Machine learning approaches for detecting depression using eeg signals

• Autor(a) principal: Oliveira, Eunice Monteiro de
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.14/37977
• topic: Convolutional neural network; Depression; Discrete wavelet transform; EEG signals; Machine learning; Depressão; Rede neural convolucion; Sinais EEG; Transformada de wavelet discreta; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Médica

The burden of neurological disorders continues to grow, as an estimated 264 million people currently suffer from depression worldwide. Due to the stigma surrounding mental illness and to the standard diagnostic approach being so human-intensive, depressed individuals are less likely to seek help. Additionally, the diagnosis’ results are dependent on the doctor’s experience. When patients are misdiagnosed, the search for physical explanations of symptoms further increases the medical care cost, which a lot of people are not able to support. Hence, the search for a cost-effective, objective, and less human-intensive diagnostic method of depression happens to be crucial. The present study focused on developing a tool capable of detecting indicative patterns of depression and automatically discriminate patients with depression through EEG signal analysis. With resource to 1D Discrete Wavelet Transform, a multiband analysis of the signals was performed per EEG channel. After the feature extraction and the feature selection processes, the obtained features fed 25 Machine Learning models and a convolutional neural network (CNN). The three classifiers with the best performance were Linear Discriminant Analysis, Cubic Support Vector Machine, and the designed CNN, with an overall classification accuracy of 94.8%, 93.9%, and 94.9%, respectively. Through these three classifiers, the comparison between depressed subjects and healthy controls reached an accuracy of 100% on several channels. The results obtained by the classifiers alongside an analysis through topographic maps lead to conclude that there is a difference in the frequency of brain waves between the two groups, with a strong incidence in the frontocentral, central, and parietooccipital regions of the scalp. Although EEG signal analysis cannot yet be applied as a diagnostic tool for depression, the findings in this study remain relevant from a theoretical point of view.