Is Wrist Accelerometry Suitable for Threshold Scoring? A Comparison of Hip-Worn and Wrist-Worn ActiGraph Data in Low-Active Older Adults With Obesity

Abstract

Background

Hip- and wrist-worn ActiGraph accelerometers are widely used in research on physical activity as they offer an objective assessment of movement intensity across the day. Herein we characterize and contrast key structured physical activities and common activities of daily living via accelerometry data collected at the hip and wrist from a sample of community-dwelling older adults.

Methods

Low-active, older adults with obesity (age 60+ years) were fit with an ActiGraph GT3X+ accelerometer on their nondominant wrist and hip before completing a series of tasks in a randomized order, including sitting/standing, sweeping, folding laundry, stair climbing, ambulation at different intensities, and cycling at different intensities. Participants returned a week later and completed the tasks once again. Vector magnitude counts/second were time-matched during each task and then summarized into counts/minute (CPM).

Results

Monitors at both wear locations similarly characterized standing, sitting, and ambulatory tasks. A key finding was that light home chores (sweeping, folding laundry) produced higher and more variable CPM values than fast walking via wrist ActiGraph. Regression analyses revealed wrist CPM values were poor predictors of hip CPM values, with devices aligning best during fast walking (R² = 0.25) and stair climbing (R² = 0.35).

Conclusions

As older adults spend a considerable portion of their day in nonexercise activities of daily living, researchers should be cautious in the use of simply acceleration thresholds for scoring wrist-worn accelerometer data. Methods for better classifying wrist-worn activity monitor data in older adults are needed.

The widespread adoption of accelerometry to assess human movement represents an important step forward in the study of physical activity in aging (1). Where self-report measures suffer from recall bias (2–4) and typically quantify average duration of physical activity and sedentary behaviors, accelerometers provide objective insight to both quantity and patterning of daily movement behaviors. The resulting data have shaped public health recommendations: The current United States physical activity recommendations (5) now underscore the importance of not only structured bouts of moderate-to-vigorous physical activity, but also the accumulation of such activity in bouts of any length coupled with the value of frequent movement to interrupt sitting.

Over the last 2 decades, accelerometry assessment has saturated the physical activity research space (see (6) for a historical overview). The mostly widely used family of devices in physical activity research are those produced by ActiGraph (Actigraph, Pensacola, FL) (7), and many methods for studying physical activity via accelerometry were guided by the early use of ActiGraph devices. Initially, these were designed to be worn at the hip and summed acceleration across a single vertical axis. Triaxial accelerometers later allowed for the capture of movement across 3 axes, which are often summarized into a single vector magnitude (VM) metric. An important limitation of single axis accelerometry is that acceleration across a single axis can be affected by device orientation (eg, the device becomes rotated during wear such that the vertical axis is no longer vertical). This is remedied by VM, which captures acceleration regardless of orientation. Many early papers simply reported acceleration “counts”: a proprietary ActiGraph metric capturing overall volume of daily activity. Later, to assist interpretation of activity behaviors, a number of research teams employed treadmill protocols to identify acceleration thresholds that approximately correspond to meaningful metabolic equivalent (MET) cut points (ie, 3.0 METs delineating moderate intensity movement, 6.0 METs for vigorous intensity movement) (8). Recognizing the impact of age, fitness, resting metabolic rate, the presence of disease, and functional states on the relationship between hip acceleration and metabolic output, other research teams quickly developed a host of population-specific cut points, including for older adults (9,10).

