Interpretable Language Model for Closed-Loop Type 1 Diabetes Control

Short summary

Researchers present LLM-T1D, a system that distills a reinforcement-learning insulin controller into fine-tuned LLaMA 3.1 8B and Qwen3 8B models for Type 1 Diabetes management. The LLM controllers achieve 73.5% Time in Range on the FDA-approved UVA/Padova simulator while providing plain-language explanations for dosing decisions. Formal safety verification guards against hallucinated outputs, addressing the trust gap that black-box RL systems face in clinical settings.

• •LLM-T1D distills RL insulin controller into LLaMA 3.1 8B and Qwen3 8B for interpretable T1D management
• •Achieves 73.5% Time in Range on FDA-approved UVA/Padova simulator with formal safety verification
• •Plain-language explanations aim to improve patient and clinician trust over black-box RL systems

Generated with AI, which can make mistakes.