Aplicação de redes neurais artificiais na identificação de insuficiencia cardíaca utilizando análise sonora da voz

• Autor(a) principal: Firmino, João Vitor Lira de Carvalho
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFPB
• Texto Completo: https://repositorio.ufpb.br/jspui/handle/123456789/24500
• topic: Insuficiência cardíaca; Diagnóstico; Vozes; Análise de sinais; Redes neurais artificiais; Heart failure; Diagnosis; Voices; Signal analysis; Artificial neural networks; CNPQ::ENGENHARIAS::ENGENHARIA MECANICA

Heart failure is a disease that disables the heart from properly pumping blood to nourish the entire body. Currently, the main diagnostic methods for this pathology are performed clinically through the measurement of B-type natriuretic peptide (BNP). As cardiovascular diseases are the main causes of premature death, the development of new technologies to identify these diseases is of great importance. Thus, this research presents the development of an identification system for the vocal distortions caused by heart failure in an individual. For the development of the software, the voices of 142 individuals were collected, separated by sex and age. Among these 142, 84 voices of people already diagnosed with heart failure were collected at the Heart Institute of Sao Paulo University (INCOR – USP) and at the Metropolitan Hospital of Joao Pessoa. On the other hand, the voices of the other 58 healthy individuals were collected in an extra-hospital environment. Furthermore, the device used for recording the voices was the PX440 digital audio recorder, produced by Sony. To analyze the collected data, the following techniques were applied to extract and select the characteristics of the signals: statistical analysis, fast Fourier transform, discrete wavelet transform and mel-cepstral analysis. By using these techniques, it was selected features to feed the artificial neural networks (ANNs) developed for each sex. Using the established architectures of the networks, an overall efficiency of 96.7% was achieved for both ANNs. In order to guarantee the usability of the created system, a computer application was developed. Using this software, values of 91.86%; 88.1% and 92.1% were obtained in the calculation of accuracy, sensitivity and specificity, respectively. Therefore, the heart failure identification system showed promising results that need to be further studied in order to improve the developed tool.