Richer database context improves LLM-generated EHR phenotyping algorithms

Detailed database context enabled LLMs to generate executable type 2 diabetes (T2DM) OMOP-based phenotyping SQL with recall ≥0.99 in synthetic data and >0.90 in UNC EHR.

• Why it matters: Limited prompt context led to SQL that often failed to run and missed cases, creating a gap for reliable EHR phenotyping without local, dataset-specific tuning.
• What they did: Eight LLM-based systems generated T2DM SQL under progressively richer prompts using synthetic and UNC Health EHR data (both OMOP-based), then compared outputs to the eMERGE reference using executability, F1, PPV, and recall.
• The result: Limited-context recall was as low as 0.01, while high-context outputs reached ≥0.99 recall at synthetic L4 and >0.90 at UNC L3; however, UNC PPV stayed substantially lower than recall.