Classification of Error Related Potential (ErrP) in P300-Speller

Abstract

P300-Speller is one of the most popular paradigm for constructing Brain Computer Interface (BCI) system that allows subjects to type letters by focusing on a specific target on a computer screen. When BCI system recognises a different command than the subject’s intentions, an Error-Related Potentials (ErrP) occurs from the subject’s brain as a response. Researchers aim to build classifiers for detecting ErrPs for single subjects. The aim of this work is to build a transfer learning classifier for detecting ErrPs across multiple subjects. We propose two different ensemble approaches for ErrP detection; Random Forest (RF) and ensemble linear Support Vector Machines (SVMs). The effect of different parameters and methods in a pre-processing stage are studied in order to find the best combination for increasing the detection sensitivity and specificity among different subjects. We obtain 68% Area Under Curve (AUC) at F3 electrode across multiple subject by using the ensemble linear SVM. We show that the F3 and C2 are the best electrodes for detecting ErrP. We also show that it is possible to extract the most useful features from centro-frontal electrodes by using 30 PCA components. We obtain 78% (AUC) by using RF with 32 features. To support our work, we compare our results with the Linear SVM classifier where our results were superior. We concluded that both RF and ensemble linear SVM can cope with the heterogeneity among different subjects.