A New Approach For Character Prediction Using P300 ERP

Abstract

In this study, a P300 ERP based character prediction algorithm is designed using dataset II of BCI Competition III. The problem of inter-subject variability has been addressed. The main challenge is to improve the performance of predicting characters using signals from the different subjects for learning and testing. Thesis study includes, a pattern recognition based approach using energy of wavelet coefficients and a statistical classifier. For the extraction of wavelet features, each single epoch was decomposed into five-octaves using the discrete wavelet transform (DWT). The optimal feature set was selected using F-score feature selection method. In order to effectively classify targets (P300) and non-target (non-P300) events Adaboost algorithm was used. Finally, a character prediction algorithm was designed using the row and column information of targets. Multiple channel data was used to perform this task. The performance of the proposed method has improved remarkably for different subjects up to 70% as compare to the best algorithm of BCI competition III which was 26%. However, for same subjects the performance was slightly degraded.