Prediction of acute kidney injury in intensive care unit patients based on interpretable machine learning.

Objective: Acute kidney injury (AKI) poses a lethal risk in intensive care unit (ICU) patients, where early detection is challenging. This study was to establish a prediction model for AKI 24 hours in advance for ICU patients and to help clinicians monitor patients at an early stage by key features.

Methods: In this study, the Medical Information Mart for Intensive Care (MIMIC) databases were used to construct a dataset of critically ill patients. Predictive models were constructed using five machine learning algorithms based on MIMIC-IV data, and the best predictive model was selected by multiple model evaluation metrics. MIMIC-III data were used for external validation. We conducted an interpretability analysis of the model using SHapley Additive exPlanations (SHAP) to clarify key features and decision-making mechanisms.

Results: A total of 18,186 patient data were included in this study. The analysis combining calibration and decision curves demonstrated that the eXtreme Gradient Boosting (XGBoost) exhibited superior performance among the five algorithms, achieving an area under the receiver operating characteristic curve of 0.88. Interpretability analysis based on the XGBoost model showed diuretic use, mechanical ventilation, vasopressor use, age, and antibiotic use as the most important decision factors of the model. The SHAP summary plot was used to illustrate the effects of the top 19 features attributed to the XGBoost.

Conclusions: The XGBoost algorithm can predict the occurrence of AKI more accurately. Interpretative analysis of the model reveals the mechanisms of key features, and reflects the individual differences between patients, providing an important clinical reference.