A new artificial intelligence system using deep learning achieved 64% time in target blood glucose range by automatically adjusting insulin doses based on real-time glucose readings, meals, and activity levels in a study of 12 people with type 1 diabetes. According to Gram Research analysis, this personalized AI approach kept average glucose at 80 mg/dL with prediction errors under 10 mg/dL, suggesting the system learned to adapt insulin dosing to individual patterns better than fixed insulin schedules. However, this is early research requiring larger clinical trials before real-world use.
Researchers developed an artificial intelligence system that learns how each person’s body responds to insulin and adjusts doses in real-time. Using data from continuous glucose monitors, meal information, and activity levels, the AI system kept blood sugar in the safe range 64% of the time—better than traditional fixed insulin schedules that don’t adapt to individual differences. This personalized approach could help people with type 1 diabetes avoid dangerous blood sugar swings while reducing the mental burden of constant dose calculations.
Key Statistics
A 2026 research study published in JMIR Diabetes found that a deep learning AI system maintained blood glucose in the target range (70-180 mg/dL) 64% of the time for 12 people with type 1 diabetes using personalized, real-time insulin adjustments.
The AI-based insulin dosing system achieved a mean glucose level of 80.06 mg/dL with a prediction accuracy error of 9.85 mg/dL, demonstrating effective personalized glucose regulation in the 12-person study.
A reinforcement learning model trained on 8 weeks of continuous glucose monitoring data learned to make insulin dose recommendations that kept blood sugar within clinically safe ranges for the majority of the evaluation period, with a reward score of 10 indicating successful learning.
The Quick Take
- What they studied: Can an artificial intelligence system that learns from experience manage insulin dosing better than traditional fixed insulin plans by adapting to each person’s unique body and lifestyle?
- Who participated: Twelve people with type 1 diabetes who wore continuous glucose monitors and tracked their meals and exercise for 8 weeks. The AI learned from their real glucose patterns, insulin doses, and activity data.
- Key finding: The AI system kept blood sugar in the safe target range (70-180 mg/dL) about 64% of the time, with an average glucose level of 80 mg/dL, suggesting it learned to make personalized insulin adjustments effectively.
- What it means for you: If this technology advances to real-world use, people with type 1 diabetes could have an AI assistant that automatically adjusts insulin doses based on their meals, exercise, and stress—reducing dangerous blood sugar swings and the mental load of constant calculations. However, this is early research and needs testing in larger groups before becoming available.
The Research Details
Researchers created an artificial intelligence system called a Deep Q-Network (DQN) that learns through trial and error, similar to how a person learns to ride a bike. The AI was trained using 8 weeks of real data from 12 people with type 1 diabetes, including their blood sugar readings from continuous monitors, insulin doses they took, and their physical activity.
The AI system looked at blood sugar patterns over a 2-hour window—chosen because rapid-acting insulin peaks within 90-120 minutes and the body takes time to respond to meals or exercise. This timeframe captures both immediate and delayed effects on blood sugar. The system learned to recommend insulin doses (like 0.5, 1, or 1.5 units) that would keep blood sugar between 70-180 mg/dL, which doctors consider safe.
The researchers rewarded the AI when it kept glucose in the target range and penalized it for extreme highs or lows. Over time, the AI learned which insulin doses worked best for each person’s unique patterns of eating, exercise, and stress.
Traditional insulin plans use fixed doses that don’t change based on what you eat, how much you exercise, or your stress level. This one-size-fits-all approach often leads to blood sugar swings—sometimes too high (hyperglycemia), sometimes too low (hypoglycemia). An AI system that adapts in real-time could prevent these dangerous swings and reduce the constant mental effort required to manage type 1 diabetes.
This is an early-stage study with a small group (12 people) and 8 weeks of data, so results are promising but not yet proven for widespread use. The study used real patient data, which is a strength. However, the research hasn’t been tested in a larger, diverse population or compared head-to-head against current best practices. The AI was trained and tested on the same dataset, which could mean it performs better in the study than it would in real life with new patients. More research is needed before this becomes a clinical tool.
What the Results Show
The AI system achieved a mean blood glucose level of 80.06 mg/dL, which is within the healthy target range. It kept blood sugar in the safe zone (70-180 mg/dL) approximately 64% of the time during the evaluation period. The system received a reward score of 10, indicating it successfully learned to regulate glucose through personalized insulin adjustments.
The prediction accuracy was strong: the average prediction error was 9.85 mg/dL (mean absolute error), meaning the AI’s glucose forecasts were typically off by less than 10 points. The slightly higher root mean square error (12.39 mg/dL) suggests the system occasionally made larger errors, but these were rare.
These results demonstrate that the AI learned to make insulin dosing decisions that adapted to each person’s unique patterns. Rather than using a fixed insulin schedule, the system adjusted doses based on real-time glucose readings, meal intake, and activity levels—mimicking what an experienced diabetes educator might do, but in real-time.
