007 and Choi_2011 in this paper [29–31]. These two options aim to identify patterns in the count data that match with pre-defined non-wear-time patterns with different algorithms. One notable difference is that Troiano_2007 defines a non-wear period as at least 60 consecutive minutes of zero count values whereas Choi_2011 uses a minimum of 90 consecutive minutes. Thus, Choi_2011 may identify shorter non-wear-time than Troiano_2007. The second criterion is referred to as “vector magnitude” with two options. The default option, i.e., “No”, denoted as requests that only the counts on the vertical axis of an accelerometer is used as input data for the selected algorithm above. The alternative option, i.e., “Yes”, uses a Euclidean-type metric combining triaxial counts [11, 23, 24].
Validation domains 2: To apply minimum wear-time requirement.
A minimum length of wear-time is necessary to reliably assess PA patterns at the daily or person level [29, 32, 33]. First, a “valid day” needs to exceed a minimum wear-time threshold (e.g., 10 hours/day). A more stringent threshold for defining a valid day (e.g., 12 hours/day) will result in fewer valid days and potentially a smaller sample size in the analytic dataset. Next, a study participant may need to have a minimum number of valid days to be included in analysis. This requirement is due to the high inter-day variations in a person’s PA. Researchers may be interested in a person-level outcome, e.g., whether a person meets the national guideline of weekly MVPA time. Ideally, they may wish to have at least one entire week of valid days to assess this outcome for a participant. However, due to participant non-compliance, the actual data collected are usually much less than ideal. The minimum number of valid days (e.g., an entire week) may screen out participants with too few valid days and could potentially bias the sample by only including the most adherent participants. Lastly, the minimum wear-time requirements can also specify a minimum number of valid weekdays and weekend days to capture greater variability in PA levels across the week.
For this domain, we compare three options: minimum time for a valid day (6, 8, 10, 12 hours of valid wear-time), minimum valid days for a participant (1, 3, 5 valid days), and minimum valid days during weekends (0 or 1 valid day), where 0 means no requirement for a valid weekend day.
Scoring domain 1: MVPA intensity threshold.
After processing the validation domains for valid wear-time and minimum time, the count data are processed to differentiate activity intensity levels, which are more meaningful and interpretable for analyses. Scoring domain 1 contains pre-defined thresholds on accelerometer counts to define intensity levels; we focused on MVPA given this intensity level is associated with multiple health benefits and is the focus of national PA guidelines. The three options to identify adult MVPA are built in ActiLife, referred to as V_1952, V_2020, and VM3_2690 hereafter. Specifically, V_1952 defines threshold for MVPA at 1,952 counts per minute on the vertical axis [34], V_2020 uses 2,020 counts per minute on the vertical axis [29], and VM3_2690 defines MVPA as 2,690 counts per minute based on the vector magnitude tri-axial counts [24]. The threshold for V_2020 is a more stringent requirement than V_1952. The VM3_2690’s cutoff value is based on vector magnitude counts and not comparable with the other two options that are based only on the vertical axis.
Scoring domains 2: Modified bouts.
Bouts are a consecutive period of time during which the PA intensity level is consistent. Arguably, PA times accrued by bouts are robust to temporary changes in a person’s PA levels or abrupt measurement errors. While all three pre-processing domains discussed previously are mandatory, this domain is optional depending on one’s interest in bouts. Modified bout definitions can also be customized in ActiLife: length of a bout, the logic to break an ongoing bout, and the MVPA threshold to define a bout. In this paper, we used the default logic to break a bout and the default minimum length of 10 minutes [29]. We considered a single criterion with three options in this domain, corresponding to three options under scoring domain 1. When MVPA intensity levels are defined by a certain threshold, bouts should also be based on the same threshold.
Statistical analysis design
We tested various pre-processing scenarios using the design of a numerical experiment, where a criterion with multiple options can be seen as a design factor with multiple factor levels in an experiment. The full-factorial experiment design has a total of 864 scenarios by enumerating all combinations of options in all criteria across the four domains. In this paper, we elected to use a partial factorial design with a subset of scenarios. The partial factorial design was preferred for two reasons. First, the empirical evidence from the partial factorial design with a few representative scenarios was sufficiently strong to show the influence of a single criterion. Second, many scenarios in the full factorial design are unrealistic. For example, some conceptual scenarios can use one cutoff for defining MVPA intensity level but a different cutoff for defining MVPA bouts. These counterintuitive scenarios are neither practical nor interpretable. Our partial factorial design selected a few reasonable and representative scenarios. We first studied the influence of validation domain 1 (valid wear-time), by specifying options in the other three domains. Next, we studied the influence of scoring domain 1 and 2 by specifying the options in validation domains 1 and 2. Scoring domains 1 and 2 were studied together, because both scoring domains should always use the same MVPA cutoff values. Lastly, we studied the influence of validation domain 2 (minimum wear-time), by fixing options in the other three domains. Table 3 lists all the selected scenarios in details.
