Diagnose your disease from coughs thanks to machine learning
The system was trained on a large dataset of 313 million two-second-long audio clips.
When you hear a person coughing, you will probably run away, but for this machine sound of coughing can be helpful.
According to a study published last year, cough is one of the most common medical complaints, accounting for as many as 30 million clinical visits per year. Up to 40% of these complaints result in a referral to a pulmonologist.
How the sound of coughing can help in the future?
According to the preprint, which has not yet been peer-reviewed, published earlier this month, healthy acoustic sounds such as coughs contain signals that could potentially help monitor health and disease.
However, these signals are underexplored in the medical machine-learning community. Furthermore, current systems are narrowly trained and focus on a single task.
In the study, a team led by Google scientists presented Health Acoustic Representations (HeAR), a scalable self-supervised learning-based deep learning system using masked autoencoders. This system was trained on a large dataset of 313 million two-second-long audio clips.
“We establish HeAR as a state-of-the-art health audio embedding model on a benchmark of 33 health acoustic tasks across 6 datasets,” the team wrote in the study.
Yael Bensoussan, a laryngologist at the University of South Florida in Tampa told Nature that traditionally, they have been using a lot of supervised learning in medicine, which is great because they have a clinical validation. “The downside is that it really limits the data sets that you can use, because there is a lack of annotated data sets out there,” he noted.
How do they do it?
Through an automated process, the team extracted over 300 million short sound clips of coughing, breathing, and throat clearing from publicly available YouTube videos. Then, each sound clip was converted into a visual representation of sound called a spectrogram. To help the model learn to predict the missing part, the team blocked segments of the spectrogram. That contributes to its adaptability.
HeAR was adapted to detect COVID-19 and tuberculosis and to include characteristics such as if the person smokes. The HeAR model achieved scores of 0.645 and 0.710 for COVID-19 detection, depending on the dataset used, indicating its performance on a scale where 0.5 represents random prediction and 1 represents perfect accuracy. The model scored 0.739 for tuberculosis detection.
The pandemic experience
In a study published four years ago, the scientists presented an AI-based preliminary diagnosis tool for COVID-19 using cough sound via a mobile app named AI4COVID-19. This tool is inspired by their independent prior studies that showed cough to be a good test medium for diagnosing a variety of respiratory diseases using AI.
“Building on the insights from the medical domain knowledge, we propose and develop a tri-pronged mediator-centered AI-engine for the cough-based diagnosis of COVID-19, named AI4COVID-19,” they wrote in the paper.
Ali Imran, one of the study’s authors and an engineer at the University of Oklahoma in Tulsa who also leads the development of AI4COVID-19, told Nature that his team plans to obtain approval from the US Food and Drug Administration (FDA) so that the app can eventually move to market.
He is currently seeking funding to conduct the necessary clinical trials. There are no FDA-approved tools that provide diagnosis through sounds at present.
The scientists are hopeful that their work will enable and accelerate further health acoustics. “Our goal as part of Google Research is to spur innovation in this nascent field,” added Sujay Kakarmath, a product manager at Google in New York City who worked on the project.
SHOW COMMENT ()
