Electronic health record (EHR) foundation models are promising tools for scalable clinical prediction across health systems, but they run into a tokenization bottleneck: the same clinical concept can be represented by different ontology codes depending on the local system. This makes it hard for a model trained on one set of codes to generalize to another.
The authors propose a discrete neural interlingua—an intermediate representation that is agnostic to the specific ontology. Instead of forcing every system to adopt the same codes, the model would map varied inputs into a shared discrete space, allowing predictions to be made without being tied to a particular vocabulary.
The abstract outlines the problem and the proposed solution but does not include experimental results, so the claims are about the architecture's potential rather than demonstrated performance. If the approach holds up, it could ease the deployment of EHR models across institutions with different coding practices.