Acute kidney injury severity in ICU patients: Developing and evaluating a data-driven analysis of clinical covariates using machine learning

BACKGROUND: Acute kidney injury remains a major cause of morbidity and mortality in critically ill patients. Existing classification systems rely on serum creatinine and urine output thresholds and do not fully capture the complex, multifactorial nature of acute kidney injury severity in the ICU. OBJECTIVE: To develop and validate a machine learning model using high-dimensional clinical data to identify covariates associated with the severity of acute kidney injury and enhance early risk stratification. METHODS: We performed a secondary analysis of 2,281 patients from the ICU in the MIMIC-IV database. Acute kidney injury severity was staged (0-3) based on serum creatinine criteria. After rigorous data cleaning, imputation, and feature selection, an XGBoost model was trained. Model performance was assessed, and gain-based metrics were used to identify the most important predictors of acute kidney injury severity. RESULTS: The XGBoost model demonstrated favorable discriminative performance, with an overall accuracy of 77.6% and a mean absolute error of 0.241. It achieved high precision and recall for stage 0 (precision = 0.917) and stage 3 (recall = 0.837), but showed decreased performance for stage 2. The top 10 covariates associated with increasing acute kidney injury severity were chronic kidney disease, sepsis, total bilirubin, norepinephrine, vasopressin, furosemide, phenylephrine, phosphate, potassium, and platelet count. CONCLUSIONS: Our machine learning-based model effectively identified key covariates of acute kidney injury severity in ICU patients using routinely collected clinical data. The findings provide a foundation for early identification, risk stratification, and targeted intervention strategies for acute kidney injury in critical care settings.