Feature Selection and Blind Source Separation in an EEG-Based Brain-Computer Interface

Efforts to develop a brain-computer interface based on the scalp-recorded electroencephalogram (EEG) have progressed substantially over the past decade. However, most EEG-based BCI systems require subjects to perform tasks that do not directly map to simple binary commands such as “yes” or ”no”. Furthermore, successful BCI implementations often require extensive biofeedback training over many weeks or even months. We have developed a “direct” protocol for a “yes”/”no” EEG-based BCI involving a single session. The overarching research question is whether and to what extent advances in signal processing and data mining methods can provide for betterthan-chance classification of the EEG in such a direct BCI task. We developed and evaluated a feature selection system for optimizing the classification accuracy in the noisy EEG domain. The system uses a modified genetic algorithm (GA) to search the space of feature subsets in a 180-dimensional feature space. Feature subsets are evaluated with a support vector machine (SVM) classifier and 10-fold cross-validation. The system finds feature subsets that provide better classification than both full feature sets and random feature subsets. The system also demonstrates that spectral features composed from blind source separation transformations of the data lead to better classification than spectral features of the original time series. We demonstrate the heightened risk of overfitting the data that accompanies the extensive search of the feature subset space, and provide a preliminary approach for measuring and mitigating the overfitting issue. The results suggest that feature selection can be used to improve the performance of even a direct, cognitive EEG-based BCI.