Multi-Objective Optimization of Feature Selection procedure for EEG signals classification

EEG is one of the biomarkers adequate for memory load assessment. Feature Selection (FS) routines for electroencephalogram (EEG) signals have been extensively studied in the past years. Current research is often based on machine learning algorithms. This paper investigates the impact of a new evolutionary approach to Multi-Objective Optimization (MOO) of FS routine for memory load classification using EEG. The main contributions of this study consist in: i) the analysis of candidate feature vectors regarding the accuracy of classification and parsimony; ii) the development of a multi-objective Genetic Algorithm (GA) with an improved selection of the preferred final solution, able to optimize the accuracy and the parsimony of any associated classifier. The feature vectors are classified using Support Vector Machine (SVM) and Random Forests (RF) classifiers. The suggested algorithms are verified on two benchmark datasets, optimizing feature selection for memory load level recognition using EEG data. The testing scenarios motivate the necessity of MOO for EEG signal feature selection. The superior performance of the proposed method recommends it as a preprocessing phase in EEG complex applications.