An artificial intelligence-powered voice analysis method demonstrated potential as a noninvasive screening tool for type 2 diabetes, according to research presented at the European Association for the Study of Diabetes (EASD) Annual Meeting in Madrid.
The study found that the artificial intelligence (AI) model detected type 2 diabetes (T2D) with 66% accuracy in women and 71% accuracy in men using brief voice recordings combined with basic health data.
Methods
Researchers analyzed voice recordings from 607 adults participating in the Colive Voice study. Participants, both with and without diagnosed T2D, provided approximately 25-second voice recordings reading a few sentences of provided text via smartphone or laptop. The study collected demographic and health data including age, sex, body mass index (BMI), and hypertension status.
Two AI techniques were employed. The first was a method for capturing up to 6,000 vocal characteristics. The second was a deep-learning approach focusing on 1,024 key features. The AI algorithm analyzed various vocal features, including changes in pitches, intensity, and tone, to identify differences between individuals with and without diabetes.
Additional Results
Women with type 2 diabetes had a mean age of 49.5 versus 40 years in patients without diabetes. The mean BMI was 35.8 versus 28.0 kg/m², respectively. Men with type 2 diabetes had a mean age of 47.6 versus 41.6 years in patients without diabetes. The mean BMI was 32.8 versus 26.6 kg/m², respectively. Improved performance was seen in women aged 60 years and older and in patients with hypertension.
There was a 93% agreement with the American Diabetes Association (ADA) T2D risk assessment tool The performance of the best models was grouped by several diabetes risk factors including age, BMI, and hypertension, and compared to the ADA tool for T2D risk assessment.
Discussion
"This study is the first step towards using voice analysis as a first line, highly scalable T2D screening strategy," said lead author Abir Elbeji from the Luxembourg Institute of Health. The study's findings were significant given that approximately 240 million adults worldwide have undiagnosed diabetes, with T2D accounting for about 90% of these cases. Around half of adults with diabetes are unaware of their condition due to general or non-existent symptoms.
Additional Data and Statistics
The study used an average of 25 seconds of participants' voices along with basic health data. Participants with T2D were generally older and more likely to be living with obesity compared to those without T2D.
The voice-based algorithms showed good overall predictive capacity. The model's performance was particularly strong in females aged 60 years or older and in individuals with hypertension.
Co-author Guy Fagherazzi, MSc, PhD, HDR, noted next steps in research will specifically target early-stage T2D cases and pre-diabetes. While the findings were promising, further research and validation would be necessary before the approach could potentially become a first-line diabetes screening strategy.
The authors declared no conflicts of interest.