Automated detection of gait events using inertial sensor signals and a discrete wavelet transform approach

Abstract

Detection of gait cycle events is a crucial step toward an effective evaluation and rehabilitation of pathologies or injuries in human locomotion. Recently, methods based on the Discrete Wavelet Transform (DWT) have been useful for this detection due to their robustness and the wide variety of options for analyzing and decomposing signals in time/frequency domains, as well as their ability to extract relevant features embedded in the signals. In this study, a detection method of main gait cycle events, using the Wavelet Symlets and Daubechies families, was developed. These events are the heel-strike (HS) and the toe–off (TO). Inertial signals were acquired by three different devices: a G–WALK (reference equipment), an Apple Watch (AW), and a noncommercial device based on Inertial Measurement Units (IMUs). The dataset was obtained from six–minute walking tests performed by 22 healthy subjects. First, the dataset was processed, and then the signals were synchronized regarding the reference system. Subsequently, the signals were decomposed into 6 levels using sym4 and db5 Wavelets to obtain multiple perspectives of the signals. Then, using automatic threshold techniques and symmetric windows, it was possible to detect HS and TO events. Finally, the IMUs–based system obtained a 94.398 % of recall, 100 % of precision, and 97.117 % of F_1–score, with absolute values delays in the detection between 10–20 ms. In contrast, the AW system performance was 90.168 %, 100 %, and 94.828 % for recall, precision, and F_1–score, respectively, with absolute values delays in the detection of 10–28 ms.