Title: Zero-shot automated insulin delivery for type 1 diabetes via dynamic physiology-aware reinforcement learning

Abstract

Insulin therapy in type 1 diabetes requires constant dose adjustment based on blood glucose, meals, physiological states, and physical activity. This demanding self-management imposes a substantial burden and increases dosing-error risk, underscoring the need for automated insulin delivery (AID) systems that reduce user intervention. However, many current systems depend on fixed, individualized parameters and may not fully adapt to rapid or unobserved physiological changes. We developed the Dynamic Physiology-Aware Reinforcement learning Controller (DPARC), a zero-shot insulin optimizer that infers latent physiological dynamics from recent continuous glucose monitoring (CGM) and insulin-delivery history without prior personalization, carbohydrate announcements, or preset subject-specific parameters. DPARC uses a rolling 24-hour CGM and insulin-history window, but closed-loop operation can begin after 1 hour of observed data by initializing unobserved history with neutral normalized padding and progressively replacing it with observations. In silico, a single frozen DPARC policy adapted within 1 hour, improved time in range compared with a total daily insulin-conditioned reinforcement learning baseline, and approached the upper-bound performance of a fully personalized model under stochastic unannounced meals with randomized timing, carbohydrate amounts, absorption variability, and meal skipping. In supervised porcine studies under unannounced meals, DPARC maintained high time in range without manual configuration, supporting large-animal feasibility while prospective human evaluation is needed before clinical efficacy can be established. Learned latent representations correlated with physiological markers including insulin sensitivity and plasma insulin concentration, supporting physiological alignment and explanatory anchors. Collectively, these findings support DPARC as a preclinical proof-of-concept zero-shot AID framework for future supervised human evaluation.