Switching kalman filters for BCI data segmentation

Abstract

As a subject performs a specific task for brain–computer interfacing (BCI), the task-specific cortical activity gives rise to scalp electroencephalographic (EEG) data that may be said to represent a sequence of task-related mental states. Continuous segmentation of this data into intervals, each representing a task-related mental state, and its classification is desirable. One approach that has not been extensively studied in the literature makes use of on-line switching filters that switch-in the most appropriate filter for each interval. This work investigates the application of switching Kalman filters to segment and classify EEG BCI data. This study was applied to imagined hand movement whereby each trial at least consists of a transition between a no-movement and a movement state. In, expert models were similarly used to segment specific sleep states in EEG data. Our objective is to develop this approach making it applicable to BCI. We discuss the advantages of this approach and identify a number of issues that make this approach challenging to BCI. Our preliminary results for EEG data of Subject K3 from BCI Competition III, Dataset 3a, show that continuous-time segmentation of the data and its labelling into movement and no-movement data is obtained with at least 74% accuracy. Rather than reporting ensemble average results, as is typical in most BCI literature, we provide results on a single trial basis to specifically highlight the inter-trial.