An international study, Colive Voice, presented at the European Association for the Study of Diabetes (EASD) 2024 conference, shows that patients with type 2 diabetes (T2D) have different voice characteristics compared with healthy controls of the same age and gender. These results “open up possibilities for developing a first-line, noninvasive, and rapid screening tool for T2D, feasible with just a few seconds of voice recording on a smartphone or during consultations,” explained the study’s principal investigator Guy Fagherazzi, a diabetes epidemiologist at the Luxembourg Institute of Health, in an interview with the Medscape French edition.

How did the idea of detecting diabetes through voice come about?

During the COVID-19 pandemic, we began analyzing voice recordings from patients with chronic diseases. We wanted to find solutions to assess people’s health remotely, without physical contact. We quickly realized that this approach could be extended to other diseases. Because my main research focus has always been diabetes, I looked into how voice characteristics might correlate with diabetes. Previous studies had indicated that patients with diabetes have distinct voices compared with the general population, and this insight formed the starting point.

What mechanism could explain why patients with T2D have different voice characteristics?

It’s challenging to pinpoint a single factor that would explain why patients with T2D have different voices from those without diabetes. Several factors are involved.

Some biological mechanisms, especially those affecting the vascular system, influence symptoms in people with metabolic diseases such as diabetes. For example, people with T2D have more frequent cardiorespiratory fatigue. Obesity and overweight are also key factors, as these conditions can slightly alter vocal parameters compared with people of normal weight. Hypertension, common in patients with T2D, adds to the complexity.

Neurologic complications can affect the nerves and muscles involved in voice production, particularly the vocal cords.

SUGGESTED FOR YOU

Therefore, respiratory fatigue, neuropathies, and other conditions such as dehydration and gastric acid reflux, which are more common in patients with diabetes, can contribute to differences in voice.

These differences might not be noticeable to the human ear. That’s why we often don’t notice the link between voice and diabetes. However, technological advancements in signal processing and artificial intelligence allow us to extract a large amount of information from these subtle variations. By analyzing these small differences, we can detect diabetes with a reasonable degree of accuracy.

In your study, you mention that voice tone can indicate diabetic status. Could you elaborate?

Yes, voice tone can be affected, though it’s a complex, multidimensional phenomenon.

Patients who have had diabetes for 5-10 years, or longer, tend to have a rougher voice than those without diabetes of the same age and gender. In our study, we were able to extract many voice characteristics from the raw audio signal, which is why it’s difficult to isolate one specific factor that stands out.

Is there a difference in voice changes between patients with well-managed diabetes and those whose disease is uncontrolled?

The roughness of the voice tends to increase with the duration of diabetes. It’s more noticeable in people with poorly controlled diabetes. Our hypothesis, based on the results we presented at the EASD conference, is that fluctuations in blood sugar levels, both hypo- and hyperglycemia, may cause short-term changes in the voice. There are also many subtle, rapid changes that could potentially be detected, though we haven’t confirmed this yet. We’re currently conducting additional studies to explore this.

Why did you ask participants to read a passage from the  Universal Declaration of Human Rights?

We used a highly standardized approach. Participants completed several recordings, including holding the sound “Aaaaaa” for as long as possible in one breath. They also read a passage, which helps us better distinguish between patients with and those without diabetes. This method works slightly better than other sounds typically used for analyzing diseases. We chose this particular text in the participant’s native language because it’s neutral and doesn’t trigger emotional fluctuations. Because Colive Voice is an international, multilingual study, we use official translations in various languages.

Your research focuses on T2D. Do you plan to study type 1 diabetes (T1D) as well?

We believe that individuals with T1D also exhibit voice changes over time. However, our current focus is on T2D because our goal is to develop large-scale screening methods. T1D, typically diagnosed in childhood, requires different screening approaches. For now, our research mainly involves adults.

Were there any gender differences in the accuracy of your voice analysis?

Yes, voice studies generally show that women have different vocal signatures from men, partly owing to hormonal fluctuations that affect pitch and tone. Detecting differences between healthy individuals and those with diabetes can sometimes be more challenging in women, depending on the condition. In our study, we achieved about 70% accuracy for women compared with 75% for men.

The EASD results focused on a US-based population. When can we expect data from France?

We started with the US because we could quickly gather a large number of patients. Now, we’re expanding to global and language-specific analyses. French data are certainly a priority, and we’re working on it. We encourage people to participate — it takes only 20 minutes and contributes to innovative research on noninvasive diabetes detection. Participants can sign up at www.colivevoice.org

Study Overview: Colive Voice

The Colive Voice study analyzed voice recordings from participants speaking for 25 seconds using their smartphones or laptops. Algorithms were trained and validated separately for men and women, assessing accuracy, specificity, sensitivity, and the area under the curve (AUC).

The study included 323 women (162 with and 161 without T2D) and 284 men (142 with and 142 without T2D). Participants with T2D were generally older and more frequently had obesity than those without the condition.

Two AI techniques were used to analyze various vocal characteristics, such as pitch, intensity, and tone. One method captured up to 6000 detailed vocal features, and the other focused on a refined set of 1024 key characteristics using deep learning.

The algorithm showed promising overall predictive accuracy, with an AUC of 75% for men and 71% for women. It accurately predicted 71% of men and 66% of women with T2D. The model performed even better among women aged 60 years or older (AUC, 74%) and patients with hypertension (AUC, 75%).

Guy Fagherazzi, a diabetes epidemiologist, heads the Deep Digital Phenotyping laboratory and the Department of Precision Health at the Luxembourg Institute of Health. His research focuses on integrating new technologies and digital data into diabetes research. He has declared no relevant financial relationships. 

This story was translated from the Medscape French edition using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.