Electroencephalogram analysis using artificial neural networks

Abstract

Through electroencephalography it is possible to detect the electrical signals created by interconnected neurons that create synaptic connections. This technique has been very important in the detection of neurological disorders such as epilepsy. Characterized by temporary changes in the bioelectric function of the brain, epilepsy causes seizures that affect awareness, movement, or sensation. Artificial neural networks (ANN) provide alternative models for detection, classification and prediction of samples by analyzing the electroencephalogram from the structure of the data, which determine the topology of the network. This article proposes the implementation of a system based on ANN to analyze, classify and process signals from an epileptic training model. In particular, the database has samples with recorded brain activity in healthy patients, patients who controlled the crisis and patients that still recorded oscillations in the signals emitted by the brain activity. After applying the fast Fourier transform, these signals were integrated into a matrix using three types of threshold and selecting the input data of an ANN for training and validation. Two methods of learning are considered: multilayer neural networks with classic validation (back propagation) and neural networks using leave one out crossed validation (LOOCV), for which the mean square error (MSE) and the amount of errors threshold are calculated in order to compare the results obtained and to find the method that provides the best results. Both networks were trained using a hybrid method based on simulated annealing and conjugate gradient. Finally, the analysis of ANN as classification systems through the two methods in operation is presented, achieving satisfactory results that show the application as a tool to support the medical diagnosis for the detection of this disorder.