The widespread use of accelerometers in activity research is still in its infancy and researchers continue to think deeply about measurement and burden associated with the device (1). For instance, as hip ActiGraphs are often worn on elastic belts, researchers feared reduced compliance especially over longer assessment periods if participants found the devices unattractive or uncomfortable to wear (11), though it should be noted that compliance rates do not typically differ by wear site (12). Perhaps more importantly, wrist wear allows for estimation of sleep quantity and quality (13), and so researchers interested in capturing both sleep and waking activities may select wrist wear to prevent participants from having to move the device from one wear location to another. However, there are several persistent challenges in the collection of accelerometer data at the wrist. At present, there are few established moderate-to-vigorous intensity cut points for scoring wrist accelerometer data in older adults. As a result, some have elected to use general adult cut points, likely contributing to underestimates of moderate-to-vigorous activity time. For both wrist- and waist-worn devices, cut-point calibration is often done via treadmill walking or running. Overground walking produces a different acceleration profile compared with treadmill walking, and as such treadmill-derived acceleration is often lower than overground walking acceleration (14). Additionally, many older adults spend relatively few minutes daily in locomotor movement and considerably more time in activities involving a high degree of arm movement and relatively low metabolic output. This has the potential to inflate estimates of time spent in physical activity and estimates of metabolic output derived from wrist accelerometry. For instance, the Bureau of Labor Statistics reports that men aged 65–69 reported engaging in 18 minutes daily of purposeful physical activity, and up to 190 minutes daily on home care (eg, food prep and chores). Those aged 70 and older reported 18 minutes of purposeful activity and 174 minutes of home care. By contrast, women aged 65–69 reported nearly 260 minutes of home care daily and just 12 minutes in purposeful physical activity. Those aged 70 or older engaged in 234 minutes of home care and just 6 minutes of purposeful activity (15). The impact of such activities on classification of light intensity and moderate-to-vigorous intensity physical activity (LPA and MVPA, respectively) collected via wrist- and hip-worn accelerometry is as yet unknown in older adults.

Given that accelerometry is often viewed as a key objective measure of physical activity that informs public health messaging, it is important to understand how both purposeful physical activity and common activities of daily living are represented on wrist- and waist-worn monitors. Thus, the purpose of this study was to characterize and contrast key structured physical activities (overground walking, stationary cycling) and common activities of daily living via accelerometry data collected at the hip and wrist from a sample of community-dwelling older adults. We recruited individuals engaging in low levels of purposeful physical activity, as they are often the target population for large physical activity interventions leveraging accelerometry (16,17).

Method

Participant Characteristics

Participants were (body mass index [BMI] 30–45 kg/m²) community-dwelling older adults aged 60+ years who were low-active (ie, self-reported engaging in less than 2 days/week of structured physical activity for at least 20 minutes). We recruited for an even sex distribution and a spread of physical function, with half our individuals having a short physical performance battery (SPPB) (18) score > 9. We excluded individuals requiring an assistive device to walk or had cognitive impairment as indicated by a Montreal Cognitive Assessment (19) score of less than 22. All participants provided informed consent and the study was approved by the Wake Forest School of Medicine Institutional Review Board (#00054432).

Study Procedures

Eligible participants attended 2 testing appointments separated by 1 week. At the baseline testing session, participants completed a series of questionnaires including demographics and a medical history, provided height and weight, and completed the SPPB. Next, participants completed 2 ratings of perceived exertion (RPE) calibration tasks. First, study staff explained the Borg RPE scale (20) before tasking the participant with walking at a leisurely pace (RPE of 11/20). Once the participant settled into a pace, they continued for 30 seconds as the researcher used a manual step counter to quantify a steps/minute cadence. This was repeated at a moderate intensity (RPE of 13) and vigorous intensity (RPE of at least 15). We selected this procedure to gain insight into accelerometer data likely to be captured during real-world overground walking rather than treadmill walking (14). We included self-selected slow, moderate, and vigorous walking as that use of arms differs by walking speed such that lower relative speeds walking involve minimal arm movement (21), thus potentially affecting the quality of data collected at the wrist. This was next repeated on a cycle ergometer wherein participants pedaled against a light resistance (30 Watts), pedaling at a cadence corresponding to an RPE of 11, and cadence was noted. Wattage was increased, holding cadence steady, until the participant felt they were working at an RPE of 13, and wattage was noted. We elected not to have participants cycle at a vigorous intensity to minimize burden and to prioritize vigorous walking.

