Towards harmonised accelerometer-derived physical activity: a comparability study across wrist, thigh, and hip wear locations
Wearable devices are widely used in epidemiological studies, but differences in wear locations make data harmonisation and pooled meta-analyses challenging. We aimed to (i) assess the comparability of measures of physical activity and sedentary behaviour derived from wrist, thigh, and hip accelerometers in adults during free-living conditions, and (ii) derive and validate calibration equations to enable harmonisation across wear locations. We analysed accelerometer data from UK adults from the SMART Work and Life (SWAL) study (concurrent wrist- and thigh-worn devices), Australian adults from the Raine Study Generation 1 26-year follow-up (concurrent wrist- and hip-worn devices), and an independent sample of UK adults wearing devices at all three locations. From each wear location, we derived overall activity (mean acceleration), step count, peak 1-minute cadence, moderate-to-vigorous physical activity, light physical activity, and sedentary behaviour. Cross-location comparability was assessed using correlation coefficients, intraclass correlation coefficients (ICCs), Bland-Altman analyses, and equivalence testing. Calibration equations between wear locations were derived using standardised major axis regression and were externally validated in the independent sample. Analyses included valid data from 658 SWAL participants (mean age 44.8 years; 72% women) and 901 Raine Study participants (mean age 56.7 years; 57% women). Cross-location correlations were strong, ranging from 0.66 to 0.89 between the wrist and thigh, and from 0.66 to 0.90 between the wrist and hip, with approximately linear relationships across the observed range of values. Absolute agreement was generally higher for wrist-thigh comparisons (ICCs 0.49–0.89) than for wrist-hip comparisons (ICCs 0.35–0.75), with larger systematic biases in the latter. In external validation in 42 UK adults (mean age 38.4 years; 64% women), calibrated thigh- and hip-derived estimates showed moderate-to-good agreement with wrist-derived estimates (ICCs 0.53–0.97 for wrist-thigh and 0.55–0.88 for wrist-hip comparisons), though confidence intervals for some measures extended into the poor agreement range. Accelerometer-derived measures from wrist, thigh, and hip locations are strongly correlated. These findings provide evidence to support pooled meta-analyses across wear locations and the development of location-agnostic processing software.