Comorbidity Burden and Length of Stay After Myocardial Infarction: A Retrospective Study Using Large Language Model-Assisted Chart Abstraction in a Tertiary Cardiac Hospital
Vanessa Zi Kun Lim, Chee Tang Chin, Fang Yee Chee, Yee How Lau, Jie Yin TAN, Heng Lee Hendy CHUA, Weiting Huang, Jonathan Jiunn Liang Yap, Siang Jin Terrance Chua, Khung Keong Yeo, Anders Olof Sahlén · The American Journal of Cardiology · 2026
Patients' length of stay (LOS) during admission for myocardial infarction (MI) represents a closely tracked outcome metric for Cardiology services, which has declined over time. Charlson Comorbidity Index (CCI) is a validated tool for assessing overall disease burden, but its abstraction is laborious. Large language models (LLMs) have emerged as an attractive option for chart abstraction, with limited uptake so far. We studied n = 6,129 MI admissions at a tertiary cardiac hospital from May 2018 to August 2024. An LLM was used to derive CCI from clinical case notes with manual validation by expert auditors in 10% of patients (R-squared 0.81). Fine and Gray survival analysis was performed to study time-to-discharge in mixed survival models with clustering on individual patients and in-hospital death as competing outcome, adjusting for covariates not already included in CCI (e.g., age, race, smoking, MI type, atrial fibrillation, revascularization). We found an increase in CCI over time (0.09 (CI 95 : 0.06 – 0.12) points per year) due to more diabetes, hypertension, renal failure, lymphomas and metastatic solid tumors) and a concomitant rise in mean LOS by 6.5 (CI 95 : 3.9 – 9.1) hours annually. CCI emerged as the strongest predictor of LOS (p < 2e-16), with no significant effect by calendar year to suggest a residual secular trend after full adjustment (p = 0.17). These results validate the use of an LLM for chart abstraction and indicate that rising patient complexity is reversing the earlier downtrend in LOS in MI patients.