Artificial intelligence system for the automatic detection of Alzheimer disease through electroencephalographic signals

• Autor(a) principal: Araújo, Teresa Guerra Mendonça de Sousa de
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.14/36513
• topic: Alzheimer disease; Nonlinear multiband analysis; Electroencephalographic signals; Classic machine learning; Deep learning; Doença de Alzheimer; Análise não linear multibanda; Sinais eletroencefalográficos; Machine learning clássico; Deep learning; Domínio/Área Científica::Ciências Médicas::Biotecnologia Médica

Alzheimer’s Disease (AD) stands out as one of the main causes of dementia. This neurodegenerative disease is characterised by the deterioration of human cognitive functions – the accumulation of toxic substances in the brain causes the progressive death of neuronal cells. Worldwide, AD represents around 65% of all dementia cases, affecting mainly elderly people. This disease is composed by four evolutionary stages and the asymptomatic period can last up until 20 years. With respect to the researcher’s community, this topic remains a huge challenge since it is crucial to create a tool to assist the diagnosis in the early stages with the aim of halting the disease progression. In this way, the main purpose of this dissertation is to develop a system that would be able to differentiate each disease stage. Thereby, a nonlinear multiband analysis of the Electroencephalographic Signals (EEG) was performed enabling to study its behaviour and to extract several features from each study group. After a feature selection per electrode, it was executed, by means of Classic Machine Learning (ML) and Deep Learning (DL) methods, the data classification through a process of leave-one-out cross validation. The maximum accuracies obtained were 78.9% (C vs MCI), 81.0% (C vs ADM), 84.2% (C vs ADA), 88.9% (MCI vs ADM), 93.8% (MCI vs ADA), 77.8% (ADM vs ADA) and 56.8% (All vs All). Considering the topographic maps, it can be concluded that the central and parietal brain regions are the ones that present the most significant differences when the study groups were discriminated. In conclusion, it can be stated that entropy features are the most relevant and that DL did not over performed Classic ML results. Regarding the state of the art with the same EEG database, the proposed method outperforms by 2% in the binary comparison MCI vs ADM. This improvement reflects the performance of this powerful tool in detecting AD.