Glonce

EHR WM world model forecasts patient trajectories with a 30x smaller footprint

researchThe Glonce editorsarXiv🔥54Glonce rating54Impact42Novelty55Relevance90Surprise25Players8Credibility72

A new arXiv paper introduces EHR WM, a predictive world model for electronic health records (EHR). It represents a patient's history as a latent trajectory conditioned on actions. Instead of modeling clinical events only through autoregressive token prediction, the framework explicitly separates the patient's state from the observed clinical actions, and learns how the state evolves through a transition operator that depends on both.

At the core is what the authors call a state dependent linearization predictor. It propagates local changes in the latent patient state, which the authors say enables time resolved prediction, multi step rollout and action conditioned simulation.

The evidence comes in four parts. First, generated latent trajectories preserve the downstream predictive signal, the temporal structure and the subgroup level statistical relationships found in the representations produced by the encoder. Second, the authors decode free running rollouts back into the original clinical feature space and compare future clinical measurements with observed values, using task matched forecasting baselines. They say this shows the learned dynamics keep predictive information beyond agreement in latent space. Third, across binary classification and time to event tasks, EHR WM achieves competitive, though task dependent, performance relative to substantially larger EHR foundation models, while using a 30x smaller parameter footprint. Fourth, perturbing the observed action sequences produces corresponding changes in the generated trajectories, which the authors present as showing that the learned dynamics are sensitive to actions.

The authors conclude that the results support world modeling as a complementary framework, not a replacement, for learning explicit, time indexed representations of patient evolution from longitudinal EHR.

Key facts

  • EHR WM is a world model for electronic health records that represents patient histories as action conditioned latent trajectories.
  • It separates patient states from observed clinical actions and learns their evolution with a state and action conditioned transition operator, rather than relying only on autoregressive token prediction.
  • On binary classification and time to event tasks, the authors report competitive, though task dependent, performance versus substantially larger EHR foundation models, with a 30x smaller parameter footprint.
  • Free running rollouts decoded into the original clinical feature space are checked against observed values using task matched forecasting baselines.
  • Perturbing observed action sequences changes the generated trajectories accordingly, which the authors present as action sensitivity of the learned dynamics.

Why it matters

Most EHR models treat a patient record as a stream of tokens to predict. This paper takes a different route: it models the patient as a state that changes over time, with clinical actions as a separate input. That makes it possible to roll a trajectory forward in time and to ask how it changes when the actions change. The reported 30x smaller parameter footprint, against substantially larger EHR foundation models, is the practical headline. The authors frame the approach as complementary to existing methods, not a replacement.

Who it affects

The paper is aimed at researchers building models on longitudinal electronic health records, especially those working with foundation models for EHR and with forecasting of patient outcomes. Teams that care about model size may find the 30x smaller footprint relevant. The source does not discuss clinicians or patients directly.

How to use it

This is a research preprint, not a product. The abstract describes tasks the model supports: time resolved prediction, multi step rollout and action conditioned simulation, evaluated on binary classification and time to event tasks. The source text mentions no code or data release, so there is nothing to download or deploy from the information given.

How solid is it

The claims come from the authors' own abstract. The performance claim is hedged by the authors themselves: competitive, though task dependent, not uniformly better than the larger foundation models. The source gives no concrete metric values, no dataset names, patient counts or EHR sources, and does not name the foundation models compared against or their parameter counts. The abstract names no authors or institutions, and no submission date, venue or code release is mentioned in the source text.

Risks and caveats

No clinical deployment, clinical validation or patient-outcome benefit is claimed. The action conditioned simulation is evaluated only as the sensitivity of generated trajectories to perturbed action sequences, which shows that the model responds to actions, not that its simulated outcomes are clinically correct. The 30x figure is a size comparison, not an accuracy result, and the accuracy comparison is described only qualitatively.

“these results support world modeling as a complementary framework for learning explicit, time indexed representations of patient evolution from longitudinal EHR”

— From the paper's abstract