Comparison of Generic and Subject-Specific Training for Features Classification in P300 Speller

Designing subject-adaptive brain-computer inter-face (BCI) systems using event-related potential (ERP) paradigm is a challenging problem for BCI researchers, as ERP response to the same visual stimuli varies from one human to another. In this paper two different training approaches, subject-specific training (SST) and generic training (GT), are proposed. The first approach employs training classifiers for each subject separately, while the second approach shuffles all the data from different subjects and train classification model on merged data. The proposed approaches are tested for three features classification algorithms: support vector machine (SVM), k-nearest neighbours (kNN) and linear discriminant analysis (LDA). It has been found that the proposed GT approach is very efficient for training kNN classifier, reaching averagely 98% accuracy, while it does not have any noticeable improvements when using LDA. SVM classifier turned to be non-efficient for classification of the target ERP component while using both training approaches.