Supporting diabetes patients in an uncertain world

Kyungbok Lee wants to make a model that helps patients with diabetes compare options for what to eat – even without perfect data.

Living with diabetes can often mean making a lot of complex decisions to ensure your blood sugar levels stay in a safe range.

At lunchtime, for example, in addition to thinking about what to eat for lunch, you may also have to think about what you had for breakfast and when, what you might have for dinner and when, how much exercise you've gotten, when you last took insulin or another medication, and what your most recent blood glucose level was measured at.

Navigating the complexity of these choices every day can be both complicated and exhausting. And many of the tools that try and help people make those decisions aren't useful in every scenario. For example, you can buy a cookbook of "diabetes-friendly" recipes with detailed nutritional and portion information—but what happens when you go out to eat and the restaurant doesn't know how many grams of carbohydrates are in each dish?

Kyungbok Lee, a post-doctoral researcher at the UNC Gillings School of Global Public Health, is now working on a project to help adolescents with Type 1 diabetes make decisions about what to eat.

He's developing a statistical model that, one day, could form the basis of an app or program where patients could predict their future blood glucose levels based on what and when they eat. Users would be able to input potential lunch options, for example, like having a slice of pizza now or a salad in half an hour. The model would then output estimates for how each of those choices would affect the user's blood glucose levels over the next few hours, allowing them to compare the options.

Through this work, Lee is also trying to help answer a bigger and even more complex question in healthcare, with applications far beyond diabetes or nutrition: How do you design tools that help patients and clinicians make decisions when real-world data are messy or incomplete and uncertainty remains, while being honest about what those data can and cannot support?

Lee studied mathematics and statistics at Seoul National University and came to UNC in 2024. Currently, he's working in the lab of Michael Kosorok, a biostatistician at Gillings and Deputy Director of the Carolina Pragmatic Trials Laboratory (CAPSL) at the North Carolina Translational and Clinical Sciences (NC TraCS) Institute.

Lee says his work aims to "bridge the gap" between theoretical statistical ideas and real-world clinical problems. Nowadays, researchers have developed a variety of analytical tools, like statistical models or machine learning models, that could help inform patient care, scientific studies, or public health policy. But these tools may need large, high-quality datasets to be useful, and clinicians and researchers often only have small, uneven, or imperfectly measured datasets with missing information.

Lee is working to design a model that works with these limitations—and uses data that will be realistically available to clinicians and researchers. Currently, he's focused on tracking and understanding how blood glucose levels change over time. He's using data from patients with continuous glucose monitors (sensors that provide a 24/7 estimate of blood glucose levels) and analyzing how those glucose levels fluctuate in relation to things like meals, exercise, and insulin doses.

By understanding those fluctuations, the model could predict how blood glucose levels change in response to different food choices. But to really make this a powerful tool, you also want to make sure the model is still useful when the data it has isn't perfect—because this kind of model will likely not always have perfect data to work with.

For one, if you want a complete record of someone's day, that person will need to spend a lot of time inputting data on meals, insulin, physical activity, and other parts of daily life.

"It's kind of annoying, inputting every single behavior of your everyday life," Lee says.

What's more, even if people are willing to spend all that time inputting data, they might miscalculate exactly how much sugar, protein, fiber, and fat is in a slice of pizza, or how many calories they burned on a walk. People may also purposely skip some data, like if they don't want their doctor to know they had a bowl of ice cream.

Lee's challenge is to build a model that can still provide useful decision support, even with incomplete or inaccurate data. In addition, the model must understand that people aren't always going to want to choose the healthiest possible option for a meal.

"If the model just keeps recommending that people eat vegetables all the time, then people won't follow the recommendation," Lee says.

Lee is thinking about different ways to strike this balance. For one, he's designing the model so the user can assess a variety of options for meal choices and times. This is closer to how people make decisions in real life than simply assessing how eating one type of food, right now, will affect their blood glucose levels. In addition, Lee is consulting with diabetes experts to ensure his model is built on a strong clinical foundation, and to understand why certain kinds of data may be missing, hard to measure accurately, or less reliable in real-world applications.

And at the end of the day, Lee says, human judgment is still vital—and this model is not meant to make decisions for anyone. Instead, he says, the model can simply give patients and providers more insight into various options, so they make a choice with their own judgment and intuition.

"What we're trying to do here is provide rigorous evidence, so that when people are trying to make a decision, the model can help them make that decision."

"You're not trying to invade the expertise of a clinician or a nutritionist," Lee says. "What we're trying to do here is provide rigorous evidence, so that when people are trying to make a decision, the model can help them make that decision."

While the model Lee is developing is currently just a research idea, the work he's doing now could eventually help create a flexible and adaptable tool. Some people, for example, may want a model that minimizes the risk of their blood glucose levels dropping too low, while others want a model that helps their blood glucose levels stay as stable as possible, he says.

And on a larger scale, by figuring out how to make less-than-perfect datasets more useful, Lee's work could help patients and providers apply cutting-edge statistical tools to all kinds of real-world healthcare challenges, beyond just diabetes and nutrition—because many real-world healthcare challenges are just a little messier and more disorganized than the textbooks might have you believe.

"A lot of cool statistical or machine learning ideas look great in textbooks, where the assumptions are cleaner and the data are easier to work with," Lee says. "Real clinical data are often a lot messier than that."