AI could detect diabetes from voice recordings
A new artificial intelligence system has shown it can identify people at higher risk of type 2 diabetes by analysing just 20 seconds of speech, potentially creating a new route for early screening.
Artificial intelligence could help detect type 2 diabetes by analysing a short voice recording, according to new research suggesting the technology may offer a faster way to identify people who need further testing.
Researchers from AI health technology company Thymia and RMIT University in Melbourne developed a speech analysis tool that can recognise patterns linked to the condition from a 20-second audio clip.
The AI model was trained using more than 63,000 voice recordings from over 21,000 people in the UK and US.
Researchers tested the system using recordings of people reading Aesop’s fables and found it could identify those who reported having type 2 diabetes with an accuracy rate of 80 per cent.
The findings are being presented at the European Association for the Study of Diabetes conference in Milan and included data from 7,319 people in Britain.
Scientists believe changes in speech, including increased hoarseness and reduced breath control, may provide clues about the effects of diabetes on the body.
The researchers said the technology could "open a new route" for diabetes screening, with recordings potentially collected through smartphones, apps or telephone calls.
Currently, type 2 diabetes is diagnosed through blood tests that measure average blood sugar levels over the previous two to three months.
Testing is available for people experiencing symptoms or through routine NHS health checks for those aged between 40 and 74.
The AI system performed consistently across different age groups and genders, but researchers said its accuracy was lower among black patients, who were less represented in the training data.
Performance was also reduced among people with conditions including heart disease, high blood pressure and obesity, as these can create similar vocal patterns.
A second analysis looked at 801 people who completed home diabetes blood tests within three months of their voice recording.
In this group, the AI model identified higher-risk individuals 75 per cent of the time.
Giedre Cepukaityte, a research scientist at Thymia, said the study represented the largest real-world investigation of speech-based type 2 diabetes screening so far.
However, she stressed that the technology would not replace traditional blood tests.
Dr Lucy Chambers from Diabetes UK said AI tools could help identify more people who may benefit from testing, but added that systems must be rigorously assessed to ensure they work effectively for everyone.
Researchers said the next stage will involve testing the technology in clinical settings and examining how it performs across different communities.
