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 critical clinical challenge. Serum lactate is a cornerstone biomarker of metabolic dysfunction, yet its predictive limitations-particularly in patients without severe hyperlactatemia-are well recognized. The lactate-to-albumin ratio (LAR), a composite mixed-unit index integrating markers of acute metabolic dysfunction and systemic inflammation, has emerged as a promising predictor; however, its incremental discriminative advantage over lactate had not been formally tested in a large multicenter cohort using paired statistical methodology.

OBJECTIVE

This study aims to determine whether LAR provides statistically significantly higher prediction of 28-day mortality than lactate alone in adult intensive care unit (ICU) patients with sepsis, using threshold effect analysis, restricted cubic splines, DeLong test, and 9 interpretable machine learning models.

METHODS

We conducted a retrospective analysis of 3637 adult patients with sepsis from the multicenter eICU Collaborative Research Database (eICU-CRD; 208 hospitals, United States, 2014-2015). The primary outcome was 28-day all-cause in-hospital mortality among patients surviving the initial 48-hour ICU admission period. We used multivariable logistic regression (LR), Cox proportional-hazards regression, threshold effect analysis, restricted cubic spline modeling, DeLong test for area under the receiver operating characteristic curve (AUC) comparison, and machine learning models evaluated with Shapley additive explanations (SHAP) for interpretability. The cohort was divided 70/30 (stratified) into training and held-out test sets; the Synthetic Minority Oversampling Technique was applied exclusively within the training partition to prevent data leakage.

RESULTS

LAR consistently demonstrated stronger and more stable associations with mortality than lactate across all subgroups. DeLong test confirmed statistically significantly higher AUC for LAR: 28-day hospital mortality (AUCLAR=0.646, 95% CI 0.623-0.670 vs AUClactate=0.617, 95% CI 0.593-0.641; Z=6.37; P<.001; ΔAUC=0.029) and 28-day ICU mortality (AUCLAR=0.642 vs AUClactate=0.621; Z=3.71; P<.001). A nominally significant Acute Physiology and Chronic Health Evaluation IV (APACHE IV) × LAR interaction (hospital mortality, P for interaction=.02) indicated stronger LAR prognostic effects in lower-severity patients (APACHE IV≤70), representing within-biomarker effect modification requiring prospective validation. Among 9 machine learning models for ICU mortality, LR, random forest (RF), and gradient-boosting decision tree (GBDT) achieved the 3 highest AUCs (0.727, 0.726, and 0.725); Light Gradient Boosting Machine (LightGBM) demonstrated the best calibration (Brier score 0.096, the only model below the null Brier of 0.101 at the natural prevalence of 11.4%); GBDT achieved the highest precision-recall AUC (0.293). SHAP identified LAR among the top 10 predictive features in 3 of 4 models for hospital mortality (RF rank 4, LR rank 7, and LightGBM rank 8) and 1 of 4 for ICU mortality (RF rank 4).

CONCLUSIONS

LAR demonstrates statistically significantly higher discrimination than lactate alone for 28-day sepsis mortality prediction. LAR may offer greater prognostic utility in patients without severe hyperlactatemia, a population in whom early risk stratification may be particularly relevant.