C54-31 Beyond the Obesity Paradox: Sepsis Outcomes Revealed Through Local LLM-Based EMR Mining

Introduction Large Language Models (LLMs) offer a solution to unstructured Electronic Medical Records (EMRs) by converting free-text into structured data. Sepsis remains a major health concern, contributing to morbidity and mortality, with disproportionally higher rates in underserved populations. A well-described but controversial phenomenon, “obesity paradox” suggests higher body mass index (BMI) may be protective in sepsis. Overweight and obese individuals have been associated with reduced mortality in sepsis; normal BMI and underweight are associated with higher mortality in many studies, contributing to the obesity paradox. LLM-based extraction offers a novel pathway to systematically evaluate these complex associations and provide critical insights for guidance. Methods We reviewed over 400 EMRs, focusing on 182 patients. Each record spanning 200 pages. Using optical character recognition (OCR) and LLM-based extraction, we identified 60 clinical variables on a secure GPU (Graphic Processing Unit)-equipped server compliant with HIPAA. Outcomes analyzed were in-hospital mortality using bivariate tests and regression models, adjusting for key covariates like SOFA score, age, and sex. Random spot checks of extracted variables were performed for assessment of data reliability. Results Among the 182 patients, a multivariable logistic regression was performed to ascertain the effects of BMI category, illness severity (SOFA score), age, and sex on the likelihood of in-hospital mortality. The model was statistically significant (p < 0.001). After adjusting for these covariates, SOFA score (adjusted Odds Ratio [aOR] = 1.19, 95% CI [1.10, 1.29], p < 0.001) and age (aOR = 1.03, 95% CI [1.00, 1.06], p = 0.028) were significant independent predictors of mortality. Most notably, patients with Class III obesity had significantly higher odds of mortality compared to those with a normal BMI (aOR = 2.57, 95% CI [1.08, 6.28], p = 0.032). A strong trend was observed for underweight patients (aOR = 2.89, 95% CI [0.83, 10.60], p = 0.099), demonstrating a U-shaped distribution. The analysis also identified predictors of mortality: number of pressors (Odds Ratio [OR] 3.012), cardiac arrest (OR 3.57), bilirubin (OR 4.023), FiO2 (OR 3.565), and respiratory rate (RR max) (OR 2.222). Conclusion Our LLM-enabled analysis found no protective effect for obesity. However, it identified disease severity markers as the most robust independent predictors of mortality. This study confirms the viability of using a secure HIPAA-compliant, locally deployed LLM for extracting unstructured data from EMRs, a critical innovation for disadvantaged health systems. Future directions will focus on prospective validation and real-time extraction for clinical decision support.