Classification of respiratory sounds using improved convolutional recurrent neural network

Abstract

Currently, auscultation using a stethoscope is performed for the diagnosis of respiratory diseases. Auscultation is a simple and non-invasive diagnostic method; however, the diagnostic results depend on the experience of the doctor, thereby rendering quantitative diagnosis difficult. Therefore, we herein propose a new automatic classification method based on deep learning algorithms for respiratory sounds to support the diagnosis of respiratory diseases. The proposed method comprises two stages. First, a spectrogram is generated by applying a short-time Fourier transform to the respiratory sound data. Subsequently, the obtained spectrogram is classified into normal and abnormal (three classes: crackle, wheeze, and both) respiratory sounds using an improved convolutional recurrent neural network. By classifying the respiratory sounds using the proposed method, the following results are obtained: sensitivity, 0.63; specificity,0.83; average score, 0.73; harmonic score, 0.72. Furthermore, the proposed method yields better accuracy compared with other methods.