Multimodal foundation models have shown strong performance on a range of clinical tasks, according to a new arXiv preprint. But the paper's authors argue that most of these systems operate on individual samples in isolation, ignoring the broader context of a patient cohort and how clinical data change over time.
The proposed model, called CODA (Cohort- and Drift-aware), is intended to address that gap. The abstract states that existing approaches "fail to model the interplay…" — the sentence is cut off in the source — but the title and framing make clear that CODA incorporates cohort-level information and temporal drift into its reasoning.
Because the source is only a truncated abstract, no quantitative results or method details are available here. The significance is in the framing: clinical AI that reasons about a single patient without accounting for the population it belongs to, or for shifts in data over time, may miss important patterns.