Researchers developed MOSAIC, a two-phase agentic large language model framework for severity phenotyping, using type 2 diabetes as a proof-of-concept. The system synthesizes clinical evidence and reasons over Electronic Healthcare Records to capture multidimensional disease severity.
- Evaluated on a synthetic cohort of 5,086 patients against three algorithmic ground truths (DCSI, DiSSCo, Cooper) and mortality outcomes.
- Open-weight MOSAIC matched the proprietary pipeline with a weighted kappa of 0.773.
- The framework spans domains absent from comparators, including biomarker-based glycaemic staging and social determinants of health.
- Agent-based tiers showed significant separation of all-cause mortality (log-rank p < 0.001) and diverged from deterministic execution (kappa = 0.428).
The study demonstrates that agentic LLM systems can generate and apply clinically meaningful severity phenotypes from structured EHR data.