Abstract
The emerging trend of Body Sensor Networks (BSNs) has excited research for enabling continuous and quantitative monitoring of human physiology, kinesiology, psychology, neuropsychology etc. In particular, one important type of BSN -- inertial BSNs -- has shown promising opportunities of monitoring human gait for various medical applications. These applications include: studying and monitoring gait pathology or degradation, assessing the efficacy of orthoses or prosthetics, and evaluating medical intervention for a gait-manifested neuropathy by comparing gait difference before/after the intervention. For such medical applications, inertial BSNs could not only function as a convenient and economical tool for quantitative medical observation, but also enable out-of-lab and long-term monitoring assessment, providing additional rich information for medical research and challenging clinical decisions beyond traditional, in-lab and/or qualitative clinical observation and further prompting medical research.
However, to demonstrate the feasibility of this technology, retrieving accurate and useful information (i.e., gait features) from inertial BSNs is essential and requires a combination of biomechanical knowledge and data processing innovation. Extracting spatial information (the basis of many gait features from traditional camera-based motion capture) from inertial BSN is particularly challenging due to sensor noise and integration drift, and existing gait feature extraction methods tend to target small subset of features for specific applications and deployment scenarios.
This work develops a systematic approach for gait feature extraction from inertial BSNs for medical applications. Techniques to tackle the challenge of tracking accurate spatial information in order to provide accurate kinematic information are introduced. Given the accurate kinematic information, linear analysis of extracting time domain gait features is applied. This work then explores the opportunity of leveraging biomechanics and applying nonlinear analysis in order to refine signal processing and enhance data separability and interpretability from extracted features. Also, machine learning approaches are explored for automating the process of feature selection and systematic error correction. Lastly, case studies for applying these techniques to various medical applications are discussed to exemplify the systematic approach. Overall, this systematic approach is aimed to conquer dominant challenges residing in inertial BSNs by extracting accurate gait features, transforming the extracted features into valuable medical knowledge, and enabling medical research requiring longitudinal gait analysis.
