People with Parkinson’s (PwP) experience gait impairment that can significantly reduce their quality of life. Wearable devices can be used in gait rehabilitation to improve mobility by providing external feedback such as visual, auditory and tactile cues to guide PwP. Current systems provide cues at a fixed distance, or a fixed pace, based on the user’s height, preferred cadence, or preferred location. However, this cueing strategy is limited as it does not account for habituation, where cues become less salient over long-term use. In addition, people with progressive degenerative diseases, such as Parkinson’s Disease (PD), experience motor fluctuations and can respond to cues differently based on their medication state and their daily condition. Current cueing mechanisms do not address symptom fluctuation and habituation, and there is limited research on adapting cues to account for inter and intra-personal variability in cue responsiveness.

This thesis proposes an adaptive cue-provision framework that can provide personalized assistance based on the individual’s real-time response. The framework monitors the user’s gait performance and their response to the provided cues and utilizes the information to build an individualized cue-response model. The cue-response model is then used in an online optimization algorithm to provide feedback to improve gait performance. The thesis develops and evaluates two iterations of the framework. The first iteration focuses on cadence as the gait performance metric and the second on stride length and cadence. Each iteration consists of developing new gait monitoring algorithms, expanding the cue-response model states, and increasing the complexity of the cost function during optimization. Each framework is also validated with healthy participants and PwP.

In the first iteration of the framework, we utilize the Canonical Dynamical System to estimate cadence and a Gaussian Process (GP) to model the cadence adaptation as a function of the provided cues. The framework is first validated with healthy adult participants and benchmarked against two additional cueing approaches. The study demonstrates that the adaptive framework is not significantly different from the state-of-the-art fixed cue strategy in changing the person’s cadence once the GP model is learned. The framework is then tested in a case study with 4 PwP. The study results show that the selection of the target cadence used during optimization impacts the participant’s gait performance. When the selected cadence is too fast, PD participants are unable to achieve the target regardless of the cueing approach. With the appropriate target, the adaptive framework demonstrates similar efficiency, de-adaptation, and robustness compared to the fixed approach.

In the second iteration of the framework, a new stride length estimation algorithm is developed. The GP formulation is changed to a multi-output GP, where the potential correlation between the two gait performance metrics, stride length and cadence, can be captured. We also introduce sparsity into the GP to reduce the computation complexity. Finally, the cost function is augmented with the target for stride length, as well as a term that penalizes rapid cue changes. We compare the framework performance in a study with a healthy older adult control group and PwP. The results show that the adaptive approach is helpful during the multi-tasking condition.