Aprendizado de máquina como ferramenta para o prognóstico de pacientes em coma usando sinais eletroencefalográficos no espectro de 1 a 100 Hz

• Autor(a) principal: Ramos, Camila Davi
• Data de Publicação: 2022
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFU
• Texto Completo: https://repositorio.ufu.br/handle/123456789/34045; http://doi.org/10.14393/ufu.te.2022.57
• topic: Eletroencefalografia; Quantificação; Processamento digital de sinais; Prognóstico do coma; Classificação; Aprendizado de máquina; Electroencephalography; Quantification; Digital signal processing; Prognosis of the coma; Classification; Machine learning; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA::BIOENGENHARIA::PROCESSAMENTO DE SINAIS BIOLOGICOS; Engenharia elétrica; Eletroencefalografia; Coma - Prognóstico; Aprendizado do computador

The electroencephalographic signal is relatively simple in terms of acquisition. Widely used in Intensive Care Units, the neurological monitoring of clinically comatose patients through this sign has been found in the international literature as a helpful exam to the prognosis of patients in coma. The latter can infer two outcomes for the patient according to the level of consciousness measured: good or bad. The poor outcome can be death from clinical causes or even related to brain death. Being able to distinguish the outcome of the comatose patient using tools of the electroencephalographic signal at frequencies from 1 to 100 Hz from mathematical models is the main objective of this study. EEG records of comatose patients were considered in this research and divided into three possible outcomes. In this context, the methodology of this study addresses quantitative calculations of the EEG signal of these patients, both in the time domain and in the frequency domain, taking into account for the latter two distinct ranges of analysis: the first considering the clinical signal, from 1 to 30 Hz; the second considering the spectrum from 1 to 100 Hz. In addition to this quantitative information, data such as the patient's level of consciousness during the examination and the etiology of the coma were considered as input variables, combined, to generate prognostic models from machine learning tools. We used the data segmented into training and test data. Also, we used cross-validation with five folds, as well as the hold out tool executed five times. Then, we obtained several confusion matrices and the main performance measures were calculated: accuracy, sensitivity, and specificity. The macro F-score measure was also calculated as a differential since no study with this measure was found in the literature. The binary classifiers considered in the creation of the models were logistic regression, support vector machine, and k-nearest neighbors. The results found in this study demonstrate that the electroencephalographic signal measured in comatose patients has the potential to be used as an attribute for prognostic models of coma. The best mean accuracy value obtained here was equal to 0.80, differentiating patients with clinical death outcomes from patients with death due to brain death; and, regardless of the data set considered, the mean accuracy was 0.68. It was also observed that, in some situations, inserting analysis of gamma and super gamma rhythms improves the performance of prognostic models, particularly when considering the active and death classes due to brain death.