Study outcomes for measuring sensitivity in pre-processing scenarios
To examine the potential influence of different pre-processing scenarios in the analysis of accelerometry data, we examined multiple outcomes. First, we examined changes in sample size after going through the two validation domains. We examined the number of participants with sufficient valid wear days, the total number of valid days, and the average number of valid days per participant, all of which reflect the concept of effective sample size in a hierarchical model.
Second, we examined the absolute MVPA level. This outcome is directly affected by the two scoring domains but may also vary by the two validation domains. For each scenario, we calculated the mean and standard deviation (SD) of the MVPA time at the person-day level. We also calculated the percent of participants meeting national MVPA guidelines of 150 minutes/week by multiplying the daily MVPA by 7 and compared with the guideline. Except for studying scoring domain 2 (modified bouts), the daily MVPA outcome always refers to all time above the selected MVPA threshold. Wherever applicable, we conducted paired t-tests to assess the statistical significance of the difference in daily MVPA between comparable scenarios.
Third, we examined the regression coefficients from a multiple regression model. The model regressed the daily MVPA time versus two predictors: age in years and gender (female versus male), whose coefficient estimates serve as our study outcomes. Other covariates in the regression models included body mass index (BMI) and average wear-time per day. However, these latter two covariates were mostly not significant in the models and excluding them did not affect the findings. We did not further adjust for other possible covariates because we were not interested in the actual estimate of any association but instead to observe their variations by pre-processing scenario.
Results
Influence by validation domain 1 (valid wear-time)
Table 4 shows the results for studying the influence of varying validation domain 1 (valid wear-time), holding options of the other three domains constant. The four scenarios give generally similar results. Using the same vector magnitude option, the Choi_2011 algorithm always yielded a slightly larger sample size compared to using Troiano_2007. For example, when not using the vector magnitude option, Choi_2011 resulted in 449 eligible participants, 2,443 valid person-days, and an average of 5.57 valid days per eligible participants. By contrast, Troiano_2007 yielded 433 eligible participants, 2,343 valid person-days, and an average of 5.41 valid days per eligible participant. However, within the same algorithm, using the vector magnitude option gives slightly larger sample sizes than using the vertical axis only. For example, using the Troiano_2007 algorithm, with the vector magnitude, the sample size was 443 eligible participants whereas it was 433 with the vertical axis (similar results observed for valid person-days and average valid days per eligible participant).
MVPA times are all minutes above threshold a.
The modest differences in sample sizes by changing the criteria for domain 1 do not yield notable changes in the outcomes of MVPA time (Table 4). The mean daily MVPA time is between 25.66 and 26.07 minutes per day, and the SD of daily MVPA time is between 20.68 and 21.06 minutes. Differences across the four scenarios are very small. Age is consistently negatively related to daily MVPA time: one year increase in age is associated with a reduction of 0.33 minutes in daily MVPA time (p<0.001). Women have on average 10 fewer minutes per day of MVPA than men (p<0.001).
Influence by scoring domain 1 (MVPA intensity threshold)
Table 5 shows the three scenarios comparing the influence of varying MVPA intensity threshold when the other criteria are held constant. The V_2020 option resulted in less average daily MVPA time than V_1952 (24.35 vs. 26.07 minutes per day, respectively, p = 0.03) and a smaller SD (20.24 vs. 21.06 minutes per day, respectively). Consequently, V_2020 resulted in fewer people meeting the MVPA guideline (46.9%) than V_1952 (49.9%). V_2020 also had slightly smaller absolute values in regression coefficients than V_1952 (coefficients of age: -0.31 vs. -0.33, and coefficients of gender: -10.3 vs. -10.9). Despite the consistent pattern, the differences between V_1952 and V_2020 in our study outcomes seem to be small or very small.
MVPA times are all minutes above threshold. (n = 433 eligible patients, 2,343 valid person-days) a.
