LLMs Enhance Interpretable Type 1 Diabetes Control.

Automating insulin delivery for Type 1 Diabetes (T1D) using Artificial Pancreas Systems (APS) powered by Reinforcement Learning (RL) shows great promise, but their "black-box" nature hinders trust from patients and clinicians. A new method, LLM-T1D, addresses this by integrating the precision of RL with the explainability of Large Language Models. This creates a more transparent and reliable insulin pump controller. The approach involves training an expert RL system and then distilling its knowledge into fine-tuned LLMs, specifically LLaMA 3.1 8B and Qwen3 8B models. This results in a controller that not only surpasses the RL system's performance in blood sugar management but also provides clear, human-understandable explanations for its decisions. Tested on an FDA-approved simulator, the LLM controllers achieved excellent time in range (73.5%) while maintaining formal safety verification against hallucinations.