Artificial intelligence is creating new opportunities to study metabolic health by allowing researchers to analyze large-scale physiological and biological data, identify patterns of metabolic dysfunction, and develop approaches for earlier disease detection. 

 

Ahmed Metwally, PhD, has built his research career at this intersection of computer science and metabolic health. Dr. Metwally is a staff research scientist at Google, where he leads the Metabolic Health AI research program. His research focuses on developing models that leverage large-scale physiological and behavioral data to enable the early detection and personalized treatment or prevention of cardiometabolic diseases. He holds a PhD in biomedical engineering and an MS in computer science, both from the University of Illinois at Chicago, and completed his postdoctoral work at Stanford University’s Snyder Lab. 

 

Dr. Metwally recently spoke with AACE Endocrine AI Editor-in-Chief Johnson Thomas, MD, FSCE, FEAA, about his latest research on the use of AI in metabolic health. In this first video, Dr. Metwally discusses his study published in Nature Biomedical Engineering on using continuous glucose monitoring and machine learning to predict metabolic subphenotypes of type 2 diabetes. 

 

 

Dr. Thomas: Could you give us a brief overview of the study, including the questions you are trying to answer and the main findings? 

 

Dr. Metwally: Type 2 diabetes is a very complex disease. It has big genetic components and also has lifestyle components, but all of that, at the end of the day, affects our physiology, and then we end up having high glucose in our bloodstream. And that's kind of the diagnosis of diabetes. The question is: How do we end up with that phenotype? Is it just one single process or multiple processes? 

 

What we found—or what has actually been known in the field—is that type 2 diabetes can develop through mainly four different phenotypes. We call them subphenotypes. 

 

The first is insulin resistance, and that can be muscle insulin resistance, which means your muscles are not responding to the insulin that's being secreted by the pancreas. The second one is hepatic insulin resistance, which means your liver is not responding to the insulin that is produced by the pancreas. The third, which is probably the most familiar, is beta-cell dysfunction: the pancreas is not secreting enough insulin in the first place.

 

The fourth is incretin deficiency. The incretin hormones, GLP-1 and GIP, are receiving a lot of attention now because of drugs that target them for weight loss. But they are naturally occurring hormones secreted when we eat. As food enters the intestine, GIP and GLP-1 are released and signal the pancreatic beta cells to stimulate insulin secretion.

 

In some people, the pancreatic beta cells are functioning well, and the liver and muscles are insulin sensitive, but the signal is not reaching the pancreas. The GIP and GLP-1 signals are not coming in the first place, and that is what we call incretin deficiency.

 

So those are the four metabolic phenotypes: muscle insulin resistance, hepatic insulin resistance, beta-cell dysfunction, and incretin deficiency. One of the main goals of the study is: Can we study that? Can we quantify these different subtypes in people who may be developing diabetes? 

 

If you select people at random and measure their HbA1c, fasting plasma glucose, or 2-hour oral glucose tolerance test (OGTT) results, many would have normal glycemic levels—meaning they are neither prediabetic nor diabetic. But if you look at insulin resistance, beta-cell dysfunction, or incretin deficiency, you may find that one of these systems is not functioning as it should. That may mean they are on a path toward developing diabetes.

 

And if they are going to develop diabetes, it's probably from this pathway first. If they want to tackle that early on, they probably need to optimize lifestyle that's really focused on one of these subphenotypes. 

 

We found that most people who have muscle insulin resistance are also having hepatic insulin resistance. Although it's not one-to-one, they are very correlated. You can see people who are developing type 2 diabetes—mostly 70% are coming from the insulin resistance system first—but there are people who are developing type 2 diabetes mostly because of pancreatic beta-cell dysfunction. And there are people who completely have incretin deficiency, and those are the people you want to give GLP-1 and GIP. 

 

So, we quantified that. We showed that using gold-standard tests. Those are not scalable tests. Those are not things that everyone can do. You cannot do that in most blood laboratories or clinical laboratories. Those are tests that, for each one of these subphenotypes, cost thousands of dollars to be done, probably need five to six hours to be completed, and have to be done by a specialist endocrinologist to finalize them. Those are the things that need to be done in research units. 

