Abstract
OBJECTIVE: To determine whether the lactate dehydrogenase-to-albumin ratio (LDAR) predicts 28-day mortality in critically ill patients with gastrointestinal (GI) malignancies and to quantify its incremental value in machine learning (ML) models. METHODS: We conducted a retrospective cohort study in Medical Information Mart For Intensive Care IV (2008-2019). Adults with GI malignancies who met Sepsis-3 within 24 h of their first intensive care unit (ICU) admission were included. LDAR was computed from laboratory results obtained within 24 h. The primary outcome was 28-day all-cause mortality. Associations were estimated with multivariable logistic regression; nonlinearity was evaluated using restricted cubic splines. Multiple ML classifiers (AdaBoost, XGBoost, random forest) were trained with a 7:3 split and 5-fold cross-validation. Discrimination (area under the curve-AUC), clinical utility (decision-curve analysis), and interpretability (Shapley additive explanation-SHAP) were assessed. Prespecified subgroup analyses stratified by metastatic status and infection site were performed. Covariates included age, sex, weight, vital signs, Sequential Organ Failure Assessment, and metastatic status. Missingness was handled with multiple imputation by chained equations. RESULTS: Among 1177 patients, 28-day mortality was 48.4%. Higher LDAR was independently associated with death; the adjusted odds ratio for Q4 vs Q1 was 5.15 (95% CI 3.52-7.61; P < 0.001). Spline analyses showed a steep risk rise above log2(LDAR) = 5.314 (≈40). Effects were directionally consistent across subgroups with no significant LDAR × metastatic interaction (P = 0.084) and borderline heterogeneity by infection site (P = 0.051). The best ML model, AdaBoost, achieved an AUC of 0.835 and yielded the largest net clinical benefit across clinically relevant thresholds. SHAP ranked LDAR among the most influential predictors. Sensitivity analyses using alternative exposure codings produced concordant estimates. Model calibration was acceptable. CONCLUSION: LDAR is a simple, interpretable, and independently prognostic biomarker for 28-day mortality in ICU patients with GI malignancies. Incorporating LDAR into ML models improved discrimination and decision benefit over conventional severity scores, supporting LDAR-based early risk stratification. External multicenter validation and evaluation of dynamic LDAR trajectories are warranted, and transparent analytic reporting.