Classification de l'état mental humain par programmation génétique sur des signaux EEG

Abstract

The advances in the development of Brain-Computer Interfaces(BCI) have been increasing in recent years, mostly because the level ofconvergence from multi-disciplinary techniques has evolved. The electroencephalography(EEG), a brain recording method studied in thisthesis, allows the construction of BCIs, however the signals are rathercomplex to process, which requires methodologies that efficiently extractpatterns from them. This thesis explores two directions: first, a systemis proposed for the epilepsy seizures recognition using a combinationof signal processing methods for an efficient feature extraction; second,it explores the usage of a meta-heuristic algorithm, namely GeneticProgramming (GP), as an alternative in the design of BCIs. Nonetheless,there is currently open-issues in GP that this thesis also explores: is therea more efficient search methodology in the exploration by GP?; what isa proper representation depending on the studied problem?; which arethe most adequate search operators?. For the first topic, a thoroughlystudy is presented by introducing a memetic GP applied to regressionproblems. Then, it is extended by adapting it to classification problems.The results are positive; GP is greatly benefited from the combinationof a general and a Local Search (LS) methodology. The last two topicsare studied simultaneously in the development of a recognition systemfor mental states using EEG. A GP version (+FEGP) is proposed thatevolves feature extraction models by using specialized search operators,individuals representation and fitness function. The results show thatthe combination of these reaches a state-of-the-art accuracy for the particulartask of mental states recognition.