Abstract:
The gold standards for gait analysis are instrumented walkways and camera based motion capture systems which are highly accurate, but require costly infrastruc ture and are only available in hospitals and specialized gait clinics. Hence, a mobile and pervasive alternative suitable for non-hospital settings is a clinical necessity. Us ing wearable inertial sensors for gait analysis has been explored in the literature with promising results; however, the majority of the existing work do not consider realis tic conditions where data collection and sensor placement imperfections are imminent. Moreover, some of the typical underlying assumptions are not compatible with patho logical gait, decreasing the accuracy. To overcome these challenges, we propose a wear able inertial sensor-based gait analysis system that builds upon the state-of-the-art gait analysis methods. Our system copes with various data collection difficulties and relaxes some of the assumptions invalid for pathological gait. The system is able to extract a rich set of standard gait parameters and produce average stride profile visualizations, easily interpretable by clinicians. To validate the success of our system and highlight its clinical applicability, we collected a gait dataset from more than 60 neurological disorder patients and conducted a feature selection study to identify the key gait pa rameters in neurological disorders. To demonstrate how the extracted gait parameters can be used for a higher level inference problem, we also introduce an automated fall risk assessment solution, exploiting deep learning methods. The achieved classification accuracies outperform the existing solutions. As a final contribution, we present the design and evaluation of a computing and communication architecture that shows how the fall risk assessment methods can be transformed into a pervasive healthcare service that can handle numerous users concurrently under realistic conditions.