Modelos de decodificação do EEG para detecção de correlatos neurais de transtornos mentais

• Ano de defesa: 2023
• Autor(a) principal: Alves, Lorraine Marques
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Doutorado em Engenharia Elétrica; Centro Tecnológico; UFES; Programa de Pós-Graduação em Engenharia Elétrica
• Palavras-chave em Português: Transtornos mentais; EEG; Microestados; Teoria de Redes Complexas; Processamento de Linguagem Natural; Engenharia Elétrica
• Link de acesso: http://repositorio.ufes.br/handle/10/12278

The electroencephalogram (EEG) is an important source of signals for assisting in the diagnosis of mental disorders, and new research is constantly exploring ways to use these signals to improve the quality of medical diagnosis. Diseases such as schizophrenia, epilepsy, attention-deficit/hyperactivity disorder (ADHD), depression, dementia, and alcoholism, among others, are examples of how the application of EEG reading and decoding techniques can be of great value in supporting medical diagnosis. Studies of brain dynamics using EEG have revealed that global neural activity can be described by a limited number of scalp topographic electric potential maps, called microstates. This work proposes new methodologies for decoding EEG signals and their application to public databases for the detection of schizophrenia, depression, and ADHD by exploring approaches applied to EEG microstates addressing the problem of binary classification (disorder vs. healthy control). In addition, a third approach was proposed to solve a multiclass classification problem for the simultaneous detection of schizophrenia, depression, and dementia. The proposed methodologies are based on complex network theory and natural language processing. Both proposals allow an understanding of how the brain signals of an individual with one of the mentioned mental disorders differ from those of a healthy person. The complex network theory allowed the determination of important topological characteristics of the constructed microstate networks, resulting in an average accuracy of 100.0% for schizophrenia and depression, and for ADHD, the average accuracies were 99.44% (ADHD vs. healthy) and 98.61% (ADHD-I vs. ADHD-C vs. healthy). The application of natural language processing on symbolic sequences of microstates revealed the importance of the information contained in a window of neighborhoods of the microstate symbolic sequence in characterizing patients with mental disorders, resulting in an average accuracy of 100.0% for schizophrenia, 98.47% for depression, and for ADHD, the average accuracies were 99.38% (ADHD vs. healthy) and 98.19% (ADHD-I vs. ADHD-C vs. healthy). A third approach derived from natural language processing allowed the solution of a classification problem involving multiple disorders, resulting in an average accuracy of 99.19% (schizophrenia vs. depression vs. dementia vs. healthy). These proposals contribute to the assistance in the process of diagnosing mental disorders, being a promising tool in the development of AI-based psychiatry.