Classification of respiratory pathology from pulmonary acoustic signals based on respiratory cycle segmentation and two-stage classification

Abstract

Auscultation is the process of listening to the internal sounds of the body using a stethoscope. This process provides vital information on the present state of the internal organs, such as the heart, lungs and the gastrointestinal system. Auscultation is subjective and prone to be not reliable. However computerized respiratory sound analysis is more effective and reliable. This thesis discusses the development of a computerized decision support system (CDSS) to detect respiratory pathology using pulmonary acoustic signals. The pulmonary acoustics signals were collected from 72 subjects to develop the CDSS. In order to develop the CDSS tool, three different methodological frameworks were proposed to determine the most effective classification of respiratory pathology. The recorded pulmonary acoustics signals were filtered to remove noise and other artifacts followed by respiratory cycle segmentation. In this work, the respiratory cycle segmentation is performed by using Fuzzy Inference system. Parametric (Mel-frequency cepstral coefficients (MFCC) and Auto-regressive model (AR)) and Nonparametric (wavelet packet transform (WPT) and Stockwell transform (ST)) representations of features were extracted. The features extracted were dimensionally reduced using principal component analysis and a statistical analysis was performed to determine the significance level of the feature vector using One-way ANOVA. Observations showed that the extracted features were statistically significant with p < 0.05. In the classification stage various nonlinear classifiers such as k-nearest neighbor (k-nn), support vector machines (SVM) and extreme learning machine (ELM) were implemented to classify the respiratory pathology from respiratory sounds. In the classification, extreme learning machine performed better than k-nn and support vector machine classifier for all the frameworks. Experimental results showed that ST based feature extraction performed well with ELM classifier with third framework. The ST based features and ELM classifier with third framework was validated using a new set of data comprising of 48 subjects and the system was found to be reliable with mean classification accuracy of 96.63%, 97.57% and 98.48% for classifying (normal, continuous lung sounds and discontinuous lung sounds), (wheeze and rhonchi) and (fine crackles and coarse crackles) respectively. After successful validation a CDSS tool was developed using the ST based features and ELM classifier with third framework