SIGNAL PROCESSING APPROACH IN P300 EVENT DETECTION AND CLASSIFICATIONFROM SINGLE TRIAL EEG FOR ATTENTION ASSESSMENT STUDIES

Abstract

To detect P300 Event Related Potential(ERP) in Brain Computer Interface (BCI) experiments is a
challenging task because of its poor Signal to Noise Ratio (SNR) and trial to trial variability. The second major
challenge is the choice of electrodes that varies from subject to subject which involves more computation time
for better classification accuracy in a dense electrode array system. The proposed work is intended to improve
the extraction and classification of P300 from the standard mid-line electrodes FZ, Pz
, Cz
 that are subject
independent with improved signal to noise ratio and reduced computation complexity . Five classifiers namely,
Weighted KNN(W-kNN), Quadratic Discriminant analysis(QD), Bagged Trees(BT), Guassian Naive Bayes(GNB) and Logistic Regression(LR) performance were compared in classifying the target and non target P300
ERPs from single trial. The time domain markers peak and latency of P300 and correlation coefficient obtained
by template matching are the three features used. The bagged tree classifier outperformed with classification
accuracy of 87.9% and AUC 0.95 which is comparatively good with other existing baseline approaches that uses
3 electrodes. With better pre-processing the proposed method will reduce the computational load for a portable
BCI application in attention assessment studies that uses peak and latency features.