Large Language Model-Enhanced Electronic Frailty Assessment and Prognostic Performance in Hospitalized Adults: A Retrospective Cohort Study

GPT-4o–enhanced electronic frailty assessment classified 58.8% of 616 hospitalized adults as frail (vs 35.1% standard) and remained independently associated with 90-day adverse outcomes.

• Why it matters: Standard eFIs from structured EHR data may miss functional deficits documented in clinical notes, leaving uncertainty about whether better frailty capture improves prognostic performance.
• What they did: In a retrospective cohort of adults aged 50+ from an academic medical center and community hospital (Feb 2022–Apr 2024), investigators compared a 35-deficit diagnosis-based eFI with an LLM-enhanced eFI using PT documentation extracted by GPT-4o. Frailty used a threshold of 0.20+, with 90-day composite outcome modeled via multivariable logistic regression and discrimination assessed by AUROC.
• The result: Both eFIs were linked to the composite outcome (adjusted OR 1.98 vs 2.46), but AUROC was similar (0.636 vs 0.660; p=0.12); LLM enhancement increased sensitivity (42.5% to 69.4%) while reducing specificity (73.0% to 52.7%).

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