Feature Selection for EEG Data Based on Logistic Regression

The article examines the application of dimensionality reduction methods based on logistic regression, such as stepwise selection and LASSO, when solving the problem of feature selection by which states of rest and increased cognitive load can be identified based on EEG data. The data set “EEG During Mental Arithmetic Tasks” was used for the experiment. To form a matrix of features from EEG records presented in EDF format, features in the time and frequency domains were extracted using the Python library eeglib. During the experiment, the quality of the subsets of features selected by various methods was evaluated, and an analysis was carried out for all possible pairs of features. This made it possible to recommend certain techniques for analyzing EEG data in order to select a set of features that would best classify the level of cognitive load. The study showed that stepwise selection and an iterative procedure that eliminates the correlation of input variables select subset of features that provide 90% classification accuracy. While the regularization methods (LASSO and its modification Bolasso-S) give the worst results. In addition, it is recommended to supplement the analysis with an evaluation of the internal validity of the classification.