Deep learning with 3D-second order difference plot on respiratory sounds

Abstract

The second order difference plot (SODP) is a nonlinear signal analysis method that visualizes two consecutive data points for many types of biomedical signals. The proposed method is based on analysing quantization of 3D-space which is originated using three consecutive data points in signal. The obtained 3D-SODP space was segmented into 3-10 spaces using octants, spheres and cuboid polyhedrons of which centroids are at the origin. Lung sound is an indispensable tool for respiratory and cardiac diseases. The study is focused on classifying the lung sounds from at risk level and the interior level of chronic obstructive pulmonary disease (COPD). The COPD is one of the most deadliest and common respiratory diseases which come into existence as a consequence of smoking. The smokers for a few years are qualified as at risk level of COPD (COPD-0). The 12 channels of lung sounds from the Respiratory Database@TR were utilized in the analysis of the proposed 3D-SODP quantization method. The lung sounds are auscultated synchronously from posterior and anterior sides of subjects using two digital stethoscopes by a pulmonol ogist clinician in Antakya State Hospital, Turkey. Deep Belief Networks (DBN) algorithm was preferred in the classification stage. It has a greedy layer-wise pre-training which is based on restricted Boltzmann machines and optimizes the pre-trained weights using supervised iterations. The proposed DBN model had 2 hidden layers with 270 and 580 neurons, respectively. The conjunction usage of 3D-SODP quantization features with the DBN separated the lung sounds from different levels of COPD with high classification performance rates of 95.84%, 93.34% and 93.65% for accuracy, sensitivity and specificity, respectively. The results indicate that the 3D-SODP quantization on respiratory sounds has ability to diagnose the levels of the COPD using the deep learning model. Especially, the octant-based quantization is effective on lung sounds with high generalization capability using a small number of feature set dimension. (C) 2018 Elsevier Ltd. All rights reserved.