One Patient, Two Futures: World Models for Clinical Trial Simulation

Randomized controlled trials are essential for evaluating treatment strategies, yet they are costly and time-consuming to conduct. Observational electronic health records offer an opportunity to learn patient dynamics and explore potential treatment outcomes through simulation. We present a clinical world model for simulating patient trajectories under alternative treatment strategies. The framework combines a variational autoencoder with treatment-conditioned latent diffusion, a clinician behavior policy that generates background care, and an intervention controller that applies the assigned strategy. On held-out MIMIC-IV intensive care records, the diffusion ensemble achieves lower long-horizon forecasting error than other baselines. When simulating lactated Ringer’s versus saline, the model reproduces the direction of the evaluated treatment effects reported in a randomized trial and yields odds-ratio estimates closer to the trial reference than those from two conventional observational estimators. The simulated trajectories also reproduce the trial-consistent pattern of lower chloride and higher bicarbonate under lactated Ringer’s. Together, these findings demonstrate that a world model trained on observational ICU data can generate treatment-arm simulations that capture both clinical outcome contrasts and physiological responses observed in a randomized trial, supporting further investigation of prospective trial prediction.