The VM3_2690 option, based on the vector magnitude, yielded very different results than the two options using vertical axis counts. Compared with V_1952 and V_2020, the mean and SD of daily MVPA time increased from 60 to 80% (20 more minutes per day in sample mean and 12 minutes per day in SD, approximately). Consequently, approximately an additional 25% of participants were deemed as meeting the national MVPA guidelines. The absolute values of regression coefficients of age and gender were also nearly doubled, compared with the previous two options.
Influence by scoring domain 2 (modified bouts)
Table 5 shows three scenarios for MVPA time by modified bouts. These three scenarios are the same as in Table 5, except that the MVPA outcomes are based on 10-minute bouts rather than using all minutes. Table 6 shows a pattern similar to Table 5. The two options using vertical axis counts, Bout_1952 and Bout_2020, produced almost identical MVPA outcomes and regression coefficients. Bout_2690 using vector magnitude had much larger mean and SD in daily MVPA time, as well as substantially larger regression coefficients in absolute values. Notably, the two options using vertical axis counts did not have significant regression coefficients (p>0.05), which indicates a great reduction of statistical power.
Influence by validation domain 2 (minimum wear-time)
Figs 1–3 present the results of the 24 scenarios studying the influence of varying the validation domain 2 (minimum wear-time), when other criteria were held constant.
x-axis marks the four levels of minimum wear-time per day, legends mark the three numbers of valid days, and colors differentiate requirements for valid day on weekends. These are scenarios 11–34 in Table 3.
x-axis marks the four levels of minimum wear-time per day, legends mark the three levels of minimum valid days, and colors differentiate requirements for valid weekend day. These are scenarios 11–34 in Table 3.
x-axis marks the four levels of minimum wear-time per day, legends mark the three levels of minimum valid days, and colors differentiate requirements for valid weekend day. Dashed bars are 95% CI. These are scenarios 11–34 in Table 3.
Fig 1 shows that using more stringent requirements—i.e., longer minimum wear-time per day, more total valid days, or at least one valid weekend day—resulted in substantially smaller sample sizes than options with looser requirements. The number of eligible participants ranged between 157 (with the strictest requirements) and 557 (with the loosest requirements) and the number of valid person-days ranged between 953 (strictest) and 3,225 (loosest). Not only was the range sizable, but even a seemingly small change in a pre-processing option appeared to greatly alter the sample size. For example, when requiring 12 hours of wear-time and 3 valid days, further requiring at least one valid day on weekends reduced the sample sizes from 349 participants with 1,694 valid person-days to 220 participants with 1,183 valid person-days. When requiring 5 valid days per participant and no requirements on weekends, changing from 10 hours of minimum daily wear-time to 12 hours reduced the sample size from 318 participants and 1,933 valid person-days to 190 participants with 1,127 valid person-days.
Fig 2 shows that the daily MVPA time outcome differed notably across the 24 potential scenarios. Higher minimum wear-time requirements resulted in higher mean daily MVPA times. The SD of daily MVPA time varied moderately across scenarios with a less clear pattern. Based on a crude SD of 20 minutes per day, differences in mean daily MVPA time between any two scenarios had a standardized effect size of 0 to .25 times SD. The proportions meeting national guidelines were higher in most stringent scenario (54.1%) than the least stringent (44.2%).
Fig 3 shows that the inference precision, e.g., the width of the 95% confidence interval (CI), differed substantially across scenarios, which was unsurprising given the largely different sample sizes across scenarios (Fig 1). Point estimates for age as a predictor displayed a moderate level of variation across scenarios. By contrast, point estimates for gender, a relatively stronger predictor than age, showed a small level of variation across scenarios. Significance levels for both predictors varied across scenarios. For example, the association with age was highly significant (p<0.001) in all scenarios requiring 10 hours minimum wear-time per day but non-significant (p>0.05) in the scenario requiring 12 hours minimum wear-time per day and at least 5 days with no requirements on weekends. Generally, the relatively stringent scenarios had wider CI but often bigger effect sizes, the latter partially compensating the loss of precision, compared with less stringent scenarios.
