A Microbiome-Conditioned Reinforcement Learning Architecture for Closed-Loop Glycemic Control

We present a microbiome-conditioned reinforcement learning (RL) architecture that reframesclosed-loop control as a problem of predictive metabolic modeling rather than reactiveglucose regulation. The system integrates multimodal physiological signals across multipletimescales, including continuous glucose monitoring, insulin delivery history, behavioral context,and microbiome-derived features that represent an individual’s metabolic phenotype. Aconstrained RL policy is coupled with a predictive world model and a supervisory safety layerto ensure clinically viable insulin recommendations. To enhance interpretability, we introduce the Microbiome Metabolic Index (MMI), a derivedrepresentation that summarizes microbiome-driven metabolic tendencies, including insulinsensitivity, postprandial response characteristics, and inflammatory burden. By incorporatingupstream biological context into the control loop, the proposed framework is designed to enableanticipatory insulin adjustments and individualized control strategies.This work proposes a new class of biology-aware control systems for diabetes managementand establishes a foundation for adaptive, interpretable, and clinically scalable closed-loop therapies.