LightGBM model for predicting acute kidney injury risk in ICU patients

  Objective  To develop a predictive model for acute kidney injury (AKI) in critically ill patients based on LightGBM and provide support for clinical decision.

  Methods  An open source critical care database, MIMIC-Ⅲ was used in this study. A total of 1 166 patients were included in this study with median age of 70.93 years, of which 513 cases (44.00%) were male, and 884 of them had developed AKI (75.8%). To predict the occurrence of AKI after 24 hours, the model was built utilizing predictors including laboratory tests and vital signs measured at ICU admission. LightGBM, logistic regression, and random forest models were established to predict the risk of AKI, and their predictive performances were evaluated using five-fold cross-validation.

  Results  The accuracy of LightGBM was 0.89, and the AUC was 0.92. However, the accuracy of logistic regression and random forest were 0.84 and 0.86, and their AUC were 0.75 and 0.89, respectively.

  Conclusion  LightGBM performs well in predicting AKI at 24 hours after ICU admission. The accuracy and AUC of LightGBM model are up to 0.89 and 0.92.