Impact of Reduced Sampling Rate on Accelerometer-based Physical Activity Monitoring and Machine Learning Activity Classification

Abstract

Purpose: Lowering the sampling rate of accelerometer devices can dramatically increase study monitoring periods through longer battery life, however the validity of its output is poorly documented. We therefore aimed to assess the effect of reduced sampling rate on measuring physical activity both overall and by specific behaviour types. Methods: Healthy adults wore two Axivity AX3 accelerometers on the dominant wrist and two on the hip for 24 hours. At each location one accelerometer recorded at 25 Hz and the other at 100 Hz. Overall acceleration magnitude, time in moderate-to-vigorous activity, and behavioural activities were calculated using standard methods. Correlation between acceleration magnitude and activity classifications at both sampling rates was calculated and linear regression was performed. Results: 54 participants wore both hip and wrist monitors, with 45 of the participants contributing >20 hours of wear time at the hip and 51 contributing >20 hours of wear time at the wrist. Strong correlation was observed between 25 Hz and 100 Hz sampling rates in overall activity measurement (r = 0.962 to 0.991), yet consistently lower overall acceleration was observed in data collected at 25 Hz (12.3% to 12.8%). Excellent agreement between sampling rates was observed in all machine learning classified activities (r = 0.850 to 0.952). Wrist-worn vector magnitude measured at 25 Hz (Acc25) can be compared to 100 Hz (Acc100) data using the transformation, Acc100 = 1.038*Acc25 + 3.310. Conclusions: 25 Hz and 100 Hz accelerometer data are highly correlated with predictable differences which can be accounted for in inter-study comparisons. Sampling rate should be consistently reported in physical activity studies, carefully considered in study design, and tailored to the outcome of interest.