Redes neurais convolucionais baseadas em filtros espaciais para prognóstico de pacientes em coma via sinais EEG

Abstract

In recent years, there have been significant advances in the detection and recovery of patients suffering from varying degrees of loss of consciousness. However, obtaining a reliable prognosis remains a challenge. The development of new architectures, as well as the use of existing machine learning architectures, applied to EEG exams, can benefit the task of obtaining more reliable prognoses for coma patients. Therefore, the objective of this work was to develop and evaluate deep learning models with the aim of generating reliable prognoses for coma patients using EEG data. 121 EEG tracing samples were used, which consisted of 42 samples with a favorable outcome and 79 with an unfavorable outcome. Within each record, the neurologists segmented 10 sections considered significant for preparing the patient’s prognosis. These excerpts, lasting 2 seconds each, have different temporal resolutions, with the sampling frequency of the exams varying between 100Hz and 600Hz. Two approaches to data processing were investigated. In the first, 9 out of 10 excerpts previously separated by experts were used. In the second, 9 sections were randomly selected directly from the original route, without the possibility of overlapping. In both cases, the samples went through a resampling process, standardizing the sampling frequency of the data set at 100Hz and 200Hz. The architectures chosen for the development of the models were Deep ConvNet, Shallow ConvNet and EEGNet, all mainly supported by convolutional layers for learning temporal and spatial patterns relevant to the problem in question. study. Simulations were conducted considering crossvalidation with 10 folders, and in each iteration the training set was subjected to a data augmentation algorithm performed by an implementation of the FTSurrogate technique. The evaluation metrics used were accuracy, specificity, sensitivity and F1-Score. The three architectures performed well, especially when subjected to classification of EEG excerpts selected by experts. In this context, the Shallow ConvNet architecture stood out in experiments that considered the sampling frequency of exams at 100Hz and 200Hz, while Deep ConvNet stood out when the frequency was standardized at 200Hz. The models developed and evaluated in this work showed good performance, with the potential to contribute significantly to clinical practice, as effective and reliable auxiliary tools in the prognosis of patients in a state of coma.