So good morning, I'm Marcelo Corrêa, an endocrinologist at Universal of Iowa. I want to thank Dr. Moose for the most amazing lecture, comprehensive. I think we all learned a lot today about the present and the future. So now I want to open the session for questions. And actually, I saw one question on the chat, but now I cannot see it. OK. So first question, how can we harmonize information from different medical systems in order to reduce biased data impurity? Right, so first off, thank you, Marcelo, and thank the organizers from Endocrine Society. We need much more data, and much more data coming from diverse systems, right? So it means working outside of the confines of your hospital system and the hospital system, et cetera. So I think that we're certainly going to have to look at better secure methods of preparing to preserve the data, but there are multiple methods that we can use to bring that data together. That's outside of today's scope and discussion. But certainly, the question's great and spot on, because really to make sure that each model that we're using is not just addressing a specific patient population in which it was trained upon, even after you do your validation steps, et cetera. We really need to take this prospectively. Different algorithms need to be tested on multiple patient populations. So that really is what is so far lacking in probably most of the studies that we discussed this morning, is taking it into different populations in which things were trained on. So I think that's going to be one of the first major and most comprehensive steps that we can take. Another question here. What is your experience in combining neurofeedback data with metabolic data in the view of genetic and microbiome differences? Well, you know, I'm, I don't know exactly how often that's been done in a study, but I think the sentiment here is multimodal data, right? And I think the sentiment here is it's not just genetic data. It's not just going to be biofeedback data, but it's going to be a multiple layers that will all fit together, right? And I think that we're now in an era in which our platforms can actually start to understand and start parsing through a micro, you know, microbiome data, genetic data, biofeedback data, labs, longitudinal EHR data sets, and the like. I know that there's been several studies that I showed today did take multiple layers in terms of developing their algorithms and platforms. But I, you know, I'm uncertain how many of those have incorporated biofeedback to date. But the sentiment is, the sentiment is spot on for, you know, a massive multimodal data sets. Okay, so I have a question. So if you have to put your money and bet on one single advance in nutrition science utilizing AI in the next 10 years, what would be the major development? I think it's going to be, can I hedge and go between two things, Marcelo? Absolutely. You know, I'm gonna, I think it's likely going to be now that we're seeing CGM data from two major players, reaching the masses through consumer aspects, right, as you don't need a doctor to get this, you're building up more data sets, you're understanding how your exercise, your sleep, your stress, each individual part of your diet affects your glucose fluxes, right? I think that's very exciting now to think of eating all the data to understand, all the people, you know, right now we have a huge amount of data from diabetics with CGMs or pre-diabetics with CGMs, but what about the pre-pre-diabetics or the folks that go on to develop other metabolic disease, right? And so I would want to know what does the, what is, what are the glucose levels look like on a hourly minute, whatever basis in those individuals, two years, three years, five years before they ultimately develop disease. So we can find out what the preventative steps are to hopefully intervene at that point. So that's one thing I'm very much excited about is the, the introduction of CGM to, you know, the masses. Number two, I think it's going to be in proteomics. As we start to understand, you know, specialized signaling molecules and other very important aspects of our physiology there combined with, you know, the other layers in which we've talked about. I think it's really what we've seen for proteomics and then continuous glucose monitoring. That's going to be fed in to take things really to the next level. Thanks. So I have two more questions here. So Rama is thanking you for the great presentation. Hold on just a minute. Let's address first Yusa's question. Do fellowship training programs will include AI development for certain topics? And secondly, how this will be incorporated in medical training? Yeah. I love that. The more we can get our medical trainees, even from the medical school level, resident fellowship, et cetera, involved in understanding how artificial intelligence works, how machine learning works, how these algorithms where, you know, are put together, where are their faults, right? Then we will be much better able to hopefully weed out ones that are actually not great, right? We are kind of the backstop for patient safety, right? To understand, you know, what's worth certain risks. And so I know that within our cardiology fellowship program, probably, you know, once every two months or so as part of our weekly journal clubs, we're talking about a new AI paper, whether it's looking at echocardiograms, ECG, CT scans, and the like, you know, there's been so much moving in the field. And so we not only look at these studies from, all right, what are the end points? You