Estimating the discriminative power of time varying features for EEG BMI.

摘要

Research in brain-machine interfaces requires integration of a number of separate research topics. Indeed, the ability to use observations of brain activity as a channel of interaction is a goal of a number of research communities. Progress in robust brain-machine interfaces (BMIs) requires a number of specializations: brain function, recording techniques for brain activity, computational models of brain activity, feature selection and classification, and feedback mechanisms. Modern BMIs are the integration of these desiderata.;Today, there is a growing need for robust BMIs as an assistive technology. In 2005, it was estimated that 3.17 million Americans were currently living with disabilities resulting from traumatic brain injury (TBI) (102; 71). Traumatic brain injury is the leading cause of disability in children and adults from ages 1 to 44 (1). Many of these people require alternative interaction technologies for both therapy and quality of life. People suffering from degenerative diseases are living longer and require more sophisticated interaction channels: fifty percent of people with Amyotrophic Lateral Sclerosis (ALS) live at least three years after diagnosis, twenty percent at least five years, and ten percent at least ten years (39).;We make the following claim: Sparse regularization improves components analysis in noisy, overcomplete environments; a psychophysiological analysis of mental rotation shows its applicability to BMIs. The combination of these approaches enables more robust BMI. This dissertation is an explanation and elaboration of these concepts and serves as evidence for our claims.;Current research in brain-machine interfaces is progressing from answering questions of effectiveness to questions of efficiency: what brain-machine interface (BMI) approaches facilitate robust interaction? Currently, robust interaction with BMIs is limited by problems of initiating and stopping interaction as well as the presence of artifacts and noise in sensor data. Increased efficiency for BMIs means greater accessibility for different populations. For the disabled population, more ways of indicating intention means greater accessibility for a greater range of impairments. For the general population, new methods of interaction allow for tighter, closed-loop biofeedback mechanisms, with applications such as simple desktop task control, game interaction, and remote robotic control.;Portable sensing arrays such as electroencephalography (EEG) are commonly used for applying neural signals to near real-time control tasks, because they offer minimally invasive sensing arrays for observing neural activity. EEG is particularly able to observe electrical activity of neural cells in response to stimuli with good temporal resolution. These event related potentials (ERPs) are the target signals used by BMIs, and one objective of ERP inference is to be able to identify target activity of a single trial. However, EEG is spatially sparse and sensitive to electrical noise, therefore robust inference requires effective methods for removing artifacts and segmenting target signals. Being able to factor noise artifacts allows us to recover features that better represent the underlying functional processes within the brain. The primary objective of this work is to improve signal classification of ERP data by improving noise factoring methods and by discovering novel ERP patterns.