In a new study presented by a team of scientists led by Google, the potential use of sound recordings of coughs for monitoring health status and diagnosing diseases through deep machine learning was explored.
The study focuses on processing sound signals like breathing and various types of coughs. The researchers developed a system called Health Acoustic Represents (Hear) that was trained on a massive dataset of 313 million two-second audio clips.
Using self-learning masking auto-encoders, the Hear system displayed remarkable adaptability and was able to identify diseases such as Covid-19 and tuberculosis, as well as determine if an individual is a smoker.
Hear displayed promising results, achieving accuracy rates of 0.645 to 0.710 in the detection of Covid-19, depending on the dataset used. An accuracy of 0.5 indicates a random guess, while 1 signifies ideal accuracy. The model performed even better in identifying tuberculosis, with a result of 0.739.
The developers utilized publicly available YouTube videos to extract over 300 million short sound clips, each of which was converted into a visual representation of sound known as a spectrogram. This process enabled the model to predict missing data, enhancing its adaptability.
Currently, the research team is actively working on the AI4Covid-19 mobile application, which has demonstrated promising results in accurately identifying coughs associated with the coronavirus.
The researchers are seeking funding for clinical trials to obtain approval from the US Food and Drug Administration (FDA), allowing them to bring the application to market. Currently, there are no FDA-approved tools for disease diagnosis based on audio recordings.
This study introduces new possibilities for advancements in healthcare, particularly in areas such as telemedicine and self-medication, by offering innovative approaches and expanding the utilization of artificial intelligence.