Feature computation for BCI applications: A general purpose approach using a genetic algorithm. Preliminary results

The main goal of a BCI system is to create a communication channel independent of muscles' activation. This is accomplished by recognizing specific mental states and using their detection to trigger actions in a computer controlled environment. To achieve such goal it is necessary to record brain activity, typically through EEG, and then process the recorded signal to compute features allowing the detection of the user's mental states being monitored. In recent years several paradigms for BCI have been developed, each based on different neural mechanism underlying the generation of a specific signal pattern. Several signal processing techniques allowing extraction of meaningful features are described in the scientific literature, yet these techniques may be quite diverse, often specific to both the experimental protocol and setup. Here we developed a general purpose genetic algorithm which proved able to face the problem of computing features allowing efficient trial classification from EEG signals. The algorithm was tested on three different datasets drawn from the BCI competition II and based on slow cortical potentials, motor imagery and self-paced movements, and obtained encouraging results.