This thesis focuses on the development of adaptive data-driven single channel and multichannel filtering methods for brain-computer interface (BCI) systems. Magnetoencephalography (MEG) and electroencephalogram (EEG) neuroimaging recording techniques are considered to measure neurophysiological activity. The inherent nonstationarity and nonlinearity in MEG/EEG and its multichannel recording nature require a new set of data-driven single and multichannel filtering techniques to estimate more accurately features for enhanced operation of a BCI. Empirical mode decomposition (EMD) and Multivariate EMD (MEMD) are fully data-driven adaptive techniques. These techniques are considered to decompose the nonstationary and nonlinear MEG/EEG signals into a group of components which are highly localised in the time and frequency domain. Also, it is shown that MEMD based filtering can exploit common oscillatory modes within multivariate (multichannel) data. It may be used to more accurately estimate and compare the amplitude information among multiple sources which serves as a key for the feature extraction of a BCI system. These simple filtering techniques are done at the preprocessing stage which helped to reduce the effect of the nonstationarity to a large extent across the sessions for both binary class and multi-class classification problems and identify features which are somewhat invariant against the changes across sessions. Different features such as Hjorth, bandpower, common spatial pattern (CSP), sample entropy and covariance matrix are extracted in the feature extraction stage for comparative evaluation. A novel subject specific MEMD based filtering and covariance matrix as a feature set approach is introduced to classify the multiple classes using Riemannian geometry framework. This approach helped to achieve high kappa value and classification accuracy when evaluated on BCI competition IV dataset 2a. This novel type of filtering can be applied without initial calibration and has the potential to drastically improve the applicability of BCI devices for daily use. Finally, a novel tangent space based transfer learning approach is proposed which utilizes the shared structure across multiple subjects and is an important step towards zero training time for BCI systems.
