Adaptation in p300 and motor imagery-based BCI systems

Official URL: http://risc01.sabanciuniv.edu/record=b1615031 (Table of Contents)

Brain Computer Interface (BCI) is an alternative communication tool between human and computer. Motivation of BCI is to create a non-muscular communication environment for the use of external devices. Electroencephalography (EEG) signals are analyzed for understanding the user's intent in BCI systems. The nonstationary behavior of brain electrical activity (such as EEG), caused by changes in subject brain activities, environment conditions and calibration issues, is one of the main challenges of BCI systems. Another set of challenges involves limited amount of training data and subject-dependent characteristics of EEG. In this thesis, we suggest a semi-supervised adaptation approach for P300 based BCI speller systems to address these types of problems. The proposed approach is applied on a P300 speller which also incorporates a language model using Hidden Markov Models (HMM). The estimated labels from the classifier are used to retrain the classifier for adaptation. We have analyzed the effects of this adaptation approach on BCI systems with non-stationary EEG data and small size of training data. We propose to solve both problems by updating the BCI system with labels obtained from the classifier. We have shown that such an adaptation approach would improve BCI performance around 30% for systems with limited amount of training data, and 40% for transferring the system subject-to-subject. Moreover, we have investigated the potential use of error related potential (ErrP) signals in the P300-based BCI systems. The detection and classification of ErrP signals in BCI setting are presented along with the experimental analysis of ErrP.