Journal Logo

SPECIAL COMMUNICATIONS: Methodological Advances

Comparison of Accelerometry Methods for Estimating Physical Activity


Author Information
Medicine & Science in Sports & Exercise: March 2017 - Volume 49 - Issue 3 - p 617-624
doi: 10.1249/MSS.0000000000001124
  • Open


Physical inactivity is the fourth leading contributor to death in the United States (5). Most of the evidence to support a relationship between physical activity and health as well as the guidelines to improve physical activity are derived from self-reported data (3,11,17,30). Self-reported data may be subject to bias and recall problems. Further, self-reported data do not allow researchers to explore precise intensities and timing of physical activity. A growing number of researchers have been using objective measures of physical activity, including person-worn devices such as accelerometers.

Hip, rather than wrist-worn, accelerometers have been traditionally used to assess physical activity because the trunk location was thought to provide a more accurate assessment of center of mass acceleration (4). Cut points to categorize the intensity of physical activity for hip-worn accelerometers were developed based on laboratory protocols of physical activity and energy expenditure. These cut points used proprietary “count” data from the vertical x-axis. Early studies of wrist-worn accelerometers using these “counts” found that the wrist location poorly predicted energy expenditure (25).

The technology and application of accelerometers in physical activity research is rapidly changing (19). Raw acceleration data are now available for three axes, novel computational techniques that move away from traditional count-based classification approaches are emerging (24), and several researchers have advocated placing accelerometers at the wrist to estimate physical activity because estimates are improving and more wrist-worn devices are becoming available (28). Although some of these developments seem likely to improve physical activity classification, they also present several logistical and analytic challenges.

To date, most of the new computational algorithms available for researchers have been developed for both hip and wrist-worn accelerometers using data from individuals in laboratory settings (6,23). One recent study, in the laboratory, compared different computation techniques (14). However, movement in the laboratory setting may not reflect free-living physical activity behavior because laboratory-based movements generally occur in sequences defined by the investigator and do not include outdoor activities such as driving and bicycling. Thus, it is not clear if these algorithms are applicable to epidemiological cohorts and intervention studies in free-living adults (2,15). In one study, the performance of wrist- and hip-based algorithms from both laboratory and free-living training data were compared (7). The free-living algorithms to detect behaviors such as walking and sitting performed better than the laboratory-based algorithms, and the hip algorithms performed better than the wrist algorithms when the behaviors were known from annotated observations of images collected by person-worn cameras (7). Other studies have shown little difference in hip- and wrist-based estimates of physical activity in laboratory settings (6,23).

Wrist-based accelerometers are now being used in a variety of studies to improve participant compliance, where compliance is defined as the total number of hours collected and number of people meeting valid day criteria (27,28). For example, the 2011–2014 National Health and Nutrition Examination Survey (NHANES) survey with wrist placement had substantially higher compliance than the 2003–2006 NHANES with hip placement (27). Higher compliance rates for wrist-worn devices might be partly due to the wrist location being preferred by users because it is more comfortable. Also, when participants are asked to wear the devices for 24 h·d−1, wrist-based devices are less likely to be removed at night and not replaced in the morning. In most hip-based accelerometer studies, participants have been instructed to wear the hip accelerometer during waking hours only. Some hip-based studies have improved compliance by using 24-h protocols and have shown that participants are able to wear hip-worn devices overnight (8,13). Currently, it is unclear whether the wrist or hip location results in the most complete and accurate 24-h measurement of physical activity and sedentary behavior. Compliance may be higher with wrist placement, but some wrist-based algorithms appear to be less accurate in classifying activity (7). Another reason to consider the wrist placement for 24-h activity monitoring is because it has been shown to provide a better approximation of patterns of polysomnography-defined sleep than the hip location (21), although few studies have attempted to use the hip location for sleep assessment or improve sleep estimates based on hip wear (16).

The current study studied hip- and wrist-based accelerometers in a free-living cohort of middle-age to older women across the United States, including working adults, nurses, and breast cancer survivors. Novel machine-learned (ML) algorithms were applied to the raw data and compared with traditional count cut points and laboratory-based algorithms using raw vector magnitude. This article aims to identify differences in physical activity and sedentary behavior estimates between the wear locations, wear time protocols, and data processing techniques. Analyses will inform comparisons of accelerometer data across studies where different methods are likely to be used.



