Explainable Insulin Pump Control with LLM Controllers for Type 1 Diabetes

Living with Type 1 Diabetes (T1D) is a constant balancing act, requiring patients to make complex decisions based on endless streams of data. While Artificial Pancreas Systems (APS) powered by Reinforcement Learning (RL) have shown promise in automating insulin delivery, their "black-box" nature makes it hard for patients and doctors to trust them fully. This paper presents LLM-T1D, a groundbreaking approach that combines the precision of RL with the clear, human-like reasoning of Large Language Models (LLMs) to create a more transparent and reliable insulin pump controller. By training an expert RL system and then distilling its knowledge into fine-tuned Llama 3.1 8B and Qwen3 8B models using a LoRA architecture, we developed a controller that not only matches or surpasses the RL system’s performance but also explains its decisions in plain, understandable language. Tested on the FDA-approved UVA/Padova T1D simulator, the LLM controllers deliver excellent blood sugar control while giving patients clear, data-driven insights they can trust. This hybrid system transforms a complex algorithm into an approachable "copilot," paving the way for safer, more understandable, and patient-centered AI solutions for managing chronic conditions like T1D.