Dept. of Biomedical Engineering
Oregon Health & Science University
Brain Computer Interfaces (BCIs) refer to direct interactions between human brains and computers, which offer non-muscular communication and control channels. BCIs are particularly useful in some applications, such as sensorimotor control, neuromuscular disorders rehabilitation, task-related performance augmentation, and target recognition. Modern non-invasive BCIs use electroencephalography (EEG) to measure brain activities, and use various signal processing and machine learning techniques to interpret the results. From the machine learning point of view, a BCI system can be seen as a classification system, which contains five parts: 1) pre-processing; 2) feature extraction; 3) feature selection and dimensionality reduction; 4) classification; and 5) post-processing. However, BCI applications are highly data-dependent, and no universal solution exists for all applications to solve the robustness, realtime, and nonstationarity problems. In this thesis, we propose that feature manipulations, including feature extraction, feature selection and dimensionality reduction, can solve or at least partly solve the robustness, realtime and nonstationarity problems. Our research focuses on two BCI applications: Augmented Cognition (AugCog) and single trial ERP detection. Augmented cognition (AugCog) is an embryonic concept aiming at enhancing the task-related performance of human users through computer-mediated assistance based on assessments of cognitive states in real-time during the execution of certain tasks. In this application, we develop different linear and non-linear feature selection and dimensionality reduction methods using Mutual Information (MI) as the criterion. We also develop a statistical similarity based approach for feature extraction. Single trial ERP detection aims at detecting Event Related Potential (ERP) after the stimuli onset, which can be used for target recognition. In the single trial ERP detection application, we compare different feature extraction, feature selection and dimensionality methods in the time, frequency and spatial domains. Experimental results show that the proposed methods improve the performance of BCI systems compared with our baseline systems. For each method we present, we also discuss both advantages and disadvantages, and give general guidelines in selecting different techniques for different data structures.
School of Medicine
Lan, Tian, "Feature extraction feature selection and dimensionality reduction techniques for brain computer interface" (2011). Scholar Archive. 706.