Experiments on using combined short window bivariate autoregression for EEG classification

In EEG-based classification problem, most of currently used features are univariate and extracted from single channels. However EEG signals recorded from multiple channels for a brain activity are correlated, features extracted from the EEG signals should reflect relationships among those channels. For this reason, we propose and apply a bivariate feature called Combined Short-Window BiVariate AutoRegres-sive model (CSWBVAR) for EEG classification problems. Given a pair of channels, we firstly divide each of them in to overlapping segments or short windows, and then estimate BVAR parameters for each pair of segments. CSWBVAR is formed by combining extracted BVAR parameters together with a pre-defined overlapping window parameter. We analyzed and compared CSWBVAR feature and univariate feature using the dataset III for motor imagery problem of BCI Competition II (2003). Preliminary results show that using CSWBVAR feature can improve classification accuracy up to 7% comparing with using univariate one with the same linear-kernel SVM classifier.