An Innovative Unbalanced ANOVA-Based Approach for Multi-Dimensional Feature Evaluation in Brain-Computer Interface Classification

Brain-computer interface (BCI) stands as a crucial tool in processing, and facilitatingcommunication with intellectually disabled individuals by leveraging brain characteristics through EEGsignal analysis. Effective classification, however, necessitates the use of suitable features. Commonlyemployed tools for evaluating selected features include Fisher Score, MI, and DBI. While these methodssequentially determine the best features, the BCI field predominantly relies on features derived fromthe Common Spatial Pattern (CSP) method, utilizing matrix decomposition under eigenvalues andvectors. Yet, the ordinal mode of CSP's feature selection may not consistently minimize classificationerrors for feature vectors exceeding two dimensions. To address this, we propose a measure based onANOVA statistical analysis of variance, capable of evaluating diverse features, including multidimensionalvectors. In this research, we applied this measure to BCI Competition III and Part Iva datafor simulation. With two classes, hand and leg movement imagery unlabeled data segments wereomitted. Our analysis considered variations in subjects, classes, and trials, addressing the imbalancecaused by noise and motion factors in BCI data. Utilizing UF-ANOVA variance analysis, we evaluatedfeatures extracted from the CSP algorithm. Mahalanobis Distance gauged the distance of each featurevector from the first-class distribution. UF-ANOVA results yielded three F factorial values (۲ each forclass changes, subject changes, and interaction changes), forming the basis for our criterion and indexdefinitions related to factorial class, factorial interaction, and the opposite of factorial subject. Theobtained p-values for all three cases were remarkably low, underscoring the significance ofcharacteristic data related to subject changes, hand or foot movement perception, and their interaction.This novel ANOVA-based measure demonstrates its versatility by accommodating multi-dimensionalfeature vectors, offering a robust approach to feature evaluation in the dynamic domain of BCI.

کلیدواژه ها:

UF-ANOVA variance analysis ، Multi-dimensional feature vectors ، Feature selection ، Brain-Computer Interface