In the near future, you can just record a 10-second voice clip on your smartphone to check if you have type 2 diabetes or not. A team of scientists from a Toronto-based research facility has developed an AI (artificial intelligence)-based screening model to carry out voice (acoustic) analysis of a six-to-10-second audio clip to detect whether the speaker has high blood glucose levels.
The AI model analyzes the acoustic variations in the voice sample along with some basic health details like age, sex, weight and body mass index (BMI). It then correlates these factors to the vocal anomalies induced by high blood glucose levels to make the diagnosis.
“In simple terms, we are looking at the features of voice as an instrument and how they may change in disease. Like the notes on a piano, the voice has a pitch, a strength or amplitude and a stability. Jitter represents temporal perturbation [temporary deviation] in frequency, while shimmer deals with variability of amplitude. These features are too subtle to be perceived by the human ear but are ‘loud and clear’ for a machine,” says Jaycee Kaufman, first author of the paper and research scientist at Klick Labs, Toronto, Canada, in an e-mail response to Happiest Health.
The researchers have also pointed out that voice-enabled diabetes detection has the advantages of being non-invasive, cost-effective and convenient since it can be installed as an app on your smartphone.
The research team has claimed in their recent study published in the journal Mayo Clinic Proceedings that AI-enabled diabetes prediction comes with an accuracy rate of 89 percent and 86 percent in women and men, respectively. According to them, the traditional diabetes tests, like the glycated hemoglobin (HbA1c), fasting blood glucose (FBG) and oral glucose tolerance (OGT) tests have accuracy rates of 91, 85 and 92 percent, respectively.
Voice as a biomarker for type 2 diabetes
The study points out that a sustained period of high blood glucose levels could affect the elastic properties of vocal cords, adding that chronic diabetes could also cause peripheral neuropathy and myopathy, potentially leading to nerve and muscle damage in the vocal cords. “Type 2 diabetes is associated with complications such as peripheral neuropathy, muscle weakness and edema, which have been previously shown to affect the voice. Peripheral neuropathy may damage the nerves in the larynx, resulting in hoarseness or vocal strain, and muscle weakness may be apparent in the muscles of the vocal cords or respiratory system. In addition, the swelling associated with edema may affect the elastic and vibrational qualities of the vocal cords, which could affect the pitch,” explains Kaufman.
She also adds that voice analysis technology is a rapidly growing field with a variety of applications in healthcare and medicine. Several studies have been conducted where voice and speech have been analyzed for predicting neurodegenerative conditions like Alzheimer’s and Parkinson’s.
“The production of voice is a complicated process that involves the combined effects of the circulatory system, respiratory system, muscular system, nervous system and other systems in the body. Anything that affects these systems may have an effect on the voice, which was the motivation for this work,” notes Kaufman.
Voice analysis for detecting diabetes
According to the researchers, the study had 267 participants (including 97 women) with no history of smoking or speech disorders. They were further diagnosed as nondiabetic or type 2 diabetic as per the standards prescribed by the American Diabetes Association (ADA).
Each individual was asked to self-record two phrases: ‘Hello, how are you? What is my glucose level right now?’ six times a day for two weeks. There were as many as 18,465 voice recordings, which were then analyzed to detect any anomalies according to each participant’s overall health condition, which had already been fed into the AI model.
The vocal variations of type 2 diabetes
The study mainly focused on 14 vocal features, including voice frequency, shimmer, harmonic-to-noise ratio, voice turbulence index, pitch and jitter, in the recordings before coming to a conclusion on whether the speaker has type 2 diabetes. Kaufman simplifies this to Happiest Health by stating that the main vocal features associated with the diabetes screening were the pitch and strength of the voice, as well as the associated variability of these features.
The study also stated that voice variations triggered by diabetes are gender-specific. “In women, the pitch and the variability of the pitch were affected, whereas in men, the strength and the variability of the strength were affected. We believe this difference may stem from the fact that men and women experience the complications of type 2 diabetes differently, which ultimately impacts the voice differently. Specifically, men may experience more muscle weakness associated with type 2 diabetes, whereas women may experience more edema,” elaborates Kaufman.
Future of voice analysis
As per the data from the International Diabetes Federation (IDF), there are as many as 537 million people with diabetes in the world. A recent ICMR-INDIAB study published in The Lancet stated that there are as many as 101 million people with diabetes in India, with another 236 million living with prediabetes.
Yan Fossat, vice president of Klick Labs and the principal investigator of the study, tells Happiest Health that the next objective is to come up with a quicker version of the model, which could be used for screening pre-diabetes as well. “We are hoping to get a better idea of how early we are able to detect voice changes,” he adds.
Takeaways
Voice analysis is now emerging as a biomarker for the early diagnosis of various conditions. According to a recent research study, voice modulations and variations were able to indicate if the speaker had type 2 diabetes or not when analyzed with factors like age, body weight, BMI, height, etc. High blood glucose levels affect the nerves and muscles of the larynx, which causes distinct yet inaudible variations in vocal patterns.