 

Again, we started from the very gold standard. We knew this probably would be a sample size that's not very large. It's also not going to be representative of the whole population. But we wanted to start from the gold standard, from the ground up, to see: Do we see a signal or not? And we've seen the signal. 

 

The next thing is: Can we identify those metabolic subphenotypes from OGTT, which is the test that you do when you drink a glucola drink and then you measure glucose every 30 minutes? What we have done in the research unit is that we tried to measure plasma glucose every 5 to 15 minutes to really mimic the CGM. 

 

You want to have that continuous glucose from plasma because, again, we started from a very gold standard and then relaxed our conditions one at a time. From the glucose from OGTT in the research unit, we were able to predict mainly three of those metabolic subphenotypes very well. That was muscle insulin resistance, beta-cell dysfunction, and incretin deficiency. 

 

Now we can go from thousands-of-dollars tests to maybe a test that can be done with just OGTT. But still, people have to go to a specific place to do the OGTT, and they have to have blood drawn to measure it. The next question would be: Can we actually do that experiment at home with CGM? Meaning that we put CGM on people and then we ask them to drink a glucola drink. So, we send people home with two glucola drinks, and then we designed a specific protocol: Don't exercise the night before. Sleep in your normal routine. Then you wake up in the morning, drink this glucola drink, and just stay still for two to three hours. 

 

Again, that's probably the same test that people would do somewhere else in the clinic, but this is now at your home and with CGM. Because we have the CGM, we can get the glucose values, and we know when people start to drink the glucola drink. We get a curve that's very similar to what we get from the OGTT in the clinic from plasma. Using that, we're also able to predict metabolic subphenotypes. I think we focused on insulin resistance and beta-cell dysfunction. We started with four subphenotypes, and now we are at two at home. And those are the two that—one of them, which is insulin resistance—we scaled even further in other studies. 

 

Dr. Thomas: One thing that you mentioned that's especially interesting is that you're validating an at-home oral glucose tolerance test using CGM. Was it a surprise to you how well the home protocol performed compared to the clinical setting, or did it not really perform compared to that gold standard, in your opinion? 

 

Dr. Metwally: From a classification perspective, it performed very well. From a regression perspective, if you want to predict the actual number of insulin resistance, it's not that accurate. But what we care as a screening tool is whether this person is insulin resistant or not? Also, because it's a small sample size, you don't want to go to a numerical prediction all the way for that. So, we did it as a classification problem first. I think, based on what we have seen, the performance from at home didn't surprise me because we got a glucose curve that's very similar to what we got from the clinic. 

 

I probably forgot to mention that between these two experiments, we did another experiment where people actually put the CGM on, went to the clinic, and did OGTT. So, we got the plasma comparison, and we got the one from the CGM. And I've seen that in 90% of people; they are pretty much similar. The numbers are not exactly the same, but the trend is the same and that's the most important thing that we are using. 

 

So, the performance didn't surprise me. The one thing that surprised me is I felt like we would get different curves at home because people would not follow the protocol. That was one of the hypotheses, but we didn't find that. What we found, for completeness, is that if you do the experiment one day, you may see certain fluctuations. But if they do it for two days, like we did, and we average those curves, that average curve is very similar to the one that we get from the clinic. So, as of now, the one that we put in the paper is two days. Can we do it with just one? Probably, but we probably need to do larger studies to figure out specific signals in the curve itself, even with some fluctuations. 

 

Dr. Thomas: To point to the limitations, like you already mentioned, it's a small study. What limitations should listeners keep in mind when they are reading this paper? 

 

Dr. Metwally: To predict metabolic subphenotypes from CGM, I think as of now it's doable. But with the limitations of the study that has been done—it's a small study, it's also people from the San Francisco Bay Area, mostly people who are normoglycemic and prediabetic—I think the next stage of this study is really to scale it. Do a nationwide study, first maybe in one country, and if we see the same trend, it should actually go globally and get people from completely different demographics, completely different areas, and see: Does that trend, does that performance actually stay the same across different demographics? Do we need just one model that is trained to capture all of these different changes, or do we need different models that are specialized for different demographics?  

 

Those are things that I think need to be done, and I think there are many initiatives to actually tackle this problem. Our study included about 50 to 60 people. To have greater confidence that these findings apply across populations, we need studies of thousands of people using CGM and metabolic subphenotyping.