Forecasting Patient Trajectories from the Electronic Health Records with Action-Conditioned World Models

Introduction

Electronic health records (EHR) capture a patient’s longitudinal clinical journey as a sequence of observations (labs, diagnoses, medications) and interventions (treatments, procedures). Traditional predictive models treat this sequence as a purely autoregressive stream of tokens, implicitly mixing the patient’s underlying health state with the actions taken by clinicians. The recent paper Forecasting Patient Trajectories from the Electronic Health Records with Action‑Conditioned World Models proposes a different paradigm: a world model that explicitly separates latent patient states from observed clinical actions and learns how actions drive state transitions over time.

Action‑Conditioned Latent Dynamics

The core of the approach is a state‑dependent linearization predictor. First, an encoder maps each time‑step of the raw EHR into a latent representation of the patient’s health state. A separate action encoder captures the clinical interventions at that step. The transition operator then updates the latent state using a linear function whose parameters depend on the current state, allowing the model to capture locally varying dynamics while remaining computationally tractable. Because the transition is conditioned on both the current state and the action, the model can generate counterfactual trajectories by altering the action sequence—a capability that is difficult for standard autoregressive models.

Preserving Predictive Signal Across Representations

To validate that the latent trajectories retain clinically relevant information, the authors compare three aspects:

1. Downstream predictive performance – classifiers and survival models trained on the latent states achieve accuracy comparable to those trained on the original high‑dimensional features.
2. Temporal structure – the latent trajectories preserve the ordering and spacing of events, enabling multi‑step rollouts that remain coherent over weeks or months.
3. Subgroup statistics – statistical relationships (e.g., incidence of complications within specific demographic groups) observed in the encoder‑derived embeddings are reproduced in the generated trajectories.

These checks demonstrate that the world model does not merely compress data but learns a meaningful dynamical system.

Forecasting and Simulation Results

The paper evaluates the model on two families of tasks:

• Binary classification (e.g., predicting readmission within 30 days).
• Time‑to‑event (e.g., survival analysis for disease progression).

Across these tasks, the action‑conditioned world model (EHR‑WM) matches or exceeds the performance of much larger foundation models, despite using roughly 30× fewer parameters. Importantly, when the model is run in a free‑running mode—generating future latent states without conditioning on observed data—and then decoded back to the original feature space, its forecasts remain competitive with task‑specific baselines that directly predict the next observation.

Action Sensitivity and Clinical Insight

A key advantage of the world‑model framework is its action sensitivity. By perturbing the action sequence (e.g., simulating the addition of a medication or the omission of a procedure), the generated latent trajectory changes in a predictable manner, reflecting the hypothesized impact of the intervention on patient health. This property opens avenues for in‑silico clinical trial simulations, personalized treatment planning, and what‑if analyses that are grounded in a learned dynamical system rather than static risk scores.

Practical Takeaways

• Compact yet expressive: Separating state and action reduces model size while preserving essential predictive information.
• Counterfactual simulation: The framework supports realistic scenario testing by altering action sequences.
• Compatibility with downstream tasks: Latent states can be fed into standard classifiers or survival models without loss of accuracy.
• Scalable to real‑world EHR: The linearization predictor enables efficient training on large, irregularly sampled clinical datasets.

Conclusion

Action‑conditioned world models represent a promising shift in how we learn from longitudinal health data. By explicitly modeling the causal influence of clinical actions on patient states, they provide both accurate forecasts and a mechanistic lens for exploring treatment effects. As healthcare systems continue to amass richer EHR streams, such dynamical approaches could become foundational tools for precision medicine, policy simulation, and proactive patient management.