AI Algorithms and Coughing
Erica Mannis explains how artificial intelligence has been programmed to accurately detect COVID positive patients
Artificial Intelligence (AI) has detected coronavirus in people from a recording of their cough with an incredible 98.5% accuracy. The algorithm developed by researchers at MIT has vast applications across healthcare in preventing and curbing future pandemics, and may soon be available in a free app.
Despite this algorithm being originally developed to detect Alzheimer’s in patients by applying three biometrics to respective neutral networks, it has proved full of potential even amongst asymptomatic patients. Being asymptomatic is defined as having no identifiable symptoms while still having the virus. Normally, this makes recognising you have the virus difficult; however, this new algorithm was 100% successful in detecting COVID-19 in asymptomatic persons from a slight difference in sound that cannot be distinguished by the human ear.
this new algorithm was 100% successful in detecting COVID-19 in asymptomatic persons
The first biometric was vocal cord strength, with the sound ‘mmm’ known to be an indicator of this. The network was made to listen to thousands of hours of audiobooks tasked with picking out the word ‘them’ instead of similar sounding words such as ‘the’ and then’. Next the network was trained in recognising emotion and stress in a person’s voice, associated with the neurological decline in Alzheimer’s disease. Researchers developed a speech classifier model that sorted actors’ voices into categories such as happy, calm, and frustrated. Lastly, the network was tasked with identifying changes in lung capacity and respiratory ability in coughs from a database.
These three models were integrated and overlaid with an algorithm for muscular degradation. This provides an ‘audio mask’ for weak muscles, background noise or poor recoding quality. The machine-learning algorithm, known as Res-Net 50, could identify Alzheimer’s better than any model available.
As coronavirus has similar neurological symptoms, the system only needed tweaking in order to function in the same way. To test its ability a website was set up for people to upload recordings of their forced coughs from their phones or laptops. Over 70,000 recordings have been submitted containing at least 200,000 coughs. This is the largest database of coughs known to the researchers. Within this sample 2,500 asymptomatic COVID carriers were identified without mistake.
Over 70,000 recordings have been submitted containing at least 200,000 coughs.
Currently the team are working with a company to develop an app able determine a person’s health from a recording of a forced cough. This would be highly useful to members of the public, especially students and workers, as a pre-screening technique.
AI projects like this have vast applications and could be used as constant background tools to reduce a pandemic’s effect by recognising asymptomatic carriers. This is revolutionary positive use for AI in medicine.