NOAH: Modeling the Full Patient Journey
— by Vivax
NOAH is a September 2026 arXiv preprint presenting a longitudinal, multimodal, time-aware model of the patient journey.
NOAH is a September 2026 arXiv preprint presenting a longitudinal, multimodal and time-aware model of the patient journey. Its central move is to treat the record as an ordered process—events, measurements, diagnoses and encounters—not as a pile of independent documents.
The reported corpus spans more than 559 million events, 431,000 hospital visits and 299,000 patients from MIMIC-family data. That scale makes the representation problem tangible, but it also makes provenance, missingness and site-specific documentation patterns part of the scientific result.
NOAH learns to forecast likely next clinical events and probes the patient representation on downstream tasks. This is useful infrastructure for studying temporal context, yet a forecast of what is likely to be documented next is not a recommendation about what should happen next.
The reported results are uneven in an informative way. Some event-prediction tasks are stronger than others; readmission is a weak headline endpoint, and long rollouts lose coverage as uncertainty compounds. A model can predict a chart's future structure without understanding the patient's physiology.
The boundary matters most for counterfactual claims. Retrospective records show what happened under an observed care pathway; they do not reveal the causal effect of an unchosen treatment. NOAH is not a validated clinical simulator, a virtual patient, or evidence for autonomous care.
Vivax reads NOAH as a serious representation direction: clinical world models need time, multiple modalities and explicit evaluation of missingness. The next step is governance-aware validation across institutions and workflows, with calibrated uncertainty and prospective endpoints—not a leap from a large corpus to clinical authority.