AI voice test shows promise for type 2

An artificial intelligence system can screen for type 2 diabetes by analyzing just 20 seconds of speech. The tool, developed by researchers at tech company thymia and RMIT University in Melbourne, Australia, examines vocal patterns for biomarkers linked to the condition, such as hoarseness and reduced breath control.

The system was calibrated using more than 63,000 voice recordings from over 21,000 subjects in the UK and the US. It analyzes short audio clips, which can be gathered via mobile applications or phone calls, for specific vocal changes. These changes include a rough or scratchy voice quality, often caused by high blood sugar damaging the vagus nerve or by stomach acid reflux, which is more common in people with diabetes. Reduced lung function from the disease can also lower the airflow needed for clear speech.

Study findings on accuracy and performance

In an initial evaluation involving 7,319 UK adults, the AI model correctly identified individuals who self-reported having type 2 diabetes 80% of the time. A secondary analysis tested the model on 801 participants who had taken HbA1c blood tests within three months of their speech recording. In this group, the model correctly identified type 2 diabetes 75% of the time, with an 82% sensitivity rate. The model could distinguish between low, medium, and high risk based on HbA1c results, with no low-risk classifications having diabetic or prediabetic blood results. Giedre Cepukaityte will present these findings at the European Association for the Study of Diabetes in Milan.

Limitations and equity concerns

The model performed well across sexes and ages but showed reduced accuracy for Black participants. This discrepancy is likely due to the low representation of Black participants with type 2 diabetes in the training data. Performance was similarly diminished among participants with obesity, high blood pressure, or heart disease.

Potential role in healthcare and next steps

Researchers propose the tool as a triage aid in primary care to prioritize who needs confirmatory blood tests, not as a replacement for current diagnostics like the HbA1c test. In the UK, about 30% of type 2 diabetes cases are undiagnosed, equating to more than a million people. The NHS spends £10.7 billion annually on diabetes, with around 60% spent on managing complications. While the NHS offers diabetes screening every five years to people aged 40 and older, fewer than half of eligible adults attend. Dr Lucy Chambers, head of research impact and communications at Diabetes UK, said AI-based technologies could help identify more people who may benefit from diagnostic blood tests, but it is crucial they are rigorously designed and tested. The researchers state the tool is not a replacement for blood tests but could be used for triage in primary care to prioritize who needs confirmatory testing.