A new artificial intelligence system improved blood sugar predictions for people with type 1 diabetes by including information about fats and proteins, not just carbohydrates. In a study of 351 adults analyzing over 15,600 meals, adding fat and protein data reduced prediction errors by about 1 mg/dL in the first 2 hours after eating. According to Gram Research analysis, the improvements were modest but consistent, suggesting that more detailed meal composition tracking could help refine diabetes management tools.
Researchers developed an artificial intelligence system that predicts blood sugar levels up to 4 hours after eating for people with type 1 diabetes. The study analyzed data from 351 adults and over 15,600 meals to understand how different nutrients, not just carbohydrates, affect blood sugar. According to Gram Research analysis, including information about fats and proteins alongside carbohydrates improved the AI’s ability to forecast glucose levels, especially in the first 2 hours after eating. This could help people with type 1 diabetes better manage their insulin doses and prevent dangerous blood sugar swings.
Key Statistics
A 2026 study of 351 adults with type 1 diabetes analyzing 15,601 meals found that including fats and proteins in an AI glucose prediction model reduced prediction errors from 21.02 mg/dL to 20.15 mg/dL in the first 2 hours after eating.
Research published in JMIR AI in 2026 showed that distinguishing between simple sugars and complex carbohydrates, combined with fat and protein data, improved blood glucose forecasting accuracy in the 2 to 4 hour period after meals, reducing errors from 34.73 mg/dL to 32.65 mg/dL.
A 2026 analysis of continuous glucose monitoring data from 351 people with type 1 diabetes found that artificial intelligence models used nutritional information differently depending on prediction timing, relying more heavily on carbohydrates in the first 2 hours and increasingly on fats and proteins thereafter.
The Quick Take
- What they studied: Can an artificial intelligence system predict blood sugar levels better if it knows about all nutrients in a meal, not just carbohydrates?
- Who participated: 351 adults with type 1 diabetes who wore continuous glucose monitors. The study analyzed data from over 15,600 meals these participants ate.
- Key finding: Adding information about fats and proteins to the AI model improved its ability to predict blood sugar levels. In the first 2 hours after eating, the model’s prediction errors dropped from 21.02 to 20.15 mg/dL when fats and proteins were included.
- What it means for you: This research suggests that people with type 1 diabetes might get better blood sugar predictions if they track more than just carbohydrates. However, this is early research, and the improvements are modest. Talk to your doctor before changing how you manage your diabetes.
The Research Details
Researchers used a machine learning system called a temporal fusion transformer, which is a type of artificial intelligence designed to predict future values based on patterns in data over time. They trained this AI system using real data from 351 adults with type 1 diabetes, including continuous glucose monitor readings, insulin doses, and detailed meal information.
The AI was taught to predict blood sugar levels up to 4 hours after eating. The researchers tested different versions of the AI: one that only knew about carbohydrates in meals, and others that also included information about fats, proteins, sugars, and complex carbohydrates. They split their predictions into two time windows: early predictions (0 to 2 hours after eating) and late predictions (2 to 4 hours after eating).
To understand how the AI was making its decisions, the researchers used a technique called integrated gradients. This method shows which nutrients the AI considered most important when making predictions at different times after eating.
Current diabetes management tools mainly focus on carbohydrates when predicting blood sugar changes. However, fats and proteins also affect how quickly and how much blood sugar rises after eating. By testing whether including these nutrients improves predictions, this research addresses a real gap in how diabetes management tools work. If the AI can predict blood sugar more accurately, people with type 1 diabetes could make better decisions about insulin doses.
This study used real-world data from a large group of people, which is a strength. The researchers tested their AI system carefully and used interpretability methods to understand how it worked. However, the study is observational, meaning it shows patterns in data but cannot prove that nutrients directly cause blood sugar changes. The improvements in prediction accuracy were modest, ranging from about 1 to 2 mg/dL. The researchers note that external validation with new data is needed before this tool could be used in clinical practice.
What the Results Show
When the AI model included information about carbohydrates, fats, and proteins, it made better predictions in the first 2 hours after eating. The prediction error (measured as mean absolute error) improved from 21.02 mg/dL to 20.15 mg/dL. This means the AI’s guesses were about 1 mg/dL more accurate on average.
For predictions made 2 to 4 hours after eating, a different combination worked best: the AI performed better when it knew about complex carbohydrates separately from simple sugars, plus fats and proteins. In this later time window, the prediction error dropped from 34.73 mg/dL to 32.65 mg/dL.
The AI used nutritional information differently depending on when it was making predictions. In the first 2 hours, the model relied heavily on carbohydrate information. As time passed, the model gave more weight to fats and proteins. This pattern makes sense because fats and proteins are digested more slowly than carbohydrates, so they influence blood sugar more in the later hours after eating.
When the researchers looked specifically at predictions made right at meal time, they found similar patterns. The improvements were consistent across different ways of measuring prediction accuracy.
