An Automated System towards Diagnosis of Pneumonia using Pulmonary Auscultations

Abstract

Respiratory sounds carry significant information about the condition of respiratory system. Respiratory sounds are often affected by sounds emanating from heart and other organs thus making the analysis task more complex. Pneumonia is a very common lungs disease and requires efficient diagnosis at initial stage for proper treatment. In this research, an automated system for diagnosis of Pneumonia based on auscultations is proposed. Auscultation signals are first preprocessed through Empirical mode decomposition (EMD), which decomposes original signal into its constituent components known as intrinsic mode functions (IMFs). Preprocessed signal is reconstructed by addition of only those IMFs which carry high discriminative information among healthy and Pneumonia subjects. IMFs which carry redundant and noisy data are rejected thus making preprocessing more effective. Next, characteristic features are extracted by fusion of Mel frequency cepstral coefficients (MFCC) and time domain features. Finally, Support Vector Machines (SVM) classifier is trained and tested through 5-fold cross validation. Experimental evaluation of proposed approach is performed on range of various classifiers on self-collected dataset which contains 480 auscultation signals of normal and Pneumonia subjects. SVM with Quadratic kernel achieved best classification results in terms of accuracy of 99.7%.