Brain-Computer Interfaces (BCI) employ brain imaging to enable human-machine interaction without physical control. BCIs are an alternative so that people suffering from severe or complete loss of motor control, like those with Amyotrophic Lateral Sclerosis (ALS), may have some interaction with the external environment. To transform individual intentions onto machine operations, BCIs rely on a series of steps that include brain signal acquisition and preprocessing, feature extraction, selection and classification. A viable BCI implementation is still an open question due to the great challenges involved in each one of these steps. This gap motivated this work, which presents an evaluation of the main feature extractors used to classify Motor Imagery trials, whose data were obtained through Electroencephalography (EEG) influenced by ocular activity, monitored by Electrooculography (EOG). In this sense, signals acquired by BCI Competition IV-2b, were considered. As first step the preprocessing was performed through Independent Component Analysis (ICA) together with a correlation threshold to identify components associated with ocular artifacts. Afterwards, different feature extraction approaches were evaluated: i) Frequency Subband Dyadic Three; ii) Common Spatial Patterns (CSP); iii) Common Spectral-Spatial Patterns (CSSP); iv) Common Sparse Spectral-Spatial Patterns (CSSSP); v) Filter Bank Common Spatial Patterns (FBCSP); vi) Filter Bank Common Sectral-Spatial Patterns (FBCSSP); and, finally, vii) Filter Bank Sparse Spectral-Spatial Patterns (FBCSSSP). These techniques tend to produce high-dimensional spaces, so a Mutual Information-based Feature Selection was considered to select signal attributes. Finally, Support Vector Machines were trained to tackle the Motor Imagery classification. Experimental results allow to conclude that CSSSP and FBCSSSP are statistically equivalent the state of the art, when two-sided Wilcoxon test with 0, 95 confidence is considered. Nevertheless, CSSSP has been neglected by this area due to its complex parametrization, which is addressed in this work using an automatic approach. This automation reduced computational costs involved in adapting the BCI system to specific individuals. In addition, the FBCSP, CSSP, CSSSP, FBCSSP and FBCSSSP confirm to be robust to artifacts as they implicitly filter the signals through autoregressive models.
Resumo: As Interfaces Cérebro-Computador (BCI) são sistemas que provêm uma alternativa para que pessoas com perda severa ou total do controle motor possam interagir com o ambiente externo. Para mapear intenções individuais em operações de má-quina, os sistemas de BCI empregam um conjunto de etapas que envolvem a captura e pré-processamento dos sinais cerebrais, a extração e seleção de suas características mais relevantes e a classificação das intenções. Neste trabalho, diferentes abordagens para a extração de características de sinais cerebrais foram avaliadas, a mencionar: i) Padrões Espectro-Espaciais Comuns (CSSP); ii) Padrões Esparsos Espectro-Espaciais Comuns (CSSSP); iii) CSSP com banco de filtros (FBCSSP); e, finalmente, iv) CSSSP com banco de filtros (FBCSSSP). Em comum, essas técnicas utilizam de filtragem de banda de frequências e reconstrução de espaços para ressaltar similaridades entre sinais. A técnica de Seleção de Características baseada em Informação Mútua (MIFS) foi adotada para a redução de dimensionalidade das características extraídas e, em seguida, Máquinas de Vetores de Suporte (SVM) foram utilizadas para a classificação do espaço de exemplos. Os experimentos consideraram o conjunto de dados BCI Competition IV-2b, o qual conta com sinais produzidos pelos eletrodos nas posições C3, Cz e C4 a fim de identificar as intenções de movimentação das mãos direita e esquerda. Conclui-se, a partir dos índices kappa obtidos, que os extratores de características adotados podem apresentar resultados equiparáveis ao estado da arte.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.