Multivariate modeling and diagnostic classification of pulmonary sounds

Computerized pulmonary sounds analysis has become prevalent in the recent decades as it provides means for quantitative and objective evaluation, contrary to the limited and subjective nature of stethoscope auscultation. Multi-variate methods disclose the inherent spatial information in the multi-channel measurements and enable to explore the system characteristics altered due to pathological conditions that have developed in the lungs. In this study, a multi-variate mathematical model, namely, vector auto-regressive (VAR) model, has been considered, and the optimal VAR model to represent the pulmonary sounds data is pursued through new goodness of fit criteria proposed specifically for the data and the application. The estimated model parameters are employed in classification using the $k$-nearest neighbor ($k$-NN), support vector machine (SVM), and Gaussian mixture model (GMM) classifiers, with an eventual aim to develop a diagnostic classifier for clinical setups. Various classifier schemes are experimented with different data sets in the quest for the most useful classifier design. The healthy and the pathological groups are discriminated successfully. In the classification of conditions including the healthy group and the obstructive and restrictive types of pathologies, a hierarchical framework is suggested. Generally, the healthy and the restrictive groups are discriminated more successfully than the obstructive group. GMM is generally the most competent classifier among all, however, SVM is also successful for certain feature arrangements. The improvement of the diagnostic classifier so as to make it appropriate for clinical setups is still open for exploration, especially with additional features to enhance the distinctive characteristics further as to prevent the confusion of the obstructive diseases.