Comparison of Multi-Class Methods of Features Extraction and Classification to Recognize EEGs Related with the Imagination of Two Vowels

One of the most recent applications for the Brain Compute Interfaces (BCI) is the recognition of Electroencephalograms (EEG) related with the imagination of specific words or letters, with the purpose to develop speech prosthesis or communication devices for people with neurological disorders or communication difficulties. This paper’s aim is to acquire signals related with the imagination of the pronunciation of two phonetically opposite vowels (/a/ and /u/), then process these EEGs through the multiclass features extraction methods as the common spatial patterns (CSP) in cascade and the Independent Components Analysis (ICA), with the purpose of comparing the percentage of classification obtained from these features through the Linear Discriminant Analysis (LDA) and through the Support Vector Machines (SVM) with multiclass classification technique of One Vs One. As a result, the features extracted with ICA and classified with SVM with the One vs One technique were the 24 % more accurate than the other combinations of multiclass methods; this combination of methods recognized the 80% of the signals related with the imagination of the vowel /u/ and the 70% of the signals corresponding to the imagination of the vowel /a/.