Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model

SepsisAgent

📃 Paper ｜ 🤗 SepsisAgent-4B

⚡ Introduction

SepsisAgent is a world model-augmented LLM agent for ICU sepsis treatment recommendation. It combines an LLM policy with a learned Clinical World Model that simulates patient responses under candidate fluid-vasopressor interventions. Instead of directly outputting a treatment action, SepsisAgent follows a propose-simulate-refine workflow: it proposes candidate actions, queries the world model for counterfactual patient trajectories, and refines the final prescription using both simulated dynamics and clinical priors.

The agent is trained with a three-stage curriculum: patient-dynamics supervised fine-tuning, propose-simulate-refine behavior cloning, and world-model-based agentic reinforcement learning. On MIMIC-IV sepsis trajectories, SepsisAgent improves off-policy treatment value while maintaining strong guideline adherence and low unsafe-action rates.

🧠 Method Overview

SepsisAgent uses a Clinical World Model as both an inference-time simulator and a training environment. The world model predicts action-conditioned patient evolution, while the LLM agent learns how to interpret these simulated responses for long-horizon treatment planning.

📊 Main Results

Clinical World Model Evaluation

Policy Value and Safety on MIMIC-IV

Results are reported on the 725-episode held-out test set. Higher is better for DR, WIS, WPDIS, and guideline adherence. Lower is better for unsafe actions.

SepsisAgent achieves the best DR and WPDIS scores among evaluated methods, while also obtaining the highest sepsis guideline adherence and the lowest unsafe-action rates. This indicates that the policy-value gains do not come from unsafe treatment shortcuts.

Ablation Study

The ablation shows that reinforcement learning in the Clinical World Model environment is the main driver of policy-value improvement. The final stage also improves intrinsic patient-dynamics prediction, including in-hospital mortality (IHM) and 24-hour vasopressor requirement (VR), even without simulator access during evaluation.

🎯 To-Do

• Release the SepsisAgent-4B.
• Upload the data processing scripts.

🙏 Acknowledgement

We gratefully acknowledge the MIMIC Code Repository for providing valuable reference implementations and resources for processing MIMIC critical care data. Our data processing pipeline was developed with reference to this project.

The data used in this work are derived from MIMIC-IV, a publicly available, de-identified electronic health record dataset hosted on PhysioNet.

📖 Citation

@misc{wu2026sepsisagent,
      title={Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model}, 
      author={Minghao Wu and Yuting Yan and Zhenyang Cai and Ke Ji and Chuangsen Fang and Ziying Sheng and Xidong Wang and Rongsheng Wang and Hejia Zhang and Shuang Li and Benyou Wang and Hongyuan Zha},
      year={2026},
      eprint={2605.14723},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.14723}, 
}