20 Seconds of Speech Can Flag Type 2 Diabetes – but Nearly Every Other Alert Is a False Positive
An AI tool developed by the company thymia and RMIT University is said to be able to flag possible type 2 diabetes from an audio recording of just 20 seconds. The results are being presented this week at the European diabetes congress EASD in Milan, and the researchers themselves describe the tool as a triage instrument – not a diagnosis. But the numbers behind it have a significant catch: nearly every other person the model flags as high risk has normal blood values.
How the tool was built and tested
The model was trained on 63,283 voice samples from 21,129 people in the UK and US, who themselves reported whether they had been diagnosed with diabetes (Medical Xpress/EASD). The training basis is thus self-reported diagnoses, not blood tests – and all the performance figures below come from the study's own authors, conveyed via the EASD press release. They are not peer-reviewed.
In the first evaluation, on 7,319 UK adults, the model gave a higher risk score to people who reported type 2 diabetes in 80 per cent of cases – a result the researchers themselves describe as clinically useful.
In a smaller subgroup of 801 participants who had taken an HbA1c blood test at home, the results were more mixed: the model gave a higher risk score in 75 per cent of cases, with a sensitivity of 82 per cent – meaning it identified 82 per cent of the people with type 2 diabetes in the group, according to the authors. But the proportion of false positives was 47 per cent. That means nearly half of those flagged as high risk had normal blood values.
On the other hand: none of those the model classified as low risk had blood results in the diabetic or prediabetic range. That is the most promising single figure in the material – but it, too, rests on the researchers' own data.
The caveats the researchers themselves point to
The results were weaker for Black participants. The authors explain this by noting that few Black participants reported type 2 diabetes in the dataset – an explanation that is their own assessment, not independently verified. Performance was also lower among people with heart disease, high blood pressure or obesity.
The high proportion of false positives is the most significant limitation. As a triage tool it may be acceptable: false positives are followed up with a blood test that rules them out. But it also means that a broad rollout would send many healthy people for unnecessary tests – a trade-off that some of the initial UK coverage has not spelled out. Some media have framed the tool as faster or better than blood tests, which goes further than the researchers' own positioning as a triage aid alongside blood testing.
The researchers' proposed use
The researchers suggest that GPs could use short audio recordings to triage patients: those at highest risk are offered confirmatory blood tests. Speech-based screening is to sit alongside blood testing, not replace it, the authors stress.
Giedrė Čepukaitytė, a researcher at thymia, says the model can reach people today's systems do not: "A voice sample can be taken over the phone or through an app, so we can reach far more of those who need a blood test than current solutions do, particularly those who never attend a health check. Our model opens a new pathway to diabetes screening. It is not a replacement for a blood test, and it should never stop anyone who thinks they need one from getting one," she told The Independent/PA (MSN).
The accessibility argument
The idea is aimed at a concrete gap in current follow-up: NHS health checks include diabetes screening and are offered to everyone over 40 every five years, but research cited by the study's authors shows that only around 40 per cent attend (The Mirror). A tool that requires 20 seconds of speech from a phone could, in principle, reach those who never book an appointment.
From the patient organisation's side, the tone is positive but cautious. "AI-based technologies could help identify more people who may benefit from diagnostic blood tests, but it's crucial that they are thoroughly designed and tested, so that no one slips through the net," said Dr Lucy Chambers, head of research impact and communications at Diabetes UK, to The Independent/PA.
What remains
The researchers are clear that clinical validation in real-world clinical settings is a future step, and there is so far no timeline for rollout. The results are being presented at the EASD congress in Milan (28 September–2 October 2026) and have not been published in peer-reviewed form – all performance figures remain, for now, the researchers' own. How well the model works in an ordinary GP population, with other diseases, other accents and other speech patterns, nobody yet knows.
As long as the sensitivity is high and the low-risk group looks clean, the logic as a triage tool is defensible. But before the tool can be recommended in primary care, the major weaknesses – the 47 per cent false positives and the performance gaps in subgroups – must be documented and handled independently of the study's own authors.