MON-551 - Optimization of the Results of Insulin Delivery with Artificial Pancreas by Integrating Artificial Intelligence-Based Control Systems

Background: The combination of continuous subcutaneous insulin infusion and continuous glucose monitoring (CGM) systems into artificial pancreas (AP) is a relatively new technology in the field of endocrinology, which is rapidly evolving with the introduction of artificial intelligence (AI). This review presents an update on how AI-based AP can deal with current issues in insulin therapy.

Methods: A systematic review of the literature was conducted for the period 2004 to 2025 using the search terms insulin infusion systems, AP, machine learning, and AI.

Results: AP has been an outbreak in the history of diabetes management, especially for type 1 diabetes where reliance on insulin doses is more prominent, while the affected population (mainly children and adolescents) may be non-adherent to the therapy. In contrast to its myriad benefits, the integration of AP creates many challenges to the patients and clinicians. As the system becomes more independent, patients are no longer required to monitor the function of AP or their blood glucose levels individually, hence some issues such as nocturnal hypoglycemia, insulin-induced postprandial hypoglycemia, blood glucose inaccuracy, and injection site failure may go unnoticed and be of safety concern. AI models have been developed with the hope of limiting such issues by predicting their occurrence. They have also showed outstanding results in prediction of external events such as physical activity and meal intake, with the data provided by CGM and wearable devices to minimize dramatic fluctuation of blood glucose levels and control the target blood glucose levels tightly. Moreover, AI-based AP is used in inpatient and intensive care unit settings to optimize insulin therapy. It is worth noting that the most common AI class used in the development of AP and insulin dosing has been reinforcement learning. Even though AI has shown promising results in most experiments, it cannot be yet implemented commercially on real patients due to black box nature of some AI algorithms and the fact that most research to-date has been conducted on simulators and in silico.

Conclusion: The non-traditional AI-based control systems outperform former proportional-integral-derivative control systems in managing diabetes with AP as it can be implemented not only to predict failure and blood glucose level trends, but also to minimize patient and clinician intervention. However, sufficient clinical data are not available to fully address the efficacy of such systems and their pitfalls.

*Unless otherwise noted, all abstracts presented at ENDO must not be released to the press or the public until the date and time of presentation. For oral presentations, the abstracts are embargoed until the session begins. The Endocrine Society reserves the right to lift the embargo on specific abstracts that are selected for promotion prior to or during ENDO.*