Submitted by Automa??o e Estat?stica (sst@bczm.ufrn.br) on 2017-11-01T21:17:51Z No. of bitstreams: 1 KelysonNunesDosSantos_DISSERT.pdf: 1067573 bytes, checksum: 151a98738e7e3c5b3dc97b14478bfd9b (MD5) / Approved for entry into archive by Arlan Eloi Leite Silva (eloihistoriador@yahoo.com.br) on 2017-11-07T22:24:56Z (GMT) No. of bitstreams: 1 KelysonNunesDosSantos_DISSERT.pdf: 1067573 bytes, checksum: 151a98738e7e3c5b3dc97b14478bfd9b (MD5) / Made available in DSpace on 2017-11-07T22:24:56Z (GMT). No. of bitstreams: 1 KelysonNunesDosSantos_DISSERT.pdf: 1067573 bytes, checksum: 151a98738e7e3c5b3dc97b14478bfd9b (MD5) Previous issue date: 2016-07-28 / A predi??o de eventos a partir de dados neurofisiol?gicos possui muitas vari?veis que devem ser analisadas em diferentes momentos, desde a aquisi??o e registro de dados at? o p?s-processamento dos mesmos. Assim, a escolha do algoritmo que ir? processar esses dados ? uma etapa muito importante, pois o tempo de processamento e a acur?cia do resultado s?o fatores determinantes para uma ferramenta de aux?lio de diagn?stico. A tarefa de classifica??o e predi??o tamb?m auxilia no entendimento das intera??es realizadas pelas redes de c?lulas cerebrais. Este trabalho realiza o estudo de t?cnicas de Aprendizado de M?quina com diferentes caracter?sticas para analisar seu impacto na tarefa de predi??o de eventos a partir de dados neurofisiol?gicos e prop?e o uso de comit?s de classificadores de forma a otimizar o desempenho da tarefa de predi??o atrav?s do uso de t?cnicas de baixo custo computacional / Event prediction from neurophysiological data has many variables which must be analyzed in di erent moments, since data acquisition and registry to its post-processing. Hence, choosing the algorithm that will process these data is a very important step, for processing time and accuracy of results are determinant factors for a diagnosis auxiliary tool. Tasks of classi cation and prediction also help in understanding brain cell's networks interactions. This work uses Supervised Machine Learning techniques with different features to analyze their impact on the task of epileptic seizure prediction from canine neurophysiological data and purposes using of ensembles to optimize the performance of event prediction task through computational low-cost techniques. Epileptic dogs' EEG data were preprocessed throug Fourier transform and only significant frequencies were considered (1 to 30Hz). It was applied a dimensionality reductor and then data was submitted to supervised machine learning techniques. Two scenarios were evaluated: first used raw data resulted from Fourier transform, as the second one transform these data. Algorithms evaluation was made through area under ROC curve (AUC) measure. Best results were to scenario A (a) an heterogeneous ensemble formed by a KNN, a decision tree and a bayesian classifier, scoring 0.7074 and (b) an example of decision tree evaluated in 0.687, and, for scenario B, best results were (a) a setup of decision tree which obtained 0.620 and (b) an heterogeneous ensemble composed by a KNN, a decision tree and a bayesian classifier, scoring 0.612.