Participants were recruited from four university sites (Harvard University; University of California, San Diego [UCSD]; University of Pennsylvania; and Washington University in St. Louis) involved in the National Cancer Institute–funded Transdisciplinary Research in Energetics and Cancer initiative (20). The sample included working adults (UCSD and Washington University in St. Louis), nurses (Harvard University), and breast cancer survivors (University of Pennsylvania). The institutional review boards at each university site approved the study protocol and consent forms, and all participants provided written informed consent. Study protocols specific to the collection of accelerometer data were identical across all sites, and all data used in these analyses were centrally pooled and uniformly processed at UCSD.


All participants wore the ActiGraph GT3X+ accelerometer for waking hours simultaneously on the hip and for 24 h on the wrist. Monitors were mailed to 87% of participants with clear instructions, a link to an instructional video, followed by two check-in phone calls. The remaining participants received and returned the devices in person. Because participants also wore a GPS device secured on the same belt as the hip accelerometer (to avoid the two becoming separated), they were not asked to wear the hip accelerometer for 24 h (as GPS devices have to be charged overnight and would be uncomfortable to wear in bed). Thus, we cannot compare hip and wrist data for a 24-h period. This article will therefore compare hip- and wrist-based estimates of physical activity and sedentary behavior for the minutes when both were worn during waking hours with at least 600 min·d−1 for the hip (34). The article will also compare the non–wear-defined waking hours to the waking hours from the 24-h protocol in the wrist only. Participants completed sleep logs for 7 d. Sleep was excluded from the analysis based on these self-reported sleep logs. Rewear for devices was requested if the screener determined that there were fewer than 5 d of data with 600 min of wear time or fewer than 4 d with 3000 min total wear time for either the hip or wrist accelerometer. In addition, the hip data were manually screened for less than four wear periods and human wear patterns to assess days of wear versus mailing days. The wrist day was screened for typical patterns of day and night wear, including almost continuous count values for day wear and longer series of zero counts punctuated by a few minutes of counts for night wear. Screening procedures were used during data collection to ensure compliance and quality control across sites. Once all data were collected, standardized data processing was performed at one site (UCSD). Not all participants in the final data set provided 5 d of data.

Data processing

Raw accelerometer data at 30 Hz were collected on three axes on the wrist of the nondominant hand and hip. Wear time for the hip was processed using the Choi algorithm in Actilife 6.11, which assesses 90 consecutive minutes of zero counts as nonwear and includes a 30-min small window to remove artifactual movement. For all analyses, only days with at least 600 min of hip wear time were included. All days with this criteria were included. We did not require a minimum number of days for analysis. We compared compliance proportions with the wake time criteria across wrist and hip locations. We also provide wrist-based estimates with the 24-h protocol for comparison because researchers argue that the longer wear time will also provide more opportunity to capture physical activity (27). For estimates of activities, only wake time was used. For the 24-h data, wake time was established by 7 d of self-reported sleep logs. Table 1 summarizes the comparisons made across wear time protocols, wear locations, and data processing techniques. Matching the wrist and hip by wear time allows comparisons for location only, taking out variations in wear time. Additional comparisons of wrist data with wake time established by sleep logs allow assessments of both wear time and location. This way we are able to establish independent effects of location and wear time, rather than confounding location differences by wear time. On the data presented to reflect a 24-h wrist protocol, a non–wear time procedure of 60 continuous zeros on the minute level vector magnitude data was used after sleep time was removed. This approach is emerging for wrist data as the 90-min count window used on the hip would not be appropriate on the wrist.

Summary of accelerometer-based comparisons made across wear time protocols and data processing techniques.

