Feature Selection Techniques Analysis for Identification of Cognitive and Resting States Based on EEG Data

The paper investigates the application of feature selection methods included in the Weka software package to solve the problem of classifying cognitive states and resting states by EEG data. For the experiment, the data set “EEG During Mental Arithmetic Tasks” was used. From the EEG records for 19 scalp electrodes, Hjort parameters and power in the delta, theta, alpha, and beta frequency bands were extracted. At the first stage of the experiment, classifying models were trained on the obtained feature space. At the second stage of the experiment, a study was made of the change in the accuracy of classifiers trained on the same set of algorithms on a subset of features obtained using feature selection methods. To obtain a subset of features, the attribute evaluators CfsSubsetEval and WrapperSubsetEval were used. We selected 25 machine learning algorithms included in the Weka software package, for which the classification accuracy exceeded 75% when trained on the original feature space. Feature selection using the CfsSubsetEval attribute estimator and the GreedyStepwise search method was successful for 13 algorithms. Feature selection using the WrapperSubsetEval estimator with the Logistic classifier and the GreedyStepwise search method was successful for12 algorithms.