Machine Learning–Augmented Traditional Analysis of Lactate vs Lactate-to-Albumin Ratio for Predicting Mortality Risk in Patients With Sepsis: Large-Scale Retrospective Study

BACKGROUND Effective risk stratification in sepsis remains a clinical challenge. While lactate is a cornerstone biomarker, its predictive limitations are well-known. The lactate-to-albumin ratio (LAR) has emerged as a promising alternative, but its superiority, particularly in patients without severe hyperlactatemia, has not been rigorously validated in a large cohort using advanced, transparent methodologies. OBJECTIVE To determine whether the lactate-to-albumin ratio (LAR) provides superior and more robust prediction of 28-day mortality than lactate alone in adult sepsis patients, through comprehensive analysis using threshold effects, restricted cubic splines, and interpretable machine learning models in a large multicenter cohort. METHODS We conducted a retrospective analysis of 3637 adult sepsis patients from the multicenter eICU database. The primary outcome was 28-day mortality. We employed a dual approach: 1. Traditional statistical analyses, including threshold and restricted cubic spline modeling, to compare the predictive behavior of lactate and LAR. 2. Development of nine machine learning models to assess their predictive performance. We used SHapley Additive exPlanations (SHAP) to provide full model interpretability and to quantify the precise predictive contribution of each biomarker. RESULTS LAR consistently demonstrated a stronger and more stable association with mortality than lactate, especially in patients with lactate levels