Counts per minute from the x-axis of the hip accelerometer were aggregated and divided into three groups: sedentary (0–99), light (100–1951), and moderate to vigorous physical activity (MVPA) (1952+) based on established cut points (9,18). We also used a new method called “GGIR” to classify MVPA in the hip and wrist from the raw vector magnitude of the three axes (32,33). The thresholds (wrist ≥100 and hip ≥69) to classify MVPA in this algorithm were developed on a laboratory sample of 30 adults, both men and women (18–65 yr) (10). We selected the GGIR approach as it was the only available wrist and hip algorithm for the ActiGraph GT3X+ at the time of data processing and because we had validated its performance in the same set of women who provided training data for the ML algorithm used here (7).

We also applied an ML algorithm called two-level behavior classification to predict behaviors from triaxial wrist and hip acceleration data. This classifier uses a combination of a random forest classifier and a hidden Markov model to predict four behaviors (sitting, standing, walking/running, and riding in a vehicle). In a previous study, the classifier was trained on a sample of 38 overweight or obese women (mean age = 55.2 ± 15.3 yr, body mass index [BMI] = 32.0 ± 3.7 kg·m−2) wearing two ActiGraph GT3X+ accelerometers (right hip, nondominant wrist) for seven free-living days. The women from the UCSD site were both breast cancer survivors and those at risk for breast cancer. Images captured by wearable cameras at 15-s intervals were used to classify activities of free-living study participants and used as the criterion (7).

Statistical analyses

Separate generalized estimating equations were used to make comparisons between methods while accounting for clustering within individuals (12). Comparisons were made at the day level. All methods were compared across the hip for PA (MVPA or walking/running) and sedentary time (sitting or sedentary). We also provided single-axis count data to allow readers to compare count intensities across the different behaviors and processing modes, although the count data were not used in the GGIR or ML procedures. The GGIR hip and wrist algorithms were compared across the hip and the matched wrist minutes. Then the hip-matched wrist minutes and the 24-h protocol wrist minutes were compared. The same analyses were performed for each ML behavior (hip vs wrist, wake time wrist vs 24-h wrist protocol). We also compared the number of participants who would meet the 150-min weekly physical activity guideline by the wear locations, processing method, and wear time.


Of the 402 people recruited to the study, 368 returned both wrist and hip accelerometers with valid days of wear time, and an additional three people returned accelerometers with valid data after requesting a rewear. Thirty-eight participants were not included in this analysis because they were used as the training sample used to define the ML activity recognition classifier, resulting in a data set of 333 women. Self-reported sleep logs were available for 321 participants for 1420 d, and sleep time was removed based on these logs, leading to a final data set of 321 women with 1420 d of data. The women who had sleep logs were more likely to have a lower BMI, more likely to be married, and had a higher education level (P < 0.05) compared with those without logs and those in the training sample. As shown in Table 2, participants were women with a mean age of 55 yr (SD = 9.2) and were predominantly white (~80%). Approximately 70% of the sample were college graduates and only 12 participants were Latina. Half the sample was used full time.

Demographic characteristics of women in the study sample (n = 321).

Table 3 shows the descriptive statistics for the number of hours the accelerometers were worn. It illustrates the wear time captured in the hip wake time protocol and the wrist 24-h protocol including sleep time. As expected, substantially more hours per day were collected with the 24-h protocol on the wrist. The percent of participants, however, who had wear time hours that met the 600-min criteria for wake time was very similar for the hip and wrist locations.

Compliance estimates for participants with at least 600 min of accelerometer wear time per day (n = 321), stratified by hip and wrist locations.

Table 4 shows the mean daily minutes between the hip data processing methods, between the hip and the wrist locations, and between wear time criteria across the wrist. There was a 10-min·d−1 difference in MVPA minutes between the GGIR raw vector magnitude algorithm and the single-axis cut point method based on the hip accelerometer. The ML walking/running estimates were double the MVPA estimates although median counts per minute were similar across the GGIR and the ML methods. There was almost a 3-h·d−1 difference in the hip cut point–based estimate of sedentary time and the hip ML sitting time. The mean MVPA minutes estimated using the GGIR algorithm were 20 min greater for data collected using wrist-based devices compared with devices worn on the hip. By contrast, the ML algorithm identified 10 fewer minutes of walking/running time for devices worn on the wrist, compared with the data collected using the hip accelerometers. The 24-h wrist protocol (excluding sleep time) provided only two to three additional minutes of walking or physical activity per day for the machine learning and raw vector magnitude algorithms when compared with the wrist wake time estimates derived from the hip wear time protocol. For the ML algorithm, estimates of standing still and standing moving were higher at the wrist than at the hip location and estimates of sitting were lower.

