Mitigating technical challenges in brain-computer interfaces for stroke rehabilitation

• Main Author: Nagarajan, Aarthy
• Other Authors: Guan Cuntai
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Brain-computer interfaces; Motor imagery
• Online Access: https://hdl.handle.net/10356/174103
• Institution: Nanyang Technological University

Brain-computer interfaces (BCIs) provide a means of non-muscular communication by translating brain activity into the control of external devices. Motor imagery (MI) has attracted significant attention among various non-invasive BCI paradigms using electroencephalogram (EEG) for its potential in stroke rehabilitation. However, MI-based BCIs encounter challenges in real-time applications for stroke patients, primarily due to limited reliability and robustness. Additionally, the scarce availability of clinical data impedes the development of cross-subject models for MI detection in stroke patients. Furthermore, the current MI-BCIs do not adequately facilitate the restoration of distal hand functions, which are essential for enhancing the quality of life for individuals with motor impairments. This thesis proposes solutions to address these technical challenges in BCIs for stroke rehabilitation using deep learning (DL) methods. Furthermore, a novel experimental protocol is introduced to enable clinically relevant practical applications of BCIs in stroke patients. The research begins with an extensive literature review focusing on the impact of EEG discrepancies on the performance of BCIs. The review delves into channel selection and transfer learning techniques that aim to enhance the resilience of EEG-BCIs. Recently, there has been a surge in studies investigating subject-independent models in the domain of MI-BCI. This trend is driven by the superior predictive capabilities of subject-independent models based on DL compared to subject-specific models. However, the literature review highlights a significant gap in the research, as most studies in this area have focused primarily on healthy subjects, with limited inclusion of stroke patients. Furthermore, the review encompasses relevant studies exploring MI decoding from the same limb. With the goal of selecting the optimal set of EEG channels to enhance overall classification performance in DL-based MI-BCIs, the author proposes subject-independent channel selection using layer-wise relevance propagation (LRP) and neural network pruning. Traditional approaches to channel selection have focused predominantly on subject-specific optimization, whereas subject-independent methods are essential for the utilization of DL models trained on cross-subject data. The proposed methodology not only achieves a significant reduction in the number of channels but also maintains subject-independent classification accuracy, while ensuring interpretability in terms of underlying neural mechanisms. Furthermore, in consideration of the limited availability of clinical data to train BCI algorithms, the research investigates the feasibility of employing DL models pre-trained on data from healthy individuals to detect MI in stroke patients, while also taking into account the inter-subject variability between the healthy and stroke populations. Through domain adaptation, the transfer learning approach demonstrates improved MI detection accuracy in stroke patients, surpassing subject-specific models. Interpretability analysis using transfer models determines channel relevance patterns that indicate contributions from the bilateral motor, frontal, and parietal regions of the cortex towards MI detection in stroke patients. Furthermore, the healthy-to-stroke transfer learning achieves comparable performance to stroke-to-stroke transfer learning, highlighting its potential to enhance the clinical use of BCI algorithms. Finally, a novel BCI experiment utilizing a robotic exoskeleton for unilateral hand motor attempt (MA) tasks is introduced. The focus of stroke rehabilitation is often the recovery of distal hand function. A mere act of opening and closing the hand has the potential to bring about significant enhancements in the quality of life experienced by individuals who have suffered from stroke. In this research study, MA-EEG data collected from healthy subjects is employed to develop subject-specific and subject-independent DL models. The results highlight the importance of this experiment in driving advancements in stroke rehabilitation. This thesis makes novel contributions to the field by optimizing EEG-BCIs for stroke rehabilitation through subject-independent channel selection, transfer learning from healthy to stroke populations, and a new BCI experiment for same-hand MA-EEG decoding. The findings pave the way for more reliable, applicable, and interpretable BCIs, enhancing their potential for clinical use and rehabilitation purposes.