Attentional Convolutional Neural Network for Automating Pathological Lung Auscultations Using Respiratory Sounds.

Respiratory diseases are one of the most prevailing diseases which can be limited by ensuring continuous monitoring and treatment through automating lung auscultations. In this paper, an attentional deep learning model has been proposed with a new data augmentation technique named as homogeneous padding to predict the respiratory diseases using the audio data of the ICBHI-2017 dataset. Since the dataset is imbalanced over the classes as well as the devices used to collect the data, we have performed ablation studies over the loss functions, and reported the appropriate loss function which performs better for this dataset. We have also performed experimental analysis to report the appropriate position of attention, and observed that the attention mechanism on high-level features has worked better in comparison with the low-level features. We have found that our developed model has outperformed most of the recent convolutional neural network (CNN)-based models, and the inclusion of an attention mechanism has contributed significantly to improve the accuracy of the model.