know, what's it trying to tell me about, you know, human health, but what are the methods involved in this? And I think that to understand the methods at each step, you know, again, allows us to make sure that we're doing something safely and we can start to understand where these tools, because that's what they are, start to break down. The same reason that when folks are doing their echocardiography training, they need to understand basic ultrasound technique, basic wave theory, because if you don't understand how that data is coming in and what's being done with it, then again, we fail to troubleshoot when things go awry. So I think it needs to happen early, early in training, now more than just a one week, I mean, more than just a one hour session that one could have in the evening as an interest group or whatnot, very important, and hopefully fellowships, residency programs, and medical training programs will start to incorporate this more vastly. Now, Roma asks, when do you think large language models will be able to integrate with EHR and become standard of care? Yeah, so, I mean, there's already several institutions that are bringing, you know, these large language models inside their safety net, right? I know several that have piloted these programs. I think one of the things that's going to be important is making sure that patients know when a large language model is being utilized as part of the recommendation, right? Or creating MyChart messages that are just automatically weeding through the information and helping to save time, but also providing valuable information to the patient. So I think there needs to be caveats to patients when a large language model is being used. And we saw several studies today where patients were notified, okay, this was created with assistance of large language models, and there is some skepticism these days, but I think that we'll see more and more of it happening. And it's going to likely provide better quality care, but we've seen this already happen in multiple institutions. I think that more and more as the roadmap is created for IT in smaller hospitals, medium size, et cetera, we'll see it fast, just like we did see with EHR. Great. Raul asks, what is the current regulation on clinical use and how can it improve point of care screenings? So right now, regulation is spotty, right? The EU has their own regulation. United States has its own regulation, right? I think that currently things are slow to roll out, right? I feel that I showed today a whole lot of stuff that we should be using now that we're not, just because of the uncertainty involved in the medical legal aspects or the compliance aspects, right? And again, that pathway hasn't been paved as solidly as the science is now, right? So I think that the technology and the insight that we're getting from these models is much further than the current use based on hesitancy from compliance and legal, right? I think that having models be able to eat EHR data and tell you, all right, especially our folks on statin therapy, folks on the right statin therapy, what about certain hypoglycemic agents, are there others that should be considered given a patient, you know, kidney function or whatnot? Yeah, I think that models will be able to read through our notes as well as our EHR and help us make sure we're not missing key guidelines. However, what we do know is physicians hate pop-ups, okay? And so if you start creating a clinical day for me where it's just giving me, oh, you haven't done this and you haven't done this, did you consider this? And it's a bunch of hard pop-ups, then it creates friction within that patient encounter and with our own workflow. So I think that's certainly going to have to be worked out moving forward. So I think we have time for maybe two more questions. Here, Trian asks, what about improvements in normal to exceptional performance in sports and science from an endocrinology point of view? Hmm, so improvements in, maybe you can say something. Probably sports medicine. My understanding is how AI can improve maybe performance. And one obvious answer is like the nutrition aspect of it, right? But there are aspects that you might think are important. Yeah, and I think we see this from, you know, a high-level consultant. Companies that are working with, you know, athletes and folks from an early age, folks that are, you know, probably getting VO2 max done in a laboratory setting multiple times a year, right? Really understanding precision nutrition. Folks that are tracking, you know, all their calories in, types of foods in, amount of physical exertion, heart rate, heart rate variability, sleep time, you know, all of that time, et cetera. Again, you know, it all depends on what the data inputs are and what the outputs you're looking for. So, you know, for the quantified selfers, which we saw early on in digital medicine, you know, those folks with all the rings, all the trackers, you know, they had CGMs before everyone else did. You know, I think that they're already starting to utilize this from a health and wellness perspective, as well as in the sports side of things. So, I don't think we have more time. Maybe the organizers can send more questions to you and hopefully you can respond to those questions. So, Dr. Mews, thank you very much for a great talk and great insight about this, you know, wonderful technologies. So, the future is here. Session is adjourned. Thank you.

AI in medicine

continuous glucose monitoring

medical training programs

regulatory landscape

precision wellness