Overcoming Inter-Subject Variability in BCI Using EEG-Based Identification

Abstract

The high dependency of the Brain Computer Interface (BCI) system performance on the BCI user is a well-known issue of many BCI devices. This contribution presents a new way to overcome this problem using a synergy between a BCI device and an EEG-based biometric algorithm. Using the biometric algorithm, the BCI device automatically identifies its current user and adapts parameters of the classification process and of the BCI protocol to maximize the BCI performance. In addition to this we present an algorithm for EEG-based identification designed to be resistant to variations in EEG recordings between sessions, which is also demonstrated by an experiment with an EEG database containing two sessions recorded one year apart. Further, our algorithm is designed to be compatible with our movement-related BCI device and the evaluation of the algorithm performance took place under conditions of a standard BCI experiment. Estimation of the mu rhythm fundamental frequency using the Frequency Zooming AR modeling is used for EEG feature extraction followed by a classifier based on the regularized Mahalanobis distance. An average subject identification score of 96 % is achieved.