Researchers trained a neural‑network patient world model on the MIMIC‑IV electronic health record corpus. The model learns a shared representation of patient trajectories and is frozen after training. For intensive‑care‑unit (ICU) operations, the authors linked the model’s time‑queryable predictions to shared ICU calendars, resources, and queues using scenario‑specific mappings. This single model can generate admission demand, care tasks, and discharge proposals for different simulation scenarios.

In experiments, the frozen patient world model matched baseline Markov and semi‑Markov models in aggregate ICU occupancy and in the way capacity and dispatch policies changed outcomes. The model also reproduced policy gains from observed patient trajectories. When patients entered the same support state from different prior histories, the learned representation captured a 24‑hour recovery difference of –6.4 predicted versus –6.7 observed percentage points—an effect that compact state models miss, which in turn affects simulated care demand, queue burden, and bed use.

Limitations remain. The ICU simulator was calibrated with published ranges for arrival streams, service durations, and capacities because MIMIC‑IV records clinical trajectories rather than empirical queue logs or staff schedules. Additionally, the model learns conditional associations under historical care practices, so the simulations quantify capacity and dispatch policy effects under observed care patterns, but cannot predict absolute waiting times or generalize across hospitals without further calibration.

The study demonstrates that a patient world model learned from real EHR data can serve as a shared clinical representation across multiple health system simulation scenarios, but its applicability depends on careful calibration and an understanding of its observational nature.