| spelling |
Rudolf Huebnerhttp://lattes.cnpq.br/9514309218273598Claysson Bruno Santos VimieiroEduardo Jose Lima IIhttp://lattes.cnpq.br/8706806811535099Cristina Natalia Espinosa Martínez2021-10-01T20:13:51Z2021-10-01T20:13:51Z2021-05-27http://hdl.handle.net/1843/38241https://orcid.org/0000-0003-0945-4896Epilepsy is a chronic neural disease that suffers around 50 million people in the world. Since epileptic seizures are spontaneous and, in some cases, cannot be controlled, the approach to detect or even anticipate them is highly required. This dissertation presents the development of two methods of epilepsy seizure detection based on Machine Learning. The first approach based on Feedforward Neural Network (FNN) works on a single electroencephalogram (EEG) channel and extracts the model's input parameters via Discrete Wavelet Transform (DWT) and Fast Fourier Transform (FFT). The second approach based on Convolutional Neural Network (CNN) uses 21 EEG channels, besides the model was analyzed for three input types, in the time domain, the frequency domain via FFT, and the time-frequency domain via Short Time Fourier Transform (STFT). Both FNN and CNN models make a binary classification, where 1 represents ictal activity, and 0 symbolizes a seizure-free period. Then, the models are tested on eight EEGs from Temple University Hospital Seizure Corpus (TUH-SC) database. In addition, the models' performance is evaluated via two training cases. In case 1, one model per patient is created, whereas in case 2, one training set is built, which contains several sessions with different seizure types to train a single model. Hence, analyzing the training case 1, the CNN model with input in the time-frequency domain via STFT reports the best results, achieving an accuracy of 95.3%, recall of 81.6%, the precision of 92.5%, F-score of 85.7%, and error of 4.7%. Training case 2 is applied to the CNN model with input in the time-frequency domain, given that this method was able to detect ictal activity in all patients. Obtaining an accuracy, recall, precision, F-score, and error of 78.1%, 76.9%, 66.2%, 65.4%, and 21.9%, respectively. In conclusion, both detection methods meet the objective of detecting ictal activity. However, the CNN model developed achieves better results when an individual model is created. Therefore, the CNN model was used to develop a prototype of an application that makes it easy to monitor an epilepsy patient.A epilepsia é uma doença neural crônica que afeta 50 milhões de pessoas no mundo. Devido às crises epilépticas serem espontâneas, em alguns casos não podem ser controladas. Portanto, um método ou uma tecnologia para prever e identificar os episódios é altamente requerido. Neste estudo apresenta-se o desenvolvimento de dois métodos baseados em aprendizado de máquina. O primeiro método está baseado na Rede Neural Feedforward (FNN) com backpropagation (BP) que funciona com um canal de eletroencefalografia (EEG), onde as entradas da rede são adquiridas mediante Transformada Discreta de Wavelet (DWT) e Transformada Rápida de Fourier (FFT). No segundo método aplicou-se Redes Neurais Convulsionais (CNN) usando 21 canais de EEG, sendo analisado para três entradas diferentes como domínio do tempo, domínio da frequência usando FFT e domínio do tempo-frequência aplicando-se a Transformada de Fourier em Tempo Curto (STFT). Tanto o modelo FNN quanto o CNN fazem uma classificação binaria, onde 1 representa atividade ictal e 0 simboliza um período livre de crises. Para testar ambos modelos foi usado oito EEG da base de dados Temple University Hospital Seizure Corpus (TUH-SC). Além disso, aplicou-se dois casos de treinamento para avaliar o rendimento dos modelos. Para o caso 1 criou-se um modelo treinado com o conjunto de treinamento de cada paciente. No caso 2, gerou-se um conjunto de treinamento que contém diferentes tipos de crises para treinar um modelo único. Os resultados mostraram que o treinamento 1, utilizando o modelo CNN com entrada no domínio de tempo-frequência reporta uma acurácia de 95.3%, recall de 81.6%, precisão de 92.5%, F-score de 85.7%, e um erro de 4.7%. No caso do treinamento 2, foi aplicado o modelo CNN com entrada no domínio do tempo-frequência, dado que este foi capaz de detectar atividade ictal em todos os pacientes testados, obtendo uma acurácia, recall, precisão, F-score, e erro de 78.1%, 76.9%, 66.2%, 65.4%, e 21.9%, respectivamente. Conforme os resultados obtidos e demonstrados, conclui-se que os métodos desenvolvidos atendem ao objetivo de detectar atividade ictal, sendo que o modelo CNN mostrou melhor solução proposta. Diante disso, o modelo CNN foi utilizado para desenvolver um protótipo de uma aplicação que facilita o monitoramento de um paciente com epilepsia.CNPq - Conselho Nacional de Desenvolvimento Científico e TecnológicoengUniversidade Federal de Minas GeraisPrograma de Pós-Graduação em Engenharia MecanicaUFMGBrasilENG - DEPARTAMENTO DE ENGENHARIA MECÂNICAhttp://creativecommons.org/licenses/by-nc-sa/3.0/pt/info:eu-repo/semantics/openAccessEngenharia mecânicaFourier, Transformações deEpilepsiaEpilepsy detectionConvolutional neural networkFeedforward neural networkShort Time Fourier transformFast Fourier transformElectroencephalogramEpilepsy seizure detection based on artificial neural networks using time-frequency signal processingDetecção de crises epilépticas baseado em redes neurais artificiais usando processamento do sinal em tempo-frequênciainfo:eu-repo/semantics/publishedVersioninfo:eu-repo/semantics/masterThesisreponame:Repositório Institucional da UFMGinstname:Universidade Federal de Minas Gerais (UFMG)instacron:UFMGCC-LICENSElicense_rdflicense_rdfapplication/rdf+xml; charset=utf-81037https://repositorio.ufmg.br/bitstream/1843/38241/2/license_rdfd434b2e45b27c6ef831461f4412a9d4eMD52ORIGINALEpilepsy_Detection_masterDoc_final_biblioteca.pdfEpilepsy_Detection_masterDoc_final_biblioteca.pdfDocumento Dissertaçãoapplication/pdf9426231https://repositorio.ufmg.br/bitstream/1843/38241/1/Epilepsy_Detection_masterDoc_final_biblioteca.pdf5ad62a3956312bc989395b387b213314MD51LICENSElicense.txtlicense.txttext/plain; charset=utf-82118https://repositorio.ufmg.br/bitstream/1843/38241/3/license.txtcda590c95a0b51b4d15f60c9642ca272MD531843/382412021-10-01 17:13:51.395oai:repositorio.ufmg.br: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ório InstitucionalPUBhttps://repositorio.ufmg.br/oaiopendoar:2021-10-01T20:13:51Repositório Institucional da UFMG - Universidade Federal de Minas Gerais (UFMG)false
|