An AI tool trained on tens of thousands of voice samples was able to spot type 2 diabetes about 80 percent of the time. It needed only 20 seconds of someone reading aloud.
The findings were presented at the European Association for the Study of Diabetes (EASD) annual meeting in Milan. They have not yet been published in a peer-reviewed journal.
The tool was built by the deep-tech company thymia with researchers at RMIT University in Melbourne, Australia.
It listens for subtle changes in speech that earlier work has linked to diabetes, such as hoarseness and reduced breath control.
Also read: A new weight-loss drug could cut your body weight by a quarter
How the AI listens for diabetes
The team trained the model on 63,283 voice samples from 21,129 people in the UK and US. To test it, each participant read a short passage from Aesop's fables into a microphone for about 20 seconds.
The first evaluation involved 7,319 UK adults, and 217 of them had type 2 diabetes. The model assigned those people a higher risk score 80 percent of the time.
A second, tougher test used 801 people with actual blood-test results. There, the tool correctly separated those with diabetes from those without in three cases out of four.
The picture is not spotless. The model correctly picked up about 82 percent of people who did have diabetes. But it also falsely flagged nearly half of the people who did not.
Also read: These fruits contain a surprising amount of protein
Any positive result would still need to be confirmed by a proper blood test.
Why diabetes may leave a mark on the voice
The researchers say long-running high blood sugar can damage small nerves and muscle fibers, including those around the voice box and the muscles that power breathing.
That can leave a voice sounding rougher, or running out of air more quickly than it should.
What it could mean in practice
A speech sample "can be taken over the phone or through an app," said Giedrė Čepukaitytė, a research scientist at thymia who presented the findings. She said that could open a route to people who never get to a health check today.
Also read: How your bread can stay fresh for up to 8 months
"It is not a replacement for a blood test, and it should never stop anyone who thinks they need one from getting one," Čepukaitytė said.
The CDC's National Diabetes Statistics Report puts the number of American adults living with diabetes at about 40 million, and estimates that 11 million of them have not been diagnosed.
Many early cases go unnoticed because the warning signs are often dismissed as tiredness.
The next step, Čepukaitytė said, is to test the model in clinical settings and to check that it works equally well for every group of people.
Also read: Using AI for health questions is tied to anxiety and depression in young adults
Also read: Breast cancer survivors aged 9 months slower with this weekly exercise habit
This article is made and published by Mie Hermansen, who may have used AI in the preparation.