Feature extraction from EEG signals and regularization for brain-computer interface

• Main Author: Mishuhina, Vasilisa
• Other Authors: Jiang Xudong
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Pattern recognition; Engineering::Computer science and engineering::Computer applications::Life and medical sciences
• Online Access: https://hdl.handle.net/10356/141579
• Institution: Nanyang Technological University

Brain-computer interface (BCI) is a promising technique which analyses and translates brain signals to computer commands. BCI can be used to communicate directly from a brain to a computer instead of communication through the central nervous system (CNS) and muscles. Consequently, it has received tremendous research interest in recent years. Various methods of signal processing and machine learning supported by neuroscience, have made it possible to extract useful information from the brain. This has enabled BCI to be successfully implemented in various communication and rehabilitation applications, such as control of computer cursor, robot, prosthesis, wheelchair etc. Similarly, there are several emerging applications in other fields of gaming, security, and education. At present, there are several techniques available to acquire brain signals. In our work, we use electroencephalography (EEG) that measures electric brain activity by electrodes located on the scalp. We choose EEG because it is a non-invasive and low cost technology with good time resolution. However, it has several difficulties with spatial resolution due to overlapping signals and distortion from the scalp. The goal of this research is to improve feature extraction and regularization of EEG signals using machine learning methods and hence achieve better results during the classification of the signals for motor imagery BCI (MI-BCI). MI-BCI establishes communication through signal generated from the human brain when a person is imagining the movement of some part of the body. The main motivation behind MI-BCI is to improve life of people with different neuromuscular disorders such as amyotrophic lateral sclerosis, cerebral palsy, stroke, or spinal cord injury, as well as to help them with rehabilitation. This thesis presents a detailed explanation of state-of-the-art methods of MI-BCI such as common spatial pattern (CSP) and its extensions. CSP is a widely used algorithm in BCI. However, it suffers from overfitting and sensitivity to noise and outliers. To enhance the performance of CSP, different variations of this method have been proposed in literature. However, each approach has its own limitations and does not perform well for different subjects and datasets. Therefore, in this thesis, we extend the existing approaches and propose new ones to improve classification. Firstly, a novel feature weighting and regularization (FWR) method that utilizes all CSP features to avoid information loss is proposed. The method can be applied in all CSP-based approaches. Experiments demonstrate that the proposed FWR method enhances the classification accuracy as compared to the conventional feature selection approaches. Secondly, this thesis proposes a novel approach time-frequency CSP (TFCSP) which decomposes the EEG signal into time stages and frequency components to find the robust and discriminative features for classification. The proposed algorithm yields higher classification accuracy than other state-of-the-art methods on multiple BCI competition datasets. These improvements will eventually lead BCI to real-life applications.