Day-level comparison of minutes classified by algorithms, wear locations and wear time protocols for n = 321 participants across n = 1420 wear days waking hours only.

Table 5 shows the percent of participants that would be classified as meeting physical activity guidelines. The hip cut point classified the fewest participants as meeting guidelines and the hip ML walking classified the most people. The GGIR wrist algorithm and ML wrist algorithm classified similar numbers of participants as meeting the guidelines both in the wrist wake time and 24-h protocols.

Percent of participants meeting guidelines of at least 150 min of physical activity per week (in 321 women) by data collection and processing techniques.


The current analyses compared three data processing protocols, two wear locations, and two wear time protocols. Future studies may use different techniques and understanding differences between them will inform our interpretation of population level physical activity. In our analyses, the prevalence of meeting physical activity guidelines (by MVPA or walking at any intensity) varied by up to approximately 50% depending on which data processing technique was used. Differences in meeting physical activity guidelines across body wear locations were as much as 20%. The current findings will inform consensus development for accelerometer wear and data processing protocols in future studies.

Researchers have recently called for standard applications of data processing techniques, at the same time as suggesting raw data processing should be used (35). A recent laboratory study compared available computational techniques, demonstrating that different methods had different strengths depending on the type of measure being predicted (14). The development and real-world validation of new algorithms in free-living settings is ongoing. Until we have more free-living algorithms, applied to population level cohorts, and comparisons across methods to health outcomes, it is not clear what method and wear location should be recommended. At this stage, comparisons across studies with varying methods may not be informative (28). Many large cohort studies have recently completed data collection on the hip, including some studies with 24-h hip protocols (16). These studies are looking for new techniques to improve physical activity and sedentary behavior classification. Some major population studies (e.g., NHANES), are now using 24-h wrist protocols, and many projects may decide to adopt the wrist location. There are fewer algorithms available for the wrist, and only a few studies have compared validation across the wrist and hip location (6,7,22,23). Further, a limited number of new classification techniques have been developed on free-living training data (15). This is one of the first studies to apply new classification techniques for the hip and wrist from free-living training data matched to the sample and compare the classification with a laboratory-based algorithm for the hip and wrist.

Not surprisingly, wear time was greater in the wrist location when the 24-h protocol was used compared with the wake time wear protocol in the hip location. The number of compliance days was also slightly greater in the wrist location. The percent of participants who met the hip criteria (98%) was higher than in other reported studies where participants were more compliant with the wrist than the hip protocol and preferred this wrist location (13). This level of compliance was particularly impressive given that most participants (89%) were mailed their devices. However, it is notable that the current study included multiple days of training for staff in how to motivate compliance, including through two reminder phone calls. We also requested that participants rewear devices if criteria were not met, although this did not always result in full compliance. Such efforts may not be feasible in larger cohort studies.

Researchers have argued that, in addition to improved compliance with the wrist location and 24-h protocol, the longer wear time will also provide more opportunity to capture physical activity. Although statistically different, our results do not show much gain in minutes or percent meeting guidelines (up to 4% more) with the 24-h wrist protocol, compared with the hip-derived wake time protocol on this wrist. However, longer wear time may improve estimates of sedentary behavior, especially if worn for 24 h at the hip. There were greater estimates of sitting time in the 24-h protocol on the wrist but greater estimates of sitting time with a shorter wear period on the hip. Although this article focuses on comparing the physical activity and sedentary behavior estimates of the hip and wrist locations, decisions on wear location may also include consideration of these locations to assess sleep duration and sleep quality. Sleep researchers have typically used the wrist location and algorithms for detecting sleep duration and sleep quality, and algorithms for these locations have been validated against gold standard polysomnography (36). Although there are studies showing that sleep duration can be assessed at the hip (29), few studies have assessed the hip location compared with polysomnography (36). For example, one study by Zinkhan et al. (36) found sufficient validity for the sleep duration measure at the hip but very poor performance of the sleep quality metrics. This is not surprising as they were applying wrist-validated algorithms for sleep quality to the hip. More research is required to develop hip-based algorithms that can detect sleep quality. Nonetheless, researchers should be aware of the current differences in performance of sleep variables at the wrist and hip if they are considering a 24-h protocol.

