An artificial intelligence can detect if YouTubers are infected with the omicron coronavirus variant with up to 80 per cent accuracy. Although vocal changes aren’t considered a key symptom of any coronavirus infection, the researchers behind the AI argue their results suggest a subtle “Omicron-specific laryngitis”.
Monitoring audio samples uploaded to social media could be a relatively quick, inexpensive way of tracking coronavirus cases in the community, according to the team. However, another scientist argues that an accuracy of at least 95 per cent would be required for methods like this to be useful in real-world settings.
at the National Institutes of Health, Maryland, and his colleagues analysed 93 hours of YouTube audio samples.
Advertisement
In these samples, 183 speakers said they were infected with the coronavirus when the omicron variant was dominant, 120 said they were infected with the coronavirus when a non-omicron variant was circulating, 138 said they had an upper respiratory infection that wasn’t the coronavirus and 192 didn’t mention having any respiratory infection.
These samples were processed to remove moments of silence and any noises other than the speaker talking. They were then split into segments, each lasting 2.5 seconds, some of which were randomly selected to be fed into the AI for training. It was then put to the test via the unused samples from the same audio-recording collection.
The results show that the AI differentiated the omicron audio samples from the speakers with no known respiratory infection with a sensitivity – an ability to correctly identify people with omicron – of 80 per cent. Its specificity – an ability to identify people without an upper respiratory infection – was 85 per cent.
Differentiating those with omicron compared with a non-coronavirus respiratory infection led to a sensitivity – an ability to identify people with omicron – of 70 per cent and a specificity – an ability to identify people with another respiratory infection – of 76 per cent.
The AI also detected non-omicron coronavirus infections from those with no respiratory infection with 58 per cent sensitivity and 80 per cent specificity. The researchers argue this suggests there are “limited acoustic biomarkers” in people who were infected with the coronavirus pre-omicron, which is now the dominant variant worldwide.
Non-omicron coronavirus infections were separated from non-coronavirus respiratory infections with 70 per cent sensitivity and 74 per cent specificity.
The researchers acknowledge that the speakers in the YouTube clips self-declared their infection status, rather than it being determined via viral sequencing, but they argue that this would make the results faster and more cost-effective if this tracking approach were to be rolled out.
However, the AI’s accuracy may be too low for it to be relied on as a means of monitoring coronavirus cases in the community.
“Had it showed a specificity and sensitivity crossing 95 per cent, it would have served as a welcome solution,” says at Stanford University, California.
Wood’s team write in the paper that the study was an “early effort” that warrants further research.
Reference: medRxiv, DOI:
