A reinforcement learning–based model trained on continuous glucose monitoring, meal intake, and physical activity data maintained glucose levels within the target range in simulation among people with type 1 diabetes. The model also modestly outperformed a long short-term memory model in time-in-range performance, according to a study published in JMIR Diabetes. 

Researchers designed a deep Q-network (DQN) reinforcement learning system—a form of artificial intelligence (AI) that learns from experience and feedback—to personalize insulin dosing recommendations and forecast future glucose levels using real-world data from the OhioT1DM dataset. The dataset included 8 weeks of continuous glucose monitoring, insulin pump records, meal data, and physical activity information from 12 individuals (6 men, 6 women) with type 1 diabetes who used insulin pump therapy. 

The model used a 2-hour state representation incorporating recent glucose levels, insulin doses, meal intake, and exercise variables. Insulin recommendations were limited to discrete dose options, and the reward function incentivized glucose values between 70 and 180 mg/dL while penalizing hypo- and hyperglycemia. 

In the simulation-based evaluation, the DQN model achieved a mean glucose level of 80.06 mg/dL and maintained glucose values within the target range 64% of the time. Mean absolute error was 9.85 mg/dL, and root-mean-square error was 12.39 mg/dL.  

Researchers compared the reinforcement learning model with a long short-term memory (LSTM) model. The DQN system achieved a lower root-mean-square error than LSTM (12.39 vs 12.87 mg/dL), suggesting fewer large prediction errors, while LSTM produced a lower mean absolute error (3.69 vs 9.85 mg/dL), indicating greater average prediction accuracy. The reinforcement learning model also produced a higher time in range (64% vs 62%) and a higher average reward score (39.1 vs 24.5). 

Although the performance differences were modest, the reinforcement learning approach was designed to dynamically adapt insulin recommendations based on glucose, meal, and activity data as patients' physiologic states change, according to the researchers. 

To improve interpretability, the researchers also incorporated Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME). SHAP analyses identified carbohydrate intake as the strongest determinant of glucose predictions. Higher carbohydrate intake generally increased predicted glucose levels, whereas higher bolus insulin doses tended to lower them. Exercise duration exerted context-dependent effects on glucose predictions. In the LIME analysis, meal carbohydrate intake contributed +8.27 units to a predicted glucose value of 148.71 mg/dL, while exercise duration contributed −0.77 units. 

The study had several limitations. The dataset included only 12 patients, training was limited to 10 episodes, and statistical comparisons were based on only two models. The researchers noted that larger datasets, longer training durations, additional comparator models, and external validation are needed to establish robustness and generalizability. 

The researchers reported no conflicts of interest. 

Expert Insight 

AACE Endocrine AI invited corresponding researcher Vibhuti Gupta, PhD, in the Department of Biostatistics & Data Science at The University of Texas Medical Branch at Galveston, to elaborate on the findings.  

Why does this study matter?  

Type 1 diabetes requires patients to make complex insulin dosing decisions every day while balancing meals, physical activity, stress, illness, and other factors that affect blood glucose levels. Despite advances in continuous glucose monitoring and insulin delivery technologies, maintaining optimal glucose control remains challenging, and incorrect dosing can lead to dangerous episodes of hypoglycemia or hyperglycemia. This study explores how reinforcement learning can provide personalized insulin dosing recommendations and improve glucose forecasting based on an individual's unique physiological responses. By leveraging continuous glucose, insulin, meal, and activity data, our approach moves beyond one-size-fits-all treatment strategies toward more adaptive and personalized diabetes management. Ultimately, this work contributes to the development of intelligent clinical decision-support tools that have the potential to improve glucose control, reduce complications, lessen patient burden, and enhance quality of life for people living with Type 1 diabetes.  

What data stood out to you?  

One of the most interesting findings was the adaptive learning and consideration of various factors—continuous glucose monitoring data, insulin administration records, meal information, and physical activity data—while forecasting glucose levels and providing personalized insulin dosing recommendations. Even relatively simple behavioral factors, such as activity patterns, contributed meaningful context that helped the model better capture day-to-day variability in glucose regulation. More broadly, the study reinforced the idea that personalized AI-driven approaches can move beyond prediction to support real-time treatment decisions, an exciting direction for the future of diabetes care. 

How might the findings influence clinical practice? 

Our findings demonstrate the potential of reinforcement learning–based systems to serve as intelligent decision-support tools that assist both patients and clinicians in managing Type 1 diabetes. Rather than relying solely on standardized insulin dosing guidelines, these systems can continuously learn from an individual's glucose patterns, insulin responses, meals, and physical activity to generate more personalized recommendations. In clinical practice, this could help endocrinologists identify patients who may benefit from tailored treatment adjustments, improve time spent within the target glucose range, and reduce the risk of hypoglycemic and hyperglycemic events. 

Is there anything else you'd like to say about this study?  

This work represents an important step toward developing personalized, data-driven approaches for chronic disease management. While our study focused on type 1 diabetes, the broader message is that AI can help transform large volumes of health data into actionable insights that support better clinical decisions and patient outcomes. We view reinforcement learning not as a replacement for clinicians, but as a tool that can augment clinical expertise and empower patients with more individualized guidance. It is also important to recognize that this research is an early-stage investigation using retrospective data. Further validation in larger and more diverse patient populations, as well as prospective clinical studies, will be necessary before such systems can be integrated into routine care.