Development of EEG based BCI approaches for detection of awareness in human disorders of consciousness

Electroencephalogram (EEG) based Brain Computer Interfaces (BCIs) have been successfully developed to help patients with motor disabilities but with retained cognitive abilities. In this thesis, the BCI techniques are developed for patients with severe brain injuries such as those in minimally conscious states (MCS) and vegetative states (VS). In 2006, neuroimaging based volitional imagery paradigms akin to the ones used for the development of motor imagery based BCIs revealed that a VS patient could produce neural responses indistinguishable from those produced by a healthy subject. The work presented in the thesis is inspired by this revelation and presents first attempts to develop electrophysiology based objective bedside methods to detect awareness in disorders of consciousness. The benefit of electrophysiology based methods is that they are able to register the response from the brain immediately and provide far better time resolution than imaging. As many patients either cannot undergo a fMRI scan or do not have access to it, it is believed that long term benefits to quality of life for this patient group can be better achieved at the bedside by an electrophysiological solution. In order to achieve the objectives, EEG data is collected using two BCI approaches: volitional imagery and event related potentials (ERPs) through rare/odd presentation of a target stimulus amongst a sequence of stimuli which produces high amplitude EEG wave after 300ms of its occurrence, this is called P300. Four different variants of volitional paradigms of 'imagine playing tennis' and 'spatial navigation' are used to collect data from 19 healthy subjects and the P300 speller is used to collect data from 5 healthy subjects, two MCS and two VS patients. In the case of imagery data, a channel selection scheme based on classifier performance, which also evaluates the contribution of each channel to the classification process, is used. This scheme is developed from the offline analysis of a benchmark dataset from the BCI competition III. The comparative results of algorithms for BCI imagery data analysis (time domain parameters (TDP), adaptive autoregressive (AAR) and bandpower (BP) for feature extraction and linear discriminant analysis (LDA), support vector machines (SVM) for classification) is presented to determine the feasibility of using these paradigms with patients. Consistent performance accuracy Figures for classification, in the range of 80-90%, are achieved showing that volitional tasks are distinguishable through EEG. A combination of AAR and LDA outperformed the other combinations of algorithms. The actively contributing channels, in achieving these classification results, are used to create EEG signatures for the volitional tasks. The EEG signatures indirectly signify the areas of brain activation for each of the volitional tasks and are found to be comparable to those obtained from neuroimaging. The validation of techniques is performed using a two class, 64-channel electrocorticogram (ECoG) dataset and initial data exploration was performed using principal component analysis (PCA). The derivative of the linear least fit polynomial was used as features and 64% classification was achieved on the unlabelled test data with multi layer perceptron (MLP) as the benchmark mechanism. Ten channels which actively contributed to the classification process were selected using genetic algorithms (GAs), thereby reducing the dimensionality, an important benefit when analysing multichannel, multi-trial datasets. Feature extraction techniques, which can combine spatial and temporal information such as common spatial patterns (CSP), were evaluated and 86% trials were classified correctly using MLP classification. The validation of classifier performance based channel selection produced six channels of interest, the bipolar combinations of which produced a best accuracy of 86% classification with AAR features and LDA classifier and also with TDP features and SVM classification. The P300 data recorded from the patients was investigated for a reproducible P300 response to the target letters. This is achieved by signal averaging and the analysis of square of Pearson‘s correlation coefficient (r-square). Clearly identifiable differential responses to the target letters were observed for three patients. It is believed that with auditory addition to the stimulus presentation in the stimulation procedure, training and consistency of responses, a tool for an objective method of diagnosis and assistive communication could be developed for this patient group. The BCI technology had not been used for the cognitively impaired patient groups such as MCS and VS, hence, the results of this work are new and contribute to bridging the gap between the core BCI research and its applications for patients. The objective measures of awareness developed through EEG based BCI methods will help to reduce the misdiagnosis rate, which is 43% for this patient group. The findings of the work presented in this thesis can be used to further develop an assistive communication tool for patients in this group