Análise de grafos em sinais de EEG para identificação do transtorno depressivo maior

Abstract

This work introduces a methodology based on graph theory to represent signals from patients with Major Depressive Disorder (MDD) for classifying samples as healthy or diseased. The approach revolves around representing the signal produced by the patient in a

directed graph, also known as a digraph. Spectral markers derived from the signal model show the different magnitude coefficients and transition levels. The digraph representation involves three matrices (Adjacency, Laplacian, and Directed Combinatorial Laplacian). This structure encourages convolutional neural networks and techniques such as transfer learning, as the matrices take the form of tensors. The methodology uses actual patient samples with electroencephalogram (EEG) signals and shows promising results in extracting relevant features for sample classification. The experiments reveal that the proposed spectral tensors achieve performance exceeding 91% in all classification tasks analyzed for Dataset I and over 95% for Dataset II. Additionally, integrating the spectral tensors with the ResNet architecture proves highly effective, highlighting the model’s ability to adapt well to this approach.