Clinical multimodal models often pick up on enrollment shortcuts—patterns that happen to correlate with the target condition in the training data but do not reflect true disease processes. Because these shortcuts are shared across modalities, the models appear accurate during development yet fail when applied to a new cohort. A new paper on arXiv introduces DISCERN, a method that explicitly models this problem using a latent structural causal model.
The core idea is to treat the observed physiology as generated by two separate latent components: the disease state and background factors. By identifying the causal contribution of the disease state, DISCERN aims to learn representations that are not tied to the specific enrollment context of the training data. This makes the model more robust when it encounters patients from different settings or populations.
The abstract does not yet include experimental results, so the reported contribution is the modeling framework itself. The paper argues that explicitly separating disease evidence from background variation is a necessary step toward generalizable medical multimodal learning.