Table of Contents
international research, colive voiceAnnounced in European Association for the Study of Diabetes (EASD) 2024 Conferenceindicates that the following patients exist. type 2 diabetes (T2D) have different vocal characteristics compared to healthy controls of the same age and gender. These results “open up the possibility of developing a first-line, non-invasive, and rapid screening tool for T2D that can be achieved using a smartphone or just a few seconds of audio recording during the consultation,” said the study's principal investigator. Diabetes epidemiologist Guy Fagerazzi explained. said in an interview with the French version of Medscape at the Luxembourg Institute of Health.
How did the idea of detecting diabetes through audio come about?
During the COVID-19 pandemic, we began analyzing audio recordings of patients with chronic diseases. We wanted to find a solution to assess people's health remotely, without physical contact. We quickly realized that this approach could be extended to other diseases. My main research interest has always been diabetes, so I looked at how speech characteristics correlated with diabetes. Previous research has shown that people with diabetes have a unique voice compared to the general population, and this insight served as a starting point.
What mechanisms could explain why T2D patients have different voice characteristics?
It is difficult to identify a single factor that explains why T2D patients' voices differ from those without diabetes. Several factors come into play.
Several biological mechanisms, particularly those affecting the vascular system, influence symptoms in people with metabolic diseases such as diabetes. For example, people with T2D experience cardiorespiratory fatigue more often. obesity Being overweight is also an important factor. Because these conditions can slightly change the vocal parameters compared to a person of normal weight. high blood pressureIt is a common symptom in T2D patients and adds further complexity.
Neurological complications can affect the nerves and muscles involved in voice production, especially the vocal cords.
Therefore, respiratory fatigue, neuropathy, and other conditions such as dehydration and acid reflux, which are common in diabetics, may contribute to voice differences.
These differences may not be noticeable to the human ear. Therefore, we are often unaware of the relationship between voice and diabetes. However, advances in signal processing and artificial intelligence techniques have made it possible to extract large amounts of information from these subtle changes. By analyzing these small differences, diabetes can be detected with some accuracy.
Your research states that tone of voice can indicate diabetic status. Could you please tell me more details?
Yes, tone of voice can be affected, but it is a complex and multidimensional phenomenon.
People who have had diabetes for more than 5 to 10 years tend to have a harsher voice than people of the same age and gender without diabetes. In our research, we were able to extract many speech features from the raw audio signal. Therefore, it is difficult to isolate one particular element that stands out.
Are there differences in voice changes between patients whose diabetes is well-controlled and those whose disease is uncontrolled?
As the duration of diabetes increases, the roughness of the voice tends to increase. This is more pronounced in people with poorly controlled diabetes. Our hypothesis, based on results presented at the EASD conference, is that fluctuations in blood sugar levels, both hypoglycemia and hyperglycemia, can cause short-term changes in speech. Although we haven't seen it yet, there are also many subtle and rapid changes that could be detected. We are currently conducting additional research to investigate this.
Why do you ask participants to read a passage from “?'' universal declaration of human rights?
We used a highly standardized approach. Participants completed several recordings, including holding an “ah-ah” sound as long as possible in one breath. They also read a passage that helped them better differentiate between diabetics and non-diabetics. This method is slightly better than other sounds typically used for disease analysis. We chose this particular text in the participants' native language because it is neutral and does not cause emotional fluctuations. Colive Voice is an international multilingual study, so we use official translations from different languages.
Your research focuses on T2D. Are you also planning to study type 1 diabetes (T1D)?
We believe that T1D patients also exhibit voice changes over time. However, our current focus is on T2D, as our goal is to develop large-scale screening methods. T1D is usually diagnosed in childhood, but requires different screening approaches. So far, our research has focused primarily on adults.
Were there any gender differences in the accuracy of voice analysis?
Yes, voice research generally shows that women's voices have different characteristics than men. Part of the reason is due to hormonal fluctuations that affect pitch and tone. Detecting differences between healthy people and people with diabetes can be more difficult for women, depending on the condition. In our study, the accuracy for women was approximately 70%, compared to 75% for men.
EASD results focus on U.S.-based populations. When can we expect data from France?
We started in America because we were able to quickly collect a large number of patients. We are currently expanding into global and language-specific analyses. French data is certainly a priority and we are working on it. It takes just 20 minutes to join and contribute to innovative research in non-invasive diabetes detection. Participants can sign up at: www.colivevoice.org
Research overview: Colive Voice
The Colive Voice study analyzed audio recordings of participants speaking for 25 seconds using their smartphones or laptops. The algorithm was trained and validated separately for men and women and evaluated for accuracy, specificity, sensitivity, and area under the curve (AUC).
The study included 323 women (162 with T2D, 161 without T2D) and 284 men (142 with T2D, 142 without T2D). Participants with T2D were generally older and more often obese than participants with T2D.
Two AI techniques were used to analyze various audio characteristics such as pitch, intensity, and tone. One method captured up to 6,000 detailed audio features, and the other focused on a refined set of 1,024 key features using deep learning.
The algorithm showed promising overall prediction accuracy, with an AUC of 75% for men and 71% for women. It correctly predicted that 71% of men and 66% of women had T2D. The model performed even better in women over 60 years of age (AUC, 74%) and hypertensive patients (AUC, 75%).
Diabetes epidemiologist Guy Fagerazzi heads the Deep Digital Phenotyping Laboratory and Precision Health Department at the Luxembourg Health Research Institute. His research focuses on integrating new technologies and digital data into diabetes research. He declared that he has no relevant financial relationships.
This story has been translated from Medscape French version We use several editing tools, including AI, as part of the process. A human editor reviewed this content before publication.
