SARAIVA, E. C.; http://lattes.cnpq.br/0429037888719725; SARAIVA, Eugênio de Carvalho.
Abstract:
Electroencephalography (EEG) data has been applied to various contexts, such as health
and human-computer interaction. EEG allows the production of data by monitoring the
electrophysiological activity of individuals in reaction to experiments or by recording
their physiological state. These data can be labeled for categorization and efficient
retrieval. However, the labeling process may be subject to inaccuracies and noise, often
caused by variations in EEG equipment configuration and divergent subjective opinions
between individuals. In this context, there is a need for mechanisms that contribute to
the understanding of the relationships between the data and the labels assigned to them.
Given the above, this research aims to design an approach to measure similarity between
classes of neuronal activation patterns, with application in the analysis of influencing
factors and classification of EEG data. The approach was outlined through bibliographic
review, construction of tools and acquisition of databases. Experimental evaluations
were guided by questions and hypothesis tests that involved semi-supervised learning,
statistical analysis of the influence of factors on response variable distributions, and
parameter optimization via genetic algorithms. As a result, the problem of similarity
analysis between classes of activation patterns in EEG data was formalized and an
original approach was elaborated. The approach has the following steps: data
acquisition, preprocessing, sample selection, cluster quantity calculation, input
parameters, semi-supervised analysis, and similarity analysis or classification of
neuronal activation patterns. Hypothesis tests indicated the applicability and efficiency
of the approach regarding the level of similarity of neuronal activation patterns between
classes and the classification accuracy rate when compared to other classifiers (neural
networks and support vector machine).