A speech-based AI model trained on recordings from more than 21,000 people showed promise in identifying type 2 diabetes, but researchers say larger, peer-reviewed studies are needed.
Researchers are testing whether a few seconds of ordinary speech could one day help flag type 2 diabetes, a condition that affects millions of people and is often diagnosed only after routine blood testing. According to a study posted on medRxiv, a speech-based artificial intelligence model trained on recordings from more than 21,000 people was able to identify the condition with promising accuracy, using just a 20-second sample of speech.
The preprint says the system produced an area under the curve score of 0.80 when matched against self-reported diabetes status, and 0.75 when compared with haemoglobin A1c blood test results, a commonly used marker of longer-term blood sugar levels. In practical terms, that suggests the model was often able to distinguish between people with and without type 2 diabetes, though it was not perfect and would still need further validation before any real-world use.
News reports based on the research say the model correctly identified about 82% of participants with type 2 diabetes and could also separate people into low-, medium- and high-risk groups according to HbA1c results. The researchers argue that a voice-based check could eventually complement existing screening tools, particularly because it would be quick, non-invasive and potentially easy to scale.
The idea is not that speech would replace established tests, but that it might help broaden access to screening for people who are less likely to attend regular health appointments. If confirmed in larger, peer-reviewed studies, the approach could offer a low-cost way to prompt earlier follow-up for people who might otherwise go undiagnosed.
Disclaimer: This content is for informational purposes only and is not intended to be a substitute for professional medical judgment, advice, diagnosis, or treatment.