At both appointments, participants were fit with one ActiGraph GT3X+ on their nondominant hip, and another on their nondominant wrist. Participants then completed a series of 4 task blocks, and the order of the task blocks was randomized for each participant. Each block contained 2–4 activities of daily living (Supplementary Table 1) and were balanced for level of exertion within each block. Participants completed each task for 200 seconds, allowing for 3 minutes of recording time while trimming the first and last minute of each reading. Participants were allowed to rest between tasks as needed. For walking tasks, study staff played a metronome at the same cadence the participant walked at during the calibration task. For each cycling task, participants cycled at the cadence and load noted during the calibration task. The order of tasks, stepping cadence, and cycling parameters were held constant at the second testing appointment.

Accelerometer Processing

Both wrist- and hip-worn devices were initialized using a sampling rate 30 Hz and downloaded in a 1-second epoch using ActiLife software version 6.13.4 (ActiGraph, Pensacola, FL) on the same computer in order to ensure time matching. Step and activity counts were derived from ActiLife software directly, and VM counts/second was computed across 3 axes as $\sqrt{x^2+y^2+z^2}$. A quality control assessment was performed on the data by visual inspection of the time series of VM counts/second to ensure the correctness of the start and stop times of each task versus paper records of these times. To calibrate the timestamps between the 2 devices, the same visual inspection was adopted to check the consistency of starting times of the 3 walking tasks. To ensure activity data consistency while performing a specific task, the first and last 10 seconds of activity data for each task was excluded, with the successive remainder of time being summarized into 60-second epochs for (ie, VM counts per minute [CPM]). Time durations that did not overlap for a full 60 seconds on the 2 devices were discarded.

Data Analysis

Descriptive statistics (mean, SD, and proportions) for demographic characteristics were calculated by sex. Box plots combining data from both appointments were used to visually inspect variability in VM counts by task. VM measures from each task were analyzed separately using mixed effects models to obtain overall estimates of mean VM counts, between- and within-person estimates of variance components, and intraclass correlation (correlation between 2 measures from the same person). The model contained a fixed effect representing appointment number and a random effect for participant. Minimum variance quadratic unbiased estimation (MIVQUE) estimation was used, as it has advantages relative to non-normally distributed outcomes (22). Finally, we conducted a series of mixed effects models to assess the ability of a simple linear regression to characterize the relationship between wrist- and waist-worn VM values for each task. Here, values from the wrist were used to predict the waist, with a random effect for participant used to account for the repeated measures on each participant. A test of slope = 1 evaluates the ability of the wrist-worn device to spread VM measurements in the same manner as the waist-worn device, with a slope between 0 and 1 indicating that the waist VM measures are compressed compared to the wrist, a slope = 1 indicating exact translation of wrist to waist, and a slope > 1 indicating that the wrist-worn device counts need “expansion” to translate into the waist. Similarly, we tested for an intercept = 0 to assess whether predicted waist-worn VM counts were generally high or low at the origin after accounting for the relationship captured by the slope. R² values from these models were obtained using the approach described in Edwards et al. (23). This model-based approach was not used for the home chores or sedentary behavior tasks because the waist-worn values did not have an approximate normal distribution and had little variability associated with them.

Results

Participant characteristics are displayed in Table 1. In total, 30 participants (n = 15 female) completed testing, 86.7% were college educated, and 10% were Black. Among female participants the average age was 72 ± 8.6 years, BMI was 32.4 ± 3.96 kg/m², and SPPB score was 9.6 ± 1.8. Among male participants, average age was 72 ± 5.1 years, BMI was 32.0 ± 3.5 kg/m², and SPPB score was 10.1 ± 1.6.

Table 1.

