Clinical & Hospital World Models
Clinical agents, patient-dynamics world models, hospital information-system environments, multi-role simulations, and medical skills.
Clinical & Hospital World Models moves medical AI from isolated QA to environments where agents act, observe, call tools, interact with clinical roles, and reason over changing patient or hospital states. This page intentionally excludes DentalGPT and GlobalDentBench, which are now organized under the dedicated dental foundation model direction.
Research Storyline
SepsisAgent uses a Clinical World Model to simulate patient responses under candidate ICU interventions.
Agentic-Hospital connects role simulation, mocked HIS/EMR, structured data views, APIs, CLI interfaces, governance, and rubric-based evaluation.
TwinHospital combines HIS data flow with multi-role agent simulation across registration, consultation, treatment, follow-up, and operations.
OpenClaw-Medical-Skills turns medical capabilities into callable skills for agentic workflows.
Representative Work
Combines propose-simulate-refine inference with world-model-based agentic RL for ICU sepsis treatment recommendation.
PaperDefines hospital intelligence as autonomous agents operating across human and digital hospital environments through standardized interfaces.
RepositoryBuilds an HIS data-flow-aware hospital simulation with patients, nurses, physicians, administrators, and management workflows.
RepositoryDefines workflow-aligned tasks and benchmarks that better match how clinicians actually work.
PaperA large open-source medical AI skills library for building tool-using clinical agents.
RepositoryTechnical Layers
Clinical states change after interventions, so agents need simulators that predict patient evolution under candidate actions.
HIS, EMR, labs, imaging, pharmacy, scheduling, billing, and governance systems become part of the agent environment.
Doctors, nurses, administrators, patients, pharmacists, and managers make hospital workflows more realistic than single-turn diagnosis tasks.
Clinical agents need permissions, logs, rubric-based scoring, guideline adherence, and unsafe-action analysis.