Machine Learning for Renal Replacement Therapy in Patients with Severe Acute Kidney Injury Associated with Sepsis

Background : The high risk of renal replacement therapy (RRT) in patients with sepsis-associated acute kidney injury (SA-AKI) remains unclear. Methods : This two-center retrospective study analyzed data from the MIMIC database and the eICU database. SA-AKI patients were included, and demographic data and laboratory parameters from the 5 days before diagnosis were collected as variables. The MIMIC database served as the training and validation set for developing models using multivariate logistic regression, random forest, SVM and XGBoost. Moreover, the eICU database was used as an external test set. We visualized the model using nomograms and the SHAP method, and evaluated model performance through variety of methods. Results : A total of 22,220 patients with severe SA-AKI were included in the analysis, with 1,358 (6.11%) receiving RRT. The RRT group exhibited more severe metabolic acidosis, including lower PH (7.23 vs. 7.31, P<0.001) , lower buffer excess(-8.25 vs. -3.10, P<0.001), higher lactate levels (5.07 vs. 3.03, P<0.001), more disordered electrolytes and poorer coagulation function. The six most important parameters were the APS score, SOFA score, minimum calcium level, maximum APTT, maximum BUN, and minimum buffer excess in the XGBoost model with great performance. The Area Under the Receiver Operating Characteristic Curve for the internal and external test and sets were 0.906 and 0.816, respectively. In the meanwhile, XGBoost has the best Area Under the Precision-Recall Curve。 Conclusion : Our study constructed a predictive model for initiating RRT in critically ill patients with severe SA-AKI using twocenter databases and the machine learning algorithms.