Brain computer interface for post-stroke motor rehabilitation

• Main Author: Mane, Ravikiran Tanaji
• Other Authors: Guan Cuntai
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computer applications::Life and medical sciences
• Online Access: https://hdl.handle.net/10356/146043
• Institution: Nanyang Technological University

Stroke is a devastating neurovascular emergency. As a result of brain damage, stroke patients generally suffer from many functional impairments and among them motor function deficits are most common and have the highest debilitating effects. Therefore, there is an everlasting need for more effective and efficient post-stroke motor rehabilitative interventions. In this quest, in the last decade, the technology of Brain Computer Interfacing (BCI) has emerged as one of the most effective tools for post-stroke motor rehabilitation. BCI systems establish a direct communication pathway between brain and external environment and empower patients for achieving functional recovery by repetitive activation of close loop motor circuits which are damaged as a result of stroke. Therefore, considering the potential of BCI, this thesis aims to improve the existing BCI technology with the ultimate goal of achieving better motor rehabilitation outcomes. In BCI systems, identification of users' intentions is achieved by classifying their brain activation patterns using machine learning algorithms. However, owing to the challenging nature of the brain signals, low classification accuracy remains one of the most important challenges faced by BCI. Particularly, detection of Motor Imagery (MI) from Electroencephalography (EEG) based BCI system, which is the commonly employed paradigm for motor rehabilitation, is among the most difficult tasks in the BCI domain. Therefore, this thesis first presents a novel, neurophysiologically inspired Convolutional Neural Network (CNN) named Filter-Bank Convolutional Network (FBCNet) for end-to-end MI classification from EEG data. The FBCNet is specifically designed to effectively capture the characteristic EEG activation patterns associated with MI known as Sensory-Motor Rhythms (SMR). With almost 8% higher classification accuracy over state-of-the-art methods, we show that incorporation of the neurophysiological priors in the design of deep learning architectures, as done in FBCNet, can lead to significantly more accurate BCI systems. Next, using interpretability analysis, we demonstrate that FBCNet can achieve excellent spectro-spatial localization of discriminative SMRs but it lacks the temporal localization capabilities. Therefore, we extend the FBCNet with temporal localization capabilities. With this modification, compared to the FBCNet, the extended FBCNet further improves the MI classification accuracy by 2.5%. Moving ahead, we analyze the classification performance of proposed and baseline deep learning architectures and traditional machine learning methods for MI detection in 25 chronic stroke patients undergoing three different BCI-based motor rehabilitation interventions for 2/4 weeks. Here, we show that in subject-specific multi-session classification settings, the proposed method of FBCNet can learn highly generalizable discriminative features which remain valid during inter-session classification. Also, from interpretability and various brain region classification analyses we demonstrate that for stroke patients, the most robust and generalizable signatures of MI are present in the motor region of the brain and deep learning architectures that learn to pay prominent attention to these features generally achieve higher inter-session classification performance. Lastly, we analyze the EEG data generated during BCI-based motor rehabilitation from a neuroscientific perspective with an aim to gain a more detailed understanding of the process of BCI-mediated brain recovery. Specifically, we examine several EEG-derived features acquired during two different BCI-driven upper extremity rehabilitative interventions to explore the utility of EEG features for the prognostication of rehabilitative interventions. We further extend this analysis to identify and quantify the treatment-induced neurological changes in the brain activity and explore the differences in the neurological changes associated with the different rehabilitative intervention and their relationship with the treatment characterization. The results of this analysis indicate a possibility of intervention specific prognostic and monitory biomarkers. This approach can be pursued to uniquely predict the expected response of a patient to an intervention and the intervention with the highest predicted gains may then be recommended to the patient, thereby enabling a highly personalized motor rehabilitation. Overall, the research presented in this thesis shows that EEG-based BCI systems can be used to characterize and promote recovery of stroke impaired motor functions and neurophysiologically inspired CNN architectures can significantly advance the technology of MI detection in rehabilitative BCI systems. These presented technological advancements hold a potential to significantly improve the rehabilitation outcomes of the BCI intervention, and the quality of life of stroke survivors. Furthermore, the identification of intervention specific prognostic biomarkers can aid clinicians in suggesting the most suitable intervention for any patient. Future research can use the findings presented in this thesis to design robust and more accurate BCI systems for highly personalized and holistic post-stroke rehabilitation.