Identifying interstitial lung disease with large language models on clinical notes and claims

Zero-shot large language models identify interstitial lung disease in 309 oncology patients at Mass General Brigham with higher PPV than published claims algorithms, reaching 100.0% PPV.

• Why it matters: Claims-based ILD phenotyping forces a tradeoff between PPV and sensitivity, risking misclassification that can attenuate drug-safety effect estimates.
• What they did: In 309 oncology patients with ILD diagnosis codes, 12 open-access LLMs were evaluated zero-shot using claims, EHR notes, or both, and compared against 20 published claims algorithms using PPV and yield.
• The result: Across 36 model–input evaluations, LLMs extended the PPV–yield frontier, with Qwen3-14B at 100.0% PPV (24.3% yield) using combined inputs; simulated drug-safety projection gave an observed odds ratio of 2.67 vs 1.07, implying 591 vs 27,200 patients for 80% power.

AI summary of the abstract · may contain errors · not medical advice · read the paper.