Bridging Large Language Model–Based AI and Conventional Statistics for Retrospective Mortality Risk Assessment in Intra-Abdominal Sepsis and Septic Shock

Background: Severe intra-abdominal infections are a major cause of sepsis and septic shock in intensive care units (ICUs), with high mortality and heterogeneity. Identifying variables associated with mortality is essential for risk assessment in this population. The potential of artificial intelligence (AI)-assisted analysis using large language models (LLMs) to complement conventional statistics remains underexplored. Methods: This retrospective, single-center study analyzed 172 adult ICU patients with intra-abdominal sepsis or septic shock, identified among 6674 patients over one year. Conventional statistical methods assessed mortality-related variables. Two LLMs (Google Gemini 3.1 Pro and Claude Sonnet 5) were separately prompted to examine the pseudonymized, outcome-labeled dataset in five runs using structured prompts; their narrative outputs were compared with conventional findings. Results: Overall in-ICU mortality was 60.5%. APACHE II showed the highest discriminatory capacity for mortality (AUC = 0.695), followed by admission lactate (AUC = 0.636) and SOFA score (AUC = 0.626). LLM outputs most consistently reproduced the direction of findings involving SOFA, APACHE II, and lactate, while coverage of other variables varied between models and runs. Conclusions: Conventional statistics and AI-assisted LLM analyses yielded largely convergent findings for mortality markers. However, variability across other findings supports AI-assisted analysis as an exploratory, complementary approach. Further validation in larger, independent, multicenter cohorts is required before widespread routine clinical application.