Aprendizado de padrões EEG para prognóstico precoce de pacientes em coma usando redes echo state e redes neurais convolucionais

Abstract

The electroencephalogram (EEG) exam registers the brain's electrical signals using positioned electrodes over the scalp. The signals can be used in many medicine applications, such as the prognosis of diseases. In this research we investigate recurrent neural networks and deep learning architectures for the prognosis of comatose patients (PPC) from their EEG data, which were collected in a Brazilian public hospital. We considered two formulations for such a problem. The first aims to divide the target class into two outcomes: favorable and unfavorable. The second also considers a third outcome related to brain death, which have great potential to contribute to organ donation procedures. Unlike most common approaches in the literature, which use quantifiers to extract features of the EEG signals, in this dissertation we develop learning models based on echo state networks (ESN) capable of considering the sequential patterns of the exam by processing architectures mono and multi-sequential of EEG signals. In addition, a spatial representation model that takes advantage of the position of the electrodes during the exam while maintaining temporal resolution was developed and later used in our deep learning architectures based on convolutional neural networks (CNN). Several experiments were carried out to evaluate the developed architectures. From the viewpoint of sequential pattern analysis, the results showed that our deep ESN-based approach has the best average predictive performance for two and three outcomes when compared to other developed ESN architectures. Regarding the analysis of spatial representation, the results demonstrated considerable progress as our CNNs were able to outperform state-of-the-art approaches for both two and three outcomes. Moreover, the present work extends the literature related to PPC and opens new perspectives for supporting the medical team in this task.