The widespread, undetectable transmission of COVID-19 has significantly fueled the pandemic; however, identifying individuals who are infected but show no symptoms presents a major challenge. Recent research from MIT suggests that a hidden indicator of early infection may be found within the characteristics of a person’s cough, offering a potential early warning system for the virus.

For many years, medical professionals have understood that the sound of a cough can reveal valuable information about a person’s health. Artificial intelligence models have been developed to identify conditions such as pneumonia, asthma, and even neuromuscular diseases, all of which manifest in distinct changes to a person’s cough.

Prior to the pandemic, researcher Brian Subirana demonstrated the potential for cough analysis to even predict Alzheimer’s disease – a finding that aligned with earlier research from IBM. Subsequently, Subirana considered whether this AI capability could be extended to detect COVID-19. He was not alone in this line of thought.

Subirana and his colleagues created a platform for individuals to submit recordings of their coughs, resulting in what they believe is the most extensive cough dataset available for research purposes. This collection of thousands of samples was then used to train the AI model, the details of which are published in an openly accessible IEEE journal.

The model appears to identify subtle variations in vocal intensity, emotional expression, lung function, respiratory performance, and muscular strength. This allowed it to correctly identify 100% of coughs from individuals carrying COVID-19 without symptoms and 98.5% of those with symptoms, while maintaining a specificity of 83% and 94% respectively, indicating a low rate of both false positive and false negative results.

“Our findings suggest that the mechanics of sound production are altered when someone has COVID, even in the absence of symptoms,” explained Subirana. He did, however, emphasize that while the system is effective at detecting abnormal coughs, it should not be used as a self-diagnosis tool for individuals experiencing symptoms without a known cause.

I requested further clarification from Subirana on this point.

“The tool is designed to identify characteristics that differentiate individuals with COVID from those without the virus,” he clarified in an email. “Previous studies have shown the potential to detect other conditions as well. It would be possible to create a system capable of distinguishing between numerous conditions, but our primary focus was on isolating COVID-19.”

The exceptionally high success rate reported may understandably raise concerns among those familiar with machine learning. While these models are powerful, a 100% accuracy rate is uncommon and often warrants further investigation. The findings will require validation using additional datasets and independent verification by other research teams; however, it is also plausible that a consistent, detectable pattern exists in COVID-19-related coughs that can be readily identified by a computer-based listening system.

The research team is currently collaborating with multiple hospitals to expand their dataset and increase its diversity. They are also partnering with a private company to develop an application for broader distribution of the tool, pending regulatory approval from the FDA.