Association Between Lactate-to-Albumin Ratio and 28-Day All-Cause Mortality in ICU-Admitted Acute Pancreatitis Patients: Development and Validation of a Machine Learning-Based Predictive Model Integrating Albumin-Corrected Indices.

Abstract

Patients with severe acute pancreatitis (AP) carry a high mortality risk. The lactate-to-albumin ratio (LAR) has shown prognostic value in conditions including sepsis, heart failure and acute respiratory failure. We aimed to assess LAR's predictive value for 28-day mortality in ICU-admitted AP patients and to establish a machine learning-based predictive model integrating three albumin-derived indices-LAR, red blood cell distribution width-to-albumin ratio (RAR), and blood urea nitrogen-to-albumin ratio (BAR)-for this outcome. We analysed a cohort of 380 ICU-admitted AP patients from the MIMIC-IV database (2008-2022) as the training set to evaluate LAR's prediction of 28-day mortality. Predictors were screened using the Boruta algorithm, incorporating LAR, RAR, BAR and clinical variables (e.g., age, creatinine). Seven machine learning models were developed to predict 28-day all-cause mortality. Internal validation employed a 2:1 random data split; external validation used 298 AP patients from the eICU-CRD database (2014-2015). The SHAP method elucidated prediction mechanisms. Multivariate Cox regression analysis revealed that patients in the highest quartile of LAR (>1.72 mmol/g) exhibited a significantly elevated mortality risk (adjusted HR = 8.25, 95% CI 2.57-26.54). The Boruta feature selection algorithm identified 11 independent predictors, including LAR, BAR, RAR, and age. Following 5-fold internal cross-validation, the Extratrees model demonstrated superior performance in the training set benchmark testing (AUC = 0.8546). During internal validation, the Extratrees model maintained optimal predictive capability (AUC = 0.852). Calibration curve analysis and Decision Curve Analysis (DCA) substantiated its high predictive accuracy and clinical net benefit. External validation further corroborated the robustness of Extratrees model (AUC = 0.82). A web-based computational platform (https://bssrmyy.shinyapps.io/DynNom_Model2/) has been deployed to enable online individualized risk assessment. LAR is a strong independent predictor of 28-day mortality in ICU-admitted AP patients. This study pioneers the development and validation of a prognostic prediction model for AP based on three albumin-corrected indices (LAR, BAR, RAR) and machine learning. The Extratrees model performed excellently in both internal and external validations, providing an efficient tool for precise risk stratification and early intervention in ICU-managed AP. Not applicable (retrospective cohort study).