Publisher:
ISBN:
Category :
Languages : en
Pages :
Book Description
Brain-computer interfaces (BCIs) allow individuals with disabilities to communicate and control their environment without the necessity for volitional speech or motor activation. However, most current BCIs are prone to significant performance fluctuations and incapable of adapting to their users. These shortcomings impair practical BCI usage. This thesis investigates the feasibility of a hybrid BCI that is capable of detecting and adapting to underlying changes in user mental state. The thesis comprises four major studies, each representing a step towards this ultimate goal. The first study formulates a novel signal processing algorithm for the frequency-domain analysis of electroencephalographic (EEG) recordings. The second study examines the ability to automatically detect fluctuations in three mental states that are important to BCI usage - fatigue, frustration, and attention - based on electrical activity recorded from the surface of the scalp using EEG. The third study explores the effects of each of these mental states on the online operation of a two-class EEG-BCI. The final study investigates the efficacy of two different methods - reliability prediction and adaptive classification - by which a BCI can adapt to changes in fatigue, frustration, and attention. In the first study, the novel algorithm, based on a clustering of spectral power features, was shown to compress EEG signals with less information loss than traditional frequency-domain analyses. In the second study, fluctuations in fatigue, frustration, and attention were detected with mean classification accuracies of 76.8%, 71.9%, and 86.1%, respectively. In the third study, a significant relationship between perceived frustration and BCI accuracy was uncovered and optimal regions for BCI performance were identified in several multi-dimensional representations of user mental state. In the final study, estimated mental state was used to predict the onset of low-accuracy BCI performance, with an 8% decrement in classification accuracy between the predicted high and low accuracy conditions. This study also demonstrated the ability to directly adapt BCI classification, leading to statistically significant increases in classification accuracy for roughly half of participants without significantly compromising performance for the remainder of the study population.