RESUMO
PURPOSE: To predict acute kidney injury (AKI) in a large intensive care unit (ICU) database. MATERIALS AND METHODS: A total of 30,020 ICU admissions with 17,222 AKI episodes were extracted from the Medical Information Mart from Intensive Care (MIMIC)-III database. These were randomly divided into a training set and an independent testing set in a ratio of 4:1. Data pertaining to demographics, admission information, vital signs, laboratory tests, critical illness scores, medications, comorbidities, and intervention measures were collected. Logistic regression, random forest, LightGBM, XGBoost, and an ensemble model was used for early prediction of AKI occurrence and important feature extraction. The SHAP analysis was adopted to reveal the impact of prediction for each feature. RESULTS: The ensemble model had the best overall performance for predicting AKI before 24 h, 48 h and 72 h. The F1 values were 0.915, 0.893, and 0.878, respectively. AUCs were 0.923, 0.903, and 0.895, respectively. CONCLUSIONS: Based on readily available electronic medical record (EMR) data, gradient boosting decision tree models are highly accurate at early AKI prediction in critically ill patients.
