New research presented at the Annual Meeting of The European Association for the Study of Diabetes (EASD) in Madrid (September 9–13) reveals that voice analysis could effectively identify undiagnosed type 2 diabetes (T2D). The study used approximately 25 seconds of voice recordings, combined with basic health information such as age, sex, body mass index (BMI), and hypertension status, to develop an AI model. This model achieved 66% accuracy in detecting T2D in women and 71% accuracy in men.
As reported by medicalxpress.com, lead author Abir Elbeji from the Luxembourg Institute of Health stated, “Current T2D screening methods are often time-consuming, invasive, lab-based, and costly. Integrating AI with voice technology could make testing more accessible and remove these barriers. This research represents the initial step towards using voice analysis as a scalable, first-line screening tool for type 2 diabetes.”
About half of the 240 million adults with diabetes globally are unaware they have the condition, as symptoms can be vague or absent—approximately 90% of these cases are type 2 diabetes. Early detection and intervention are crucial to prevent severe complications, making the reduction of undiagnosed T2D cases a significant public health goal.
The study aimed to create and evaluate a voice-based AI algorithm to detect T2D in adults. Researchers had 607 participants from the Colive Voice study, both diagnosed with and without T2D, provide voice recordings by reading a set of sentences through their smartphones or laptops.
Analysis showed that individuals with T2D were generally older and had higher BMIs compared to those without the condition. The AI algorithm analyzed vocal features like pitch, intensity, and tone using two advanced methods: one that examined up to 6,000 vocal characteristics and another deep-learning technique focusing on 1,024 key features.
The AI model’s performance was evaluated against established diabetes risk factors, including age, BMI, and hypertension, and compared with the American Diabetes Association (ADA) risk assessment tool. The voice-based algorithms demonstrated solid predictive ability, correctly identifying 71% of male and 66% of female T2D cases. The model was particularly effective for women aged 60 and older and those with hypertension.
The study showed 93% agreement with the ADA risk score, indicating comparable performance to the established screening tool.
Co-author Dr. Guy Fagherazzi of the Luxembourg Institute of Health remarked, “While these findings are encouraging, further research and validation are needed before voice analysis can be adopted as a primary diabetes screening method. Future research will focus on detecting early-stage type 2 diabetes and pre-diabetes.”