EEG Multi-Objective Feature Selection Using Temporal Extension

Abstract

Nowadays Electroencephalogram (EEG) devices allow the recording of signals that can be used to extract information necessary to identify different types of cognitive processes. In EEG classification, Feature Selection (FS) represents a pivotal phase, as these problems request the processing of a large amount of high-dimensional patterns. In this paper, FS has been solved by an embedded multi-objective genetic optimization procedure which evolves a population of potential solutions (subsets of features), subject to the simultaneous minimization of the misclassification ratio and number of selected attributes. Random Forests (RF) classifiers are adopted, due to their fast training and their compatibility with spread classes of very diverse patterns. The main contribution presented in this paper consists in introducing an inertial behavior to feature extraction. The available feature set is extended with features from previous time frames, and FS is performed on this extended set. In this context, the experimental analysis illustrates the impact of the temporal extension on FS. Additionally, two enhancements are proposed for the multi-objective optimization, to support an effective Pareto-ranking of the solutions in the expanded exploration search space. Thus, the number of trees in the embedded RF classifier is gradually increased, for reducing the computational load requested for the evaluation of the misclassification ratio, without impeding the exploration. Also, the preference for the minimization of misclassifications is set by introducing a dynamic objective function for describing the parsimony of the selected subset of attributes. The proposed FS is experimentally demonstrated on EEG data collected during mathematical tasks of gradual complexities.