Using Frequency Correction of Stethoscope Recordings to Improve Classification of Respiratory Sounds

Recent advancements in artificial intelligence have brought some spectacular innovations, including systems for automatic analysis of respiratory sounds. These solutions, however, usually have a very limited set of supported signal sources and it raises question if these can be adapted to work with signals that they have not been trained on. This work explores the possibility to automatically discover the characteristics of unknown source of body signals and convert them to match that of a known source. Next, it answers a question of how well the AI can perform given signals processed in this way. Our proposed method is used to adapt the ICBHI 2017 Respiratory Sounds Challenge recordings to match the expected input signal characteristic of commercially available AI solution for lung sounds analysis. The method achieves an ICBHI score of 60.99% which is an improvement of 2.7% points over the current challenge leader. The presented results show that existing AI solutions can solve congenial tasks even when fed with signals from unknown sources and even achieve superior performance to dedicated solutions. In this case the advantage of using larger and more precisely described training data during AI development outweighs potential imperfections of the audio data transformation leading to higher accuracy. In fact, the true performance of the proposed solution might be even higher because some ICBHI recordings were found to be mislabeled, which could also explain limited success of models developed based on this data.