Attention-Based CRNN Models for Identification of Respiratory Diseases from Lung Sounds

Respiratory diseases are a major global health concern, with millions of people suffering from disorders such as asthma, bronchitis, chronic obstructive pulmonary disease (COPD), and pneumonia. In recent years, machine learning and other forms of Artificial Intelligence have proven to be useful resources for resolving issues in the medical field. In this study, we examine the diagnostic utility of Convolutional Recurrent Neural Network (CRNN) models for identifying respiratory diseases using digitally recorded lung sounds. We developed two deep learning models to diagnose and classify lung diseases: a binary classification to classify COPD and non-COPD, and a multi-class classification model to classify five lung disorders (COPD, URTI-upper respiratory tract infection, Pneumonia, Bronchiectasis and Bronchiolitis) and healthy conditions. The ICBHI 2017 challenge dataset [1] was used to develop the models. The accuracy of the binary and multiclass classification models was 98.6% and 97.6%, respectively, with ICBHI Scores of 0.9866 and 0.9723.