Spatio-temporal matched filter adjustment for enhanced accuracy in brain responses classification

In this paper, we apply modified spatio-temporal matched filtering (MSTMF) to enhance electroencephalographic (EEG) signals in evoked potentials (EP) based brain–computer interfaces (BCI). Our focus is on the effective treatment of noise in the system under consideration.

The applied MSTMF is a spatio-temporal extension of generalized matched filtering, which allows for optimal enhancement of weak, repeatable signals embedded in colored Gaussian noise. However, since spontaneous EEG signals are often corrupted by high-energy super-Gaussian artifacts, which deviate from this distribution, we propose rejecting these artifacts before applying MSTMF. Particularly effective have been algorithms based on independent component analysis (ICA) and empirical mode decomposition (EMD). After artifacts rejection, performed locally within time segments they occupy, without disturbing other parts of the signal, the classification of brain responses becomes more accurate. Nevertheless, the nonstationarity of the EEG signal remains a challenge that must be addressed.

Therefore, we propose adjusting the MSTMF to the current noise properties to improve its performance in this demanding environment. This can be achieved by properly calculating the noise covariance matrix, which is necessary to determine the filter coefficients, using both the learning and currently processed signal segments.

As a result, we have developed an enhanced method based on MSTMF for improved discrimination of evoked potentials and verified its performance on two publicly available reference databases: BCIAUT-P300 (for IFMBE Scientific Challenge) and Speller (for the BCI Competition III Challenge 2004). For these databases, we have achieved overall accuracies of 92.67% and 99.5%, surpassing the reference methods presented in the literature.