Masking Covariance for Common Spatial Pattern as Feature Extraction

Abstract

Brain Computer Interface (BCI) requires accurate and reliable discrimination of EEG’s features. One of the most common method used in feature extraction for BCI is a common spatial pattern (CSP). CSP is a technique to distinguish two opposite features by computing the spatial pattern of the measured EEG channels. The existing CSP is known to be prone to the over-fitting problem. Covariance estimation is an important process in obtaining spatial pattern using CSP. The empirical covariance estimation ignores the spurious information between channels of EEG device. This may cause inaccurate covariance estimation, which results to lower accuracy performance. In this study, a masking covariance matrix is introduced based on the functionality of brain region. The addition of masking covariance is to improve the performance of CSP. Features obtained through features extraction is then used as the input to Extreme Learning Machine (ELM). Comparisons between features of conventional CSP and with the addition of masking covariance are visually observed using the collected EEG signals using EMOTIV. The performance accuracy of the proposed technique has offered slight improvement from 0.5 to 0.5567. The obtained results are then discussed and analyzed in this study. Therefore, by introducing masking covariance matrix, the performance of the existing CSP algorithm can be improved.