AI Learns to Detect Respiratory Diseases from Cough Sounds

A new study led by a team of scientists at Google has explored the potential of using cough sound recordings to monitor health and diagnose diseases through deep machine learning. The research focuses on processing audio signals such as breathing and various types of coughs.

The team developed a system called Health Acoustic Representations (HeAR), which was trained on an extensive dataset of 313 million two-second audio clips. This system uses self-supervised masked autoencoders and demonstrates impressive adaptability, being able to identify diseases like COVID-19 and even tuberculosis, as well as determine whether a person is a smoker.

HeAR has shown promising results, achieving accuracy scores of 0.645 and 0.710 in detecting COVID-19 depending on the dataset used, where a score of 0.5 indicates random guessing and 1 represents perfect accuracy. For tuberculosis detection, the model performed even better, reaching a score of 0.739.

The developers of HeAR utilized publicly available YouTube videos to extract over 300 million short audio clips, each of which was then converted into a visual representation of sound—a spectrogram. This allowed the model to learn to predict missing parts of the data, increasing its adaptability.

Currently, the researchers are actively working on a mobile app called AI4COVID-19, which has already shown promising results in accurately identifying coughs caused by the coronavirus. They are seeking funding to conduct clinical trials in order to obtain approval from the U.S. Food and Drug Administration (FDA), which would allow them to bring the app to market. At present, there are no FDA-approved tools for diagnosing diseases based on audio recordings.

This research opens up new possibilities for healthcare fields such as telemedicine and self-care, offering innovative approaches and expanding the boundaries of artificial intelligence applications.