Accelerometry analysis options produce large differences in lifestyle physical activity measurement

Abstract

Objective measurement of physical activity (PA) using accelerometers has become increasingly popular across recreational and clinical applications. However, the effects of multiple processing algorithms, filters, and corrections on PA measurement variability may be underappreciated. Objective: To examine how lifestyle PA estimates are impacted by multiple available scoring methods. Approach: Wrist-worn accelerometers (ActiGraph GT3X+) were worn by 132 adults (87 F) having various activity levels for one week. Lifestyle PA was assessed across four PA domains: daily energy expenditure (EE); active EE; moderate-to-vigorous PA (MVPA); and steps using 1–5 algorithms per domain, with/without wrist correction and low-frequency-extension (LFE). Estimates were compared to self-report (International Physical Activity Questionnaire). Main results: PA estimates differed between algorithms with variable but frequently large effect sizes (d = 0.08–1.88). The wrist correction reduced PA estimates across all domains (p < 0.05, d = 0.26–3.04) except step counts and one daily EE algorithm (d = 0.0). Conversely, the LFE increased step counts (d = 1.44, p < 0.05) but minimally affected all other outcomes (d = 0.08–0.20, p < 0.05). Correlations between objective and self–reported PA were small to moderate (ρ = 0.22–0.45) and decreased with the wrist correction. Significance: Measurement of PA using accelerometry is highly dependent on algorithm and filter selection; previously validated methods are therefore not interchangeable. Users should take caution when interpreting absolute PA estimates, and reporting standards should require detailed methodology disclosure to optimize comparisons across studies.

Influence of Epoch Length and Recording Site on the Relationship Between Tri-Axial Accelerometry-Derived Physical Activity Levels and Structural, Functional, and Hemodynamic Properties of Central and Peripheral Arteries

“…Anthropomorphic data (BH and BW), body location (hip or wrist), sex, race, and age of the participant were entered into the ActiLife software (v.6.13.4; ActiGraph, Pensacola, FL, USA) (Mueller et al, 2020 ). Data from the accelerometers was then downloaded and processed.…”

“…The same PA indices (EE, MVPA%, steps/minute, and VM CPM) were calculated for hip and wrist records. Similar to previous studies, according to the manufacturer's recommendations, for wrist-worm data, the “Worn on wrist” option (ActiLife software) was selected during data scoring (McMinn et al, 2013 ; Mueller et al, 2020 ; Nuss et al, 2020 ; Guediri et al, 2021 ). When that option is selected, ActiLife software applies “piece-wise scaling” to the collected data, enabling direct comparison between PA indices (e.g., EE) obtained from hip and wrist measurements (ActiGraph, 2021 ).…”

“…When that option is selected, ActiLife software applies “piece-wise scaling” to the collected data, enabling direct comparison between PA indices (e.g., EE) obtained from hip and wrist measurements (ActiGraph, 2021 ). To this end, the “Worn on wrist” option converts counts obtained at the wrist [“wrist counts,” (WC)] into “hip equivalent counts” (HEC), using manufacturer's calibration equations (published on the manufacturer's website): (i) WC range: 0–644, HEC = 0.5341614 * WC; (ii) WC range: 645–1,272, HEC = 1.7133758 * WC – 759.414013; (iii) WC range: 1,273–3,806, HEC = 0.3997632 * WC + 911.501184; and (iv) WC range: 3,807 to infinity, HEC = 0.0128995 * WC + 2383.904505 (Mueller et al, 2020 ; ActiGraph, 2021 ). The resultant counts (HEC) are then used to calculate EE and MVPA% using equations and cut-off points validated for data obtained at the hip or waist [e.g., Freedson VM3 ('11) formula used to calculate EE was developed from data obtained with waist-worn devices] (Sasaki et al, 2011 ).…”

“…However, signals derived from wrist accelerometers could be affected by accelerations related to arm movements not necessarily associated with higher PA-related energy expenditure (EE) (Hildebrand et al, 2014 ; Ellis et al, 2016 ). This has led some devices to implement conversion algorithms (equations or filters) that attempt to standardize the levels of accelerometer-derived PA indices (e.g., EE levels), regardless of the recording site (e.g., hip or wrist; McMinn et al, 2013 ; Mandigout et al, 2019 ; Mueller et al, 2020 ; Nuss et al, 2020 ; ActiGraph, 2021 ). However, as mentioned for epoch lengths, it remains to be defined whether the accelerometer position (even applying data correction schemes to make different recording sites comparable) determines the strength of the association between PA indices and the arterial status.…”

Associations of participation in organized sports and physical activity in preschool children: a cross-sectional study

“…However, tailoring the cut-offs to suit a different epoch may cause some problems [49]. The cut-offs of PA level classification may be most accurate when they are applied under the same condition as the calibration study that developed these cut-offs [49,50]. Therefore, accelerometer data was analyzed in 60 s epoch strictly following the epoch setting in the calibration study of PA level cut-offs [28].…”