Participant Characteristics

	Women (N = 15) Mean (SD)	Men (N = 15) Mean (SD)
Age	72 (8.6) Range = 60–86	72 (5.1) Range = 64–79
BMI	32.4 (3.96) Range = 24.6–41.0	32.0 (3.5) Range = 27.0–41.4
SPPB	9.6 (1.8) Range = 5–12	10.1 (1.6) Range = 7–12

Notes: BMI = body mass index; SD = standard deviation; SPPB = short physical performance battery.

Table 2 displays descriptive statistics (mean, between-person standard deviation, within-person standard deviation) based on average VM CPM recorded during each task, and these data are depicted in box plots in Figure 1. Average VM CPM increased with intensity during ambulatory tasks, including the 3 walking tasks and the stair climb, and the wrist device produced higher absolute VM values and greater variability relative to the waist-worn device (see Figure 1). Intraclass correlation values from waist-worn devices were higher for each task suggesting better reliability. A similar pattern was observed for the 2 cycling tasks: average VM values increased from light to moderate cycling for both wear locations. Average VM from the wrist were lower than at the hip and clustered near 0 (Figure 1), which is attributable to grasping the stationary cycle’s handlebar, thus limiting limb movement.

Table 2.

VM Recorded at the Waist and Wrist During Each Task

	Waist-Worn Device		Wrist-Worn Device
	Mean (BSD, WSD)	ICC	Mean (BSD, WSD)	ICC
Mobility/walking
Slow walking	3 001.7 (807.3, 276.1)	0.90	5 095.4 (1 664.4, 843.7)	0.80
Moderate walking	3 331.1 (963.9, 310.0)	0.91	5 993.4 (1 887.3, 970.0)	0.79
Fast walking	3 614.7 (1 109.1, 479.6)	0.84	7 032.9 (2 207.5, 1 242.6)	0.76
Stair climb	3 598.1 (974.3, 605.4)	0.72	5 823.6 (1 236.5, 1 114.1)	0.55
Cycling
Light cycling	1 317.0 (1 014.1, 992.4)	0.51	460.5 (442.6, 525.9)	0.41
Moderate cycling	1 723.8 (1 176.8, 1 064.9)	0.55	631.3 (931.6, 903.6)	0.52
Home chores
Laundry	140.8 (157.2, 123.2)	0.62	7 676.0 (1 324.9, 941.5)	0.66
Sweeping	1 277.8 (611.1, 495.7)	0.60	7 569.5 (3 549.1, 721.6)	0.96
Sedentary
Sitting still	12.1 (16.5, 56.6)	0.08	535.0 (547.4, 695.6)	0.38
Standing still	4.9 (10.2, 13.8)	0.35	511.4 (491.0, 878.7)	0.24

Notes: BSD = between-person standard deviation; ICC = intraclass correlation; VM = vector magnitude; WSD = within-person standard deviation.

Figure 1.

Box plots depicting vector magnitude (VM) counts per minute from each task.

Both wear locations produced low average VM values during quiet sitting and standing, though it is notable that the wrist-worn device tended to produce more outlying values as depicted in Figure 1. The waist-worn accelerometer provided VM values that exceeded only sitting and stationary standing. Likewise, sweeping, which entailed light whole-body movement, produced average VM values in a range between the means for stationary tasks and slow walking. By contrast the wrist-worn device produced average VM values for both home chores that were greater than those obtained during vigorous walking with a high degree of variability.

The results of the analyses focused on linear regression using wrist-worn device information to predict the waist VM are depicted in Table 3. For each task, the estimated intercepts and slopes differed greatly, with all intercepts being significantly different from 0, and all slopes significantly different from 1. All slopes were <1 indicating there is a compression of wrist-worn counts when predicting waist-worn counts, and the amount of compression differed considerably across tasks. It was notable that the R² values were quite low among these mobility tasks, with the R² values improving with increasing intensity, supporting the notion that arm use becomes more purposeful as intensity increases.

Table 3.

