CNN based Categorization of respiratory sounds using spectral descriptors

Chronic Obstructive Pulmonary Disease (COPD) is a dreadful disease which is a wide umbrella comprises of emphysema, bronchitis etc. It threatens the life of almost nearly 3 million people all over the world. The diagnosis of COPD can be detected in a better manner based on the lung sound analysis with the help of deep learning models such as convolutional neural network (CNN). In this work, the presence of COPD with different class of the sound like normal breathe sounds and abnormal breathe sounds such as wheeze, crackle and rhonchi are classified by using multi-class classifier. Spectral descriptor features from linear spectrum and MFCC from Mel spectrum are extracted. For experimentation and classification, a total of 596 lung sound signals are considered in this work. The classifier such as K-NN and decision tree are used to obtain an improved accuracy compared to binary machine learning classifier. The results indicates than an overall accuracy of 96.7% is obtained with multi-class classifiers using deep learning CNN model. The multi-class classifier results are also compared with SVM classifier.