Machine learning models for mortality prediction in critically ill patients with acute pancreatitis-associated acute kidney injury

Abstract

Background The occurrence of acute kidney injury (AKI) was associated with an increased mortality rate among acute pancreatitis (AP) patients, indicating the importance of accurately predicting the mortality rate of critically ill patients with acute pancreatitis-associated acute kidney injury (AP-AKI) at an early stage. This study aimed to develop and validate machine learning-based predictive models for in-hospital mortality rate in critically ill patients with AP-AKI by comparing their performance with the traditional logistic regression (LR) model. Methods This study used the data from three clinical databases. The predictors were identified by the Recursive Feature Elimination algorithm. The LR and two machine learning models including random forest (RF) and extreme gradient boosting (XGBoost) were developed using the ten-fold cross-validation to predict in-hospital mortality rate in AP-AKI patients. Results A total of 1,089 patients from Medical Information Mart for Intensive Care-IV (MIMIC-IV) and eICU Collaborative Research Database (eICU-CRD) were included in the training set, and 176 patients from Xiangya Hospital were included in the external validation set. The in-hospital mortality rate of the training and external validation sets was 13.77% and 54.55%, respectively. Compared to the AUC values of the LR model and the RF model, the AUC value of the XGBoost model [0.941, 95% confidence interval (CI): 0.931-0.952] was significantly higher (both P < 0.001), and the XGBoost model had the smallest Brier score of 0.039 in the training set. In the external validation set, the performance of the XGBoost model was acceptable with an AUC value of 0.724 (95% CI: 0.648-0.800). However, it did not differ significantly from the LR and RF model models. Conclusions The XGBoost model was superior to the LR and RF models in terms of both the discrimination and calibration in the training set, while whether the findings can be generalized needs to be further validated.