Overview
- Researchers presented results at the EASD meeting in Milan on Monday showing the model flagged self‑reported diabetes 80% of the time and matched HbA1c‑defined diabetes 75% of the time with a sensitivity of 82% and a false positive rate of 47%.
- The system was trained on 63,283 voice samples from 21,129 people in the UK and US and was validated using remote 20‑second recordings of participants reading Aesop’s fables.
- Accuracy fell for recordings from Black participants and for people with heart disease, high blood pressure or obesity, which researchers link to low representation and to similar vocal changes caused by those conditions.
- Authors stress the tool is intended as a fast, non‑invasive screening or triage aid to prompt blood testing rather than a diagnostic replacement, and they say further clinical validation and subgroup testing are needed.
- Several study authors are employees, co‑founders or equity holders in thymia Ltd, and researchers point to large gaps in current screening as a reason the approach could scale testing if equity and clinical questions are resolved.