Use of genetic algorithm for the selection of EEG features

4th International Conference on Mathematical Modeling in Physical Sciences

物理学;数学

Asvestas, P.^1 ; Korda, A.^2 ; Kostopoulos, S.^1 ; Karanasiou, I.^3 ; Ouzounoglou, A.^2 ; Sidiropoulos, K.^1 ; Ventouras, E.^1 ; Matsopoulos, G.^2

Department of Biomedical Engineering, Technological Educational Institute of Athens, Greece^1

School of Electrical and Computer Engineering, National Technical University of Athens, Greece^2

Institute of Communications and Computer Systems, Athens, Greece^3

关键词: Biomedical informatics;    Brain electrical activity;    Classification accuracy;    Classification algorithm;    Eventrelated potential (ERPs);    Healthy volunteers;    Multivariable functions;    Optimization techniques;   

Others  :  https://iopscience.iop.org/article/10.1088/1742-6596/633/1/012123/pdf
DOI  :  10.1088/1742-6596/633/1/012123

来源: IOP

【 摘 要 】

Genetic Algorithm (GA) is a popular optimization technique that can detect the global optimum of a multivariable function containing several local optima. GA has been widely used in the field of biomedical informatics, especially in the context of designing decision support systems that classify biomedical signals or images into classes of interest. The aim of this paper is to present a methodology, based on GA, for the selection of the optimal subset of features that can be used for the efficient classification of Event Related Potentials (ERPs), which are recorded during the observation of correct or incorrect actions. In our experiment, ERP recordings were acquired from sixteen (16) healthy volunteers who observed correct or incorrect actions of other subjects. The brain electrical activity was recorded at 47 locations on the scalp. The GA was formulated as a combinatorial optimizer for the selection of the combination of electrodes that maximizes the performance of the Fuzzy C Means (FCM) classification algorithm. In particular, during the evolution of the GA, for each candidate combination of electrodes, the well-known (Σ, Φ, Ω) features were calculated and were evaluated by means of the FCM method. The proposed methodology provided a combination of 8 electrodes, with classification accuracy 93.8%. Thus, GA can be the basis for the selection of features that discriminate ERP recordings of observations of correct or incorrect actions.