Personalized mobile gait analysis: Towards wearable-based rehabilitation in daily life

This thesis contributes to methods of evaluating an individual's health status using wearable devices by analyzing their gait patterns, namely, the unique way they walk. Many health conditions or impairments can be detected in gait patterns, and wearable devices such as Inertial Measurement Units (IMUs) have great potential for monitoring these conditions in a mobile setting. Mobile gait analysis, which captures rich walking movement information through the unobtrusive nature of IMUs, is a prominent and intensively researched topic in digital healthcare. However, there remains a lack of standards to effectively select IMU devices and evaluate gait quantification algorithms. More importantly, although a large body of studies have explored IMU-based gait analysis, the majority of existing methods lack personalization and interpretability, hindering their practical applicability for healthcare. To address the issues mentioned above, this thesis contributes guidelines, methods, and benchmark datasets to develop and evaluate personalized gait analysis tools. The first part of the thesis addresses the need for standards to effectively evaluate IMU devices and the algorithms to quantify gait patterns into spatio-temporal parameters. We proposed a step-by-step guideline for device selection, demonstrated with a series of comparative analyses using seven commercial IMUs. Moreover, we validated gait quantification algorithms using data from healthy and stroke populations and three different reference systems. The open-source algorithms and the publicly available dataset enable algorithm benchmarking in future research. Building on the methods established in the first part, the second part of the thesis focuses on developing methods to investigate personalized gait responses using IMU-derived features. We created a dataset containing gait pattern changes induced by physical and cognitive stressors common in daily-life settings. We proposed a statistical method to re-analyze population-level trials at an individual level, which revealed highly personalized gait responses. Additionally, we introduced a visualization technique utilizing transparent neural networks to generate physiologically relevant insights from raw IMU data, facilitating further investigation into the cause of the movement changes. These methods and the accompanying publicly available dataset provide personalized and insightful ways to analyze gait data obtained from wearables, which can be applied in clinical settings to improve health outcomes. Finally, the third part of the thesis combines the insights of the two preceding parts and presents a practical application in personalized stroke rehabilitation. We created a dataset of post-stroke gait patterns at two separate time points, enabling personalized evaluation of longitudinal rehabilitation progress. Personalized changes in gait patterns were visualized through foot movement trajectories and spatio-temporal gait parameters. The utilization of transparent neural network visualizations further revealed personalized gait changes in the time series sensor data. Our qualitative and quantitative analyses were evaluated and validated by domain experts, highlighting the effectiveness and applicability of IMU-based gait analysis in the context of stroke rehabilitation. To summarize, this thesis demonstrates the value of mobile sensor-based gait analysis in generating personalized health insights for research and clinical settings. It provides practical guidelines, evaluation frameworks, personalized and explainable analysis methods, and application examples in stroke rehabilitation. Together with the publicly available datasets created, the works in this thesis facilitate future method development and evaluation to enable wearable-based pervasive digital healthcare.