A Criterion to Evaluate Feature Vectors Based on ANOVA Statistical Analysis

The objective of this research is to evaluate feature vectors based on statistical analysis, focusing on application in brain-computer interface (BCI) domain. Common spatial pattern (CSP) is one of the most frequently used algorithms in BCI to extract features from electroencephalogram (EEG). However, since CSP would be a greedy algorithm by solving it through eigenvalue decomposition method, choosing features in a sequential way does not necessarily result in the minimal achievable classification error for higher than 2-dimensional feature vectors. To overcome this issue, Unbalanced Factorial ANOVA (UF-ANOVA) analysis based on linear regression has been used in order to evaluate features extracted from CSP algorithm. Finally, a criterion based on Mahalanobis distance and F distribution parameter resulted from ANOVA table is introduced to evaluate feature vectors. It is shown that proposed criterion is compatible with widely used criterions such as Fisher score (FS) and Mutual information (MI). Moreover, proposed analysis is not limited to one-dimensional feature vectors and can be applied to higher dimensions.