Mobility concerns most daily tasks (e.g., householding, shopping), affecting life quality. Gait speed, recognized as the sixth vital sign, is a key to characterize mobility. It is also a primary outcome of many clinical interventions. Monitoring gait in unsupervised free-living situations is crucial. It offers the possibility to assess purposeful gait (e.g., catching a bus) in contextual situations (e.g., socializing), multitasking conditions requiring attention, and where the activity is affected by environmental components (e.g., buildings, streets).
The Global Navigation Sattelite System (GNSS) measures real-world gait speed, but it suffers from high power consumption and is available only outdoors. Multiple Inertial Measurement Units (IMU, including accelerometer and gyroscope) worn on the body could be used to estimate speed accurately. However, it is challenging and cumbersome to wear them every day. Another alternative is to use a single IMU, where the wrist and the Lower Back (LB) are recognized as appropriate sensor locations for real-life conditions. The Wrist-mounted IMU could be integrated inside a watch, thus, increasing user satisfaction. The LB-worn IMU could capture robust gait patterns even for patients and could be used to extract gait parameters like asymmetry. The wrist-based algorithms are mostly validated in supervised situations. They significantly lose their performance in daily life. While many LB-based methods exist, they have not been fully compared to determine what algorithms and under what criteria (slow, normal, and fast walkers) lead to better performance.
This thesis primarily presents accurate wrist-worn IMU-based (with barometer) speed (including cadence and step length) estimation and gait bout detection algorithms using Machine Learning (ML). An online personalization was devised in which the GNSS was sporadically used to capture a few speed data of a person's gait to tune the speed model gradually. Biomechanically-derived features were also extracted based on acceleration intensity, periodicity, noisiness, and wrist posture. The gait bout detection algorithm was validated against a multiple-IMU-based system for healthy people in unsupervised daily life. High sensitivity, specificity, accuracy were achieved (90, 97, 96 %). The personalized speed algorithm was also validated against GNSS for healthy subjects in real-world conditions, reaching an accuracy of 0.05 and 0.14 m/s for walking and running.
Furthermore, this thesis performs cross-validation on the LB-based algorithms to investigate the best algorithms for different speed ranges. Twenty-nine algorithms were organized in a conceptual framework, improved, and implemented. A novel combination technique was also proposed. The cross-validation against an instrumented mat and a multiple-IMU-based algorithm on both healthy and patient populations offered the combined approach as an accurate and robust solution with an error of 0.10 m/s. Finally, this thesis demonstrates the feasibility of using the proposed wrist-based algorithms for long-duration monitoring of gait in a large cohort study (around 2800 subjects). Results showed that the gait speed significantly improves frailty and handgrip strength estimations.
Overall, the proposed algorithms are independent of sensor orientation, thus, easy-to-use. A single IMU offers a high battery life and comfort, perfect for long-duration outdoors/indoors monitoring.
EPFL2020
