Classificação de imagética motora em sinais eletroencefalográficos: uma abordagem baseada em interface cérebro máquina e aprendizagem profunda

Abstract

Patients with neurological trauma or people with motor disabilities can have their quality

of life improved with assisted devices controlled by brain-machine interfaces that inter- pret electroencephalogram signals to predict imaginary movements. The objective of

this final project of the course is to develop prototypes based on deep neural network architectures for cleaning features, automatically selecting the most relevant attributes through algorithms based on Evolutionary Computing and building a learning machine

for classifying imagined signals. The BCI Competition IV 2a database was used. Ba- sed on the literature results, the first experiment was carried out with the third individual,

12 configurations were tested, the feature extractors: explicit descriptors, InceptionV3 and VGG16, combined with the classified ones: J48, RF100, RF500 and SVM. The best configuration was VGG16+RF500, which was applied to all individuals with the aim of identifying the best individual for our context. For our problem, the individual with the

best performance was the second, with him the deterioration of characteristics was car- ried out with the Genetic Algorithm and Particle Swarm Optimization, the second being

more efficient and with superior classification results. Finally, the performance of the system was compared with and without feature selection, which resulted in a three-fold reduction in training time and maintenance of average accuracy at around 75%. In this way, the proposed objectives were achieved and new tests should be carried out in future projects.