Breath sound classification by using the smart phone

Abstract

Respiratory sounds are non-expensive, non-invasive, and give more information, so respiratory sound analysis is important for clinical testing. However, the accuracy of respiratory sound analysis depends on the clinician's expertise. Many studies try to develop an automation system for the classification of breath sounds. The system is the cooperation of sound processing, image processing, and neural networks. However, the systems are based on computers and the computer based systems are not easy to use in the remote area. Thus, this study proposed to develop the breath sound classify that easy to use in the remote area. Recently, the smart phone has become more powerful and more flexible than the PC, and there is a possibility of developing the breath sound classification on the smart phone. This study proposes to develop a smart phone-based respiratory sound classification. The advantage of the smart phone base system is that it is more flexible and patients can easily use it. In this study, the Android phone cooperates with the TarsosDSP sound library and Tensorflow lite. Some samples of breath sounds from the ICHBI database and an online learning website for respiratory sounds were used. The samples included normal breath sounds (136 samples), crackling sounds (111 samples) and wheeze sounds (111 samples). The experimental method, the samples of breath sounds were played with audio player software on a computer, and the electronic stethoscope was used to record the sounds. Then the breath sound classification software was used for filtering noise, recording sound, computing the spectrogram, and processing the neural network. The result found the smart phone's base respiratory sound classification system can diagnose breath sound. The accuracy for normal breath sounds was 80%, crackle sounds were 87%, and wheeze sounds were 85%. Finally, the characteristics of the breath sound spectrogram were discussed.