Section: New Results

Classification of EEG data by evolutionary algorithm for the study of vigilance states

Participants : Marie Chavent, Laurent Vézard.

The objective of this work [42] is to predict the state of vigilance of an individual from the study of its brain activity (EEG signals). The variable to predict is binary (alertness "normal" or "relaxed"). EEG of 44 participants in both states (88 records) were collected with a helmet with 58 electrodes. After a pretreatment step and data validation, a test called "test slope" was chosen. The usual methods of supervised classification (k nearest neighbors, binary classification trees, random forests, and discriminant sparse PLS) were used to provide predictions of the state of participants. The test was then refined using a genetic algorithm, which has built a reliable model (average true classification rate by using CART equal to 86.68 ± 1.87%) and to select an electrode from the initial 58. This work is in collaboration with Pierrick Legrand (EPI Alea) and Frédérique Faïta (EA 487 "Sciences cognitives et facteur humain"