Predição não invasiva de nível de oxigenação sanguíneo a partir de áudios usando redes neurais artificiais

Abstract

Respiratory insufficiency is a symptom present in several respiratory diseases, which can cause anything from discomfort to death. In order to speed up the screening process or assist health professionals, research in the health field involving technology is being encouraged. One type of technology that is widely used is artificial intelligence (AI), mainly artificial neural networks (ANN). One of the focuses of research in the area of AI is machine learning through audio, which has already proven possible to identify diseases with its help, one of these diseases being COVID-19, that it was possible to detect it through speech through the SPIRA project (CASANOVA et al., 2021b). By taking advantage of the database collected by the SPIRA project, this work aims to create an ANN model that, through regression, can identify in this case not a disease, but a symptom present in several diseases, for this the model must be capable to predict the level of blood oxygenation present in the body, which is one of the main indicators of respiratory insufficiency, with the ultimate objective of assisting the health professional in their final diagnosis decision. For this, several SpiraNet models from Casanova et al. (2021b) were trained using mel spectrogram and one using the MFCC representation, each test carried out had the objective of adding a new hyperparameter in order to arrive at the ideal model. The results demonstrate that, unlike performing a symptom classification, performing blood oxygenation regression is more complex, obtaining 26% in the results with the Pearson coefficient and 6% in the results with the coefficient of determination, which are preliminary results that indicate that there is room for progress in this task.