Discussion
Our findings highlight the presence and the lack of influence of various pre-processing criteria on accelerometry-assessed PA outcomes. Our study is also among the first to investigate this topic in a community sample of Latino adults. Our main findings include the following. First, numeric algorithms and the decision to use the vector magnitude for identifying valid wear-time did not seem to result in notable influences on all study outcomes. Second, our data suggests that MVPA intensity thresholds based on vertical axis and tri-axial counts are not comparable, although each set of thresholds was independently validated in the literature. The prevalence of meeting PA guidelines varied greatly from very low to moderately high depending on the pre-processing scenarios, which is consistent with prior reports [8, 17]. This finding is concerning since meeting PA guidelines is an important health outcome. Lastly, the requirement for minimum wear-time played an important role influencing downstream statistical analyses. Key take-aways from our analyses are that 1) minimum wear-time requirements are the most influential criteria and thus need to be reported clearly in the methods of research studies; 2) predictors with weak to moderate levels of association with MVPA outcomes, such as age, may be more influenced by the pre-processing criteria than those with strong associations, such as gender. This finding is in agreement with prior studies on the importance of wear-time requirements [18, 19]. Therefore, high requirements for wear-time are not always helpful in studying the associations between MVPA and predictors.
Given these lessons learned, sensitivity analyses based on alternative pre-processing criteria, in particular, minimum wear-time requirement settings, are highly recommended to check the robustness of the subsequent statistical estimates based on pre-processed data. Although such sensitivity analysis cannot establish the validity for any pre-processing criteria, it is informative in the sense of revealing the extent to which the subsequent statistical estimates depend on specific pre-processing choices. For example, in an imaginary study for the association between the MVPA outcome and a particular predictor, the estimated relationship changes its significance level between alternative pre-processing criteria (cf. Fig 3). The presence of this level of sensitivity should be reported as a limitation of the study. Vice versa, if the estimates remain relatively stable between alternative pre-processing criteria, the robustness of results may be claimed as a strength.
It is also worth noting key differences when using modified bouts versus not using them. Since modified bouts are a more stringent requirement to define MVPA than using all available valid wear-time, this resulted in lower MVPA. Our study suggests that MVPA outcomes based on unbouted versus bouted data are essentially two distinct concepts, each of which have their own absolute levels, SD, and associations with potential predictors. Their numeric values, statistical inference, and interpretations are not comparable, even when other pre-processing criteria are the same.
This study is subject to several limitations. First, analyses were based on a single and unique primary dataset. The sampled participants are based on an understudied population and results should be interpreted with caution. The primary data were collected by study participants’ compliance to receive, wear, and return the accelerometer devices. Failure to comply with the basic protocol (e.g., worn by more than one individual) can result in biases in the data. Second, we only used one type of accelerometer with one wear position (hip) and epoch length of 30 seconds. We were not able to study sensitivity due to different wearing position or different epoch lengths. Third, pre-processing was conducted in the proprietary software ActiLife. Despite its powerful functionality, friendly user interface, and popularity among various users, ActiLife is specialized and does not have full data processing capacity. Certain technical issues can be better addressed in a general-purpose statistical software, such as the GGIR library in R. Fourth, pre-processing criteria, in particular, minimum wear-time, can impact measurements of sedentary behavior and light PA measurements, just as they do on measuring MVPA. However, the cutoff for light PA is less well-established compared with the MVPA cutoff. The lower bound for sedentary behavior also overlaps with the algorithm for assessing valid wear-time. These important issues are beyond the scope of the current paper. Future studies are still needed to continue the investigation.
Despite these limitations, our results demonstrate that all pre-processing criteria used in analyzing accelerometry data need to be carefully considered, well understood, and well documented. Certain criteria may be preferred to alternative criteria for some scientific or other reasons. For example, one may prefer Choi_2011 to Troiano_2007 for the latter algorithm’s more comprehensive scan over time. Another example is the wear-time requirement of 12 hours for three days, which was used in the widely cited analyses for NHANES accelerometry data. Some may argue all time above threshold makes more sense than time in bouts. While we are not positioned to judge which criteria are more scientifically sound, it is clear that when comparing a study’s results with prior studies, one must be sure to only compare results generated by the same or very similar pre-processing criteria. Results from studies with very different pre-processing procedures are incomparable; improper comparisons can contribute to misleading conclusions. Sensitivity analyses are necessary to check the influence of selected and alternative pre-processing criteria in a study’s main findings. Before presenting any statistical estimates, specific pre-processing criteria that led to the conclusions need to be clearly presented.
