Invariant Common Spatio-Spectral Patterns Hohyun Cho 1 , Minkyu Ahn 1 , Sangtae Ahn 1 , Sung Chan Jun 1 1 School of Information and Communications, Gwangju Institute of Science and Technology, Gwangju, South Korea Correspondence: Sung Chan Jun, School of Information and Communications, Gwangju Institute of Science and Technology, Buk-gu, Gwangju, South Korea 500-712. E-mail: scjun@gist.ac.kr, Phone +82 62 715 2216, Fax +82 62 715 2204 Abstract. To achieve an efficient brain-computer interface (BCI), various feature extraction methods have been developed. Among them, the common spatial pattern (CSP) method and its variants have been used. It has been reported that the common spatio-spectral pattern (CSSP) method incorporating simple spectral information performs better than CSP. However, like CSP, CSSP is less robust to the non-stationarity of EEG. This motivates us to propose invariant CSSP by adding a noise suppression term to the Rayleigh coefficient of CSSP. We investigated how our proposed method is invariant to noise (eye blinking) and its classification performance was comparatively tested with empirical data from 52 subjects. The proposed method — invariant CSSP — outperformed the CSP, iCSP, and CSSP algorithms. Keywords: brain computer interface, feature extraction, CSP, iCSP, CSSP, non-stationarity 1. Introduction An EEG-based brain-computer interface directly transfers information from the brain signal to a computer or machine through an electrical pathway without limb movement. Continuous EEG data is non-stationary and variable over experiments, so BCI may not be an optimal method. A common spatial pattern (CSP) method has been used widely in BCI [H. Ramoser 2000], and several variants of CSP algorithms have recently been proposed. Among them, the common spatio-spectral pattern (CSSP) incorporating spectral information and the invariant common spatial pattern (iCSP) reducing non-stationarity have yielded relatively better performance than conventional CSP [S. Lemm 2005, B. Blankertz 2008]. In this work, we propose a variant of CSP by combining CSSP and iCSP, calling the new method the ‘invariant CSSP’. Adopting the iCSP concept, we add a reasonable noise suppression term to the Rayleigh coefficient of CSSP. 2. Materials and Methods 2.1. Experimental Data Sets, Training, and Simulation A conventional offline motor imagery experiment [H. Ramoser 2000] of left/right hand was performed with 52 subjects (26 females, mean age±std age = 24.8±3.86). Sixty-four EEG electrodes (Biosemi ActiveTwo system) were attached on the scalp in a 10-20 international system and EEG signals were collected with 512Hz sampling rates. A total of 100 trials per class (left or right hand) were collected for each subject. To collect various non-stationary EEG signals, several types of noise data, such as eye blinking, eyeball up/down movement, eyeball left/right movement, jaw biting, head movement, and rest state were recorded for each subject prior to the motor imagery experiment. EEG signals were band-pass filtered between 8-30Hz, and a time segment from 0 to 4 seconds after onset of the motor imagery cue was extracted for each trial. On these preprocessed single trials, we performed feature extractions using the CSP, iCSP, CSSP, and proposed iCSSP methods. We estimated the performance of the algorithms by means of 10-fold cross-validation using FLDA. The noise covariance incorporated into iCSP and iCSSP was estimated using 6 kinds of noise data for each subject. For simulation purposes, only pure “eye-blinking” data was extracted by the ICA method and added to each test trial in a 10-fold cross-validation with different scaling. 2.2. Invariant Common Spatio-Spectral Patterns We propose a modified version of CSSP that is invariant to non-stationarity. The proposed method (invariant CSSP) estimates the spatio-spectral filters that maximize variance for one class and at the same time minimize non-stationary noise and variance for the other class. It can be formulated as follows:         1 2 1 1 1 1 ˆ ˆ max 1 w A w w C w T T w ,         2 1 2 2 2 2 ˆ ˆ max 2 w A w w C w T T w , where     ˆ ˆ ) 1 ( ˆ   i i C A , ] 1 , 0 [   (1)