Predicting a person’s glycemic response to a meal or snack is a key part of the successful management of both type 1 (T1D) and type 2 diabetes (T2D); yet, due to the complexity of the factors involved, it can require expensive and invasive testing such as blood draws and stool samples.
In research published in the Journal of Diabetes Science and Technology, Samantha Kleinberg, PhD, Farber Chair Professor of Computer Science at the Stevens Institute of Technology in Hoboken, New Jersey, and colleagues showed that they could predict how the body would react to a meal “without the need for either personalized training data or invasive physiological data.”
They leverage data from two datasets — one from the United States with 397 individuals with T1D, and one from China including 100 people with T2D — to gather standard demographic, physiological, food consumption, and long-term glucose data using continuous glucose monitoring and detailed food diaries.
The researchers’ prediction model, developed using a machine learning algorithm designed for categorial data, achieved accuracy comparable to previous work that relied on invasive test results. Switching from macronutrient to food categories when assessing food consumption further improved accuracy.
“Our findings suggest that personalized nutrition could be used widely in diabetes management and may not require expensive or invasive data to obtain sufficient accuracy,” the team reported.
To find out more about why it’s so important to assess glycemic responses, the role of menstrual cycles on the variability of individual responses, and the findings’ implications for diabetes management, Medscape Medical News spoke with Kleinberg.
Why is it so important to predict glycemic responses to meals in people with diabetes?
For people with diabetes, predicting glycemic response is crucial. First, it can help to support daily glucose management by helping people choose foods that minimize increases in glucose. Second, it can help guide manual insulin dosing more accurately than, say, carb counting. The predictions could also be integrated into automated insulin delivery systems to make their predictions more accurate.
Do we know why there is such a large degree of variation in glycemic responses between individuals? What is driving it?
This is a key question, and we cannot yet fully answer it. In all models so far, there seems to be a ceiling on how accurately glycemic response can be predicted. However, we found two sources of variation: Menstrual cycle phase and time of day. This helps explain why someone might have different responses to a meal on different days. We suspect that there may be other factors such as stress or sleep that play a role, but we did not have the data to look at those yet.
How does menstrual cycle affect glycemic responses?
Hormones play a role in insulin sensitivity. We saw this reflected in the data with higher average glucose during the perimenstrual phase and in higher glycemic responses to meals repeated during this phase. More work is needed to understand potential individual differences in this area, but this suggests it should play a role in diabetes management.
What do the invasive microbiome data offer that made these appear necessary for predicting glycemic responses?
Microbiome has been a focus of prior works as it is linked to digestion and metabolism. However, it is expensive and requires stool samples, so it is not always practical to collect. We wanted to see how well we could predict meal responses without it and were surprised to find we could do pretty well. In ongoing work, we are now testing whether microbiome remains informative once food type is included.
Why do you think food categories led to improved accuracy vs macronutrients?
Food categories provide more information than macronutrients alone. They can tell us something about other nutrients like fiber and characteristics like processing level that also make a big difference on glycemic impact.
What are the implications of your results for diabetes management?
We are hopeful that our work will enable personalized nutrition to be a bit more accessible and scalable. It could be used to tailor dietary plans, guide insulin dosing, and potentially reduce overall insulin needs.
The main limitation is that currently an individual would need to track their meals to get a prediction of their response. We are working on ways to make that less cumbersome.
The research was supported by the National Institutes of Health and National Science Foundation. No relevant financial relationships were declared.
