Somewhere between a spoken sentence and a medical chart, researchers have proposed that the human voice may carry quiet traces of blood sugar. In a preprint that has not yet faced the scrutiny of peer review, a team linked to the company thymia describes an AI model that listens to about twenty seconds of someone reading aloud and estimates the risk of type 2 diabetes. The result is promising in the way early instruments often are: it points toward a simple, humane first question, while leaving the answer to the older and more certain instruments of the blood test.
AI Voice Screening Flags Diabetes Risk in 20 Seconds, But Results Are Preliminary and Unreviewed
A 20-second recording cannot tell anyone they have diabetes.
Mark
So what is this thing actually doing? It's not reading my words, is it?
Mimi
No. It listens to vocal qualities, like hoarseness, roughness, and how someone controls breath and voice. Twenty seconds of speech is enough for it to produce a risk score.
Mark
The headline number is 0.80. That sounds like 80% accuracy.
Mimi
It isn't. It's an AUC. Give the model one person with reported diabetes and one without, and it ranks the person with diabetes higher about 80% of the time. That's different from diagnosing 80% of people correctly.
Luke
And worth remembering that 0.80 was measured against self-reported diagnoses, not lab results. That's a softer yardstick.
Mark
Was there a harder one?
Mimi
Yes. In 801 participants with HbA1c blood tests, the AUC was 0.75. At one threshold it caught about 82% of diabetes-range cases, with a 47% false-positive rate.
Luke
That false-positive rate matters. Nearly half of people without diabetes-range results would be flagged. As a nudge toward a blood test, maybe tolerable. As anything more, no. Also, that 82% is at one chosen threshold, so it isn't a fixed property of the model.
Mark
Who's behind it?
Mimi
Researchers at thymia with academic collaborators. Several authors are employees, founders or shareholders of the company.
Luke
Which doesn't invalidate the work, but it's a reason the preprint status matters. It hasn't been peer reviewed. We don't yet know how it holds up under outside scrutiny, or why it performs worse in some groups and in people with cardiovascular disease or hypertension.
Il Polso
- A preprint AI model claims it can flag type 2 diabetes risk from just 20 seconds of spoken audio, a claim that has not yet been peer reviewed.
- Its AUC of 0.80 against self-reported diagnoses sounds impressive, but it is a ranking measure, not a count of correct diagnoses.
- Against HbA1c blood results, the model caught about 82% of diabetes-range cases while wrongly flagging 47% of people without diabetes.
- Performance drops in some groups and in people with cardiovascular disease or hypertension, raising questions about fairness and reliability.
- Developers have commercial stakes in the findings, so independent validation and broader population testing are the next essential steps.
Somewhere between a spoken sentence and a medical chart, researchers have proposed that the human voice may carry quiet traces of blood sugar. In a preprint that has not yet faced the scrutiny of peer review, a team linked to the company thymia describes an AI model that listens to about twenty seconds of someone reading aloud and estimates the risk of type 2 diabetes. The result is promising in the way early instruments often are: it points toward a simple, humane first question, while leaving the answer to the older and more certain instruments of the blood test.
The test asks very little of the person taking it. They read a short passage aloud, in this case lines from Aesop's fables, and an algorithm listens for twenty seconds. It is not paying attention to the words but to the texture of the voice: a rougher, hoarser quality, and subtle changes in how breath and voice are controlled. Researchers link these qualities to type 2 diabetes.
The model was built by the technology company thymia and academic collaborators, trained on more than 63,000 voice samples from about 21,000 people in the UK and US, and then tested separately on 7,300 UK adults. Its first measure, an AUC of 0.80 against self-reported diagnoses, means that the system tends to rank a person with diabetes above a person without it about four times in five. It does not mean the system correctly diagnosed four in five people.
A tougher test used blood results from a subgroup of 801 participants. Against the clinical threshold for diabetes, the model reached an AUC of 0.75. At one setting it caught about 82% of people in the diabetes range, but it also flagged 47% of those without, so nearly half of the people who did not have diabetes would have been wrongly alarmed.
For that reason, the developers do not propose the technology as a replacement for a blood test. They describe it as a first-pass screen: a short clip recorded on a phone, with a suggestion to seek conventional testing if the score is high. The study also found weaker performance in some groups and in people with heart disease or hypertension, and several authors have commercial ties to thymia.
The most important caveat is that this is a preprint on medRxiv, not yet peer reviewed, and its figures have not been independently checked. Still, the appeal is clear. Diabetes screening depends heavily on blood tests and on people showing up for checks. If the findings hold, a twenty-second recording might nudge more people toward testing. What comes next is peer review, wider testing, and an explanation of why the model performs unevenly.
The test asks for very little. A person reads a short passage aloud, in this case lines from Aesop's fables, and an algorithm listens for 20 seconds. What it is listening for is not the words but the texture of the voice: increased hoarseness and roughness, and shifts in how a speaker controls breath and voice. Researchers say these qualities carry a signal linked to type 2 diabetes.
The model comes from the technology company thymia and its academic collaborators. It was trained on 63,283 voice samples from 21,129 people in the UK and US, drawn from both reading tasks and free speech. The team then tested it separately on 7,319 adults in the UK, who each supplied a short recording of the fable passages.
The first test compared the model's output with participants' self-reported diagnoses. It scored an AUC of 0.80. That number is easy to misread. It means that if the system is handed one person who reports diabetes and one who does not, it will tend to give the higher risk score to the person with diabetes about 80% of the time. It does not mean the system correctly diagnosed 80% of participants.
The second test was tougher, because it used blood results rather than self-reports. A subgroup of 801 participants completed HbA1c tests, which measure average blood glucose over the preceding few months. Measured against the clinical threshold for diabetes, the voice model reached an AUC of 0.75. At one operating threshold it picked up about 82% of people whose HbA1c fell in the diabetes range, but the false-positive rate was 47%. Nearly half of those without diabetes-range results would have been flagged anyway.
That is why the technology is nowhere near replacing a blood test, and its developers are not suggesting it should. The proposal is narrower: a first-pass screen. Someone could record a short clip through a phone or app and, if the score suggested elevated risk, be advised to get conventional testing.
The study has other limits. Performance was weaker in some groups, and in people with conditions such as cardiovascular disease and hypertension. The researchers say further testing across different populations is needed. Several authors are also employees, founders or shareholders of thymia, which could benefit commercially from the technology.
The most important caveat is that the work is a preprint on medRxiv and has not been peer reviewed. Its figures are the team's own and have not been independently scrutinized.
The appeal is still plain. Diabetes screening leans heavily on blood tests and on people actually showing up for health checks. A 20-second recording cannot tell anyone they have diabetes. If the findings hold up, though, it might offer an unusually simple way to say that it may be time to get tested. What to watch for is peer review, testing in broader populations, and an explanation of why the model does worse in some groups.
Citazioni salienti
The technology is not meant to replace a blood test; it is proposed as an easily accessible first screen that could prompt conventional testing.— Paraphrase of the researchers' stated aim, as reported by Moneycontrol
Vuoi la storia completa?
Leggi l'originale su Moneycontrol ↗
Technology