Respiratory Disease Detection through Spectogram Analysis with Explainable Deep Learning

One of the leading causes of death is represented by respiratory disease. In this context, screening is really crucial in order to diagnose respiratory disorders early. In this paper, we propose an approach to automatically analyse respiratory audio recordings for disease detection, by converting respiratory audio recordings into spectrograms. We design a convolutional neural network aimed to classify respiratory audio recording into one of the following classes: Bronchiectasis, Bronchiolitis, Chronic obstructive pulmonary disease, Upper respiratory tract infection, Pneumonia or Healthy. Moreover, the proposed method provides a kind of explainability about the model decision by highlighting the areas on the spectrogram related to a certain respiratory diagnosis, to increase confidence and trustworthiness from doctors and patients: in this way, we aim to boost the development of machine learning techniques in the real-world medical context. We present an experimental analysis including 2751 different annotated respiratory audio recordings of length varying from 10s to 90s related to 126 different patients obtaining an accuracy equal to 0.849, thus showing the effectiveness of the proposed method for explainable respiratory disease detection.