Seleção de atributos e aprendizagem de máquina na detecção e classificação de patologias da laringe

• Ano de defesa: 2025
• Autor(a) principal: Pignelli, Rogerio
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Cornelio Procopio; Brasil; Programa de Pós-Graduação em Engenharia Elétrica; UTFPR
• Palavras-chave em Português: Patologia; Aprendizado do computador; Fala; Pathology; Machine learning; Speech; CNPQ::CIENCIAS SOCIAIS APLICADAS; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/38020

Voice disorders affect millions of people worldwide, compromising communication abilities and often indicating underlying pathologies. The rapid diagnosis of these disorders is crucial for improving patient quality of life. Traditional methods based on visual inspection via laryngoscopy and subjective evaluations by healthcare professionals require significant resources and time. Consequently, speech processing and machine learning techniques have been applied to assist in the pre-diagnosis of laryngeal disorders with the intent of reducing the amount of invasive procedures when not strictly required. In this work, we investigate three distinct approaches. Initially, we employ feature extraction through Wavelet-Packet decomposition for the identification of Reinke’s edema, comparing the performance of SVM and MLP algorithms. Next, optimization algorithms are applied to determine the ideal number of layers and neurons per layer in the MLP, aiming to enhance the identification of Reinke’s edema. Finally, a methodology for interpretability in voice disorder classification is proposed through the careful selection of features, with the objective of identifying the most significant contributors for discriminating among multiple pathologies. Wavelet-packet features with SVM delivered 100% accuracy in the identification of Reinke’s edema with a similar performance achieved by a PSO-optimized MLP. By applying feature reduction and explainability methods, we found 16 features that resulted in F1-score of 0.62 in the most demanding cross-database analysis. Shimmer and the second coefficient of MFCC were the most relevant features for the detection of voice disorders, aligning with clinical threshold and reinforcing the relevance of the selected features.