Periodic monitoring of breath sounds is essential for early screening of several lung diseases caused due to the inflammation of the airways by viruses such as COVID-19. These tests require external medical equipment to be used or necessitate people to visit a hospital. During the current pandemic situation like COVID-19, it is difficult for a large number of people to undergo such tests. Fortunately, smartphones are ubiquitous and their microphone can be leveraged for recording breath sound data. We present a smartphone-based solution for monitoring breath sounds from the user via the in-built microphone together with and our AI-based anomaly detection engine, for preliminary screening for lung diseases.
The two major tasks involved in this project are breath sound detection followed by anomaly detection. For the breath sound detector module, multiple machine learning algorithms are used to detect whether the recorded sound is a breath sound. Subsequently, in the anomaly detection engine, various machine learning algorithms are employed by extracting various features from the audio signal. Considering the effectiveness of deep learning in classifying images, we have used the spectrogram image generated using Fast Fourier transform with Convolutional Neural Network (CNN), an Ensembled CNN, and Gated Convolutional Recurrent Neural Network (Gated CRNN). Data for this project is obtained from various sources including the RALE database, which has sounds of wheezes, crackles, etc. For the breath detector module, we have achieved an average accuracy of about 97% and for anomaly detection, 94% is obtained. We have developed an android application supported by a cloud-based implementation allowing the use of AI algorithms. This app can effectively be used as a remote screening tool for lung diseases by using breath sounds as a biomarker.