Conclusion
Pre-processing steps for accelerometer data can influence the effective sample size and the magnitude of the MVPA outcome as well as its association with other predictors. Moderate changes in minimum wear-time can yield notably different output data and subsequently influence analyses assessing the impacts of interventions on MVPA behaviors. Processed data using vector triaxial magnitude and conventional vertical axis counts are not directly comparable. Sensitivity analyses using alternative pre-processing scenarios are highly recommended to verify the robustness of analyses for accelerometry data. Between-study comparisons should be based on the same or very similar choices in pre-processing criteria.
Supporting information
S1 File. Deidentified pre-processed analytic data are included in the file “S1 File.zip”.
https://doi.org/10.1371/journal.pone.0316357.s001
(ZIP)
References
- 1. Lee I-M, Shiroma EJ, Lobelo F, Puska P, Blair SN, Katzmarzyk PT. Effect of physical inactivity on major non-communicable diseases worldwide: an analysis of burden of disease and life expectancy. Lancet. 2012;380(9838):219–29. pmid:22818936; PubMed Central PMCID: PMC3645500.
- 2. U.S. Department of Health and Human Services. 2008 Physical Activity Guidelines for Americans. 2008.
- 3. Troiano RP, McClain JJ, Brychta RJ, Chen KY. Evolution of accelerometer methods for physical activity research. British Journal of Sports Medicine. 2014;48(13):1019–23. pmid:24782483
- 4. Kohl HW III, Fulton JE, Caspersen CJ. Assessment of physical activity among children and adolescents: a review and synthesis. Preventive Medicine. 2000;31(2):S54–S76.
- 5. Sirard JR, Pate RR. Physical activity assessment in children and adolescents. Sports Medicine. 2001;31(6):439–54. pmid:11394563
- 6. Westerterp KR. Assessment of physical activity: a critical appraisal. European Journal of Applied Physiology. 2009;105(6):823–8. pmid:19205725
- 7. Warren JM, Ekelund U, Besson H, Mezzani A, Geladas N, Vanhees L. Assessment of physical activity–a review of methodologies with reference to epidemiological research: a report of the exercise physiology section of the European Association of Cardiovascular Prevention and Rehabilitation. European Journal of Preventive Cardiology. 2010;17(2):127–39. pmid:20215971
- 8. Freedson P, Bowles HR, Troiano R, Haskell W. Assessment of physical activity using wearable monitors: recommendations for monitor calibration and use in the field. Medicine & Science in Sports & Exercise. 2012;44(S1):S1–S4.
- 9.
Ong L, Blumenthal JA. Assessment of physical activity in research and clinical practice. Handbook of Behavioral Medicine: Springer; 2010. p. 31–48.
- 10. Ainsworth BE, Caspersen CJ, Matthews CE, Mâsse LC, Baranowski T, Zhu W. Recommendations to improve the accuracy of estimates of physical activity derived from self report. Journal of Physical Activity and Health. 2012;9(s1):S76–S84. pmid:22287451
- 11. ActiGraph Software Department. ActiLife 6 User’s Manual 2012. Available from: https://s3.amazonaws.com/actigraphcorp.com/wp-content/uploads/2018/02/22094137/SFT12DOC13-ActiLife-6-Users-Manual-Rev-A-110315.pdf.
- 12. Migueles JH, Cadenas-Sanchez C, Ekelund U, Delisle Nyström C, Mora-Gonzalez J, Löf M, et al. Accelerometer data collection and processing criteria to assess physical activity and other outcomes: a systematic review and practical considerations. Sports Medicine. 2017;47(9):1821–45. pmid:28303543
- 13.
Maziewski P, Kupryjanow A, Kaszuba K, Czyżewski A, editors. Accelerometer signal pre-processing influence on human activity recognition. Signal Processing Algorithms, Architectures, Arrangements, and Applications SPA 2009; 2009: IEEE.
- 14. Lee PH. A sensitivity analysis on the variability in accelerometer data processing for monitoring physical activity. Gait & Posture. 2015;41(2):516–21. pmid:25540989
- 15. Masse LC, Fuemmeler BF, Anderson CB, Matthews CE, Trost SG, Catellier DJ, et al. Accelerometer data reduction: a comparison of four reduction algorithms on select outcome variables. Medicine & Science in Sports & Exercise. 2005;37(11):S544.