The 2-hour observation window proved effective for capturing both immediate insulin effects and delayed responses from meals or exercise. The discrete insulin dose options (0.5, 1, and 1.5 units) provided enough flexibility for personalization while remaining practical for real-world use. The system’s ability to balance multiple competing goals—keeping glucose high enough to prevent dangerous lows while keeping it low enough to prevent complications—suggests the AI understood the complex trade-offs involved in diabetes management.
According to Gram Research analysis, traditional insulin pump therapy typically keeps patients in target range 50-60% of the time, while this AI system achieved 64%. However, this comparison should be interpreted cautiously because the AI was tested on the same data it learned from. Previous research shows that personalized approaches generally outperform fixed regimens, and this study supports that principle by using machine learning to personalize in real-time rather than requiring manual adjustments.
The study included only 12 people over 8 weeks, which is a very small sample. Results may not apply to people with different body types, ages, or types of diabetes complications. The AI was trained and tested on the same patients’ data, which could make results look better than they would be with completely new patients. The study didn’t compare the AI directly against current best-practice insulin management or other AI approaches. Real-world factors like stress, illness, hormonal changes, and medication interactions weren’t fully captured in the dataset. The system hasn’t been tested for safety in actual clinical use, where mistakes could have serious consequences.
The Bottom Line
This research is promising but too early for clinical recommendations. People with type 1 diabetes should continue following their doctor’s current insulin management plan. However, this study suggests that future AI-assisted insulin dosing could be valuable, so staying informed about emerging diabetes technology is worthwhile. Confidence level: Low to Moderate (early-stage research with small sample).
People with type 1 diabetes and their families should find this interesting as a potential future tool. Healthcare providers managing type 1 diabetes should monitor this research direction. People with type 2 diabetes using insulin may eventually benefit, though this study focused on type 1. People without diabetes don’t need to apply these findings, though understanding AI in healthcare is increasingly relevant.
If this technology advances through larger clinical trials successfully, it could take 3-5 years before becoming available as a clinical tool, and possibly longer before insurance covers it. Real-world testing would need to demonstrate safety and effectiveness in diverse populations before widespread adoption.
Frequently Asked Questions
Can artificial intelligence help manage type 1 diabetes better than traditional insulin pumps?
Early research suggests AI systems that learn from individual glucose patterns could keep blood sugar in target range more often than fixed insulin schedules. A 2026 study achieved 64% time in range with personalized AI dosing, compared to typical pump therapy at 50-60%. However, larger clinical trials are needed before AI becomes standard care.
How does an AI system learn to adjust insulin doses for each person?
The AI analyzes patterns from continuous glucose monitors, meal data, and activity levels over 2-hour windows. It learns through trial and error which insulin doses work best for each person’s unique body responses. The system gets rewarded for keeping glucose in safe ranges and penalized for dangerous highs or lows, similar to how humans learn from experience.
When will AI insulin dosing be available for people with type 1 diabetes?
This technology is still in early research stages. Larger clinical trials and safety testing are needed before it could become available, likely 3-5 years at minimum. Your doctor should continue managing your insulin plan with current proven methods until AI systems are clinically validated and approved.
What data does an AI diabetes system need to work effectively?
The AI needs continuous glucose readings, insulin dose records, meal information, and physical activity data. A 2-hour observation window captures both immediate insulin effects and delayed responses from food or exercise. The more accurate and complete this data, the better the AI can learn your personal glucose patterns.
Could AI insulin dosing prevent dangerous low blood sugar episodes?
Potentially, yes. By learning individual patterns and adjusting doses in real-time, AI could reduce both dangerous highs and lows. The study showed the system balanced these competing risks effectively, but real-world testing in larger populations is needed to confirm safety benefits before clinical use.
Want to Apply This Research?
- Track the percentage of time your blood glucose stays in your target range (as set by your doctor) each week. Record this weekly percentage in your diabetes app to see if any changes in your routine correlate with better or worse control.
- Start logging meals, exercise duration, and stress levels alongside your glucose readings. This data helps identify your personal patterns—which foods spike your glucose most, how exercise affects you, and when stress impacts your control. Over time, you’ll develop intuition about your body’s responses.
- Use your continuous glucose monitor data to identify your three most common blood sugar problem times (e.g., mornings, after lunch, bedtime). Focus on understanding what triggers highs or lows during those windows. Share these patterns with your doctor to refine your insulin plan.
This research represents early-stage development of AI for diabetes management and is not yet approved for clinical use. People with type 1 diabetes should continue following their healthcare provider’s current insulin management plan. This study involved only 12 participants over 8 weeks and has not been compared directly to current standard care in a clinical trial. AI-assisted insulin dosing could have serious safety implications if errors occur, and such systems require extensive testing and regulatory approval before use in patients. Do not attempt to modify your insulin regimen based on this research. Consult your endocrinologist or diabetes care team before making any changes to your diabetes management. This article is for educational purposes only and should not be considered medical advice.
This research translation is published by Gram Research, the science division of Gram, an AI-powered nutrition tracking app.