An MFCC Features-driven subject-independent Convolution Neural Network for Detection of Chronic and Non-chronic Pulmonary Diseases

Chest auscultation, recording lung sound, is an essential procedure to diagnose abnormalities in the respiratory system. Classification of chronic and non-chronic pulmonary diseases may serve as a potential screening tool for clinicians which in turn can save time and resources for both the patients and the clinicians. Therefore, an automated analysis of chest auscultation for detection of chronic and non-chronic diseases is worth to attempt as a research problem. In this paper, we propose a convolutional neural network (CNN) architecture to classify lung diseases using Mel frequency cepstral coefficients (MFCC) of lung sounds. The MFCCs are derived by applying cosine transforms on Mel spectrograms. Finally, before testing the robustness of the model for subject independency using a publicly available dataset from ICBHI 2017, data augmentation was used to balance the data for classification. A training accuracy of 97.58% ± 2.34% and testing accuracy of 86.25% ± 2.77% were achieved from a cohort of healthy individuals (n = 26), chronic patients (n = 72) and non-chronic patients (n = 28) . The results showed that lung sounds processed with MFCC and CNN can be a potential screening tool in pulmonary disease diagnosis for low-resource settings.