Computerized analysis of pulmonary sounds using uniform manifold projection

Respiratory sounds heard through the stethoscope lend useful information for diagnostic purposes. However, the accuracy and efficiency of clinical diagnosis are constrained by subjective interpretations and the level of auscultation expertise of the physician. In this work, Fourier-based signal processing approaches and uniform manifold projection (UMAP) for data visualization, are combined for tackling the computerized analysis of respiratory sounds. The results show that differences and similarities between the signals can be identified trough patterns emerging from the information embedded in stationary and dynamical spectral features from the pulmonary acoustic recordings. The outcomes of these different perspectives evince discriminative ability between the recording devices, the chronic and non-chronic condition of patients, and adventitious sounds. In a contrasting view to the analysis based on machine learning or deep learning, whose implementation is profusely reported in scientific literature, this paper shows that the proposed approach renders an effective analysis of pulmonary sounds, with clinical correlations. Furthermore, the information visualization achieved by means of UMAP can be an attractive tool for assisting physicians during patients auscultation.