AI boosts insulin precision in diabetes care, but barriers persist

Despite consistent improvements in glycemic outcomes across simulations, real-world adoption is lagging. The study finds that while many RL algorithms demonstrate superior control of blood glucose levels - marked by improved time-in-range and reduced hypoglycemia - in silico trials, few have undergone clinical testing in diverse patient populations. This gap between lab and clinic remains a major barrier to widespread use.

A new study finds that artificial intelligence is rapidly reshaping how insulin is recommended and delivered to people with diabetes, but also warns that the most promising AI models remain under-tested outside simulation labs. The research, conducted by biomedical engineers at the University of Bern in collaboration with Maastricht University, outlines both the clinical potential and persistent limitations of AI-powered insulin systems, particularly those based on reinforcement learning algorithms.

Titled "The Role of Artificial Intelligence in Enhancing Insulin Recommendations and Therapy Outcomes", the study reviews more than 40 recent AI-driven insulin control models, comparing their performance and feasibility for closed-loop and open-loop delivery systems. Reinforcement learning (RL) models, which are capable of adapting dosing strategies to individual users based on continuous data feedback, are highlighted as one of the most promising advancements in the pursuit of personalized diabetes therapy.

Unlike traditional insulin calculators or rule-based decision systems, RL-based controllers learn over time by interacting with the user’s glycemic response data, optimizing dosage for varying factors such as food intake, exercise, illness, sleep, and insulin sensitivity. These models are being integrated with continuous glucose monitors (CGMs) and insulin pumps in hybrid and fully automated closed-loop systems, often referred to as artificial pancreas technologies.

Despite consistent improvements in glycemic outcomes across simulations, real-world adoption is lagging. The study finds that while many RL algorithms demonstrate superior control of blood glucose levels - marked by improved time-in-range and reduced hypoglycemia - in silico trials, few have undergone clinical testing in diverse patient populations. This gap between lab and clinic remains a major barrier to widespread use.

In type 1 diabetes, most of the development has centered on integrating RL algorithms with CGMs and insulin pumps. Notable examples include actor-critic and deep Q-learning models designed to optimize basal and bolus insulin administration. Several of these algorithms are capable of functioning without meal announcements or exact carbohydrate estimation, a breakthrough in reducing user burden and minimizing errors in insulin management.

The study also cites a growing interest in applying RL to type 2 diabetes. In this area, model-based RL systems have shown success in optimizing insulin regimens through simulation of patient-specific glycemic responses. One trial cited in the paper demonstrated improved long-term glycemic outcomes and fewer hypoglycemic events when compared with traditional insulin titration methods.

Key factors contributing to improved therapy outcomes include model-free online learning, which eliminates the need for pre-defined physiological models and allows the system to adapt directly to the individual’s patterns. Systems incorporating variables such as insulin-on-board, insulin sensitivity, and macronutrient-specific dose adjustments—particularly for high-fat meals or post-exercise states—also showed stronger performance.

However, the authors caution that major technical, ethical, and economic challenges must be addressed before these systems can be widely adopted. One of the primary concerns is algorithm transparency. Most reinforcement learning models are highly complex and difficult to interpret, which undermines patient and clinician trust. The study recommends designing user-friendly interfaces that can explain insulin recommendations in understandable terms, improving trust and adherence.

Access and affordability present another critical issue. Most closed-loop insulin delivery systems require expensive equipment, including CGMs and insulin pumps, which are out of reach for many people, especially in low-resource settings. The authors highlight ongoing efforts to develop cost-effective RL systems compatible with traditional devices like insulin pens and SMBG (self-monitoring of blood glucose) meters. These systems could help democratize access to advanced insulin management, particularly in underserved communities.

Patient engagement and data burden are also identified as major factors influencing long-term adherence. For many individuals, especially older adults or those with limited health literacy, the task of logging meals, insulin doses, and glucose readings is overwhelming. The study urges the development of passive data collection tools, visual dietary estimators, and smart reminders to reduce user workload while improving system performance.

Privacy and data security are additional concerns. As insulin delivery systems collect sensitive personal health information, often including eye movement, posture, and other biometric data, ensuring GDPR and HIPAA compliance is essential. The authors advocate for strong encryption, transparent data policies, and anonymized model training to safeguard patient rights.

Regulatory uncertainty further complicates deployment. AI-based medical technologies fall into a gray area under many current frameworks. While some algorithms may qualify as medical devices under FDA or EU MDR regulations, their dynamic learning nature makes certification and post-market surveillance difficult. The study calls for flexible, adaptive governance models that balance innovation with safety, along with practical implementation of ethical principles like fairness, autonomy, and accountability.

To fully realize the promise of AI-enhanced insulin therapy, the researchers conclude that collaboration across disciplines is essential. Engineers, clinicians, data scientists, and patients must work together to build interpretable, accessible, and clinically validated systems that operate safely across diverse populations and care settings. Future research should also incorporate broader factors affecting glucose control, such as sleep quality, stress, and concurrent medications, to increase the accuracy of dose predictions.