AI Model Detects Type 2 Diabetes Risk from a 20-Second Voice Sample
- An artificial intelligence model developed to assess type 2 diabetes risk from a twenty-second voice sample could help filter patients for additional blood testing, according to research presented...
- To train the artificial intelligence model, researchers used 63,283 voice samples collected from 21,129 individuals in the United Kingdom and the United States, who self-reported their diabetes diagnoses.
- While prior studies noted that patients with type 2 diabetes can experience rougher voices, hoarseness, and shifts in vocal and respiratory control, those acoustic traits had not been...
An artificial intelligence model developed to assess type 2 diabetes risk from a twenty-second voice sample could help filter patients for additional blood testing, according to research presented on August 28 at the European Association for the Study of Diabetes annual meeting in Milan. The technology was built by UK-based artificial intelligence firm thymia in collaboration with researchers from the Royal Melbourne Institute of Technology in Australia, with initial findings published on August 18 on the preprint server medRxiv.
Voice-Based Screening Model Architecture and Validation
To train the artificial intelligence model, researchers used 63,283 voice samples collected from 21,129 individuals in the United Kingdom and the United States, who self-reported their diabetes diagnoses. The performance of the system was then validated using voice data gathered while participants read an Aesop fable for approximately twenty seconds. In an initial evaluation of 7,319 British adults, the artificial intelligence model assigned higher risk scores with about an 80% higher probability to 217 participants who reported a type 2 diabetes diagnosis compared to those without a diagnosis. A second evaluation compared the model against blood test results from 801 participants who underwent glycated hemoglobin blood testing within three months of their voice recording. The model assigned higher scores to diabetes patients than non-patients in 75% of cases and successfully identified 82% of participants confirmed to have diabetes through blood work. Among 83 individuals selected from the lowest 10% of risk scores classified as low risk by the model, zero blood test results showed diabetes or prediabetes. However, the false-positive rate for misclassifying healthy individuals as positive reached 47%.
Clinical Implications and Proposed Screening Pathways
While prior studies noted that patients with type 2 diabetes can experience rougher voices, hoarseness, and shifts in vocal and respiratory control, those acoustic traits had not been adequately validated for large-scale screening until now. Managing complications like heart disease or nerve damage makes early detection vital, particularly given that the UK National Health Service spends approximately £10.7 billion annually on diabetes, with roughly 60% of that total directed toward complication management.
Because voice samples can be obtained via phone or app, they can reach far more people than current testing pathways,
the research team stated, while noting that the method represents the largest real-world voice-based screening study for type 2 diabetes conducted to date. thymia and Royal Melbourne Institute of Technology researchers emphasize that voice tests should not replace blood tests, proposing instead that smartphones or telephones could initially screen high-risk populations before blood work provides formal confirmation.