Scientists have discovered vocal patterns associated with signs of diabetes

Changes in the body's functioning are not always noticeable to humans and can manifest themselves in a variety of forms. Even the voice, which we are accustomed to perceive only as a means of communication, may contain signs related to a health condition. Researchers are studying whether artificial intelligence algorithms are capable of finding such patterns and using them to detect diseases. The features of the voice associated with type 2 diabetes are described in the Izvestia article.

What was found in the voice recordings

Researchers have studied the relationship between voice characteristics and type 2 diabetes. The results are published in the IEEE Journal of Biomedical and Health Informatics (J-BHI). The researchers analyzed 14,759 voice recordings of 461 adults. The participants recorded short sentences in English on their smartphones, selected at random.

Only voice recordings and the gender of the participants were used for the analysis. Based on them, the researchers trained models separately for women and men. This approach made it possible to search for acoustic patterns associated with type 2 diabetes without having to select specific voice characteristics in advance.

How artificial intelligence analyzed the voice

First, the recordings were converted into spectrograms, visual maps showing sound frequencies and their changes over time. They were then processed by a model that makes it possible to identify complex combinations of acoustic features directly in the audio data.

Unlike the approach in which the algorithm analyzes pre-selected parameters, the new model could take into account the relationships between different frequencies and their changes over time. In particular, we are talking about more complex combinations of characteristics that may remain unnoticeable when studying individual parameters.

Scientists talked about the connection of speech features with brain aging

How accurate were the results?

The balanced accuracy of the model was 71.3% for women and 72.4% for men. This indicator takes into account the algorithm's ability to recognize both people with diabetes and people without the disease, giving both groups the same weight. Another indicator, the area under the ROC curve, reached 0.788 for women and 0.779 for men.

The work continues the 2023 study published in the journal Mayo Clinic Proceedings: Digital Health. Then the scientists identified 14 predefined acoustic characteristics associated with type 2 diabetes. Among them were the pitch and intensity of the voice, as well as the shimmer and jitter — indicators of fluctuations in the characteristics of the audio signal.

Then the usual accuracy of the model was 89% for women and 86% for men. However, these results cannot be directly compared with the new work: the studies differed in terms of input data, analysis methods, model validation methods, and datasets. Therefore, the difference in numbers does not in itself mean that the new model performs worse.

Why voice analysis won't replace blood analysis yet

The authors consider voice biomarkers as a promising area for further study of the relationship between speech and type 2 diabetes. According to Fossat, the researchers found complex combinations of acoustic features that were not the focus of attention when using predefined voice parameters. The team intends to continue studying such patterns.

At the same time, the results show a link between the features of voice recordings and diabetes, rather than confirming that voice can already reliably diagnose the disease. Further verification is necessary for the practical application of this approach. At this stage, the study does not provide grounds to consider voice analysis as a substitute for standard diagnostic methods, including blood tests.