DESIGN AND VALIDATION OF AN INSTRUMENTED KNEE BRACE WITH ADJUSTABLE RESTRICTIONS FOR MACHINE LEARNING-BASED GAIT ANALYSIS

Abstract

Gait data can be analysed to classify which activities are performed and gauge the gait-related health of the performer. We consider health in the context of knee osteoarthritis and gait-afflicting conditions. This work captures gait from sixteen healthy participants who performed walking and sit-to-stand activities in a lab while artificially restricted at four severities, using a custom collection device and instrumented brace. A convolutional neural network classified activities with participant-averaged accuracies of 65.61% ± 8.85% and 67.21% ± 9.06% on two well-performing hyper parameter sets. Analysis found time-series features that differentiated between some restriction levels for each of the activities. Significant difference was found between swing phase of unrestricted walking activities and all other restriction levels, and for toe-off of unrestricted walking activities and moderate and severe restrictions. For sit-to-stand, significant difference was found between jerk standard deviation of negotiating sitting edge of unrestricted data and moderate and severe restrictions.