Towards the Prognosis of Patients in Coma using Echo State Networks for EEG Analysis

Abstract

The electroencephalogram (EEG) exam registers the brain's electrical signals using positioned electrodes over the scalp. The signals can be used in many medical applications, such as the prognosis and monitoring of diseases. However, EEG data also require efficient learning solutions in order to deal with their inherent noises and artifacts. In this paper, we propose simple, fast and accurate architectures of Echo State Networks (ESNs) capable of considering the sequential patterns of the exam. To be specific, we develop three ESNs based on a baseline, a single-reservoir and a multi-reservoir architectures, which are responsible for mapping temporal (or sequential) inputs into a high-dimensional space. To evaluate the ESN architectures, we addressed the problem of prognostic of patients in coma (PPC) from EEG records considering two formulations: binary and multi-class. Experiments on both scenarios considering the three ESN architectures as well as traditional techniques (without reservoir) were performed. The results revealed the potential of our single and multi-reservoir architectures over the baseline architecture and other techniques without the reservoir structure. Interestingly, the ESN single-reservoir architecture achieved the best predictive performance for both binary and multi-class formulations.