Classificação de anomalias em sons respiratórios utilizando processamento digital de sinais de áudio e redes neurais artificiais

Abstract

In this research, a machine learning model was developed to classify lung sounds, with the purpose of identifying the presence of continuous (wheezes) and discontinuous (crackles) abnormal sounds. Digital signal processing techniques were used to process audio signals and extract its features (MFCCs and Mel scale spectrograms) and a convolutional neural network was used to perform the classification in two scenarios. The data used were obtained from a dataset distributed free of charge on the internet. The first scenario is concerned with classifying the sounds present in the respiratory cycles, and the best result obtained was the detection of crackles, with 82% accuracy. The second scenario consists in classifying the patients' comorbidities, where 94% of accuracy and F1-score (macro average) were obtained. The results show that it is possible to perform the classification with precision only for some classes, which indicates that it is necessary to refine the method or data used.