We've already seen numerous ways in which AI is helping speed up medical diagnoses, and it now has a new trick: being able to spot type 2 diabetes in someone, based on just a few seconds of them talking.
Hold a conversation while emotional, intoxicated, or sleepy, you might notice one's voice can be affected by a host of biological factors; a fact that opens up the opportunity for AI to spot subtle changes that might develop with shifts in health.
We should note that the study was carried out by scientists from Klick Labs, who have a vested interest in developing and selling this AI detection technology. However, their findings have been published in a peer-reviewed journal, and are well worth taking a look at to see if type 2 diabetes detection can be improved.
"Current methods of detection can require a lot of time, travel, and cost," says Jaycee Kaufman, a research scientist at Klick Labs. "Voice technology has the potential to remove these barriers entirely."
The team asked 267 participants – some with type 2 diabetes, and some without – to record a fixed phrase six times a day into a phone app, for a period of two weeks. A total of 18,465 recordings were then processed to extract 14 different characteristics from the vocals, including pitch and intensity.
The researchers used a set of these recordings to train the AI on what a person's voice sounds like, based on factors like their sex, age, BMI, and whether they had type 2 diabetes or not. They used the remaining samples to test what the AI had 'learnt'.
Factoring in considerations like age and sex, the model was able to spot type 2 diabetes to an accuracy level of 89 percent for women and 86 percent for men.
Interestingly enough, the key vocal signals that identified type 2 diabetes were different for men and women. In men, intensity and amplitude variation were most important; in women, variations in pitch were the main giveaway.
The researchers admit larger and more diverse groups of people need to be tested to validate these results, but the early findings are positive. Right now, diagnosing type 2 diabetes requires blood to be taken, followed by a lengthy wait for analysis and a report. This method requires little more than access to a smartphone app.
While 1 in 11 adults worldwide have been diagnosed with the condition, researchers think hundreds of millions of people don't know they're living with type 2 diabetes. Being able to reduce that number would also mean being able to put treatments in place earlier, and reducing the costs of managing diabetes in the population.
"Our research highlights significant vocal variations between individuals with and without type 2 diabetes and could transform how the medical community screens for diabetes," says Kaufman.
The research has been published in Mayo Clinic Proceedings: Digital Health.