Advancing free-living gait bout segmentation using smart garments

Abstract

Gait analysis provides an optimal method for classification of an individuals fall risk. However, to ensure the gait data is non-biased, data during activities of daily living (ADL) is necessary. With ADL data, a new technical challenge arises through the issue of being able to reliably detect and classify gait bouts during free-living activities without any user intervention. Current solutions for gait bout detection, such as Python library GaitPy, attempt to solve this issue through machine learning (ML) models [1]. While GaitPy does provide promising results, it fails to accurately detect smaller gait bouts (i.e., less than 10 steps) which are usually completed during ADL and are indicative to classification of an individuals fall risk. Additionally, due to variances in sensor setups, one ML model may preform excellent for one environment, but fail to reproduce the results for a slightly different environment. The focus of the present study is to investigate the use of a ML model for ADL gait bout detection for the Myant SKIIN garments which use a triaxial accelerometer located on the waist in front of the left anterior superior iliac spine (ASIS). Being able to reliably detect gait is imperative for proper gait data collection due to the dependency the fidelity of the analysis has on data quality. After analyzing the proposed random forest model for Myant SKIIN garments, we conclude the new model outperforms GaitPy in detecting smaller bouts. These findings provide a promising outlook for the detection of gait bouts during ADL and the assessment of an individuals fall risk.