Instrumenting gait in neurological disorders: multi-modal approaches using wearables

Gait - how someone walks - is considered the ‘sixth vital sign’ of health. This is because poor gait is associated with low life satisfaction, an increased risk of falls and severe injuries. In Europe, people over the age of 65 make up more than 19% of the population, a figure projected to rise significantly in the future. Accordingly, the frequency of age associated conditions such as neurological disorders (Parkinson’s Disease or Stroke) also rise with the prevalence of neurological gait disorders increasing from 10% (60–69 years) to 60% in those > 80 years. Increased life expectancy, coupled with a growing prevalence of neurological disorders means more people will be coping with mobility loss. Therefore, understanding and evaluating impaired gait becomes essential for promoting healthy aging, managing neurological conditions, and/or improving the overall well-being of individuals. Although there are various reference technologies used in the gait analysis, the focus of this thesis is on wearable sensors (e.g., inertial measurement units, IMUs) due to numerous advantages including affordability, accessibility, and ease of use in clinical settings and beyond. This thesis initially presents a thorough literature review, exploring the development and progression of gait assessment techniques, technologies, and methods as well as limitations in the previous gait studies. It then focuses on contemporary techniques such as the use of artificial intelligence (AI) for human activity recognition (HAR) and edge computing for remote gait analysis to gain the necessary knowledge improve the limitations. Through a series of original research investigations, this thesis 1) investigates the consistencies of two different IMU algorithms during walks in different environments in a pilot study. This study reveals the differences and inconsistencies in the extracted temporal parameters; however, the underlying reasons cannot be fully understood due to the limitations of unimodal temporal parameters. Afterwards, the thesis 2) focuses on developing a framework through a multi-layer data fusion technique for multimodal gait analysis to go beyond the limitations of the unimodal approach. The study results show that there are differences in all gait characteristics, not only temporal parameters, during indoor and outdoor walking in healthy participants and a small group of stroke survivors. This highlights the significance of conducting gait studies in various environments with extended data collection, to gain deeper insights into the impacts of habitual environments. Nonetheless, data collection outside of clinical settings results in a substantial volume of unlabelled data. Data labelling, such as identifying walking bouts in a continuous wearable data stream, is essential for automatically labelling walking periods, thereby reducing the time required for offline processing. To achieve an effective data labelling, this thesis 3) develops an AI model that fuses the features of IMU and electromyography (EMG) data. In the case of limited datasets of neurological conditions, as mobility loss is a significant barrier in creating rich and diverse datasets to perform effective model training, this thesis also produce a data augmentation framework. The outcomes of HAR studies performed in this thesis show that IMU and EMG data fusion at the feature level can provide highly accurate activity classification, and data augmentation improves the performance of the AI model in limited neurological datasets. Finally, this thesis focuses on remote gait analysis due to the time-consuming and labour-intensive offline data processing. To mitigate this limitation, this thesis 4) presents a prototype edge device that can perform both real-time HAR and parameter extraction (e.g., step and stride times) without a need for data post-processing. Validation studies show that the developed device can accurately perform remote gait analysis in clinics and beyond. The comprehensive conclusions drawn from this thesis demonstrate that contemporary techniques significantly ameliorate the prevailing limitations in the domain of gait analysis. Utilising advanced methodologies, this thesis successfully addresses previous constraints, paving the way for more automated and comprehensive gait analysis