LSU Research Insights: AI Digital Twins Bring Personalized Diabetes Care Into Daily Life

For millions of people managing diabetes at home, the hardest part isn’t knowing what to do—it’s knowing what to do today.

Dr. Shiyu Li is an assistant professor in the LSU School of Kinesiology in the College of Human Sciences & Education and also director of the Health and Behavioral Intervention Technologies (HABIT) Lab.

She is working at the intersection of behavioral science, artificial intelligence, and clinical care to close that gap in diabetes management.

Her team is developing a “human-in-the-loop” digital twin—an AI-powered, personalized model that helps patients anticipate how daily choices will affect their health and turn general advice into specific, actionable steps.

“The “human in the loop” is really about trust and safety,” she said. “The AI helps us scale personalization, but the clinicians and researchers ensure that every recommendation is appropriate, safe, and actually doable in real life.”

We asked Li to speak about her work in this area.

Can you tell us a little more about your research and efforts to build a “human-in-the-loop” digital twin for diabetes care? What is the power of using AI and a digital twin in diabetes care for patients at home? Why is a “human in the loop” needed?

This has really been a team effort. We brought together engineers, nursing scientists, a physician, a dietitian, and me, a behavioral scientist, to make sure the system is not only technically sound but also clinically meaningful and realistic for patients’ daily lives.

At its core, the system is a personalized model for each patient. It learns their diet and physical activity patterns from wearable data and uses that to predict how today’s lifestyle choices may affect tomorrow’s weight and blood sugar.

Based on that, it sends daily text messages with concrete, next-day guidance: what to eat for each meal, portion sizes, nutrient balance, and step goals.

In our randomized trial, we saw promising results. The AI feedback group lost an average of 5.9 pounds over three months, compared to 3.6 pounds in the control group receiving human-written feedback, while maintaining stable glucose levels. The model achieved about 80% prediction accuracy across patient subgroups.

The “human in the loop” is really about trust and safety. The “human” is not just one person—it’s the whole team. The AI helps us scale personalization, but the clinicians and researchers ensure that every recommendation is appropriate, safe, and actually doable in real life.

What are some of the biggest challenges for diabetes patients in terms of glycemic stability at home? Why is a digital twin helpful for this?

Managing diabetes at home is challenging because blood sugar is invisible and constantly changing. A stressful day, a different meal, or even a missed walk can affect glucose levels, and often patients don’t realize it until after the fact.

“ Instead of thinking, ‘Why did my glucose spike?’ (patients) can start asking, ‘If I make this choice today, what will happen tomorrow?’ That shift toward more proactive decision-making is where a lot of the value lies. ”

On top of that, clinic visits for diabetes care are often months apart. During that time, patients are largely on their own, often with general advice like “eat healthier” or “be more active.” While well-intentioned, that kind of guidance is too broad to be truly helpful day to day.

A digital twin helps fill that gap. It doesn’t just model physiology. The digital twin model uses AI to learn patterns in lifestyle behaviors and how those connect to glucose over time. This information allows patients to move from reacting to problems to anticipating them.

So, instead of thinking, “Why did my glucose spike?” they can start asking, “If I make this choice today, what will happen tomorrow?” That shift toward more proactive decision-making is where a lot of the value lies.

What kind of feedback is helpful for diabetes patients at home, that a digital twin model can provide?

Imagine going home after a clinic visit with a handful of general guidelines—eat healthier, move more, check your glucose—and then not seeing your doctor again for 3 to 6 months. That is the reality for many people living with diabetes.

What we’ve learned is that helpful feedback needs to be specific enough to act on. That’s where a digital twin makes a difference. Because the system understands a person’s patterns, it can provide concrete guidance: what to eat for each meal the next day, in realistic portions and aligned with their dietary needs.

Even activity recommendations are personalized—for example, suggesting a short walk after dinner because it’s likely to help stabilize glucose based on that person’s data.

From a behavioral science perspective, there are three key elements: giving people a clear reason to act, removing uncertainty about how to act, and making sure they can see progress over time.

When those are in place, feedback becomes something people can actually use—not just information, but something that supports real change.

Why is precision care so important for diabetes management? Why is this classically difficult?

Diabetes is fundamentally an individual disease. Two people with the same diagnosis can respond very differently to the same meal, the same medication, or the same exercise routine.

However, the way we deliver lifestyle recommendations today is largely built around population averages. Clinical guidelines are designed for what works for most people—but in reality, the average patient does not exist.

What makes this even more challenging is that variability is not only between individuals—it also exists within the same person. Day-to-day factors like stress, sleep, meal, and physical activity can significantly shift how someone responds, even when everything else seems the same.

So we have a situation where diabetes is highly dynamic and individual, but care is largely static and generalized. That is why precision care is essential but very difficult to achieve.

What kind of nutrition feedback, advice, or guidance can an AI system provide that improves diabetes management?

What our system provides is very different from general nutrition advice. Every day, a patient receives a text message with a concrete next-day meal plan—specific meals for breakfast, lunch, dinner, and snacks, with portion sizes and nutrient balance already worked out for them.

This starts with clinical input. Our dietitian defines the nutritional boundaries to ensure everything is safe and appropriate for the patient’s condition. The system then personalizes within those boundaries, learning from each individual’s glucose responses over time.

But what the system recommends is only part of the picture. How it communicates that recommendation matters just as much. We ground our messaging in psychological and behavioral theories—supporting autonomy, building confidence, and creating a continuous feedback loop based on recent patterns.

What excites me most is the potential to extend access. Right now, this level of personalized guidance typically requires frequent clinical support. The goal is to make that kind of support available to patients in their daily lives, not just during clinic visits.

Do you foresee this eventually being used by all people to improve precision nutrition, beyond just diabetes patients?

Yes, I do. While diabetes is our starting point, the underlying technology is not specific to any one condition—it’s about understanding how an individual’s body responds to food, activity, and daily behaviors. Diabetes is a natural place to begin because the need is immediate and the outcome glucose is continuously trackable, but the same principles apply much more broadly.

There is also a significant opportunity to improve access. Today, having access to highly personalized nutrition guidance is still a privilege. It often requires specialized care and resources, which are not available to everyone. Approaches like ours have the potential to make that level of insight more widely accessible.

From a broader perspective, the most exciting aspect is prevention. Many common chronic conditions—such as Type 2 diabetes, cardiovascular disease, dementia, cancer, and obesity—are strongly influenced by daily lifestyle behaviors.

Tools that help individuals better understand their own responses to daily lifestyle behaviors could support earlier, more informed decisions and, ultimately, help reduce the risk of developing these conditions.

What is the most rewarding aspect of your work/research? What do you wish more people knew about your work/research/field?

The most rewarding aspect of this work is seeing how something as abstract as an AI model or behavioral theory can translate into small, everyday decisions that actually improve someone’s health.

In diabetes care, it’s rarely one big decision that matters. It’s the accumulation of many small choices throughout the day. Being able to support those moments in a practical way is very meaningful.

As a behavioral scientist, I also wish more people understood that these decisions don’t happen in isolation. They are shaped by the environments we live in, the options available, and the constraints people face. So simply giving more information is often not enough.

That’s the focus of our work in the HABIT Lab at the School of Kinesiology. We study how to design digital tools make healthy lifestyle behaviors easier and more sustainable.

Whether it’s through a digital twin or other approaches, the goal is to support behavior change in a way that people can enjoy and maintain. Ultimately, it’s about moving from telling people what to do to helping them actually do it.