Research Context:

Computational lung sound analysis (CLSA) is an important tool for inexpensive screening of patients supporting medical diagnosis. Most methods, however, suffer from insufficient accuracy, mainly because of the variability and lack of lung sound data. Approaching these challenges by using our recently developed high-quality multi-channel recording device and exploiting recent advances in machine learning is thus of utmost importance for clinical use.

Research questions and objectives:

Our main objective is to use deep learning methods for computational lung sound analysis to support medical diagnosis and long-term monitoring of lung damage. We aim to answer the following research questions:

• Is it possible to classify lung diseases at an early stage with high sensitivity and specificity including uncertainty measures?
• Are lung sound recording positions in a multi-channel recording setting important for specific diseases such as lung fibrosis?
• What is the best deep learning architecture for lung sound classification?
• How can we exploit transfer and continual learning to compensate for small sample size and to learn without catastrophic forgetting.

Methods:

Our research focus is two-fold: (i) We aim to record a high-quality lung sound corpus for diseases/conditions such as congestive heart failure, idiopathic pulmonary fibrosis, pneumonia, bronchitis and pleuritis using our multi-channel recording device in an observational clinical trial. We plan to make parts of the recorded anonymized database publicly available. This makes the classification results reproducible for other research groups. (ii) We will exploit recent advances of deep learning to improve automatic lung sound analysis and monitoring. In particular, we will perform neural architecture search for automatic discovery of optimal model architecture and transfer and continual learning to continually accumulate knowledge for different classification tasks and to utilize existing lung sound data more efficiently. Moreover, Bayesian deep learning will be used to provide uncertainty estimates at low costs by using randomized MAP sampling of the models. Semi-supervised learning techniques will be exploited to reduce the effort for data labeling.

Level of originality:

The fundamental innovations are the exploitation of neural architecture search, continual and transfer learning and efficient uncertainty estimation to improve machine learning models for CLSA.