LightGBM demonstrates optimal performance in predicting 28-day mortality risk in patients with sepsis-associated acute kidney injury: development and validation of an interpretable machine learning model based on MIMIC-Ⅳ and eICU databases

Sepsis-associated acute kidney injury (SA-AKI) is a life-threatening condition with high mortality. This study aims to develop and validate an interpretable machine-learning model for assessing mortality risk in SA-AKI patients, and facilitate early identification of high-risk patients to support clinical decision-making.

A retrospective cohort study was conducted on the clinical data derived from the Medical Information Mart for Intensive Care Ⅳ (MIMIC-Ⅳ) and the eICU Collaborative Research Database (eICU-CRD). A total of 24487 patients from MIMIC-Ⅳ and 13757 ones from eICU-CRD were included for analysis. Predictive variables were selected through univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. The MIMIC-Ⅳ dataset was randomly divided into training and internal testing sets in a ratio of 8:2. Ten machine-learning algorithms were used to construct mortality risk prediction models, including logistic regression (LR), random forest (RF), light gradient boosting machine (LightGBM), K-nearest neighbors (KNN), eXtreme gradient boosting (XGBoost), ridge regression (Ridge), elastic net (ENet), decision tree (DT), stacking ensemble model (Stacking), and multilayer perceptron (MLP). Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and with calibration curve and decision curve analyses. SHapley Additive exPlanations (SHAP) were adopted to analyze the contribution and directional effect of each variable on prediction. External validation was performed using SA-AKI patients meeting inclusion criteria from the eICU-CRD database.

A total of 31 predictive variables were selected through univariate analysis combined with LASSO regression, and 10 machine-learning models for mortality risk prediction were constructed. The LightGBM model achieved an AUC value of 0.817 (95%CI: 0.803 to 0.830) in the internal testing set and 0.735 (95%CI: 0.725 to 0.745) in the external validation set based on the eICU-CRD database, outperforming other models. After comprehensive comparison of discrimination, calibration, and decision curve analyses across all models, the LightGBM model demonstrated optimal performance, suggesting favorable predictive efficiency and clinical applicability in mortality risk prediction for SA-AKI patients. SHAP analysis revealed that the most important variables for 28-day mortality risk included sequential organ failure assessment (SOFA) score, Charlson comorbidity index, Glasgow coma scale (GCS) score, mean body temperature, mean blood urea nitrogen, mean serum sodium, mean respiratory rate, mean creatinine, Oxford acute severity of illness score (OASIS), and red blood cell distribution width (RDW).

For SA-AKI patients, the mortality risk prediction model based on the LightGBM algorithm demonstrates optimal performance. The SOFA score, Charlson comorbidity index, and other variables are important indicators for predicting 28-day mortality in these patients, providing a reference for early identification of high-risk patients.