Availability of numerous signal features greatly enhances the potential of using complex machine learning techniques to accurately estimate physical activity and sedentary behavior. These techniques are becoming increasingly popular, as they provide improved estimates as compared to the traditional activity count cut-points [3,4]. Another potential advantage of using raw acceleration is increased inter-monitor output equivalency through elimination of proprietary signal processing specifications used to derive activity counts. For example, activity counts from ActiGraph? (ActiGraph? Inc., Pensacola, FL, USA) monitors are not the same as those from the Actical (Phillips Respironics, Andover, MA, USA) monitor due to manufacturer specific signal processing [5].

While raw accelerometry is a possible solution for inter-monitor output equivalency, several sensor and digital signal processing specifications need to be similar between monitors to ensure equivalency.Two activity monitors used in physical activity research are the ActiGraph? GT3X+ and GENEA (Unilever Discover, Colworth, UK). These monitors have a dynamic range of ��6 g and allow users to collect raw acceleration at various sampling frequencies ranging from 10 to 160 Hz at 10 Hz increments. Currently, the GENEA is commercially unavailable, however, it is the only activity monitor that has been calibrated with an open-source machine-learning technique to predict the type of physical activity and sedentary behavior from raw acceleration [3].

Raw acceleration from the GT3X+ is currently being used in the National Health and Nutrition Examination Survey to obtain nationally representative physical activity and sedentary behavior estimates [6]. There is no evidence examining the equivalency of raw acceleration outputs from these monitors and whether an algorithm developed on one monitor can be applied to data from the other to produce similar activity type recognition accuracy. Thus, the purposes of this study were: (1) To compare mean vector magnitude, which is a computed metric of triaxial raw acceleration from both monitors during mechanical shaker testing at various oscillation frequencies and (2) to determine if there is an interaction-effect in predicting activity type when a prediction model developed on Carfilzomib one monitor is applied to data from another monitor.

We compare activity type recognition accuracy rates when a model developed using the GT3X+ is applied to GT3X+ and GENEA data, and vice versa.2.?Methods2.1. Mechanical Shaker TestingMechanical shaker testing was performed using an orbital shaker (VRW International, Radnor, PA, USA; Advanced Orbital Shaker, Model 10000-2) that produces controlled oscillations between 0.25 and 4.2 Hz. Oscillation radii can be adjusted between 1.27 and 5.7 cm.