At the European Association for the Study of Diabetes’s annual meeting in Milan, scientists reported that artificial-intelligence analysis of a speech sample lasting only a few seconds can quickly identify individuals with type 2 diabetes non-invasively.
AI analyzes speech patterns
A new study—conducted on a scale never before attempted—suggests that artificial intelligence can analyze brief voice clips to identify potential signs of diabetes quickly, without invasive methods, and on a broad scale. Scientists at the advanced technology firm thymia created an AI system designed to spot speech patterns associated with the disease. The team built the model using 63,283 voice recordings from 21,129 individuals across the UK and US, all of whom confirmed whether they had received a diabetes diagnosis.
To assess the model’s accuracy, researchers tested it on 20-second audio clips where participants recited one of Aesop’s fables. The first test involved 7,319 UK adults, of whom 67% were women and 45.8% were at least 40 years old. Among them, 217 participants had previously been diagnosed with type 2 diabetes.
The AI system correctly assigned higher risk scores to those with type 2 diabetes 80% of the time, a level deemed clinically meaningful, compared to individuals without the condition.
Validation against blood tests
The model performed well across different sexes and ages. However, performance was lower on recordings from Black participants. This was likely due to the low number of Black participants reporting type 2 diabetes, say the authors. Performance was also lower on recordings from people with heart disease, high blood pressure or obesity. These conditions often coincide with type 2 diabetes and may cause similar vocal changes.
The second phase of testing included 801 participants who conducted at-home HbA1c blood tests within three months of recording their voices. The HbA1c test, which tracks average blood sugar over two to three months, serves as the standard diagnostic tool for type 2 diabetes and can also detect prediabetes, where sugar levels are raised but not yet severe enough for a full diagnosis.
Here, the speech model gave a higher risk score to people with type 2 diabetes than to those who did not have the condition, based on HbA1c tests, 75% of the time. The sensitivity was 82% and the false positive rate was 47%. None of the people the model classed as low risk had blood results in the diabetic or prediabetic range.
Pathways for implementation
Giedrė Čepukaitytė, Research Scientist at thymia, who will be presenting the findings at EASD, says: “This is the largest real-world study of speech-based screening for type 2 diabetes to date which also checks the model’s predictions against blood test results as well as against what people reported about their own diagnosis. Those flagged up as higher risk by the model had blood results to match. This has the potential to change what screening looks like.
A speech sample can be taken over the phone or through an app, so we can reach far more of the people who need a blood test than current pathways do, particularly those who never get to a health check. Our model opens a new route to screening for diabetes. It is not a replacement for a blood test, and it should never stop anyone who thinks they need one from getting one. Our next step is to test…”
