Unraveling the brain: a quantitative study of EEG classification techniques

• Ano de defesa: 2021
• Autor(a) principal: Alípio, Lênon Guimarães Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Classificação de EEG; EEG classification; Fourier transform; Machine learning; Neural networks; Redes neurais; Transformada de Fourier; Wavelets
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45133/tde-01072021-132416/

The problem of EEG Classification, where one tries to identify neural conditions through electroencephalographic signal analysis, has been gathering increasing attention from the scientific community with the recent advances in EEG technology and Big Data/Machine Learning techniques. However, much of the current research on this topic presents significant methodological flaws, such as non-optimization of models hyperparameters, data leakage between train and test datasets, and poor choice of comparison baselines, among others, which render many of the obtained results dubious. Thus, it is not clear what are the state-of-the-art methods for the EEG Classification problem today, nor how they compare to one another. In this dissertation, we tackle this problem by, first, surveying methods proposed in the scientific literature which claim to achieve state-of-the-art performance while still adhering to data science and statistical guidelines that can sustain such a claim. Then, we make a quantitative comparison of these methods on four different EEG datasets. Of the 11 methods studied, we show that those based on Fourier Transforms, Wavelet Transforms, and Hjorth Parameters are the ones with the best overall performance, and can that they can be used as a strong baseline against which any new methods and analyses hereafter proposed in the EEG Classification field should be compared.