We observed large differences in physical activity behavior estimates across hip and wrist locations, and the direction of these differences varied by the way the data were processed (e.g., using machine learning to derive estimates of physical activity behaviors, i.e., walking, or using the GGIR approach to derive physical activity intensities). For example, the ML algorithm classified more minutes of walking/running (at any intensity) for hip-worn accelerometers compared with accelerometers worn at the wrist. However, GGIR-derived estimates of MVPA for hip-worn accelerometers were lower than estimates from wrist-worn devices. Without a “gold standard,” it is not possible to confirm which position or processing approach is more accurate; however, accuracies reported in the literature are higher for the ML algorithms than GGIR algorithms (ML accuracy is estimated at more than 80%, compared with 59% for GGIR) (7,10). Further, in the same women who provided training data for the ML methods applied here, we tested the GGIR algorithm against image annotations and reported similar estimates of MVPA across locations (17 and 18 min·d−1 on hip and wrist), whereas the hip ML algorithm always performed better than the wrist location (7). This suggests that when applied to potentially more “noisy” free-living data, the GGIR (developed in the laboratory in adult and older men and women) may overestimate MVPA.

Both the vector magnitude and the data ML algorithms classified many more participants being active 150 min·wk−1 than the currently widely used cut points. These findings support other studies showing that the current hip-based cut points may underestimate common forms of physical activity (including walking) in older cohorts (15). Current population level data with hip-based accelerometer cut points suggest less than 5% of adults and less than 3% of older adults meet physical activity guidelines (26). In this sample of healthy middle-age and older adult, women the cut points classified more than 20% of the sample as meeting guidelines. This suggests that this sample may be particularly active and the results may not be generalizable. Comparatively, the more complex data processing techniques detected more minutes of physical activity and walking per day. New techniques may present a more positive picture for public health. Currently, the physical activity guidelines are for moderate intensity physical activity, which the GGIR raw vector magnitude algorithm identifies. The GGIR MVPA estimates from the wrist identified more people meeting guidelines than the hip location. Depending on wear location, population estimates of MVPA could vary.

The behavior-based ML algorithm identifies all locomotion, regardless of intensity. Although the count data were not used in the ML algorithm, for comparison, median counts per minute for the walking classification were 1070, suggesting that most walking occurs below the 1952 MVPA threshold. The behavioral algorithms identified more than double the amount of walking than either the laboratory-based algorithm or the cut points identified as MVPA minutes. Given the previously validated higher accuracy of the hip-based free-living algorithm to detect walking (7), these results suggest that the hip location will detect more walking behavior than the wrist location.

Although current evidence demonstrates a dose response effect of exercise intensity, suggesting that higher intensity exercise provides the most benefits, these conclusions were drawn from studies that lacked objective information about total walking. With the surgeon general's new report on walking (31), it is important to understand the health effects of walking both at any intensity and at higher intensities. Longer durations than the current guidelines may be recommended if higher intensities cannot be achieved. Accurate behavior-based algorithms can further these investigations. If we wish to simplify public health guidelines with a “walking” message, the behavior-based algorithms demonstrate that 74% of this sample walks for 150 min·wk−1. If the relative difference in estimates were applied to national data sets, we might see almost 20% of the population meeting walking guidelines. Clearly, there is still great improvement needed in population levels of PA. More accurate estimates will hopefully better reflect actual changes over time with continued public health efforts to improve PA levels. Many efforts will include built environment improvements that will focus on walking; hence, accurate population estimates are needed.

