JMIR Publications

ABSTRACT

Clinical foundation models are increasingly trained on longitudinal electronic health records, learning continuous, high-dimensional patient representations that are not organized around the diagnostic vocabulary. Yet these systems are still built, evaluated, and governed as if the disease label were the natural unit of machine reasoning: the diagnosis is the privileged prediction target and the unit in which the model is expected to reason. In this Viewpoint we argue that this inherited assumption should be reversed. A disease label is a compressed, human-compatible abstraction whose usable resolution was bounded not by biology alone but by what clinicians and institutions could reliably name, teach, remember, and share. Foundation models relax that constraint, because the representation used to reason need no longer be human-readable: a machine can reason over a higher-dimensional latent patient state and render a named diagnosis only when a clinician, payer, regulator, or registry requires one. We therefore separate three things the label conflates—the internal representation a model reasons over, the clinical decision it is optimized against, and the human- and institution-facing code it renders—and reframe diagnosis as an external interface, a projection from that internal representation into a human-compatible code, rather than the substrate of machine reasoning. The claim is empirical, not rhetorical, and we hold it to a falsifiable test: comparing label-based against foundation-model latent representations, under matched data and compute, on outcomes defined outside the diagnostic coding system—treatment response, trajectory, dose, timing, and toxicity. We specify one such test in BCR–ABL-positive chronic myeloid leukaemia. The boundary condition is explicit: where a label is already a sufficient statistic for the decision, a richer representation buys nothing.