Wearable sensor-based rehabilitation exercise assessment for post-stroke rehabilitation

This thesis focuses on the use of wearable sensors (WS) and machine learning (ML) algorithms in post-stroke rehabilitation assessment. The conventional approach to rehabilitation involves subjective clinical assessments and frequent therapy sessions, which are time-consuming, costly, and often limited in availability. To address these limitations, WS have emerged as a portable and cost-effective solution, enabling patients to perform rehabilitation exercises at home. These sensors provide quantitative data on patients' movements, allowing for continuous monitoring and assessment. Additionally, ML algorithms offer the potential to enhance the accuracy and efficiency of rehabilitation assessment by processing the data collected from WS. The research presented in this thesis first aims to analyse recent developments in WS-based post-stroke rehabilitation assessment, identify limitations in the field, and propose state-of-the-art ML algorithms to improve assessment performance. The primary motivation is to provide a more comprehensive, personalised, and objective evaluation of motor function and mobility, leading to improved rehabilitation outcomes and quality of life for stroke survivors. Chapter 2 provides a comprehensive literature review that examines the current state-of-the-art in post-stroke rehabilitation assessment, specifically focusing on the utilisation of wearable sensors and machine learning techniques. The review encompasses a thorough examination of commonly employed sensors, targeted body limbs, outcome measures, study designs, and machine learning approaches. Furthermore, the review highlights the limitations encountered by researchers in the field, particularly pertaining to the accuracy of assessment algorithms and the availability of data. Subsequent chapters in this thesis address these identified limitations by proposing innovative solutions. Chapter 3 presents an approach aimed at enhancing the accuracy of assessment algorithms by adapting widely used computer vision algorithms to the time-series domain. This adaptation enables more precise and reliable analysis of the collected time-series data, thereby improving the assessment process. In Chapter 4, a novel methodology is introduced, which involves the transformation of time-series data into images and the subsequent utilisation of computer vision algorithms for assessment purposes. Furthermore, a linear interpolation methodology is implemented to adjust the size of the encoded images, allowing for an increase or decrease in dimensions. A comprehensive comparative analysis is then conducted to evaluate the impact of image size on the performance of the assessment algorithm. Finally, Chapter 5 introduces a novel algorithm that generates heterogeneous and realistic data, which serves to enhance the rehabilitation assessment process. By generating synthetic data that closely resembles real-world scenarios, this algorithm addresses the limitation of limited data availability, ultimately leading to more robust and accurate assessments. The contributions of each chapter provide insights into the current state-of-the-art in WS-based rehabilitation assessment, algorithm optimisation, data encoding techniques, and data augmentation strategies. The findings of this research aim to advance post-stroke rehabilitation outcomes and contribute to a more accurate and personalised assessment for stroke survivors.