Inteligência Artificial Aplicada a Análise de Eletroencefalografia

• Ano de defesa: 2018
• Autor(a) principal: Duarte, Júlia Bertelli
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Engenharia Mecânica
• Palavras-chave em Português: Eletroencefalografia; Electroencephalography; Epilepsia; Epilepsy; Transformada Wavelet; Wavelet Transform; Transformada de Hilbert Huang; Hilbert Huang Transform; k-NEO. Entropia; k-NEO; Análise Envelope; Entropy; Análise Discriminante Linear; Envelope Analysis; Análise Discriminante Quadrática; k-Nearest Neighbor; WkNN; Rede Neural Probabilística; Perceptron multicamadas; Algoritmos Genéticos; Algoritmos Genéticos.; Evolução Diferencial; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA
• Link de acesso: https://repositorio.ufu.br/handle/123456789/22949; http://dx.doi.org/10.14393/ufu.te.2018.815

The technological development of electronics, linked to growth of scientific knowledge on the causes of human diseases in the late twentieth century, made it possible to develop new equipment and diagnostic and therapeutic techniques in Medicine. An artificial intelligence system must be able to stock knowledge, applied this knowledge and acquire new knowledge through experience. The physics spend a large time in neurological exams analysis and most of all are not an anomaly. To assist the specialist, we want to develop a program, based on artificial intelligence, capable of separating the brain signals into signs with and without anomalies. Electroencephalogram signs are used in this work because of the great need for programs to aid in their diagnosis. To do this, an available database of Bern-Barcelona are used, which consists of signs with and without the presence of ictal (signal event caused by an epileptic seizure). For the evaluation of the developed processes, we used the metrics of sensitivity, specificity, accuracy, positive predictive values, negative predictive values and Mathew correlation coefficient. We used six classification procedures (kNN, WkNN, LDA, QDA, PNN and MLP-BP) and two optimization algorithms (genetic algorithm and differential evolution) to test an application of these in the classification of focal and non-focal signals. For calculation of symptoms, statistical analysis such as a classical statistical analysis (RMS value and kurtosis, for example), Wavelet and Hilbert-Huang Transforms, analysis envelope and entropy and k-NEO calculations are used. From a pre-analysis, the parameter that most repeat as possible symptoms as RMS value, crest factor and kurtosis, with 15, 14 and 10 occurrences respectively. The best results observed is using WkNN, with differential evolution, which resulted in an accuracy of approximately 84%.