Descritor de características em componentes espectrais baseado em dígrafo para detecção de doenças neurodegerativas

Abstract

This research presents a method for describing biomedical signals based on constructing a directed graph, which also suggests processing the signal's spectral coefficients. This processing is based on the transitions of the coefficients identified in terms of the different levels of magnitude using the spectral marker chosen. In this context, four different spectral markers are proposed to analyze the frequency components and are based on different local points in the signal. The methodology has a simple and adaptive implementation. It can represent one or more signals from the same patient in the same structure, making it suitable for single-channel and multichannel applications. In the context of neurodegenerative diseases, three applications were chosen from among the most common pathologies. At first, the method extracts characteristics from patients' voice signals to classify healthy patients in relation to those with Huntington's disease; in this application, a dataset with voice records was used. In addition to the statistical test based on analysis of variance, measures of the correctness of the confusion matrix were also considered. The second application evaluates the method in a multichannel scenario, particularly for electroencephalogram signals for classifying healthy patients about those with Alzheimer's disease. Five different datasets were used, the first two for binary classification (healthy and sick). At the same time, the last three had samples with other neurodegenerative diseases and were considered for the multiclass scenario. Finally, the third application investigates the method's performance in representing patients with Parkinson's disease. In this scenario, two datasets were used that had already been evaluated in previous studies with healthy and sick patients. In all cases, as a result of signal processing, the methodology produces a three-dimensional tensor made up of three different perspectives of the digraph. The spectral tensor combined with deep neural networks obtained equal or better results when compared to expert approaches in each application, with hit rates above 95% in all eight datasets considered. This shows that the directed graph can capture relevant information, and the tensor structure makes a promising combination with deep neural networks feasible and offers an alternative to extracting attributes associated with neurodegenerative diseases.