The Problem of Late Diagnosis
Type 2 diabetes often progresses silently, and many people only discover the disease when complications arise. In the United Kingdom, about one third of cases remain undiagnosed, which increases healthcare system costs with treatments for sequelae.
Training the Voice Tool
Researchers from the company thymia, in partnership with RMIT University in Melbourne, Australia, fed an AI model with tens of thousands of audios from volunteers from the United States and the United Kingdom. Each participant read a short fable, allowing the algorithm to capture subtleties such as mild hoarseness and changes in breathing during speech.
Accuracy Obtained in Tests
- 80% accuracy in classifying people with diabetes just by voice.
- 82% agreement when compared directly with blood tests.
The system does not replace laboratory tests, but serves as an initial screening to indicate who should undergo analyses with greater urgency.
Next Steps and Applications
The technology will be evaluated in real clinics in the coming months. In the future, screening could occur through apps or phone calls, expanding access in regions with little medical infrastructure and allowing earlier interventions.
The definitive diagnosis, however, will continue to depend on confirmatory laboratory tests.