Importantly, the behavior-based algorithms we applied to this cohort were able to distinguish sitting from standing. As studies demonstrate that excessive sitting time may be related to poor health, it is important to have more accurate estimates of sitting time. Cut point estimates of sitting are likely to include standing (which is not a sedentary behavior) shown by the median counts per minute of 96 for the ML algorithm–defined standing category. Although ActivPAL devices (placed on the thigh) are becoming the gold standard in sitting intervention studies, there are many large prospective epidemiological cohorts with high-quality measures of health outcomes that have hip accelerometer data (16). With improved measures of sitting and standing, we can more accurately estimate the health effects of prolonged sitting (i.e., sitting without breaks). Current accelerometer measures of breaks based on the 100 counts per minute cut point have been shown to underestimate total sedentary time compared with criterion measures (15). In these analyses, the difference between the 100 count and ML sitting time was almost 3 h·d−1. Such a magnitude of difference may have effects on health outcomes. In addition to classifying standing time that falls within the 100-count threshold (15), the ML algorithm identifies vehicle time separately (as its pattern is different and the counts can exceed 100). After processing, we combine the vehicle minutes with the sitting (not in vehicle) minutes. Although the cut point approach was not designed to specify vehicle time, this is clearly a sedentary behavior that contributes to inactivity levels. New methods that can distinguish categories of sitting can improve our understanding of such behaviors with health. Importantly, sitting can be detected in both hip and wrist locations. Some researchers believed that hand movement during sitting would prevent detection of sitting; however, the patterns created by this type of movement are different from other activities. We saw almost an hour more sitting time from the hip algorithm than the wrist, and this combined with the higher physical activity detection in the hip may weigh into researchers decisions to adopt the hip or wrist location.


The current study was a convenience sample of middle-age and older women, including cancer survivors and nurses from multiple locations across the United States who may not represent the population as a whole. Physical activity was higher in this cohort than other population studies so our results may not be generalizable. With an average daily total count of 232,549, our sample would be in the 50% percentile for age (55 yr) of the range recently reported by Bassett et al. (1). Given that this study compared methods, we would not expect a differential bias across techniques or body placement. Unfortunately, the current study did not have a gold standard for comparison (e.g., person-worn camera data or in person observations) that would allow us to establish which technique provided a more accurate estimate of the behaviors and intensities examined. We, however, compared two algorithms that have been validated against known behaviors, in this study's female training sample, and this information helped us interpret the current results. Further, the women in this sample (mean age = 55 yr) were intentionally very similar to the women used in the machine learning training and validation phase (mean age = 55 yr). BMI, however, differed between the training and the application sample (32 vs 27), which may have affected the performance of the algorithm. The GGIR laboratory algorithm was developed on adults, both men and women (mean age 34), and the 1952 cut point on young adults (men and women). The differences in estimates may partly due to the fact that cut points and GGIR data processing techniques were not validated in middle-age and older women. Some women in the sample were breast cancer survivors, but cancer status was not known for all participants, limiting our ability to consider this in our analyses. Finally, we were not able to compare 24-h protocols across the hip and wrist because our participants also wore GPS devices on the same belt as the hip accelerometer. They could have not worn this bulkier device overnight and it had to be charged overnight. Because of our need to have matched GPS and accelerometer data for other study purposes, we did not allow participants to remove the GPS from the accelerometer belt. Future studies with combined GPS and accelerometer devices that could be worn overnight might prove fruitful.


Results of device-based measurement of physical activity in middle-age and older women vary greatly by data processing techniques and device placement. In our analyses, the prevalence of physical activity and walking/running varied by almost 50% depending on which data processing technique was used. Differences across body wear locations were as much as 20%. Thus, efforts to standardize processing methods, or include standard estimates regardless of the choices made in a particular study, are vital. Otherwise, comparisons across studies will remain extremely challenging. Although concrete wear location and processing recommendations cannot be made at this stage, this study demonstrates differences in methods that need to be considered when comparing estimates of physical activity and behavior across studies using different device placement and processing techniques. Where possible, researchers should report estimates using multiple processing techniques so that more informed comparison can be made until a consensus is established. As more wrist-based studies with 24-h protocols are published, we need to understand differences in prevalence rates that may be due to the data collection and processing techniques, not necessarily population differences.

