Comorbidity Burden and Length of Stay After Myocardial Infarction: A Retrospective Study Using Large Language...

Area of Science:

• Cardiology
• Health Informatics
• Medical Data Science

Background:

• Patient length of stay (LOS) for myocardial infarction (MI) is a key cardiology metric that has been declining.
• Charlson Comorbidity Index (CCI) quantifies disease burden but is labor-intensive to abstract from patient records.
• Large language models (LLMs) show potential for automated clinical chart abstraction.

Purpose of the Study:

• To investigate the trend of patient complexity using CCI in MI admissions.
• To assess the impact of patient complexity on MI LOS.
• To validate the use of LLMs for deriving CCI from clinical notes.

Main Methods:

• Retrospective study of 6,129 MI admissions (May 2018 - August 2024) at a tertiary cardiac hospital.
• LLM used for CCI derivation from clinical notes, with 10% manual validation (R-squared 0.81).
• Fine and Gray survival analysis performed to model time-to-discharge, adjusting for covariates and competing risks.

Main Results:

• CCI increased by 0.09 points/year, driven by conditions like diabetes, hypertension, and metastatic cancers.
• Mean MI LOS increased by 6.5 hours annually, strongly correlated with higher CCI (p < 2e-16).
• LLM-derived CCI accurately reflected patient complexity, and rising complexity explained the reversal of LOS downtrend.

Conclusions:

• Increasing patient comorbidity burden is a primary driver of rising MI length of stay.
• LLMs offer a validated and efficient method for clinical chart abstraction, including CCI calculation.
• Automated abstraction tools are crucial for understanding evolving patient complexity and healthcare outcomes.