Decision tree algorithm applied to MIMIC-III database for the prediction of acute kidney injury in ICU patients

Abstract

Objective Acute kidney injury (AKI) is one of the most common complications and fatal factors in intensive care unit (ICU). Accurate prediction of AKI risk and identification of key factors related to AKI can provide effective guidance for clinical decision-making and intervention for patients with AKI risk. Methods A total of 30 020 patients in ICU (including 17 222 AKI patients and 12 798 Non-AKI patients) were selected from the public database MIMIC-III in this study, and basic information, physiological and biochemical indicators, drug use, and comorbidity during their stay in ICU were collected. All patients were randomly divided into training sets and independent testing sets according to the ratio of 4:1, and logistic regression, random forest, and lightgbm were applied to construct models for AKI predication in three time points including 24 h, 48 h and 72 h, respectively. The 10-fold cross validation was used to train and validate various models to predict the occurrence of AKI, and obtain important features. Furthermore, 24 h prediction models were used to predict AKI every 24 h during the 7-day window. Results lightgbm achieved the best performance with AUC values of 0.90, 0.88, 0.87 for 24 h, 48 h, and 72 h prediction, respectively, and F1 values were 0.91, 0.88, and 0.86. In prediction of every 24 h, the success rates of identifying AKI patients were 89%, 83%, and 80% in one day, two days and three days in advance, respectively. It was found that the length of stay in ICU, body weight, albumin, systolic blood pressure, bicarbonate, glucose, white blood cell count, body temperature, diastolic blood pressure and blood urea nitrogen played vital roles in predicting AKI for ICU patients. Using only 24 important features, the models could still achieve prominent prediction performance. Conclusions Based on basic information, physiological and biochemical indicators, drug use, and comorbidity, machine learning methods can be adopted to effectively predict AKI risk for ICU patients at several time points, and determine the dominant factors relative to AKI.