Feed

Acute kidney injury (AKI) is a frequent and severe complication in intensive care unit (ICU) patients with respiratory failure, associated with high mortality, prolonged hospitalization, and substantial healthcare burden. Conventional risk scores, such as SOFA and APACHE II, are not optimized for AKI prediction in this heterogeneous population. This study aimed to develop and validate an early AKI prediction model using machine learning. We analyzed 10,780 adult ICU patients with unspecified respiratory failure from the MIMIC-IV database, of whom 53.96% developed AKI according to KDIGO criteria. Ten supervised learning algorithms were trained using predictors from the first 48 h of ICU admission, with each model independently selecting its 15 most informative features via recursive feature elimination. Extreme gradient boosting (XGBoost) achieved the best performance (AUC 0.9023; accuracy 0.8247; sensitivity 0.8077; specificity 0.8386; precision 0.8040; negative predictive value 0.8419; F1-score 0.8058; Brier score 0.108). SHAP analysis identified creatinine_max, length of hospital stay, BUN_max, preexisting renal disease, and urine output as the most influential predictors. Leveraging routinely available early ICU data, this model enables accurate and interpretable AKI risk stratification. With external validation and integration into electronic health records, it could support proactive prevention and individualized management of AKI in critically ill respiratory failure patients.