Análise de dados de eletroencefalograma para diminuição do número de canais

Abstract

Nowadays the number of people that suffer from epilepsy is around 50 million. Of these people, about 70 percent can live seizure-free, controling them with medication. However, thinking about the remaining 30% who still cannot control their seizures, it is necessary to find a method to decrease the number of electrodes needed to identify the beginning of a seizure, in order to make possible the use of a daily wear device that warns the individual when a seizure is coming. Taking this into consideration, this work aims to verify the impact of the decrease in electrodes on the inference of the moment of seizure. For this study, a dataset from the medical center of the American University of Beirut was used with data from the electromagnetic waves, generated by the brain of an epileptic patient, captured by an electroencephalogram. Their data is then adapted and broken down into training and test data to be used by a learning machine, called a random forest, in the Pyhton language. In it the data with all 19 leads were imputted, and from the obtained results, only the 9 most relevant leads were selected to be re-applied to the machine. The results found in both tests showed that the difference in accuracy between them was 3.47%, and the confusion matrices generated were similar to each other. Were separated 1 second windows (500 lines) of the data and took their averages; then, the same procedure was done with the new data obtained. As a result, not only the accuracy increased compared to the first test (with original data and 19 channels), but the result obtained with fewer channels was the highest ever obtained, an accuracy of 91.32%, contrary to the initial expectations of the work. In principle, it is expected that the accuracy of the system will be greater with the use of a larger number of electrodes, since this increase corresponds to a greater amount of information, but one must observe the objective of what one wants to achieve and make compromises if necessary. It was known beforehand that accuracy would suffer; however, it was not expected that with the average and few channels, accuracy would be higher than that obtained using the average and all channels. However, for the purpose of this work, both results - higher and lower accuracy - were satisfactory and the goal was achieved.