Adaptive classification for brain computer interfaces

In this paper we evaluate the performance of a new adaptive classifier for the use within a Brain Computer-Interface (BCI). The classifier can either be adaptive in a completely unsupervised manner or using unsupervised adaptation in conjunction with a neuronal evaluation signal to improve adaptation. The first variant, termed Adaptive Linear Discriminant Analysis (ALDA), updates mean values as well as covariances of the class distributions continuously in time. In simulated as well as experimental data ALDA substantially outperforms the non-adaptive LDA. The second variant, termed Adaptive Linear Discriminant Analysis with Error Correction (ALDEC), extends the unsupervised algorithm with an additional independent neuronal evaluation signal. Such a signal could be an error related potential which indicates when the decoder did not classify correctly. When the mean values of the class distributions circle around each other or even cross their way, ALDEC can yield a substantially better adaptation than ALDA depending on the reliability of the error signal. Given the non-stationarity of EEG signals during BCI control our approach might strongly improve the precision and the time needed to gain accurate control in future BCI applications.