Processamento e análise de sinais digitais vozeados para o pré-diagnóstico de patologias laríngeas
Carregando...
Data
2022-01-12
Autores
Satake Junior, Edson Haruyuki
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Deficiências vocais continuam a afetar parcelas significativas da população mundial, no entanto os processos clínicos tradicionais são comumente invasivos e submetem pacientes a possíveis traumas. Este trabalho, desenvolve um método computacional, utilizando análises acústicas e técnicas de processamento de sinais, que permite detectar a presença de patologias laríngeas por meio de sinais digitais de voz, e os classificar em saudáveis ou patológicos. São utilizadas para a discriminação dos sinais as características de fator de perturbação direcional (DPF), perturbação média relativa (RAP) e fator de jitter (JF), enquanto uma máquina de vetores de suporte (SVM) e um algoritmo K-vizinhos mais próximos (KNN) são utilizados como classificadores. Foram utilizados 136 sinais de voz, cuja quantidade de sinais saudáveis e patológicos são iguais, e estes correspondem a casos de Edema de Reinke. Por fim, os testes mostraram uma acurácia global média de até 67% e máxima de 85%, para a SVM, e média de 74% e máxima de 88%, para a KNN. Enquanto a acurácia média de detecção de patologias alcançou 70% e máxima de 82%, para a SVM, e média de 65% e máxima de 88%, para a KNN.
Voice disorder continue to affect significant portions of global population, however the traditional clinical processes are commonly invasive and submit patients to potential trauma. This work, develop a computational method, using acoustic analysis and signal processing techniques, which allow to detect the presence of laryngeal pathologies through digital voice signals, and classify them into healthy or pathological. The characteristics of directional perturbation factor (DPF), relative average perturbation (RAP) and jitter factor (JF) are used for signal discrimination, while a support vector machine (SVM) and a K-Nearest Neighbors algorithm (KNN) are used as classifiers. 136 voice signals were used, whose quantity of healthy and pathological signals are the same, and these correspond to cases of Reinke’s Edema. Lastly, the tests showed an average global accuracy of 67% and maximum of 85%, for SVM, and average of 74% and maximum of 88%, for KNN. While the average accuracy of pathologies detection reached 70% and a maximum of 82% for SVM, and average of 65% and maximum of 88% for KNN.
Voice disorder continue to affect significant portions of global population, however the traditional clinical processes are commonly invasive and submit patients to potential trauma. This work, develop a computational method, using acoustic analysis and signal processing techniques, which allow to detect the presence of laryngeal pathologies through digital voice signals, and classify them into healthy or pathological. The characteristics of directional perturbation factor (DPF), relative average perturbation (RAP) and jitter factor (JF) are used for signal discrimination, while a support vector machine (SVM) and a K-Nearest Neighbors algorithm (KNN) are used as classifiers. 136 voice signals were used, whose quantity of healthy and pathological signals are the same, and these correspond to cases of Reinke’s Edema. Lastly, the tests showed an average global accuracy of 67% and maximum of 85%, for SVM, and average of 74% and maximum of 88%, for KNN. While the average accuracy of pathologies detection reached 70% and a maximum of 82% for SVM, and average of 65% and maximum of 88% for KNN.
Descrição
Palavras-chave
Processamento de sinais, Detecção de patologias, Patologias laríngeas, Deficiência vocal, Aprendizado de máquina, Máquina de vetores de suporte, K-vizinhos mais próximos, Cepstro, Signal processing, Pathology detection, Laryngeal pathologies, Voice disorder, Machine learning, Support vector machine, K-nearest neighbors, Cepstrum