The study involved 607 adults with and without type 2 diabetes mellitus. They were asked to read several sentences from the proposed material and make 25-second voice recordings.

The researchers also collected information about the health status, age, gender, body mass index (BMI) and blood pressure of the volunteers. This was required to develop and test an AI model capable of detecting diabetes mellitus.

The AI algorithm analyzed various vocal characteristics such as pitch, intensity and tone of voice to identify differences between people with and without diabetes. The researchers noted that two advanced methods were used in the development: one allowed for up to six thousand detailed vocal characteristics, and the other, a more complex deep learning approach, focused on 1024 basic ones.

Voice algorithms have demonstrated good overall predictive ability, correctly detecting the presence of diabetes in men in 71% of cases, and in women in 66% of cases. It is noted that the model coped with the task best by analyzing the voice recordings of women aged 60 years and older, as well as people with hypertension.

The scientists noted that, despite the promising results, the development requires improvement. They hope that in the future, the AI algorithm they have created can be used for early diagnosis of diabetes mellitus.