DeepRespNet: A deep neural network for classification of respiratory sounds

Abstract

Respiratory sounds convey significant information about the pulmonary status. This study proposes a deep learning-based framework to create an automatic, non-invasive, diagnostic method of categorizing pulmonary sounds. A labelled database of pulmonary sounds has been collected using an electronic stethoscope and audio recording instrument. Two deep learning architectures, 1D DeepRespNet and 2D DeepRespNet are proposed in this work that were trained and evaluated with normalised 1-D time series and 2-D spectrograms of acoustic signals of six types of lung sounds, respectively. The models were highly optimized to yield superior performance on the considered dataset. Experimental results demonstrate that the 2D DeepRespNet model trained with spectrogram-based representations yields higher accuracy of 95.2% on the test data as compared to the 1D DeepRespNet trained on the time-series data. The proposed model may be deployed on a single board computer or integrated into a smartphone to develop a standalone diagnostic tool to accurately and objectively classify abnormal lung sounds with low time complexity.