Smart Diagnostic System for Classification of COPD and Healthy Subjects from Lung Sound Analysis

World Health Organization statistics depicts pulmonary abnormalities as one of the fatal disease classes. Chronic Obstructive Pulmonary Disease (COPD) is a prominent death cause in both adults and infants. In 2019, it was ranked at the 3rd position which contributed around 3.5 million fatalities to the worldwide mortality rate. The main objective of the research work is to devise non-invasive techniques to classify COPD and healthy subjects from Lung sound analysis by keeping in view about the WHO’s Global Alliance against CRDs (GARD) vision "a world in which all people breathe freely". The scheme is designed through Digital Signal processing and machine learning methods. The lung sound (LS) of healthy and COPD subjects are used from publicly available ICBHI’s LS database. Lung sound data of 35 Normal and 631 COPD subjects are used in the research. To denoise the acquired LS signal, the wavelet decomposition technique is applied at the preprocessing stage for denoising. Daubechies5 is used as the mother wavelet. The capability of Mel Frequency Cepstral Coefficients (MFCC) and various time/spectral domain features of acquired signals are investigated in the classification of the normal and COPD subjects. Imbalanced classes are balanced through the ADASYN technique. Various classification models have been experimented after performing the feature engineering, but K-nearest neighbors (KNN) classification method has outperformed with the accuracy of 97.20% on Mel Frequency Cepstral Coefficients as key features. 10-Folds Cross Validation (CV) technique has validated the system accuracy of the proposed model. The simulation was performed on MATLAB 2019a software.