Background Acute kidney injury (AKI) has a high incidence rate in intensive care unit (ICU) and becomes more prevalent in the elderly population. Therefore, finding the risk factors of AKI for the aged in ICU and establishing an early prediction model is helpful for clinical decision making.
Objective To establish an early prediction model of AKI for elderly patients in ICU by machine learning.
Methods Clinical data about elderly patients aged ≥65 years who were admitted to the ICU of the First Medical Center, Chinese PLA General Hospital from January 2018 to December 2021 (split into a training set 80% and internal testing set 20%) were collected. The outcome variable was whether AKI occurred. A total of 55 predictive variables including demography, vital signs, laboratory tests, and comorbidities were included in this paper. Multivariable predictive model was developed using machine learning models like Decision tree, Random forest, Logistic regression, XGBoost, and LightGBM. Area under Receiver Operating Characteristic Curve (AUC), accuracy, sensitivity, specificity and F1 value were used to evaluate the model predictive performance and choose the optimal model.
Results A total of 968 elderly patients were enrolled, of whom 304 were AKI patients (31.4%). The AUC of LightGBM model was 0.887, which was higher than that of Decision tree model, Random forest model, Logistic regression model and XGBoost model (0.795, 0.850, 0.853 and 0.875, respectively). The top 5 features selected by LightGBM model were Chalson's comorbidity index (CCI), fluid intake, serum creatinine (Scr), N-terminal pro-brain natriuretic peptide (NT-proBNP), Lactic acid (Lac).
Conclusion An early prediction model for AKI in the advanced from ICU is established by using LightGBM with good predictive efficiency, which is helpful for early clinical prediction and prevention of AKI in ICU.
DownLoad: