Final Activity Report Summary - BRAINCOM (Increasing the Information transfer of EEG-based brain-computer Interfaces.)
A key challenge of EEG-based Brain-Computer Interfaces is increasing the information transfer rate and ensuring that only the signals are generated by the brain and not by peripheral organs like muscles or eye movements.
Typically, the EEG data is processed in several steps, first some spatial filters (e.g. rereferencing, bipolar, Laplace, CSP) is applied, then feature are extracted, which are then classified. The classical BCI based on motor-imagery uses features characterising the temporal autocorrelation (e.g. bandpower, or adaptive autoregressive (AAR) parameters) in combination with linear discriminant analysis (LDA) for classification.
Within the present work, it was possible to improve the classification performance (on an average from 21 recordings) by more the 6 % and increasing the mutual information from 0.196 to 0.34 bits/trial. This improvement was obtained by several means: (i) a stable version of the RLS algorithm could be obtained enabling a replacement for Kalman filtering, (ii) the variance of the innovation process has been included as additional features.
Also some work on removing ocular artefacts (signals caused by eye movements) has been performed. In a recent work, a fully automated method based on regression analysis for reducing ocular artefacts was validated. Alternative approaches suggest methods based on independent component analysis (ICA). Often, ICA methods are only semi-automated but there are some suggestions how to make them fully automated. These methods have been implemented and compared. The manuscript is currently in preparation. The software algorithms have been incorporated into "BioSig - an open source software library for biomedical signal processing".
Typically, the EEG data is processed in several steps, first some spatial filters (e.g. rereferencing, bipolar, Laplace, CSP) is applied, then feature are extracted, which are then classified. The classical BCI based on motor-imagery uses features characterising the temporal autocorrelation (e.g. bandpower, or adaptive autoregressive (AAR) parameters) in combination with linear discriminant analysis (LDA) for classification.
Within the present work, it was possible to improve the classification performance (on an average from 21 recordings) by more the 6 % and increasing the mutual information from 0.196 to 0.34 bits/trial. This improvement was obtained by several means: (i) a stable version of the RLS algorithm could be obtained enabling a replacement for Kalman filtering, (ii) the variance of the innovation process has been included as additional features.
Also some work on removing ocular artefacts (signals caused by eye movements) has been performed. In a recent work, a fully automated method based on regression analysis for reducing ocular artefacts was validated. Alternative approaches suggest methods based on independent component analysis (ICA). Often, ICA methods are only semi-automated but there are some suggestions how to make them fully automated. These methods have been implemented and compared. The manuscript is currently in preparation. The software algorithms have been incorporated into "BioSig - an open source software library for biomedical signal processing".