Researchers from thymia and RMIT University have developed an AI model that screens for type 2 diabetes in 20 seconds by analyzing vocal changes like hoarseness and breath control. Validated on recordings of participants reading Aesop's fables, the tool achieved 75% to 80% accuracy in identifying high-risk individuals. While performance was lower for Black participants and those with obesity, the technology offers a scalable, noninvasive triage tool to help reach over a million undiagnosed people in the UK.
How diabetes affects the voice
- ▪Type 2 diabetes is linked to vocal changes such as increased hoarseness, roughness, vocal strain, and reduced control of breath and voice while speaking.
- ▪High blood sugar can harm the vagus nerve that controls voice box muscles, while stomach acid reflux and reduced lung function in diabetes patients also cause vocal changes.
Training and validation of the model
- ▪The AI speech model was trained on 63,283 voice samples from 21,129 people in the United Kingdom and United States who reported their diabetes diagnosis status.
- ▪Researchers validated the AI speech model using 20-second remote recordings of participants reading aloud from one of Aesop's fables.
Accuracy of the AI model
- ▪In a subgroup of 801 participants who took home HbA1c blood tests, the AI speech model correctly assigned higher risk scores to diabetic individuals 75% of the time.
- ▪In an evaluation of 7,319 adults in the United Kingdom, the AI speech model assigned higher risk scores to self-reported type 2 diabetes patients 80% of the time.
Limitations in model performance
- ▪The AI speech model showed lower performance on recordings from Black participants, likely due to their low representation and low reporting numbers in the study.
- ▪The AI speech model showed lower performance on participants with heart disease, high blood pressure, or obesity, which can cause vocal changes similar to type 2 diabetes.
Undiagnosed diabetes in the UK
- ▪Approximately 30% of diabetes cases in the United Kingdom are undiagnosed, and fewer than half of eligible adults attend standard NHS health check screening appointments.
- ▪Dr Lucy Chambers of Diabetes UK stated that more than one million people in the United Kingdom live with undiagnosed type 2 diabetes, missing crucial early treatment.
Research and presentation
- ▪Researchers at the technology company thymia and RMIT University developed an AI model that can screen for type 2 diabetes in 20 seconds using speech recordings.
- ▪Giedrė Čepukaitytė of thymia will present the research findings at the European Association for the Study of Diabetes meeting in Milan, Italy, between September 28 and October 2, 2026.
Accessing the screening tool
- ▪The AI speech-screening model is intended to sit alongside standard blood testing to triage high-risk patients for further clinical testing, rather than replacing blood tests.
- ▪The AI speech model can screen patients using voice recordings collected remotely over the phone or through a mobile application.
Debatable claims
- ▪AI speech analysis is too unreliable to serve as a clinical triage tool
- ▪App-based medical screening does more harm than good for public health
- ▪A high false-positive rate is acceptable in preliminary public health screening tools
- ▪Healthcare systems should delay deploying AI screening tools until demographic performance gaps are resolved
Story comments
Loading comments…