A new preprint on arXiv presents a predictive world model for electronic health records (EHR) that frames patient histories as action-conditioned latent trajectories. Rather than treating clinical events as a sequence of tokens to be predicted one by one, the model learns a latent representation of the patient's state and how interventions or actions shift that state over time.
The authors position this as a departure from standard autoregressive approaches, which model clinical events solely through next-token prediction. By conditioning on actions, the model aims to capture the dynamics of care—how treatments and decisions influence future health states—potentially enabling more realistic forecasts of patient trajectories.
The abstract is brief and does not include experimental details or quantitative results, so the reported benefits are not yet verified. As a preprint, the work has not undergone peer review, and the full methodology and evaluation remain to be examined.