Construction and validation of prognostic models for sepsis associated acute kidney injury patients with moderate to high SOFA score

Background Sepsis-associated acute kidney injury (S-AKI) is a prevalent complication in critically ill patients and is generally associated with elevated mortality. The objective of this study was to develop and validate an interpretable prognostic prediction model for S-AKI patients with moderate to high SOFA scores using machine learning methods. Methods The data for the training cohort were obtained from the Critical Care Medicine Information Repository IV (MIMIC-IV) database, version 3.0, for modelling purposes, while the validation data were obtained from the eICU Collaborative Research Database (eICU-CRD). The least absolute shrinkage and selection operator (LASSO) regression method was used to determine predictors of mortality. The predictive model was constructed with seven machine learning algorithms. The predictive effects were assessed using receiver operating characteristic curves (ROC) and decision curve analysis (DCA). Finally, the SHapley Additive exPlanations (SHAP) method was employed to visualise the characteristics of the model. Results In the course of the study, 2,728 MIMIC-IV patients were included, of whom 806 died; in addition, 3,309 eICU-CRD patients were included, of whom 521 died. A model incorporating 17 variables was constructed, with XGBoost demonstrating optimal performance, yielding an area under the curve (AUC) of 0.783 (95% confidence interval [CI]: 0.749-0.817) in the internally validated model and an AUC of 0.704 (95% CI: 0.680-0.728) in the externally validated model. The analysis revealed significant contributors to the model, including biochemical markers such as blood urea nitrogen (BUN), international normalized ratio (INR), age, partial thromboplastin time (PTT), and vital signs (respiratory rate, heart rate, and temperature). SHAP-based summary plots were utilised to elucidate the comprehensive impact of the XGBoost model. Conclusion A machine learning model was constructed in order to predict 28-day mortality in S-AKI patients with intermediate- to high-risk SOFA scores. The potential of this model as a clinically reliable tool was then validated. The utilisation of this model is anticipated to facilitate an enhancement in the comprehension of the risk of mortality in S-AKI patients among clinicians.