Escolha de parâmetros aplicados a modelos inteligentes para o incremento da qualidade do aprendizado de sinais de EEG captados por dispositivo de baixo custo

Abstract

Since the creation of the first electroencephalography (EEG) equipment at the beginning of the 20th century, several studies have been carried out based on this technology. More recently, investigations into machine learning applied to the classification of EEG signals have started to become popular. In these researches, it is common to adopt a sequence of steps that involves the use of filters, signal windowing, feature extraction and division of data into training and test sets. The choice of parameters for such steps is an important task, as it impacts classification performance. On the other hand, finding the best combination of parameters is an exhaustive work that has only been partially addressed in studies in the area, particularly when considering many parameter options, the progressive growth of the training set and data acquired from low-cost EEG equipment. This thesis contributes to the area by presenting an extensive research on the choice of parameters for processing and classifying of EEG signals, involving both raw signals and specific wave data collected from a low-cost equipment. The EGG signals acquisition was done with ten participants, who were asked to observe a small white ball that moved to the right, left or remained stationary. The observation was repeated 24 times randomly and each observation situation lasted 18 seconds. Different parameter settings and machine learning methods were evaluated in classifying EEG signals. We sought to find the best parameter configuration for each participant individually, as well as obtain a common configuration for several participants simultaneously. The results for the individualized classifications indicate better accuracies when using data from specific waves instead of raw signals. Using larger windows also led to better results. When choosing a common parameter combination for multiple participants, the results indicate a similarity to findings when looking for the best parameters for individual participants. In this case, the parameter combinations using data from specific waves showed an average increase of 8.69% with a standard deviation of 4.02%, while the average increase using raw signals was 7.82% with a standard deviation of 2.81%, when compared to general average accuracy results. Still in the case of the parameterization common to several participants, the maximum accuracies using data from specific waves were higher than those obtained with the raw signals, and the largest windows appeared among the best results.