Smartwatch data can be used to assess early diabetes risk

The data your smartwatch already collects could soon help detect a warning sign of type 2 diabetes.

Hidden in the heart rate, sleep and daily activity patterns captured by everyday wearable devices are subtle clues that, when combined with routine health data and analyzed with artificial intelligence (AI), may indicate insulin resistancethe researchers report on March 16 in Nature.

An estimated 20 to 40 percent of U.S. adults live with insulin resistance, which occurs when the body’s cells stop responding properly to the sugar-metabolizing hormone insulin, a key early event in the progression to type 2 diabetes. However, most affected people are unaware of the condition because its diagnosis usually requires specialized tests that are not part of routine medical care. This means that doctors usually detect the problem only after blood sugar levels have already started to rise, at which point metabolic damage may already be underway.

Catching it earlier could open the door to “timely lifestyle interventions,” says David Klonoff, an endocrinologist at Mills-Peninsula Medical Center in San Mateo, Calif., who runs the nonprofit Diabetes Technology Society and was not involved in the research. These include dietary changes, increased physical activity and weight loss, especially through the use of successful products. GLP-1 drugs, all of which have been shown to slow or even reverse the metabolic shift towards disease.

“If we can identify people who are insulin resistant, we can change the entire trajectory of diabetes,” says Ahmed Metwally, a bioengineer at Google Research in Mountain View, California.

Some researchers have proposed using arm-worn sensors to do this. Yet these devices cost hundreds of dollars per month and are used primarily by people who already have diabetes, limiting their usefulness for large-scale screening. In contrast, smartwatch-based approaches rely on devices millions of people already wearsaid Klonoff.

“This study establishes a scalable method… for early detection of metabolic risk,” he says.

The new system, developed by Metwally and colleagues, draws on smartwatch data collected over tens of millions of hours from 1,165 people wearing either Fitbit devices or Pixel watches, both sold by Google or its subsidiaries. Machine learning algorithms sifted through this data, along with routine laboratory measurements such as cholesterol tests and demographic factors like age, to detect patterns linked to insulin resistance.

The most predictive factors came from clinical and demographic data, rather than signals from the smartwatch itself. Using only measurements taken from routine lab tests and basic health data, such as fasting blood sugar, body mass index, and blood lipid count, Google’s model was able to distinguish people with insulin resistance from those without it about 76% of the time.

But performance increased to around 88% with the addition of smartwatch data feeds.

Such readings are not perfectly reliable – sleep estimates, for example, are known to vary in accuracy between devices and users – but even these imperfect signals added useful information to the model. Resting heart rate was particularly informative, although daily steps and sleep duration also contributed to the predictive power.

Ultimately, Metwally imagines a future in which wearable electronics discreetly detect early signs of metabolic disease in millions of people. And others in the field see an equally promising approach.

“This paper convincingly demonstrates that consumer wearable data contains substantial metabolic information relevant to the prediction of insulin resistance,” says Giorgio Quer, director of artificial intelligence at the Scripps Research Translational Institute in La Jolla, Calif., who was not involved in the research.

“The ability to continually, longitudinally, and passively monitor metabolic health through wearable devices, particularly when powered by [AI] models, represents an exciting opportunity towards a more personalized and scalable digital medicine model,” he says.