Fourth International BCI Meeting 2010

Asilomar Conference Center, Carmel, California

May 31 - June 3, 2010

Birgit Baernreuther-1, Thorsten O. Zander-1, Jessika Reissland-1, Christian Kothe-1, Sabine Jatzev-1, Jessika Reissland-1, Matti Gaertner-1, Sebastian Welke-1, Scott Makeig-2
1-Berlin Technical University, Human-Machine Systems, Berlin, Germany
2-Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA

Background and Objective: Physiology-based passive BCIs (pBCIs) allow access to information about the user states, responses, or intentions that is difficult to infer from their behavior and/or the behavioral context (Zander et al., 2008). Physiological information that is correlated with externally inaccessible or covert aspects of user states, responses, or intentions may be highly useful for optimizing Human-Machine Interaction (HMI) systems. It has been shown that, for example, the perception of a machine induced error is well detectable by a pBCI and can consequently be used to enhance HMI (Zander et al., 2008). Reissland et al. (2009) showed that even a complex covert aspect of user behavioral intentions, bluffing, could be detected by a pBCI using EEG data. Here, we decomposed the scalp data from the bluffing experiment using ICA to get a better insight into the underlying brain processes and to attempt to increase the bPCI performance.

Methods: Under a collaboration between Team PhyPA (Berlin) and the Swartz Center for Computational Neuroscience (UCSD), six pairs of subjects played a computerized version of a common German dice game during which high-density EEG was recorded from each player. This game features a series of moments in which one player has to decide whether to bluff or to quit. The EEG data recorded during these decision periods was used to extract features including decision-making (Walton et al., 2004) and intention to deceive the opponent (Fang et al., 2003). Using a linear classifier based on automatically selected features (based on Blankertz et al., 2002) from 128 channels of EEG data (or the same processes decomposed into 128 maximally independent component processes) we attempted to predict whether the player would decide to bluff or not. An extended infomax ICA decomposition was computed for each subject's data (Makeig et al., 1996) and the resulting independent components were categorized into brain and non-brain (artifact) processes. Bluffing classification accuracy was estimated using only information from brain or artifact processes, respectively.

Results: Using the time-domain EEG channel data directly, we achieved a classification accuracy of 81.4% (6.5%) across all subjects, estimated in offline analyses by a 10x10-fold cross-validation (biased generalization). This result indeed shows that it is possible to derive covert user intentions from single-trial EEG recorded in a social game context. The ICA-based classification results will be presented at the meeting.

Discussion and Conclusions: This study will contribute to the understanding of how classification results can be improved by focusing on a subset of independent EEG component processes, and should give new insights concerning the relationship between brain and non-brain components and pBCI classification. We note that information from non-brain (artifact) processes contained in EEG-based BCI data can potentially be used enhance classification accuracy in a biologically principled manner.

Return to Current Abstracts