Multi-channel k-NN classifications of respiratory sounds

Abstract

Our main interest in this study is to compare two different feature sets derived from respiratory sounds for optimum classification where multi-channel classification algorithm with each channel weighted equally is used. Two class recognition problem made of healthy and pathological sound data is addressed. The performance of our classifier is based on how well it differentiates between healthy and pathological sounds. For this purpose, parallel recording from 12 microphones placed on the posterior chest were used to extract two different group of sets of features for classification. Respiratory sounds of pathological and healthy subjects were analyzed via frequency spectrum and autoregressive (AR) model parameters. Since due to the physiology of the lungs, the transmission characteristics and therefore the spectral characteristics differ for respiratory sounds heard at different locations on the chest, separate reference libraries were built for each microphone location. Each subject is represented by 13 channels of respiratory sound data of a single or multiple respiration cycles depending on applied feature extraction methodology. Two reference libraries, pathological and healthy, were built based on multi-channel respiratory sound data for each channel and for each respiration phase, inspiration and expiration, separately. A multi-channel classification algorithm using k nearest neighbor (k-NN) classification method was designed. Performances of the two classifiers using quantile frequencies and AR model parameters as feature sets, are compared separately for inspiration and expiration phases.