Latest News

Vocal Diabetic Clues Could Help Improve Dismal Diagnostic Rate

By Deborah Borfitz

November 2, 2023 | A private biomedical engineering and biomarker research lab in Canada has succeeded in measuring what type 2 diabetes does to the body as manifested through the voice and it is “opening a new lens” to the identification of at-risk individuals and potentially the assessment of therapeutic interventions, according to Yan Fossat, vice president of Toronto-based Klick Labs. In their latest published study using smartphone-recorded voice segments predictive of the disease, Klick researchers were fascinated to discover that predictive accuracy was tied to pitch-related acoustic features for women and intensity-related acoustic features for men, which corresponds to the differing complications they experience. 

The team is now putting final touches on the protocol for a planned prediabetes study to see if the same sex differences can be seen with the precursor health condition, he says. Vocal biomarkers of hypertension are separately being investigated that could be combined with those for diabetes or prediabetes to increase the precision of the prediction methodology.

Thanks to machine learning and the ubiquity of cellphones, the promise of digital health is starting to come to fruition in ways that scientists “could only dream of” a decade ago, says Fossat. As described in an article that published recently in Mayo Clinic Proceedings: Digital Heath (DOI: 10.1016/j.mcpdig.2023.08.005), the Klick Labs team built an algorithm around 14 acoustic features previously linked with type 2 diabetes that are highly predictive of disease risk. 

In an age- and body mass index (BMI)-matched dataset, the model had 75% accuracy for women and 70% accuracy for men. But when applied to an unmatched dataset—reflective of what is seen in real-world scenarios—the figures were respectively 89% and 86%," notes Klick research scientist Jaycee Kaufman.  

That means it outperforms the gold-standard blood glucose screening test, with an accuracy of about 85%, says Fossat. “We see this largely as a screening tool; it is not meant to be a diagnostic. It’s an algorithm that ingests voice and spits out a score of probability of a person having diabetes.” 

The research builds on a prior study proposing that mathematical modeling be used in lieu of gold standard screening methodologies that produce discrepant results across patients of different backgrounds and may be unreliable when evaluating how glucose changes over time (npj Digital Medicine, DOI: 10.1038/s41746-020-0283-x). Earlier this year, the Klick Labs also published results from a registered study highlighting the mathematical approach in the classification of glucose homeostasis (Mayo Clinic Proceedings: Digital Health, DOI: 10.1016/j.mcpdig.2023.02.008). 

Sex-Based Differences

The predictive model described in the latest study was trained on six- to 10-second recordings from people with and without type 2 diabetes—together with data on their age, sex, height, and weight—before being used on 267 participants in India who were part of the larger study looking at the relationship between voice and glucose control. Subjects were asked to record a phrase into their smartphone six times daily for two weeks, producing more than 18,000 recordings for analyzing differences between the cohorts with and without the disease. 

Thickening of the vocal cords has been associated with diabetes. Nerve and muscle degeneration, both of which affect the voice, are also among long-term complications of the disease, says Kaufman. 

In terms of disease prediction, the metrics of variation that mattered most in the voice related primarily to pitch for females and intensity for males, she adds. The reason for the discrepancy is not entirely clear. 

But investigators have some ideas. Diabetes is probably not a “one-size-fits-all” problem based purely on blood glucose level, although it is treated as such, Fossat says. “We’re talking about entirely different features for men and women.” 

The sex-based differences are reflected in the disease experience, which for men is highlighted by muscle weakness and atrophies and for women by edema or higher water content in the body, notes Kaufman. These two distinct complications are reflected in the voice, respectively, as a lowering of intensity or pitch. 

Exploding Field

The use of vocal biomarkers in disease risk prediction is an exploding if relatively new field that began as an effort to improve the detection of neurodegenerative disorders such as Alzheimer’s and Parkison’s disease, says Fossat. Klick Labs is attempting to expand their use to diabetes and hypertension as well as other cardiometabolic diseases and women’s health issues. 

“Almost anything in the body that has some propensity to affect the voice is an area of study for us,” he adds. Coupled with machine learning, voice biomarkers can add considerable speed to the conduct of biomedical research.   

Klick Labs also hopes to get involved in clinical trials where the effects of diabetes treatment are compared in men and women based on the vocal pitch and intensity biomarkers, says Fossat. “The current blood biomarkers just measure glucose or A1C; they don’t necessarily measure what this is doing to your body.” 

Current detection methods involve a trip to the doctor’s office, which helps explain why almost one in two adults (240 million people) living with diabetes worldwide aren’t even aware they have the condition, says Kaufman, citing statistics from the International Diabetes Foundation. About 90% of the diabetic cases are type 2 diabetes. 

How any sort of screening app makes its way to market will depend on the reason people in a particular locale are going undiagnosed, Fossat says. Klick Labs might variably be presenting its technology to consumers, healthcare providers, and government agencies.  

In Canada, the app might function as a self-screening tool, much like the Canadian Diabetes Risk Questionnaire (CANRISK), says Fossat. It was developed by the Canadian Task Force on Preventive Health Care in partnership with the Public Health Agency of Canada and is used by citizens to get a score letting them know how likely they are to have type 2 diabetes. “In other countries where access is a big problem, it could be something used much more as a medical screening tool.”