A Novel Method based on Regularized Logistic Regression and CCA for P300 Detection using a Reduced Number of EEG Trials
Keywords:brain computer interfaces, even related potential, p300, canonical correlation analysis, regularized logistic regression
The P300 event-related potential (ERP) is an electroencephalography (EEG) signal evoked by external stimuli,characterised by a positive deflection at the 300 ms after presentation of an interesting stimulus for the user. In literature, P300-based brain-computer interface (BCI) has been implemented to translate the subject's intent for restoring communication and control functions. In order to detect the ERP of the EEG signals and taking into account low signal-to-noise ratio (SNR) of the P300 signal, it have been widely used methods to average multiples trials, which permit to diminish the random noise. However, the signal averaging technique requires to process multiple EEG trials, thus the ITR (Information Transfer Rate) of the P300-based BCI is significantly reduced. An open challenge of P300-based BCI systems focuses on recognizing ERP signals (as target and no-target) using a reduced number of trials with an enough classification accuracy. In this work, we propose a novel method based on regularized regression logistic using as features the coefficients obtained from a canonical correlation analysis(CCA), as a method for detection of visual P300 ERPs using a reduced number of EEG trials.The proposed method was evaluated with a freely available EEG dataset, and was compared with two widely used methods: mean-Amplitude LDA and stepwise LDA, using only 1,2,3 and 4 trials for processing. As results, proposed novel method outperforms other standard algorithms for P300 detection, providing accuracies of 83.5%(1 trial), 84.7%(2 trials), 85.6%(3 trials) and 86.3%(4 trials) and Area Under Curve (AROC) of 0.657, 0.728, 0.745 y 0.753 respectively.