Confirm virtual-to-real world mapping of CIDT-derived DRL control policy

Determine whether the virtual-to-real world mapping of the Critical Illness Digital Twin (CIDT) is adequate for the discovered deep reinforcement learning (DRL) control policy for sepsis by experimentally validating that the simulation-derived policy translates to real biological systems and patient trajectories. Specifically, assess the mapping adequacy of the CIDT’s Innate Immune Response Agent-Based Model (IIRABM)-based, DDPG-trained control policy in an appropriate in vivo sepsis model that mirrors clinical complexity and variability.

Background

The paper proposes a Critical Illness Digital Twin (CIDT) that uses the Innate Immune Response Agent-Based Model (IIRABM), a Model Rule Matrix (MRM) framework with genetic algorithms and active learning for uncertainty conservation, and simulation-based deep reinforcement learning (DRL) to discover complex multimodal control policies aimed at curing sepsis.

While the authors demonstrate robust improvements in simulated recovery rates across diverse parameterizations, they explicitly acknowledge that they currently cannot confirm the adequacy of mapping from the virtual CIDT to real-world biology for the discovered control policy. They outline a future in vivo testing platform integrated with CIDT-driven DRL training to iteratively refine and validate this mapping against clinically relevant complexity.