GMM and LDA Applied to Lung Diseases Detection

Abstract

This study presents experimentally tested methods, which can be used for a quantitative assessment of respiratory sounds as the indicators of pulmonary disorders. In particular, conducted experiments considered both normal and abnormal lung sounds (LS). As a part of the RALE Database, signals were recorded from healthy subjects and those with respiratory disorders. Current medical practices including evaluation of respiratory diseases often involve qualitative and frequently subjective auscultation. However, the application of quantitative signal analysis methods could improve the assessments of these diseases. In particular, we utilized acoustic evaluation methodologies based on the MFCC (Mel Frequency Cepstral Coefficients) acoustic vectors representation, GMM (Gaussian Mixed Models), and LDA (Linear Discriminant Analysis). To assure the validity of determined class models representing diagnostic classification, the LS signals were cross validated within sequential sets of respiratory cycles for a given subject as well as cross correlated within the specific groups of subjects representing particular conditions of normal or given class of abnormal pulmonary functions. Higher order MFCC vectors, including 9, 10 and 11 Gaussian mixtures, resulted in improved classification of the LS attributes, reached up to 98 % of efficiency recognition. This documented automated classification of LS makes it suitable for a more efficient mass screening of respiratory disorders. In particular, the presence of peculiar sounds such as crackles and wheezes lead to more robust models thus reflecting the useful applicability of the presented diagnostic tool. These techniques can assist in broader analysis, identification, and diagnosis of pulmonary disorders manifested by peculiar auscultatory findings.