A universal, accurate intensity-based classification of different physical activities using raw data of accelerometer

Accelerometers are increasingly used for objective assessment of physical activity. However, because of lack of the proprietary analysis algorithms, direct comparisons between accelerometer brands are difficult. In this study, we propose and evaluate open source methods for commensurate assessment of raw accelerometer data irrespective of the brand.

Twenty-one participants carried simultaneously three different tri-axial accelerometers on their waist during five different sedentary activities and five different intensity levels of bipedal movement from slow walking to running. Several time and frequency domain traits were calculated from the measured raw data, and their performance in classifying the activities was compared.

Of the several traits, the mean amplitude deviation (MAD) provided consistently the best performance in separating the sedentary activities and different speeds of bipedal movement from each other. Most importantly, the universal cut-off limits based on MAD classified sedentary activities and different intensity levels of walking and running equally well for all three accelerometer brands and reached at least 97% sensitivity and specificity in each case.

Irrespective of the accelerometer brand, a simply calculable MAD with universal cut-off limits provides a universal method to evaluate physical activity and sedentary behaviour using raw accelerometer data. A broader application of the present approach is expected to render different accelerometer studies directly comparable with each other.