DiSSCo

MOSAIC: Multi-LLM agentic framework for clinical severity phenotyping in EHR data

MOSAIC is a two-phase agentic LLM framework for disease severity phenotyping applied to type 2 diabetes, evaluated on a synthetic EHR cohort of up to 4,886 patients. The system incorporates domains absent from traditional algorithmic comparators—including glycaemic staging, beta-cell function, and social determinants of health—and shows open-weight models matching proprietary pipelines (weighted kappa 0.773). Agentic classification diverged meaningfully from deterministic rule execution of the same rubric (kappa 0.428), suggesting genuine reasoning beyond fixed rules. The work provides early evidence that agentic LLM systems can generate clinically meaningful severity phenotypes from structured EHR data.