SepsisAgent: LLM Agent with Clinical World Model for ICU Treatment
A new agent named SepsisAgent has been developed by researchers, featuring a large language model (LLM) enhanced with a world model specifically for recommending sepsis treatments in the ICU. This innovative system employs a Clinical World Model to predict patient reactions to various fluid-vasopressor interventions, utilizing a propose-simulate-refine approach prior to finalizing any treatment. Initial tests indicated that merely accessing the world model led to variable decision-making performance from the LLM, which prompted the need for specialized training. SepsisAgent undergoes a three-phase training process: supervised fine-tuning of patient dynamics, behavior cloning through propose-simulate-refine, and reinforcement learning based on the world model. The research utilizes sepsis trajectories from MIMIC-IV for training, and the findings can be found on arXiv with the identifier 2605.14723.
Key facts
- SepsisAgent is a world model-augmented LLM agent for sepsis treatment recommendation.
- It uses a Clinical World Model to simulate patient responses under fluid-vasopressor interventions.
- The workflow is propose-simulate-refine before committing to a prescription.
- World-model access alone yielded inconsistent LLM decision performance.
- Training involves a three-stage curriculum: supervised fine-tuning, behavior cloning, and reinforcement learning.
- Training data comes from MIMIC-IV sepsis trajectories.
- The paper is on arXiv with identifier 2605.14723.
- Sepsis management requires sequential treatment decisions under rapidly evolving patient physiology.
Entities
Institutions
- arXiv