Performance analysis of a Principal Component Analysis ensemble classifier for Emotiv headset P300 spellers

Abstract

The current trend to use Brain-Computer Interfaces (BCIs) with mobile devices mandates the development of efficient EEG data processing methods. In this paper, we demonstrate the performance of a Principal Component Analysis (PCA) ensemble classifier for P300-based spellers. We recorded EEG data from multiple subjects using the Emotiv neuroheadset in the context of a classical oddball P300 speller paradigm. We compare the performance of the proposed ensemble classifier to the performance of traditional feature extraction and classifier methods. Our results demonstrate the capability of the PCA ensemble classifier to classify P300 data recorded using the Emotiv neuroheadset with an average accuracy of 86.29% on cross-validation data. In addition, offline testing of the recorded data reveals an average classification accuracy of 73.3% that is significantly higher than that achieved using traditional methods. Finally, we demonstrate the effect of the parameters of the P300 speller paradigm on the performance of the method.

P300-based applications for interacting with smart mobile devices

“…Unlike traditional approaches, we apply PCA to the data recorded on each individual channel of the training data to find the principal components of each channel [5]. The training data are then projected on the selected principal components.…”

“…In addition, we reduced the number of trials for each symbol to 10 trials compared to 15 trials which we used in our previous studies [4,5]. For each of the 6 subjects, we recorded a labeled dataset of 20 characters as training data and an unlabeled online testing dataset of 12 symbols (i.e.…”

“…We used an intensification time of 100 ms and an inter-intensification time of 300 ms. A time segment of 90 samples that corresponds to approximately 700 ms was then extracted for each epoch. These parameters were chosen based on our previous performance analysis of data recorded using the Emotiv EPOC neuroheadset [5].…”

“…In CAR filter, the mean of all channels samples at the same time instant is subtracted from each sample to eliminate noise affecting all channels at this time instant. A moving average window of 13 samples was then applied to each channel to smooth the recorded data [5]. As an additional step in the cross-subject generalization testing, we normalized the data of each subject separately by, first, computing the mean and standard deviation of the signals recorded on each channel.…”

“…In both applications, we utilize a principal component analysis (PCA) ensemble classifier which we have recently introduced that has low computational complexity and achieves acceptable classification accuracy [4,5]. In this paper, we use PCA for feature extraction and Fisher's Linear Discriminant for classification.…”