ICU planning no longer rebuilds states per question: one frozen patient world model drives admission, care and discharge simulations
Synopsis
ICU resource planning has long relied on compact state-transition models hand-specified for a single policy question, so a new question means redefining states. A patient world model trained on MIMIC-IV and then frozen encodes irregular electronic health records as state-event sequences, returns predictions at any interval the simulator requests, and connects through scenario-specific mappings to shared bed and staff queues. Across admission, care-dispatch and discharge simulations it agrees with task-fitted Markov and semi-Markov comparators on directional responses to capacity and dispatch policies, and reproduces the dispatch gains obtained by replaying observed trajectories.
Interpretation
One frozen model of patient dynamics can supply admission, care-dispatch and discharge simulations at once, so a new question requires changing only external mappings and resource rules rather than rebuilding the patient state space. Health system simulation has typically hand-specified compact clinical states, transitions and service mechanisms for a single policy question, discarding history not retained in those states and often forcing a state-space redesign or transition re-estimation for a new question. On patient-level partitions of MIMIC-IV, the model and task-fitted Markov and semi-Markov comparators give the same direction of response to expanding beds, prioritizing acute tasks and expanding transfer slots, and the model closely tracks the dispatch gains of replayed observed trajectories.
The model writes each patient history as a sparse state-event grid and queries future clinical distributions at continuous intervals requested by the simulator, so several patients can advance on a shared event clock. Existing health trajectory generators advance one patient's timeline asynchronously from one emitted event token to the next, which does not support advancing trajectories to arbitrary times on a shared clock or mapping clinical states to bed and staff demand. The model is trained over a clinical vocabulary of 64 state variables and 25 event types with dual-axis attention across variables and time, and once frozen is coupled through scenario-specific mappings that turn admission proposals into bed requests, organ-support events into care tasks and live-exit proposals into downstream service requests.
Compact states drop history such as entry direction, while the learned representation separates patients stepping up into single-organ support from those stepping down from dual support. Markov and semi-Markov baselines conditioned on current support state, dwell time and hours since ICU admission do not encode entry direction, and produce identical paired trajectories and zero endpoint contrast for matched patients. Across 404 matched held-out patient pairs, the model's upward-minus-downward recovery difference is a stated number of percentage points against a stated observed value, and that contrast propagates into different care-task requests, queue burden and occupied bed time.
Perspective
The demonstrated setting is retrospective adult ICU records from a single medical center in MIMIC-IV covering 2008 to 2022, with arrival streams, service durations and capacities set from published ranges and scenario designs rather than measured queue logs or staff schedules; conclusions are therefore limited to differences driven by patient trajectories under identical arrival, service and staffing rules.
Like Markov and semi-Markov models fitted to observational records, the model learns conditional associations under historical care practices rather than causal effects; the simulations quantify how capacity and dispatch policies shape admission access, queue waiting and resource utilization under observed care patterns, not how delayed care alters patient survival.