Artificial intelligence can now assess a person's risk for type 2 diabetes from a 20-second voice sample, offering a potential tool to screen candidates for further blood testing before complications develop.
Voice Data Training and Validation
The research team trained the AI model using 63,283 voice recordings collected from 21,129 individuals in the UK and the US, where participants self-reported their diabetes status. To test its performance, investigators asked trial participants to read an Aesop’s fable for roughly 20 seconds. In an initial evaluation of 7319 UK adults, the AI model assigned higher risk scores to the 217 individuals who reported a type 2 diabetes diagnosis at roughly an 80% higher probability than to those without a diagnosis.
A second evaluation compared 801 participants against glycated hemoglobin blood test results taken within three months of their voice recording. The model assigned higher scores to diabetes patients than non-patients 75% of the time, demonstrating an ability to spot prediabetes alongside established cases. Among participants confirmed to have diabetes via blood tests, the AI flagged 82% as high risk. Among the bottom 10% of participants classified as low risk, none of the 83 individuals tested positive for diabetes or prediabetes on their blood panels, indicating the tool could help filter out individuals who do not require immediate follow-up testing.
Implementation Challenges and Future Screening Plans
Despite its potential for wide accessibility, the model registered a false-positive rate of 47%, incorrectly classifying healthy individuals as high risk. Because of this rate, the researchers propose using smartphone or telephone voice collections merely to flag high-risk candidates for subsequent blood test confirmation rather than letting voice tests replace standard diagnostics.
The study represents the largest real-world investigation into voice-based screening for type 2 diabetes to date. Because voice samples can be gathered via phone calls or smartphone apps, the research team notes that the approach could eventually reach significantly more people than conventional screening pathways currently allow.