This work was supported by the NCI Centers for Transdisciplinary Research on Energetics and Cancer (Transdisciplinary Research in Energetics and Cancer) (grant nos. U01 CA116850, U54 CA155496, U54 CA155626, U54 CA155435, and U54 CA155850) and the National Institutes of Health (grant nos. UM1 CA176726 and R01 ES017017). Dr. James was supported by the Harvard Cardiovascular Epidemiology Program (grant nos. T32HL098048 [NHLBI] and K99CA201542 [NCI]). Dr. Mitchell was supported by Award Numbers F32CA162847 (NCI) and K01HL123612 (NHLBI). Dr. Marinac was supported by Award Number F31CA183125 (NCI). The opinions or assertions contained herein are the private ones of the authors and are not considered as official or reflecting the views of the National Institutes of Health. Results of this study are presented clearly, honestly, and without fabrication, falsification, or inappropriate data manipulation. None of the authors have conflicts of interest, and the results of this study do not constitute endorsement by the American College of Sports Medicine.


1. Bassett DR, Troiano RP, McClain JJ, Wolff DL. Accelerometer-based physical activity: total volume per day and standardized measures. Med Sci Sports Exerc. 2015;47(4):833–8.
2. Bastian T, Maire A, Dugas J, et al. Automatic identification of physical activity types and sedentary behaviors from triaxial accelerometer: laboratory-based calibrations are not enough. J Appl Physiol (1985). 2015;118(6):716–22.
3. Blair SN, Brodney S. Effects of physical inactivity and obesity on morbidity and mortality: current evidence and research issues. Med Sci Sports Exerc. 1999;31(11 Suppl):S646–62.
4. Chen KY, Acra SA, Majchrzak K, et al. Predicting energy expenditure of physical activity using hip- and wrist-worn accelerometers. Diabetes Technol Ther. 2003;5(6):1023–33.
5. Danaei G, Ding EL, Mozaffarian D, et al. The preventable causes of death in the United States: comparative risk assessment of dietary, lifestyle, and metabolic risk factors. PLoS Med. 2009;6(4):e1000058.
6. Ellis K, Kerr J, Godbole S, Lanckriet G, Wing D, Marshall S. A random forest classifier for the prediction of energy expenditure and type of physical activity from wrist and hip accelerometers. Physiol Meas. 2014;35(11):2191–203.
7. Ellis K, Kerr J, Godbole S, Staudenmayer J, Lanckriet G. Hip and wrist accelerometer algorithms for free-living behavior classification. Med Sci Sports Exerc. 2016;48(5):933–40.
8. Fairclough SJ, Noonan R, Rowlands AV, Van Hees V, Knowles Z, Boddy LM. Wear compliance and activity in children wearing wrist- and hip-mounted accelerometers. Med Sci Sports Exerc. 2016;48(2):245–53.
9. Freedson PS, Melanson E, Sirard J. Calibration of the Computer Science and Applications, Inc. accelerometer. Med Sci Sports Exerc. 1998;30(5):777–81.
10. Hildebrand M, van Hees V, Hansen BH, Ekelund U. Age group comparability of raw accelerometer output from wrist- and hip-worn monitors. Med Sci Sports Exerc. 2014;46(9):1816–24.
11. Holmes MD, Chen WY, Feskanich D, Kroenke CH, Colditz GA. Physical activity and survival after breast cancer diagnosis. JAMA. 2005;293(20):2479–86.
12. Horton NJ, Fitzmaurice GM. Regression analysis of multiple source and multiple informant data from complex survey samples. Stat Med. 2004;23(18):2911–33.
13. Huberty J, Ehlers DK, Kurka J, Ainsworth B, Buman M. Feasibility of three wearable sensors for 24 hour monitoring in middle-aged women. BMC Womens Health. 2015;15:55.
14. Kate RJ, Swartz AM, Welch WA, Strath SJ. Comparative evaluation of features and techniques for identifying activity type and estimating energy cost from accelerometer data. Physiol Meas. 2016;37(3):360–79.
15. Kerr J, Patterson RE, Ellis K, et al. Objective assessment of physical activity: classifiers for public health. Med Sci Sports Exerc. 2016;48(5):951–7.
