An AI detects type 2 diabetes from voice
Researchers from Thymia and RMIT University trained an AI to spot signs of type 2 diabetes in the voice. Tested on 20-second recordings, it achieves promising results but remains less reliable for certain profiles.
An artificial intelligence can detect signs of type 2 diabetes from a 20-second voice recording, according to work presented at the European Association for the Study of Diabetes congress in Milan. The tool, developed by Thymia with RMIT University in Melbourne, could serve as a first screening filter, as many cases still go undetected. In France, 23% of diabetic people were unaware of their condition in the Esteban survey by Santé publique France.
A voice that reveals more than it seems
The model is based on an observation already noted by research: type 2 diabetes is sometimes accompanied by subtle changes in speech, such as hoarseness, a rougher voice, or less control of breathing. To train the AI, researchers used 63,283 voice samples from 21,129 people in the UK and the US, whose diabetes status was known. Then, the tool was tested on short recordings, making it easier to imagine in a remote screening pathway. Almost too simple, actually.
In the first evaluation, conducted on 7,319 adults in the UK aged 40 or over, 217 reported type 2 diabetes. Among them, 80% obtained a risk score higher than that of non-diabetic individuals. The authors describe the results as "clinically useful," although the test was less effective for Black people, who were less represented in the training cohort, as well as for patients suffering from heart disease, hypertension, or obesity.
A broader screening, but not a diagnosis
The second test involved 801 participants who recorded their voice before undergoing, within three months, a glycated hemoglobin test, or HbA1c. This test measures average blood sugar over the past two to three months and serves as a reference for diagnosing diabetes in many countries. The AI correctly identified 82% of people with type 2 diabetes and distinguished three risk levels. However, 47% of participants were wrongly classified as at risk: false positives, therefore.
This margin of error limits the use of the system as a standalone tool. However, its value lies in its accessibility: a voice sample can be collected by phone or via an app, without immediate laboratory testing. Giedrė Čepukaitytė, a research scientist at Thymia, believes the innovation "has the potential to transform screening" and specifies: "A voice sample can be collected by phone or via an app, which allows us to reach many more people needing blood analysis than current pathways, especially those who never go for a health check-up. Our model opens a new path for diabetes screening. It does not replace a blood test and should never prevent anyone who thinks they need one from having it."
Key points
- 63,283 voice samples were used for training.
- 20 seconds of speech are enough for an initial screening.
- 80% concordance in a UK test.
- 47% false positives in the second trial.
- Thymia and RMIT University conducted the study.
- Presented at the EASD congress in Milan, late September.
By the numbers
- 3.8 million people — treated for diabetes in France in 2023.
- 589 million adults — living with diabetes worldwide, according to the IDF.
- 853 million adults — global projection for 2050, according to the IDF.
- £10.7 billion — annual NHS budget for diabetes in the UK.
- 23% — share of French diabetics unaware of their condition, according to Esteban.