freedomintelligence/sepsisagent

SepsisAgent is a research project from an academic and industry lab that builds an AI agent to recommend treatment for sepsis patients in intensive care units. Sepsis is a life threatening condition, and choosing the right fluid and medication doses for a patient is a high stakes decision. Rather than having a language model guess a treatment directly, this project pairs the language model with a separate learned Clinical World Model, which simulates how a patient is likely to respond to a proposed treatment before that treatment is chosen. The system works through what the authors call a propose, simulate, and refine process. The language model first proposes candidate treatment actions, the Clinical World Model simulates what would likely happen to the patient under each option, and the language model then refines its final recommendation using both that simulated outcome and general clinical guidelines. The agent is trained in three stages: first learning to predict how patients change over time, then learning the propose, simulate, refine behavior by imitation, and finally being fine tuned with reinforcement learning inside the simulated environment. The README reports results on real intensive care unit data from the MIMIC IV dataset, a large publicly available and de identified hospital records collection. According to the tables shown, SepsisAgent outperforms several comparison methods, including a general purpose reasoning model, on measures of treatment value while also scoring highest on following medical guidelines and having the lowest rate of unsafe treatment decisions. The repository includes runnable code: a script to run the Clinical World Model on its own, and a script to run the full agent using a trained model hosted on Hugging Face. A single anonymized example patient case is included so the whole pipeline can be tried end to end, since the full hospital dataset cannot be redistributed. There is an accompanying academic paper describing the method in full. This is a research codebase intended for people working in machine learning or clinical informatics, not a tool for direct patient care.