Using accelerations of single inertial measurement units to determine the intensity level of light-moderate-vigorous physical activities: Technical and mathematical considerations

Abstract

Quantifying physical activity and estimating the metabolic equivalent of tasks based on inertial measurement units has led to the emergence of multiple methods and data reduction approaches known as physical activity metrics. The present study aims to compare those metrics and reduction approaches based on descriptive and high order statistics. Data were obtained from 147 young healthy subjects wearing inertial measurement units at their wrist or ankle during standing, walking and running, labeled as light, medium or vigorous activities. The research question was, first, if those metrics allowed differentiating between light, moderate, and vigorous physical activities, and, secondly, what was the relationship with the metabolic equivalent of the task performed. The results showed that each metric differentiated the level of activity and presented a high correlation with the metabolic equivalent of the task. However, each metric and data reduction approach demonstrated its specific statistical characteristics related to the localization of the sensors. Our findings also confirm the absolute necessity to detail explicitly all calculus and post processing of metrics in order to quantify the level of activity by inertial measurement units.

Introduction

Physical activity improves the general well-being (Kubota et al. 2017). However, for an efficient physical activity program, the optimal physical activity intensity has to be determined. Planning adequate physical activity and then prescribing an optimal intensity for those is a challenge for health care policies (WHO, 2010). The relevance of on-body sensors to monitor physical activity is well-established (Marschollek, 2016) and inertial measurement units (IMUs) have become a popular technical solution (Rault et al. 2017) within their constraints (Kerr et al. 2017).

Although multiple studies have dealt with physical activity monitoring (Marin, 2016), there has been little investigation of the mathematical and technical rationale regarding the methodology used to establish the correspondence between the inertial sensors signal (commonly acceleration) and the physical activity measure itself. The technical rationale or the mathematical background are most of the time superficial, not demonstrated and, in most cases, only a final output without any technical background is provided. In addition, as far as the technical implementation is concerned, there is no consensus as to sensor placement or input data prerequisites (Nez et al. 2018). Often various positions are suggested for the use of the sensor (Storm et al. 2015) even though the hip is already recognized as providing better results than the wrist (Cleland et al., 2013). Hardware heterogeneity (Storm et al., 2015) is another well-known issue, potentially causing different results (Nez et al. 2016). An example is the varying sampling rate according to the hardware of the sensor (Yang and Hsu, 2010) and the difference between continuous and intermittent data collection. Some recent studies have demonstrated the influence of sampling frequency on PA estimation (Brønd and Arvidsson, 2016).

Regarding the mathematical background, the most common outputs are “activity count”, identified over a specific period of time (Yang and Hsu, 2010). However, the exact calculation of such parameters is often missing, and leaves room for speculation. Most of the time, this activity count derives from acceleration measurements of the IMU and analytical approaches have been proposed (Chen and Bassett, 2015). The acceleration signal is rectified and integrated by a user-defined epoch. An “activity” is counted once the resulting signal exceeds a certain threshold. However, this threshold is often undocumented.

To quantify movement performance various metrics based on IMU have been proposed in the literature (Lepetit et al., 2018, Lepetit et al. 2019). Some are easy to determine, such as the vector magnitude minus one (Van Hees et al., 2013), others, however, require numerous calculi (Fradet and Marin, 2016). After this metric computation, regression equations are proposed to estimate energy expenditure (Rothney et al. 2010) or cut-points are proposed for sedentary time and physical activity intensity classification (Migueles et al., 2017). The level of activity as “light”, “moderate” or “vigorous” is based on Metabolic Equivalent of Task (MET) (Ainsworth et al., 2011) even though the relationship between IMU based metrics and MET is questionable (Yang and Hsu, 2010).

Previous elements showed that the use of a single IMU to qualify and classify the intensity of physical activity did not find consensus in terms of set up and mathematical rationale for IMU deduced metrics. Consequently, IMU’s users are faced with multiple options to place the IMU on the body as well as with various calculations to score the level of activity. In addition, in some propriety systems, this calculation is hidden. In this context, the aim of this paper is (i) to provide direct comparison of the efficiency of various IMU deduced metrics according to IMU localization to quantify and differentiate light-moderate-vigorous physical activities, and (ii) to assess the relationship between those metrics and the MET.