Comparison of tri-axial accelerometers step-count accuracy in slow walking conditions

Highlights

• Three accelerometers were evaluated in four minutes of treadmill walking at three slow speeds.

• Step count was estimated using PAA and SAX.

• The estimated step counts were compared directly to observer counted steps.

• Findings showed that device has a significant effect on the accuracy of the estimated count.

• Our results highlight the inaccuracy of the hardware itself as opposed to simply the algorithm.

Abstract

Accelerometers have shown great promise and popularity for monitoring gait. However, the accuracy of accelerometers for gait analysis in slow walking conditions is largely unknown. In this study, we compared the accuracy of three accelerometers recommended for gait analysis – Axivity AX3, APDM Opal, and the Actigraph wGT3X-BT, by holding the step-count algorithm constant. We evaluated device accuracy in four minutes of treadmill walking at the speeds of 0.9 m/s, 1.1 m/s, and 1.3 m/s. We constructed a symbolization of the gait data to count the steps using Piecewise Aggregate Approximation and compared the estimated step counts with observer counted steps from video recordings. Our results highlight the variation between the performance of devices – the Axivity AX3 provides more accurate step counts than the other two devices. In this, we provide evidence for future scientific teams to make decisions on selecting accelerometers which can more accurately measure steps taken at slower walking speeds, and suggest ways to improve the design of algorithms and accelerometers.

Introduction

Affordable accelerometer-based Body Worn Monitors (BWM) have been shown to be a promising tool for gait monitoring [1], [2], [3]. However, BWM may not be accurate when assessing physical activity levels in slow walking populations [4], [5], [6]. Some pedometers and accelerometers have been shown to underestimate the step counts at slow walking speeds [4], and the accuracy of step count estimation decreased as the speed became slower [5].

There are two factors that may lead to the decreased accuracy of step counts from accelerometer systems in slow speeds. First, the algorithms adopted by the companies to analyze the data generated by the accelerometers’ hardware may not be able to accurately estimate steps. Attention has been focused upon designing algorithms for average or fast walking speeds [8], [9], [10]. For example, many algorithms identify the cyclic nature of walking based on the stride frequencies and symmetric gait [11], [12], which may be violated by altered gait, including slower speeds. The thresholds in these algorithms are often set to deprioritize the low-energy slow walks in order to control the accuracy for normal speed [13].

Secondly, the accelerometers’ hardware may have a decreased sensitivity to detect steps when walking at a slow speed. For instance, the kinematics, such as stride dimensions, temporal components and displacement of body segments, vary between slow and normal speed [7]. These kinematic factors are important for the accelerometers’ hardware to detect body movement and collect physical activity data. However, we have limited knowledge on the sensitivity of the raw data from the accelerometers in capturing these changes in slow walking populations. Whereas pedometer accuracy is solely reliant on the hardware configuration, the extent of an accelerometer’s inaccuracy due to its hardware sensitivity is more difficult to determine. As new scientific teams are beginning to utilize these accelerometers to collect raw accelerometer data for purposes of their own analysis [14], [15], not knowing the accuracy of the hardware poses a problem for these researchers.

In this paper, we address problem of determining an accelerometer’s hardware accuracy by evaluating the hardware accuracy of three recommended accelerometers – Axivity AX3, APDM Opal and ActiGraph wGT3X-BT – in four minutes of treadmill walking at the speeds of 0.9 m/s, 1.1 m/s, and 1.3 m/s. We do this by holding the step detection algorithm constant and estimating the step counts based on the raw data collected by the accelerometer’s hardware in these slow-walking conditions. We construct a symbolization of the gait data to count the steps using Piecewise Aggregate Approximation and compare the estimated step counts with observer counted steps from video recordings. Our algorithm is validated by comparing to the accuracy of the propriety algorithm of the ActiGraph wGT3X-BT. We therefore make two contributions with this paper: (1) our comparison allows us to identify the accelerometers that are more or less effective in the measurement of step counts at slow walking speeds and (2) we shed some light on the contributing factors to the variation in step count found between different accelerometer systems.