Machine Learning/AI on Wearable Sensor Signals to Monitor the Clinically Relevant Episodes of Gait

The use of wearable sensors to remotely monitor walking as a measure of physical activity and mobility has become popular and today includes consumer devices such as smartwatches or wristbands. However, such devices and algorithms focus on (daily) step counts or walking speed and fail to account for context-relevant data when evaluating gait function. The ability or avoidance to walk over various demanding surfaces such as stairs, slopes or cobblestones defines a very patient centric and clinically meaningful mobility outcome and promises to give more specific insights to monitor disease progression or to personalize rehabilitation. This work presents machine learning (ML) and deep learning (DL, artificial intelligence) approaches and models to process wearable sensor signals to classify the clinically relevant surfaces of walking. In this context, the presentation addresses questions such as signal types (inertial sensors, plantar pressure signals), signal preprocessing (e.g. gait cycle segmentation), different model architectures including batch normalization, transferability between independently collected datasets including class reduction, single versus multiple sensor set-ups and other variables which can influence the performance of such classification methods. In addition methods on how to collect ground-truth labeled data to further investigate system performance in real-world settings are discussed.