A program that uses easily available data and no more than 10 seconds of speech is capable of identifying whether someone has diabetes seven times out of eight, a study has found. Better still, it should soon be possible to get it as an app on any smartphone, providing a cheap and accessible option for people with limited access to medical facilities.
A team at Klick Labs had 267 people who had recently undergone standard testing for Type 2 diabetes record a short phrase on their phones six times a day for two weeks. They then searched for acoustic differences between those who had tested positive and negative.
Combining the presence or absence of identified features in the voice prints, and the participants' age, sex, height, and weight, an artificial intelligence (AI) model predicted the individuals’ status. It proved 86 percent accurate for men, and 89 percent for women.
“Voice synthesis is a complex process that relies on the combined effects of the respiratory system, the nervous system, and the larynx. Anything that affects these systems can influence the voice,” the researchers write. Most people may not be able to identify these changes, certainly not reliably, but computers can perform more subtle analyses.
The most powerful predictive tools were pitch and the variation in pitch between times the phrase was recorded. Some other methods added to the accuracy of the predictions for only one sex, with “perturbation jitter” (no, we don’t know what it means either) being predictive for women and “amplitude perturbation quotient shimmer” (ditto) for men.
Long-term high blood sugar can damage the peripheral nerves and muscle fibers, which can produce voice disorders, previous studies have shown. Even temporary spikes in blood glucose, which diabetes can exacerbate, have been proposed as affecting the elasticity of vocal cords, although this has yet to be proven. More indirectly, anxiety and depression can alter people’s voices, and diabetes may contribute to both.
“Our research highlights significant vocal variations between individuals with and without Type 2 diabetes and could transform how the medical community screens for diabetes,” said first author Jaycee Kaufman in a statement. “Current methods of detection can require a lot of time, travel, and cost. Voice technology has the potential to remove these barriers entirely.”
Diabetes rates are skyrocketing worldwide, creating one of the most serious health crises the world faces, and exacerbating many other problems. Late diagnosis is a major contributor, since many interventions that prevent, or greatly mitigate, diabetes’ consequences come too late.
The International Diabetes Federation estimates almost half the adults with diabetes are not aware of their status. Inevitably, the proportion is highest in the countries with healthcare systems least able to cope with what will come. The authors of the study quote an estimate that undiagnosed diabetes will cost the world $2.1 trillion per year by 2030.
Several tests are used to identify Type 2 diabetes, but all require collecting blood samples, which some people avoid, and visiting healthcare providers and laboratory testing.
“Voice technology could revolutionize healthcare practices as an accessible and affordable digital screening tool,” said co-author Yan Fossat.
Klick Labs intends to repeat the study with a larger sample, as well as seeking ways to improve the accuracy further and exploring whether detection can include length of time being diabetic. They are also interested in extending the work to other conditions including prediabetes and high blood pressure.
The study is open access in Mayo Clinic Proceedings: Digital Health.