Feature selection and blind source separation in an EEG-based brain-computer interface (Q2502801)

Summary: Most EEG-based BCI systems make use of well-studied patterns of brain activity. However, those systems involve tasks that indirectly map to simple binary commands such as ``yes'' or ``no'' or require many weeks of biofeedback training. We hypothesized that signal processing and machine learning methods can be used to discriminate EEG in a direct ``yes''or ``no'' BCI from a single session. Blind source separation (BSS) and spectral transformations of the EEG produced a 180-dimensional feature space. We used a modified genetic algorithm (GA) wrapped around a support vector machine (SVM) classifier to search the space of feature subsets. The GA-based search found feature subsets that outperform full feature sets and random feature subsets. Also, BSS transformations of the EEG outperformed the original time series, particularly in conjunction with a subset search of both spaces. The results suggest that BSS and feature selection can be used to improve the performance of even a `direct' single-session BCI.