EHRWorld: A Patient-Centric Medical World Model for Long-Horizon Clinical Trajectories

TL;DR

The paper tackles the challenge of modeling long-horizon clinical trajectories with world models. It shows that naive LLMs struggle to maintain consistent patient states under sequential interventions, identifying substantial error drift. To address this, it introduces EHRWorld and the EHRWorld-110K dataset, employing a causal sequential training paradigm and set-based, dual-mode (inquiry/intervention) generation to simulate dynamic disease progression. Across extensive experiments, EHRWorld significantly outperforms baselines in long-horizon stability, accuracy on clinically sensitive events, and inference efficiency, underscoring the practical value of causally grounded, temporally evolving clinical data for robust medical world modeling.

Abstract

World models offer a principled framework for simulating future states under interventions, but realizing such models in complex, high-stakes domains like medicine remains challenging. Recent large language models (LLMs) have achieved strong performance on static medical reasoning tasks, raising the question of whether they can function as dynamic medical world models capable of simulating disease progression and treatment outcomes over time. In this work, we show that LLMs only incorporating medical knowledge struggle to maintain consistent patient states under sequential interventions, leading to error accumulation in long-horizon clinical simulation. To address this limitation, we introduce EHRWorld, a patient-centric medical world model trained under a causal sequential paradigm, together with EHRWorld-110K, a large-scale longitudinal clinical dataset derived from real-world electronic health records. Extensive evaluations demonstrate that EHRWorld significantly outperforms naive LLM-based baselines, achieving more stable long-horizon simulation, improved modeling of clinically sensitive events, and favorable reasoning efficiency, highlighting the necessity of training on causally grounded, temporally evolving clinical data for reliable and robust medical world modeling.

Figures (6)

• Figure 1: Overview of patient simulation challenges and performance evaluation. The upper and middle panels illustrate a clinical scenario where a standard LLM-based simulator fails to infer implicit physiological states or correctly update patient status following medical interventions. In contrast, the proposed EHRWorld model maintains logical consistency and robustness. The lower panel presents the trajectory of relative error for Chloride levels across eight simulation rounds. We compare our approach against GPT-5.2, demonstrating that our model significantly constrains the rate of error propagation, resulting in a widening performance gap that highlights robustness in long-horizon simulations.
• Figure 2: Overview of the EHRWorld framework. (a) Data construction pipeline. Raw EHR records from MIMIC-IV are processed into the EHRWorld-110K dataset by integrating static patient context with longitudinal clinical events. (b) The EHRWorld model. At each step, the model conditions on the current patient state and a set of clinical actions, generates outcomes via a dual-mode mechanism for inquiries and interventions, and updates the interaction history for sequential trajectory simulation.
• Figure 3: Performance stability analysis across global and dynamic clinical metrics. EHRWorld-14B demonstrates superior robustness, exhibiting the smallest performance degradation in S@25 accuracy and SMAPE error compared to general foundation models.
• Figure 4: Qualitative case study comparison. We visualize the ground truth versus model predictions for a 69-year-old patient. Colors denote relative error ($\Delta$) severity: precise ($\Delta \le 10\%$), acceptable ($10\% < \Delta \le 20\%$), and deviation ($\Delta > 20\%$). EHRWorld-14B shows superior stability, keeping most metrics within low-error margin.
• Figure 5: Distributions of demographics and clinical event statistics for the processed EHRWorld-110K dataset. The histograms illustrate patient age, length of stay (LOS), and event counts across diverse modalities (e.g., Lab Events, Medications, Procedures). Additionally, we plot Event Intensity (total events divided by LOS) to represent the density of clinical activities. Inset boxes report descriptive statistics (maximum, mean, and median) for each metric. Note that the y-axes for LOS and event counts use a logarithmic scale to visualize the long-tail distributions.