Evaluation of Intercepts and Slopes for Waist-Worn Device Versus Wrist-Worn Device

	Intercept (95% CI; p for test vs 0)	Slope (95% CI; p for test vs 1)	R 2
Mobility/walking
Slow walking	2 862 (2 479, 3 245) p < .0001	0.022 (-0.028, 0.071) p < .0001	0.5%
Moderate walking	2 778 (2 345, 3 211) p < .0001	0.088 (0.045, 0.131) p < .0001	9.8%
Fast walking	1 867 (1 302, 2 432) p < .0001	0.242 (0.178, 0.307) p < .0001	25.0%
Stair climb	1 543 (1 024, 2 062) p < .0001	0.340 (0.265, 0.417) p < .0001	35.3%
Cycling
Light cycling	1 317 (869, 1 764) p < .0001	-0.17 (-0.479, 0.143) p < .0001	0.8%
Moderate cycling	1 465 (972, 1 958) p < .0001	0.233 (0.044, 0.421) p < .0001	4.2%

Note: CI = confidence interval.

Discussion

The stated aims of this study were to characterize common structured physical activities and activities of daily living via waist- and wrist-worn ActiGraph devices, and to determine the extent to which (and under what circumstances) data produced from wrist- and waist-wear locations are comparable. We found that the waist-worn accelerometer provided VM values corresponding approximately to the relative energy cost expected of each task. We observed a predictable increase in acceleration as individuals increased their walking or cycling intensity from a self-selected slow to moderate and fast speeds. Movement levels were low during standing chores and increased from a stationary task (folding laundry) to one requiring light ambulation (sweeping). The hip-worn ActiGraph also recorded acceleration during stationary cycling at levels that were lower than slow walking. This is not unexpected: in the case of stationary cycling, cadence and resistance each affect intensity and so accelerometer data scored via simple thresholds are unlikely to effectively capture cycling activity.

Data collected on the wrist produced several findings with implications for the use of wrist accelerometers in physical activity research. As with the waist monitor, we observed an increase in acceleration as individuals moved from self-selected slow to moderate and fast walking, albeit with greater variability and poorer reliability. However, we also observed that translating between waist and wrist devices is not a simple task: the monitors had poor alignment during home chores and cycling. Even during overground ambulation, the devices aligned best during the more vigorous ambulatory task, where an individual can be expected to engage their arms purposefully in walking. There was poor alignment during slow walking when an individual is likely to have little arm movement.

Perhaps most importantly from the perspective of assessing longitudinal changes in older adults’ physical activity levels, we observed the greatest discrepancy between waist- and wrist-worn ActiGraph data during common low-intensity home chores involving the use of the arms (standing and folding laundry, sweeping). Recall that contemporary processing techniques classify activity time using acceleration cut points. The waist-worn ActiGraph classified light-intensity home chores (average VM 140.76 CPM) between sedentary (average VM 8.48 CPM) and slow walking (average VM 3 001.73 CPM) tasks. By contrast, a wrist accelerometer classified home chores as more intense than vigorous walking. The impact of differentially classifying these common light tasks as LPA (waist) versus MVPA (wrist) can be illustrated with 2 theoretical scenarios (see Figure 2).

Figure 2.

Example daily distributions of sitting, light intensity physical activity, and moderate to vigorous intensity physical activity.

In the top panel we have displayed common time use among older adults across 16 waking hours: 10 hours spent awake, 4 hours spent in light-intensity chores (which would be categorized as LPA via cut point-based scoring on a waist ActiGraph and MVPA on a wrist-worn ActiGraph), and 2 hours spent in other light ambulation. Ideally, an activity intervention prescribing 30 minutes of walking would draw this time from a participant’s sedentary time. This substitution is depicted in the second panel of Figure 2, to the far right.

Since either wear location can be expected to differentiate between low-intensity movement periods and ambulation, both would report a change of −30 sedentary minutes, 0 LPA minutes, and +30 MVPA minutes. However, reducing time spent in nonexercise activity while increasing exercise time is common in response to structured exercise interventions among older adults (24,25). Should an older adult reduce time spent in light home chores as they increase time spent exercising, data from a waist-worn ActiGraph processed using cut points would report a change of 0 minutes of sedentary time, −30 minutes of LPA, and +30 minutes of MVPA. Data from a wrist-worn device would indicate this substitution would result in no change in any activity category.

