Attribute Selection for EEG Signal Classification Using Rough Sets and Neural Networks

This paper describes the application of rough sets and neural network models for classification of electroencephalogram (EEG) signals from two patient classes: normal and epileptic. First, the wavelet transform (WT) was applied to the EEG time series in order to reduce the dimensionality and highlight important features in the data. Statistical measures of the resulting wavelet coefficients were used for the classification task. Employing rough sets, we sought to determine which of the acquired attributes were necessary/informative as predictors of the decision classes. The results indicate that rough sets was able to accurately classify the datasets with an accuracy of almost 100%. The resulting rule sets were small, with an average cardinality of 6. These results were confirmed using standard neural network based classifiers.