Against blood tests, the model detected 82% of diabetes cases but also flagged 47% of participants who did not have the condition.
Researchers at health-technology company thymia tested an AI model that uses short voice recordings to flag possible type 2 diabetes. The study, presented at the European Association for the Study of Diabetes meeting in Milan, is a preprint that has not been peer-reviewed.
The researchers propose offering the voice test by phone or through an app to help reach people who miss routine health checks. Those flagged would be referred for blood testing.
Earlier studies have linked diabetes to changes in voice quality and breath control. The team trained its model on recordings from 21,129 people and tested it on 20-second recordings of 7,319 UK adults reading a short passage aloud. A subgroup of 801 also took blood tests; 35 met the blood-test threshold for diabetes. In that subgroup, the researchers reported 82% sensitivity and a 47% false-positive rate.
The researchers compared the model with QDiabetes, an established questionnaire for estimating future diabetes risk, which they used here to identify current cases. The questionnaire was better at ranking self-reported cases above non-cases. In the blood-tested group, differences between the two tools were not statistically significant.
Performance was lower among Black and Asian participants, although the small number of diabetes cases in those groups makes those estimates uncertain. It also fell among people with obesity, high blood pressure or heart disease. The model has not yet been tested in a clinic.
All seven authors work for thymia, and four hold equity in the company. The study received no external funding.