Continuous Glucose Monitoring Meets AI to Decode Diabetes Complexity

Researchers used machine learning and glucose monitor data to uncover unique diabetes subtypes, paving the way for tailored treatments and better outcomes.

Study: Prediction of metabolic subphenotypes of type 2 diabetes via continuous glucose monitoring and machine learning. Image Credit: Ljupco Smokovski / Shutterstock

Diabetes has long been lumped into two categories—Type 1, which often appears in childhood, or Type 2, which is associated with obesity and typically develops later in life. However, scientists have learned that not all patients with Type 2 diabetes are the same, with some patients varying in body weight, age of onset, and other characteristics.

Now, researchers from Stanford Medicine have developed an artificial intelligence-based algorithm that uses data from continuous blood glucose monitors to parse four of the most common Type 2 diabetes subtypes: muscle insulin resistance, β-cell dysfunction, hepatic insulin resistance, and impaired incretin action.

"It's a tool that people can use to take preventative measures. If the levels trigger a prediabetes warning, for instance, dietary or exercise habits could be adjusted," said Michael Snyder, PhD, a professor of genetics who co-led the study. Snyder is the Stanford W. Ascherman, MD, FACS Professor in Genetics.

Roughly 13% of the U.S. population, some 40 million people, have been diagnosed with diabetes, and 98 million have prediabetes, so a widely accessible technology that pinpoints diagnostic details would be a game changer for diabetes care, Snyder said.

"The majority of people with diabetes have Type 2, and they're just called 'Type 2,'" said Tracey McLaughlin, MD, a professor of endocrinology. "But it's more complex than that, and there are different underlying physiologies that lead to the condition."

There's been a growing movement to subclassify Type 2 diabetes, which accounts for 95% of all diabetes, to better understand the risk of having other related conditions, such as cardiovascular, kidney, liver, or eye complications, and to identify the underlying physiology of individuals' diabetes. "This matters, because depending on what type you have, some drugs will work better than others," McLaughlin said. "Our goal was to find a more accessible, on-demand way for people to understand and improve their health."

The technology could have been helpful for Snyder years ago when he learned he was prediabetic. "When I found out I was on my way to becoming diabetic, I increased my muscle mass, which is one of the common ways to help decrease sugar in the blood, but it had no effect. That's because I'm not traditionally insulin resistant," he said. His variety of Type 2 diabetes stems from something called a beta cell deficiency, which means that the cells that produce insulin don't function the way they should.

A recent study validates the algorithm's predictive power for different subtypes. It demonstrated an area under the curve (AUC) of 95% for muscle insulin resistance, 89% for β-cell dysfunction, and 88% for impaired incretin action.

A paper detailing the research was published in Nature Biomedical Engineering on Dec. 23. McLaughlin and Snyder are co-senior authors. Ahmed Metwally, PhD, a former postdoctoral scholar at Stanford Medicine who is now a research scientist at Google, is the lead author.

Delineating Details of Diabetes

Currently, diagnosing diabetes is based solely on the glucose level in the blood and can be made through a simple blood draw. "But those tests reveal little about the biology underlying high blood sugar," McLaughlin said. "Understanding the physiology behind it requires metabolic tests done in a research setting, but the tests are cumbersome and expensive and not practical for use in the clinic."

However, continuous glucose monitors, available over the counter, can test for high blood sugar and compile more detailed information about the physiologic subtypes driving glucose dysregulation.

Insulin, a hormone made in the pancreas, regulates glucose levels, or sugar, in the bloodstream by encouraging cells to absorb it and use it as energy. If the pancreas does not make enough insulin, known as insulin deficiency, blood glucose rises. Insulin resistance, a standard diabetes marker, occurs when cells don't respond to the cues from insulin, which also results in a buildup of blood glucose.

Type 2 diabetes can also result from a defect in the production of incretin, a hormone released by the gut after eating that stimulates insulin secretion from the pancreas, or by insulin resistance in the liver. Each of these four physiologic subtypes of diabetes might require distinct therapies, emphasizing the need for precision medicine approaches.

Testing the Algorithm

McLaughlin and Snyder wondered whether an everyday gadget like a continuous glucose monitor could produce data with hidden signals relating to the different subtypes of diabetes. The device, which users attach to their upper arms, measures the rise and fall of blood sugar levels in real-time. People who drink glucose drinks often show a spike in their blood glucose, but the level and pattern of those spikes vary from person to person.

In a study of 56 participants in three cohorts, including 21 with prediabetes, researchers applied an artificial intelligence-powered algorithm to identify patterns within peaks and dips that corresponded to different subtypes of Type 2 diabetes. The study incorporated a rigorous validation process, including both clinical and at-home settings.

Participants who used the continuous glucose monitors also underwent the oral glucose test performed at a doctor's office. "People have looked at that for decades and have found certain parameters that indicate insulin resistance or beta cell dysfunction, which are the main drivers of diabetes," McLaughlin said. "But now we have the monitors, and you can get a much more nuanced picture of the glucose pattern which predicts these subtypes with greater accuracy and can be done at home."

The CGM-based tool detected these subtypes accurately, which exceeded traditional methods like HOMA-IR.

Broadening Accessibility

In addition to higher-resolution data for people with diabetes or prediabetes, using the monitor has other perks. "Even if a person with insulin resistance doesn't develop diabetes, it's still important to know," McLaughlin said, "because insulin resistance is a risk factor for a variety of other health conditions, like heart disease and fatty-liver disease."

McLaughlin and Snyder plan to continue testing the algorithm with people diagnosed with Type 2 diabetes. They hope that the technology's broad availability will boost access to care, even when patients cannot make it to a doctor's appointment.

"We also see this technology as valuable health care tool for people who are economically challenged or geographically isolated and can't access a health care system," McLaughlin said.

Stanford's Department of Genetics and the Department of Medicine supported the work.

The National Institutes of Health funded this study (grants R01 DK110186-01, U01-DK105535, U01-DK085545, UM1DK126185, and 2T32HL09804911), the Stanford PHIND center, the Stanford Diabetes Research Center, the Wellcome Trust, Stanford Lifestyle Medicine and the American Diabetes Association.