Developing artificial intelligence models that extract clinically relevant information from continuous glucose monitoring (CGM) data remains challenging. Glucose patterns reflect overlapping physiological and behavioral influences, recordings often contain gaps, and the clinical assessments needed to train models for specific metabolic tasks can be difficult to obtain.

A new study from Google Research, reported in a preprint under peer review, explored whether a foundation model called GlucoFM could learn reusable patterns from CGM recordings without clinical labels, then apply that information across different metabolic research tasks.

Researchers developed GlucoFM to analyze slower glucose trends and short-term fluctuations separately before combining them. In retrospective evaluations, the model outperformed comparison models on average across different metabolic classification tasks. It also improved predictions of glucose responses to meals when combined with recent glucose measurements, meal nutrition, and patient information.

The findings support further investigation of a shared modeling approach for multiple metabolic tasks, although the study did not assess longitudinal clinical outcomes, treatment response, or prospective clinical use.

In his conversation with AACE Endocrine AI Editor-in-Chief Johnson Thomas, MD, FSCE, FEAA, Ahmed Metwally, PhD, of Google’s Metabolic Health AI research program, discusses how GlucoFM works and how it compares with other prediction methods.

(The following interview transcript has been edited for length and clarity.)

Dr. Thomas: As endocrinologists, we use CGM every day, so this research is very exciting. Can you summarize it?

Dr. Metwally: We developed a foundation model for CGM. In this case, that means condensing a large amount of CGM data and looking at how the compressed representation can reflect a person's physiology.

If you have a good representation from the CGM data, you can use it to predict different metabolic states and potentially achieve better performance than if you use the original CGM data by itself.

And because you're learning from many people with different demographics, age, and phenotypes, that representation should be generalizable enough instead of having to use different models. That's why we call it a foundation model.

What we have done is go beyond CGM as any other time series data where you get the CGM data and pass it into a transformer to get a compressed representation.

Before we pass the CGM data into the transformer, we incorporate what we know about the physiology. We know that your glucose levels may sometimes be steady. At other times, there may be a lot of fluctuations. When they’re steady, it's probably related to your state. Are you insulin resistant? Do you have diabetes or not? When glucose fluctuates, it's probably related to an event. Maybe you are eating something specific. Maybe you are stressed. Maybe you are doing hard exercise.

What we have done is take that signal and separate it into two parts—the one that represents state and the one that represents events. And we did that using a signal processing filter. So now you have two different signals, and you combine them again and then pass the combined signal to the transformer. This gives the model physiologically meaningful information.

So, when you get the embedding—the compressed representation—it’s a richer representation than if you just used the original CGM data.

We tested this on prediction for insulin resistance, diabetes risk, β-cell dysfunction, and hyperlipidemia. Across the tasks evaluated, you get better performance than other models that treat CGM data as normal time series data.

The other application we've tested is prediction of the postprandial glycemic response. To do that, you use prior information like CGM data, the nutrients and calories of the meal you are going to eat, your demographics, and your state. We've seen that this information, plus the glucose embedding, shows the best performance.

Those are the major applications we tested, and we focused on the ones that we think will have some clinical utility. What we're doing with GlucoFM is also completely passive compared with other prediction methods, like going to a clinic to have plasma taken or being put on CGM, or doing at-home CGM with a strict testing protocol.

We take one day of CGM data and don't ask you to do anything. We do GlucoFM, get the embedding, and then use that to predict insulin resistance. You can see how many years it's been to go from where we started to where we are now.