- 16. Evenson KR, Terry Jr JW. Assessment of differing definitions of accelerometer nonwear time. Research Quarterly for Exercise and Sport. 2009;80(2):355–62. pmid:19650401
- 17. Zenko Z, Willis EA, White DA. Proportion of adults meeting the 2018 physical activity guidelines for Americans according to accelerometers. Frontiers in Public Health. 2019:135. pmid:31231627
- 18. Banda JA, Haydel KF, Davila T, Desai M, Bryson S, Haskell WL, et al. Effects of varying epoch lengths, wear time algorithms, and activity cut-points on estimates of child sedentary behavior and physical activity from accelerometer data. PloS one. 2016;11(3):e0150534. pmid:26938240
- 19. Herrmann SD, Barreira TV, Kang M, Ainsworth BE. Impact of accelerometer wear time on physical activity data: a NHANES semisimulation data approach. British journal of sports medicine. 2014;48(3):278–82. pmid:22936409
- 20. Orme M, Wijndaele K, Sharp SJ, Westgate K, Ekelund U, Brage S. Combined influence of epoch length, cut-point and bout duration on accelerometry-derived physical activity. International Journal of Behavioral Nutrition and Physical Activity. 2014;11:1–8.
- 21. Miller GD, Jakicic JM, Rejeski WJ, Whit‐Glover MC, Lang W, Walkup MP, et al. Effect of varying accelerometry criteria on physical activity: the look ahead study. Obesity. 2013;21(1):32–44. pmid:23505166
- 22. Choi L, Ward SC, Schnelle JF, Buchowski MS. Assessment of wear/nonwear time classification algorithms for triaxial accelerometer. Medicine & Science in Sports & Exercise. 2012;44(10):2009. pmid:22525772
- 23. Howe CA, Staudenmayer JW, Freedson PS. Accelerometer prediction of energy expenditure: vector magnitude versus vertical axis. Medicine & Science in Sports & Exercise. 2009;41(12):2199–206.
- 24. Sasaki JE, John D, Freedson PS. Validation and comparison of ActiGraph activity monitors. Journal of Science and Medicine in Sport. 2011;14(5):411–6. pmid:21616714
- 25. Keadle SK, Lyden K, Hickey A, Ray EL, Fowke JH, Freedson PS, et al. Validation of a previous day recall for measuring the location and purpose of active and sedentary behaviors compared to direct observation. International Journal of Behavioral Nutrition and Physical Activity. 2014;11:1–11.
- 26. Derose KP, Cohen DA, Han B, Arredondo EM, Perez LG, Larson A, et al. Linking churches and parks to promote physical activity among Latinos: Rationale and design of the Parishes & Parks cluster randomized trial. Contemporary Clinical Trials. 2022;123:106954.
- 27. Perez LG, Cohen DA, Seelam R, Han B, Arredondo EM, Castro G, et al. Church contextual factors associated with Latinx physical activity and park use. Family & Community Health. 2022;45(3):163–73. pmid:35536714
- 28. Riebe D, Franklin BA, Thompson PD, Garber CE, Whitfield GP, Magal M, et al. Updating ACSM’s Recommendations for Exercise Preparticipation Health Screening. Medicine & Science in Sports & Exercise. 2015;47(11):2473–9. Epub 2015/10/17. pmid:26473759.
- 29. Troiano RP, Berrigan D, Dodd KW, Masse LC, Tilert T, McDowell M. Physical activity in the United States measured by accelerometer. Medicine & Science in Sports & exercise. 2008;40(1):181. pmid:18091006
- 30. Troiano RP. A timely meeting: objective measurement of physical activity. Medicine & Science in Sports & Exercise. 2005;37(11):S487.
- 31. Choi L, Liu Z, Matthews CE, Buchowski MS. Validation of accelerometer wear and nonwear time classification algorithm. Medicine & Science in Sports & Exercise. 2011;43(2):357.
- 32. Rich C, Geraci M, Griffiths L, Sera F, Dezateux C, Cortina-Borja M. Quality control methods in accelerometer data processing: defining minimum wear time. PloS one. 2013;8(6):e67206. pmid:23826236
- 33. Herrmann SD, Barreira TV, Kang M, Ainsworth BE. How many hours are enough? Accelerometer wear time may provide bias in daily activity estimates. Journal of Physical Activity and Health. 2013;10(5):742–9. pmid:23036822
- 34. Freedson PS, Melanson E, Sirard J. Calibration of the computer science and applications, inc. accelerometer. Medicine & Science in Sports & Exercise. 1998;30(5):777–81. pmid:9588623