AI listens for diabetes in a 20‑second voice clip
Researchers have unveiled an artificial‑intelligence system that can infer the presence of type‑2 diabetes from a brief spoken recording. The model, created by the tech firm thymia in partnership with RMIT University, was fed more than 63,000 voice samples from over 21,000 volunteers in the United Kingdom and the United States. Participants self‑reported whether they had received a diabetes diagnosis, allowing the algorithm to learn subtle acoustic signatures linked to the disease.
From training to real‑world testing
After the extensive training phase, the team conducted a prospective study involving 7,319 British adults. Each person narrated a short story for roughly twenty seconds, either over the phone or via an online platform. The AI assigned a risk score to every speaker based on patterns it had identified – such as a huskier timbre, reduced breath control, or a slightly rougher articulation, traits previously associated with metabolic disturbances.
Performance metrics: promising yet imperfect
When the algorithm’s predictions were compared with the participants’ self‑declared health status, it correctly ranked the diabetic individual higher than the non‑diabetic counterpart in eight out of ten cases – an 80 % accuracy rate. To obtain a more objective benchmark, 801 volunteers also underwent a standard blood test measuring average glucose levels over recent months, the gold‑standard for diagnosing diabetes.
Against this laboratory reference, the AI correctly identified 82 % of the true diabetic cases. However, it also raised alarms for 47 % of the participants who were actually free of the disease, yielding a specificity of just 53 %. In other words, nearly half of the healthy individuals would be flagged for further testing.
Potential role as a screening filter
Proponents suggest the voice test could serve as an initial sieve in primary‑care settings. A practitioner might ask a patient to record a short utterance; those with high risk scores would then be referred for a confirmatory blood test. The approach could broaden outreach, especially for people who rarely attend routine health checks, because recordings can be captured via a simple phone call or a mobile app.
Nevertheless, experts caution against premature adoption. The relatively low specificity means many false positives, which could strain resources and cause unnecessary anxiety. Moreover, the model performed less reliably for certain demographic groups, such as Black participants, likely due to under‑representation in the training data.
Why the test isn’t a replacement
Clinicians emphasize that a finger‑prick glucose measurement remains far more dependable. The AI‑driven voice analysis should be viewed as a complementary tool, not a substitute for established diagnostic procedures. Until the algorithm’s sensitivity and specificity improve, especially across diverse populations, its utility will likely stay confined to research environments.
Future work will focus on expanding the dataset, refining the acoustic features, and integrating socioeconomic variables to reduce bias. If these hurdles are overcome, a quick, non‑invasive voice check could become a valuable first step toward earlier detection of type‑2 diabetes.
Source: https://scientias.nl/ai-hoort-diabetes-type-2-aan-je-stem-geen-vervanging-voor-een-bloedtest/
