Abstract
One of the important areas of brain-computer interface (BCI) research is to identify event-related potentials (ERPs) which are spatial-temporal patterns of the brain activity that happen after presentation of a stimulus and before execution of a movement. One of the important ERPs is the P300 which is an endogenous component of ERPs with a latency of about 300 ms which is elicited by significant stimuli (visual, or auditory). Various machine learning-based classifiers have been used to predict the P300 events and relate them to the human intended activities. However, the vast majority of the employed techniques like Bayesian linear discriminant analysis (BLDA) and regularized fisher linear discriminant analysis (RFLDA) are black box models which are difficult to understand and analyse by a normal clinician. In addition, due to the inter- and intra-user uncertainties associated with the P300 events, most of the existing classifiers need to be trained for a specific user under specific circumstances and the classifier needs to be retrained for different users or change of circumstances. In this paper, we present an interval type-2 fuzzy logic-based classifier which is able to handle the users' uncertainties to produce better prediction accuracies than other competing classifiers such as BLDA or RFLDA. In addition, the generated type-2 fuzzy classifier is learnt from data via genetic algorithms to produce a small number of rules with a rule length of only one antecedent to maximise the transparency and interpretability for the normal clinician. We also employ a feature selection system based on an ensemble neural networks recursive feature selection which is able to find the effective time instances within the effective sensors in relation to given P300 event. We will present various experiments which were performed on standard data sets and using real-data sets obtained from real subjects' experiments performed in the BCI laboratory in King Abdulaziz University. It will be shown that the produced type-2 fuzzy logic-based classifier will learn simple rules which are easy to understand explaining the events in question. In addition, the produced type-2 fuzzy logic classifier will be able to give better accuracies when compared to BLDA or RFLDA on various human subjects on the standard and real-world data sets.