Hybrid and model based approaches for new BCI spellers

Abstract

Electroencephalography (EEG) based brain-computer interfaces (BCIs), due to their non-invasive, portable and temporal resolution properties, are widely used in the field of neural engineering. In order to make BCI paradigms more practical and feasible for real life applications, new approaches are being tested such as hybrid BCIs and model based BCIs. In the first phase of this study, a novel hybrid speller BCI is proposed, incorporating P300 and code-modulated visual evoked potential (c-VEP) paradigms, with the objective of improving the spelling accuracy and information transfer rate (ITR), compared to individual P300 and c-VEP paradigms. Moreover, fusion techniques have been applied in order to effectively combine the information of P300 and c-VEP at the score level. We have implemented and compared two different approaches, linear discriminant analysis (LDA) and maximum probability estimation (MPE), in order to identify which one works best for this hybrid BCI. The proposed BCI consists of 36 targets presented as 6x6 matrix on screen with a refresh rate of 120 Hz. Seven healthy subjects participated in experiments where each subject performed a training session followed by five test sessions. The P300 and c-VEP signals are obtained by using bandpass filters of 0.5-6 Hz and 6-41 Hz respectively, on the raw hybrid data. For P300, stepwise linear discriminant analysis (SWLDA) is performed on training data from all the 10 EEG channels to obtain the feature vector. For c-VEP, canonical correlation analysis (CCA) is performed on training data to obtain the reference templates for all 36 symbols. In comparison with the accuracy and ITR values of c-VEP alone, that is without simultaneously making use of the P300 data obtained during the hybrid experiments, MPE-based hybrid improved only by 1.1% and 2.1 bits/min, on average, respectively, whereas the values worsened by 12.3% and 19.8 bits/min in the case of LDA-based hybrid. Moreover, the statistical tests on the mean accuracy and ITR values of all the subjects showed that the results of MPE-based hybrid and of c-VEP alone are not statistically different (p=0.293). Although the MPE-based hybrid is not statistically better than the c-VEP alone, it can be highly effective if the primary goal is to only increase the accuracy, using a range of improvements in P300 methods as discussed in conclusion. However, it would not be useful if the purpose is to increase the speed of the speller since the individual c-VEP paradigm, when optimized for timing, has the capability of giving an average ITR of 114.9bits/min or higher, on its own. In the second phase of this study, model based c-VEP BCI is implemented, aimed at improving the training time compared to the case where all the targets are assigned arbitrary pseudorandom binary sequences and training is required for all the symbols separately. For this purpose, moving average model has been implemented to simulate the responses for c-VEP visual stimulation patterns, for 60Hz and 120Hz monitor refresh rate respectively. The average of the correlation between measured response and modeled response for 60Hz and 120Hz is 0.357 and 0.396 respectively. The average accuracy and ITR obtained for model based c-VEP BCI is 87.1% and 76.4 bits/min for 60Hz respectively and 82.1% and 72.4 bits/min for 120Hz respectively. Modeling results suggest that it is possible to perform a training on a single visual stimulus pattern and achieve a good fit model.