Classificação de alto nível baseada em meta-grafos de grafos de visibilidade para detecção do Autismo

Abstract

This study explores advanced classification techniques for the detection of Autism Spectrum Disorder (ASD) using ATR-FTIR spectroscopy applied to saliva samples. The proposed platform is fast, sustainable, and enables non-invasive sample collection. The resulting ATR-FTIR spectrum is a high-dimensional sequence that encodes molecular interactions, yet its sequential nature is often overlooked in related studies. While traditional ATR-FTIR approaches rely on low-level classifiers based on physical attributes, more recent high-level methods incorporate topological and structural features. However, these high-level methods face challenges in the graph construction phase, particularly in accurately capturing the sequential relationships within the data. To address this limitation, we propose VisG2, a meta-graph construction method that represents spectra as visibility graphs and computes their similarities to build a high-level structure. Experiments with real ASD data demonstrate that VisG2 outperforms both traditional and modern classifiers such as SVM, CNN and Transformers, as well as other commonly used graph construction techniques. Furthermore, when combined with low-level methods, VisG2 forms a hybrid classifier that further enhances the performance of the original high-level approach. The results confirm VisG2’s effectiveness in capturing sequential relationships and highlight its potential for ASD detection and other applications involving sequential data.