Classification of Respiratory Sounds Based on Hybrid Convolutional Recurrent Neural Network

Nearly 8 million people suffer and die from respiratory diseases every year. Therefore, to reduce the number of deaths which are caused by the diseases, early detection and early treatment of respiratory diseases are required as global issues, and several techniques have been proposed until now. Currently, the ICBHI (International Conference on Biomedical and Health Informatics) 2017 Challenge Dataset has been released for research on respiratory sound analysis, and respiratory sound classification methods using this dataset have been proposed worldwide. The authors also proposed a respiratory sound classification method using an improved CRNN (Convolutional Recurrent Neural Network), which is a combination of CNN and RNN (Recurrent Neural Network) with some modifications. However, it was still difficult to classify by image features alone due to noise such as voice in the respiratory sound data. To overcome this problem, we try to classify breath sounds automatically using a deep learning model that considers the features of the raw breath sound data. To extract the sound features of the raw respiratory sound data, we use a 1D-CRNN, which is a 1D-CNN reconstruction of an improved CRNN proposed in a previous study of ours. Then, it is combined with the deep features obtained by our previous improved CRNN (2D-CRNN) for the final classification. The proposed method achieves AUC (Area Under Curve) of 0.92, sensitivity of 0.75, specificity of 0.86, and ICBHI score of 0.80 based on the ROC (Receiver Operating Characteristic) analysis, respectively, which are the highest values compared to the other methods under the same experimental conditions.