While this illustration is simplified to aid in interpretation, an inflated accelerometer signal due to upper body movement is an important and unpredictable source of bias in the study of physical activity among older adults. Older adults spend approximately 4 hours daily in home chores and relatively few minutes in moderate-to-vigorous ambulation (26). These chores can be of light, moderate, or vigorous intensity, and the type of work around the home will vary by season. Older adult women typically engage in more chores around the home relative to their male peers, and time-use shifts by age and employment status (15). It is also notable than a common use of ActiGraph accelerometers is to estimate energy expenditure. Current widely used energy expenditure equations use either vertical axis or VM counts as a predictor of metabolic output, with equations fit based on treadmill walking (27). As such, the inflated accelerometer signal during low-intensity activities requiring upper body movement would contribute to overestimates of energy expenditure.

Taken together, these results should be cause for caution in the use of acceleration intensity threshold-based metrics collected via wrist-worn accelerometers as a measure of physical activity behavior (28,29). Given the wide array of arm movements recorded by a wrist accelerometer, it is unsurprising that the use of a single acceleration threshold is a poor match for estimating metabolic output. We would like to emphasize, however, that with the application of suitable scoring techniques, the tendency for wrist accelerometers to capture upper body movement is a key feature for researchers studying individuals with limited mobility, including many older adults. Indeed, many individuals engage in upper body activities ranging from light (eg, folding laundry) to vigorous (eg, hand cycling) intensities that are poorly captured by a waist accelerometer. Similarly, these data may provide valuable insight into patterns of nonambulatory activity occurring across the day (30,31). A likely path forward lies in the use of machine learning techniques for the identification of physical activity types and intensities. These have demonstrated reasonable classification of a variety of movement types at the waist and wrist (32), and when combined with other physiological signals (eg, heart rate) may allow for accurate measurement of a broader array of behaviors, such as cycling or resistance training. This multisensory machine learning-driven approach has been widely adopted in the commercial wearables space, though the ways in which these models are developed and updated are often proprietary. GGIR (33)—a large-scale open-source accelerometer processing project that to date has focused on calibration, wear time, and threshold processing algorithms—offers a useful model for how complex machine learning models can be disseminated in a transparent fashion.

An important strength of this study was the inclusion of participants with varying levels of physical functioning and an even distribution of sex, as this helps to avoid limited generalizability from an overly homogenous sample. One potential limitation was our relatively small sample size, though it is similar in size to other widely cited accelerometer validation studies (eg, (9,34–37)). Because wrist accelerometers are being used at an increasing pace in research on older adults, it is critical to the integrity of research findings that investigators are aware of the concerns identified in the current study. We stronger encourage additional work on the validation of wrist-worn accelerometers on varied subpopulations of older adults studied over extended periods of time.

Conclusion

Accelerometers have become staple tools in the study of physical activity and health in aging. Given their wide use, understanding how both hip- and wrist-worn monitors capture common activities of daily living is important for insuring accurate data collection and interpretation. Our results highlighted several important limitations to the use of wrist-worn devices. As researchers are likely to continue to use wrist-worn accelerometers to monitor physical activity behaviors, methods for better classifying wrist-worn activity monitor data in older adults are needed.

Funding

This work was supported by the Wake Forest University Claude D. Pepper Older Americans Independence Center (P30-AG21332).

Conflict of Interest

None declared.

Author Contributions

J.F., M.E.M., and W.J.R. led the design of the study. J.F. oversaw study completion and manuscript preparation. M.E.M. and S.-H.C. completed data processing and analysis and participated in manuscript preparation; C.D. and K.K. participated in protocol development, data collection, and data processing; and W.J.R. participated in manuscript preparation.