To diagnose either type 2 diabetes or pre-diabetes, clinicians typically rely on a lab value known as HbA1c. This test captures a person’s average blood glucose levels over the previous few months. But HbA1c cannot predict who is at highest risk of progressing from healthy to prediabetic, or from prediabetic to full-blown diabetes.

Now, scientists at Scripps Research have discovered that artificial intelligence can use a combination of other data—including real-time glucose levels from wearable monitors—to provide a more nuanced view of diabetes risk.

The new model, described in Nature Medicine, uses continuous glucose monitor (CGM) data alongside gut microbiome, diet, physical activity and genetic information. It flags early signs of diabetes risk that standard HbA1c tests may miss.

“We showed that two people with the same HbA1c score can have very different underlying risk profiles,” says co-senior author Giorgio Quer, the director of artificial intelligence and assistant professor of Digital Medicine at Scripps Research. “By bringing in more data—how long glucose spikes take to resolve, what happens to glucose overnight, what the food intake is, and even what’s happening in the gut—we can start to tell who’s on a fast track to diabetes and who isn’t.”

“Ultimately, the goal of this work is to get a better understanding of what is driving diabetes progression and how we can intervene earlier in the clinic,” adds co-senior author Ed Ramos, the senior director of digital clinical trials at Scripps Research.

While some variation in blood sugar is completely normal—especially after eating—frequent or exaggerated glucose spikes can be a sign that the body is struggling to manage sugar effectively. In healthy individuals, blood sugar typically rises and falls smoothly. But in people at risk of diabetes, these spikes can become sharper, more frequent or slower to resolve, even before routine lab tests like HbA1c pick up a problem. The new study shows that tracking these day-to-day dynamics provides a much more detailed view of a person’s metabolic health, and might help identify trouble earlier.

The findings are a result of a multi-year, digital research program called the PRediction Of Glycemic RESponse Study (PROGRESS). The study used social media outreach to enroll more than 1,000 people from across the U.S. in a fully remote clinical trial. Participants included people with diagnoses of either pre-diabetes or diabetes, as well as healthy individuals.

For ten days, they wore Dexcom G6 CGMs, tracked their meals and exercise, and sent in samples of their blood, saliva and stool for testing. The researchers also had access to participants’ electronic health records, which included previous lab values and diagnoses made by medical practitioners.

“This was a really pioneering effort in the remote clinical trial space,” says Ramos. “We had to design a study that participants could complete entirely on their own—from applying sensors to collecting and shipping biological samples—without ever visiting a clinic. That level of self-guided participation required a completely different kind of infrastructure than usual.”

Using the data, the researchers trained an AI model to distinguish people with type 2 diabetes from healthy individuals.

One of the clearest signals of diabetes risk that the researchers found was the time it took for a blood sugar spike to return to normal. In people with type 2 diabetes, it often took 100 minutes or more for blood sugar to decrease after a spike, while healthier individuals returned to baseline much faster. The study also found that people with a more diverse gut microbiome and higher activity level tended to have better glucose control, while a higher resting heart rate was linked to diabetes.

Importantly, the AI model didn’t just detect risk in people with already elevated HbA1c. When applied to pre-diabetic individuals, it found that some looked metabolically similar to those with diabetes, while others resembled healthy individuals, despite having similar lab values. This level of granularity could help clinicians personalize treatment—focusing on lifestyle changes or early therapies for patients with the highest risk of disease progression.

While the current study was a snapshot in time, the researchers are continuing to follow participants to see whether the model’s predictions translate to real-world disease progression. They also validated the model using a separate set of patient data from Israel, strengthening its potential for broader clinical use.

The team envisions future versions of the model being used by clinicians, or even individuals using CGMs at home, to assess metabolic risk and monitor how daily choices affect diabetes.

“Ultimately, this is about giving people more insight and control,” says Quer. “Diabetes doesn’t just appear one day—it builds slowly, and we now have the tools to detect it earlier and intervene smarter.”