In a paper in the Journal of Diabetes Science and Technology, researchers at Stevens Institute of Technology offer a new approach: a data-sparse model capable of accurately predicting individual glycemic responses with no need for blood draws, stool samples, or other unpleasant testing. The key to their approach? Keeping track of what people actually eat.

“It might sound obvious, but until now most research has focused on macronutrients, such as grams of carbohydrates, instead of the specific foods that people are eating,” explains Dr. Samantha Kleinberg, Farber Chair Professor of Computer Science. “We’ve shown that by analyzing food types, it’s possible to make highly accurate predictions with far less data.”

Dr. Kleinberg’s team studied two datasets that include both detailed food diaries and continuous glucose monitor data for almost 500 people with diabetes (both types 1 and 2) based in the United States and China. Using existing food databases and ChatGPT, they classified each meal according to macronutrient content and also leveraged the structure of foods (so meats are more similar to each other than to cheeses), enabling them to differentiate between nutritionally equivalent foods.

By training an algorithm using both nutritional data and food features, plus a few demographic details, the team was able to predict each individual’s glycemic response to each food with virtually the same levels of accuracy found in prior studies that included detailed microbiome data and other hard-to-collect information.

“We still don’t know why including the food features makes such a big difference,” Dr. Kleinberg says. It’s possible that food information is a proxy for micronutrients that drive glycemic responses, or that the physical properties of certain foods lead people to eat or digest them differently. “What’s clear, though, is that when it comes to blood sugar, there’s more at work than just macronutrients,” Dr. Kleinberg says.

By focusing on food types, the team was also able to explore individual variations in glycemic responses. “Because people eat the same meals again and again, the data gives us visibility into the way that individual responses to specific foods change over time,” Dr. Kleinberg explains. The team found that including data about menstrual cycles in their model accounted for much of the intra-subject variation, suggesting that shifting hormone levels could play an important role in mediating individual glycemic responses.

The team’s model also accurately predicts glycemic response for both U.S. and Chinese populations — an important finding, since microbiome-based models often struggled to deliver accurate results across different cultural contexts. “We don’t need data on a specific regional population to be able to make predictions there,” Dr. Kleinberg explains.

The new model is also powerful enough to predict an individual’s glycemic responses based on demographic data, without customized training on food logs or other personalized data. As a result, clinicians could potentially use the model to offer nutritional advice during an initial meeting with a patient, without the need for laborious food logging or intrusive testing. “We can offer better recommendations if we have more data, but we can get very good results with no personalized information at all,” Dr. Kleinberg explains. “That means we can give patients useful advice right away — and hopefully that will motivate them to keep going.”

Next, the team plans to refine their model using larger datasets, and to explore whether adding microbiome data increases their model’s accuracy. “That’s the big question, because if food information alone gives us everything we need, there might be no need to collect stool samples or do other tests,” Dr. Kleinberg says. “That could make personalized nutrition more affordable and accessible for everyone.”