The study found that distinguishing between simple sugars and complex carbohydrates was helpful for later predictions (2 to 4 hours after eating) but not as important in the first 2 hours. This suggests that the type of carbohydrate matters more as time passes after a meal. The AI’s interpretability analysis showed that the model learned to weight nutrients differently across the prediction horizon, suggesting it captured some real patterns about how nutrients affect blood sugar timing.
Most current diabetes management systems focus primarily on carbohydrate counting because carbohydrates have the fastest and largest effect on blood sugar. This study builds on previous research suggesting that fats and proteins also matter by showing that including them in an AI prediction model can improve accuracy. However, the improvements are modest, which aligns with the fact that carbohydrates remain the dominant factor in blood sugar control. This research doesn’t contradict existing diabetes management approaches but suggests they could be refined.
The study has several important limitations. First, the improvements in prediction accuracy are small, ranging from 1 to 2 mg/dL. While statistically meaningful, these differences may not be clinically significant for all patients. Second, this is a retrospective study using existing data, so it cannot prove that nutrients directly cause blood sugar changes. The AI learned patterns in the data, but these patterns don’t necessarily reflect the actual physiology of how nutrients affect blood sugar. Third, the study included only adults with type 1 diabetes who were already using continuous glucose monitors, so results may not apply to all people with type 1 diabetes. Finally, the researchers emphasize that external validation with new data from different populations is needed before this tool could be used in real clinical practice.
The Bottom Line
If you have type 1 diabetes, continue following your doctor’s current meal management plan. This research suggests that tracking fats and proteins alongside carbohydrates might help, but the evidence is not yet strong enough to recommend changing your routine. Talk to your healthcare team about whether more detailed meal tracking could help you. Confidence level: Low to Moderate. This is early research that needs further testing before it can be recommended for clinical use.
People with type 1 diabetes who struggle with unpredictable blood sugar spikes after meals may find this research interesting. Healthcare providers developing diabetes management tools should pay attention to these findings. People with type 2 diabetes or those who don’t use continuous glucose monitors should not assume these findings apply to them. This research is not relevant for people without diabetes.
If this AI tool were to be developed for clinical use, it would likely take several years of additional testing and validation. Any improvements in blood sugar control would probably be gradual, developing over weeks to months as you learn to use the tool effectively.
Frequently Asked Questions
Does tracking fat and protein help control blood sugar in type 1 diabetes?
Research suggests it may help slightly. A 2026 study found that including fat and protein information improved AI blood sugar predictions by about 1 mg/dL. However, carbohydrates remain the primary factor. Talk to your doctor about whether detailed tracking would benefit your specific situation.
How long after eating does fat affect blood sugar?
According to a 2026 study of 351 people with type 1 diabetes, fats and proteins have more influence on blood sugar in the 2 to 4 hour period after eating compared to the first 2 hours. Carbohydrates affect blood sugar more quickly.
Can artificial intelligence predict my blood sugar better than carb counting?
AI systems that include comprehensive meal data show modest improvements over carbohydrate-only predictions, reducing errors by 1 to 2 mg/dL. However, this research is early stage. Current carbohydrate counting remains the evidence-based standard for diabetes management.
Should I change how I manage my type 1 diabetes based on this research?
Not yet. This is early research that needs external validation before clinical recommendations can be made. Continue following your doctor’s current plan. Discuss these findings with your healthcare team if you’re interested in more detailed meal tracking.
Does this research apply to type 2 diabetes?
This study focused specifically on type 1 diabetes, which involves absolute insulin deficiency. The findings may not apply to type 2 diabetes, which involves different physiological mechanisms. Consult your doctor about what applies to your situation.
Want to Apply This Research?
- In addition to logging carbohydrates, track the grams of fat and protein in your meals for 2 weeks. Record your blood sugar readings at 2 hours and 4 hours after eating. Compare patterns between high-fat meals and low-fat meals to see if timing of blood sugar peaks differs.
- When logging meals in your diabetes app, start entering fat and protein amounts alongside carbohydrates. Many apps already have this feature. This creates a more complete picture of your meals and could help you notice patterns in how different foods affect your blood sugar timing.
- Set up alerts in your app to remind you to check blood sugar at 2 hours and 4 hours after meals. Over 4 to 8 weeks, review your data to identify which meal compositions lead to the most stable blood sugar. Share these patterns with your healthcare team to refine your insulin dosing strategy.
This article summarizes research findings and is not medical advice. The improvements in blood sugar prediction shown in this study are modest and the research is early stage. Do not change your diabetes management plan based on this article. Always consult with your healthcare provider or endocrinologist before making changes to how you manage type 1 diabetes, including meal tracking or insulin dosing. This study has not yet been validated in clinical practice, and the findings should not be interpreted as proven physiological effects of individual nutrients.
This research translation is published by Gram Research, the science division of Gram, an AI-powered nutrition tracking app.