Feature Engineering for an Efficient Motor RelatedEcoG BCI System

Invasive Brain Computer Interface (BCI) systems through Electrocorticographic (ECoG) signals require efficient recognition of spatiotemporal patterns from a multi-electrodes sensor array. Such signals are excellent candidates for automated pattern recognition through machine learning algorithms. The importance of these patterns can be highlighted through feature extraction techniques. However, the signal variability due to non-stationarity is ignored while extracting features, and which features to use can be challenging to figure out by visual inspection. In this study, we introduce the signal split parameter to account for the variability of the signal and increase the accuracy of the machine learning classifier. We use genetic selection, which allows the selection of the optimal combination of features from a pool of 8 different feature sets. Genetic selection of features increases accuracy and reduces the BCIs prediction time. Along with Genetic selection, we also use a reduced signal length, which leads to a higher Information Transfer Rate. Thus this approach enables the design of a fast and accurate motorrelated EcoG BCI system.