16. Lee IM, Shiroma EJ. Using accelerometers to measure physical activity in large-scale epidemiological studies: issues and challenges. Br J Sports Med. 2014;48(3):197–201.
17. Manson JE, Rimm EB, Stampfer MJ, et al. Physical activity and incidence of non-insulin-dependent diabetes mellitus in women. Lancet. 1991;338(8770):774–8.
18. Matthews CE, Chen KY, Freedson PS, et al. Amount of time spent in sedentary behaviors in the United States, 2003–2004. Am J Epidemiol. 2008;167(7):875–81.
19. Matthews CE, Hagströmer M, Pober DM, Bowles HR. Best practices for using physical activity monitors in population-based research. Med Sci Sports Exerc. 2012;44(1 Suppl 1):S68–76.
20. Patterson RE, Colditz GA, Hu FB, et al. The 2011–2016 Transdisciplinary Research on Energetics and Cancer (TREC) initiative: rationale and design. Cancer Causes Control. 2013;24(4):695–704.
21. Ray MA, Youngstedt SD, Zhang H, et al. Examination of wrist and hip actigraphy using a novel sleep estimation procedure. Sleep Sci. 2014;7(2):74–81.
22. Rosenberger ME, Haskell WL, Albinali F, Mota S, Nawyn J, Intille S. Estimating activity and sedentary behavior from an accelerometer on the hip or wrist. Med Sci Sports Exerc. 2013;45(5):964–75.
23. Staudenmayer J, He S, Hickey A, Sasaki J, Freedson P. Methods to estimate aspects of physical activity and sedentary behavior from high-frequency wrist accelerometer measurements. J Appl Physiol (1985). 2015;119(4):396–403.
24. Staudenmayer J, Zhu W, Catellier DJ. Statistical considerations in the analysis of accelerometry-based activity monitor data. Med Sci Sports Exerc. 2012;44(1 Suppl 1):S61–7.
25. 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.
26. Troiano RP, Berrigan D, Dodd KW, Mâsse LC, Tilert T, McDowell M. Physical activity in the United States measured by accelerometer. Med Sci Sports Exerc. 2008;40(1):181–8.
27. Troiano RP, McClain JJ. Objective measures of physical activity, sleep, and strength in U.S. National Health and Nutrition Examination Survey (NHANES) 2011–2014. In: Proceedings of the 8th International Conference on Diet and Activity Methods Rome, Italy: NHANES. Rome, Italy: 2012. p. 24.
28. Troiano RP, McClain JJ, Brychta RJ, Chen KY. Evolution of accelerometer methods for physical activity research. Br J Sports Med. 2014;48(13):1019–23.
29. Tudor-Locke C, Barreira TV, Schuna JM Jr, Mire EF, Katzmarzyk PT. Fully automated waist-worn accelerometer algorithm for detecting children's sleep-period time separate from 24-h physical activity or sedentary behaviors. Appl Physiol Nutr Metab. 2014;39(1):53–7.
30. U.S. Department of Health and Human Services. In: Physical Activity Guidelines Advisory Committee Report. Washington (DC): U.S. Department of Health and Human Services; 2008. pp. 1–683.
31. U.S. Department of Health and Human Services. In: Step it Up! The Surgeon General's Call to Action to Promote Walking and Walkable Communities. Washington (DC): Department of Health and Human Services, Office of the Surgeon General; 2015. pp. 1–56.
32. van Hees VT, Fang Z, Langford J, et al. Autocalibration of accelerometer data for free-living physical activity assessment using local gravity and temperature: an evaluation on four continents. J Appl Physiol (1985). 2014;117(7):738–44.
33. van Hees VT, Gorzelniak L, Dean León EC, et al. Separating movement and gravity components in an acceleration signal and implications for the assessment of human daily physical activity. PLoS One. 2013;8(4):e61691.
34. 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.
35. Wijndaele K, Westgate K, Stephens SK, et al. Utilization and harmonization of adult accelerometry data: review and expert consensus. Med Sci Sports Exerc. 2015;47(10):2129–39.
36. Zinkhan M, Berger K, Hense S, et al. Agreement of different methods for assessing sleep characteristics: a comparison of two ActiGraphs, wrist and hip placement, and self-report with polysomnography. Sleep Med. 2014;15(9):1107–14.


© 2017 American College of Sports Medicine