Interpretable Respiratory Sound Analysis with Ensemble Knowledge Distillation

Chronic respiratory diseases are one of the leading causes of death in the world. Respiratory sounds are an important indicator to diagnose the most diseases related to respiratory system. Many recent works have focused on the analysis of adventitious sounds. Unfortunately, these approaches cannot analyze respiratory sounds in real time during auscultation and lack an easily trusted model by doctors. In this paper, we propose a novel respiratory sound analysis framework with interpretable ensemble knowledge distillation. In our work, multiple teacher models will be trained to learn lung sounds from different sources, and then they will apply the learned knowledge to guide the student model training through knowledge distillation to make our model more powerful in predicting accuracy and efficiency. Meanwhile, our model is interpretable and reliable, and its process of prediction will be approximated by the decision tree regularization. Experiments demonstrate the effectiveness of our method on the respiratory sound database.